This PDF file contains the front matter associated with SPIE Proceedings Volume 9414, including the Title Page, Copyright information, Table of Contents, Introduction (if any), and Conference Committee listing.

Inflammatory muscle disease is a group of rare idiopathic conditions that cause progressive skeletal muscle weakness. Diagnosis typically requires an assortment of clinical tests, including magnetic resonance imaging (MRI) of the thigh muscles to assess fat infiltration and inflammation. We hypothesise that features from multi-spectral MRI can accurately predict patient diagnosis, without the need for additional tests. A novel method is presented using computer-extracted features of the T1-STIR bivariate histogram to detect disease. The dataset comprised of 78 image-pairs from 8 patients with inflammatory muscle disease symptoms and 61 image-pairs from 9 control cases with no disease. T1 and STIR slices were co-registered and the background discarded. A feature vector was designed to measure the distribution of standardised intensity values (e.g. standard deviation, kurtosis, gini) for the muscle, fat, and leg regions of the bivariate histogram. Feature dimensionality was reduced using a combined leave-one-out and k-folds cross-validation method to select the most important features. A Bayes network was trained to predicted patient diagnosis on a per-slice basis, 10- fold cross-validated. The system attained 92% sensitivity and 82% specificity (ROC area 0.93). These results support the hypothesis that accurate diagnosis of inflammatory muscle disease is possible using MRI alone, without the need for additional clinical tests, with the potential benefit of faster diagnosis and better care for patients with this group of rare conditions.

Segmentation of the sternum in medical images is of clinical significance as it frequently serves as a stable reference to image registration and segmentation of other organs in the chest region. In this paper we present a fully automated algorithm to segment the sternum in low-dose chest CT images (LDCT). The proposed algorithm first locates an axial seed slice and then segments the sternum cross section on the seed slice by matching a rectangle model. Furthermore, it tracks and segments the complete sternum in the cranial and caudal direction respectively through sequential axial slices starting from the seed slice. The cross section on each axial slice is segmented using score functions that are designed to have local maxima at the boundaries of the sternum. Finally, the sternal angle is localized. The algorithm is designed to be specifically robust with respect to cartilage calcifications and to accommodate the high noise levels encountered with LDCT images. Segmentation of 351 cases from public datasets was evaluated visually with only 1 failing to produce a usable segmentation. 87.2% of the 351 images have good segmentation and 12.5% have acceptable segmentation. The sternal body segmentation and the localization of the sternal angle and the vertical extents of the sternum were also evaluated quantitatively for 25 good cases and 25 acceptable cases. The overall weighted mean DC of 0.897 and weighted mean distance error of 2.88 mm demonstrate that the algorithm achieves encouraging performance in both segmenting the sternal body and localizing the sternal angle.

Degenerative changes to the cervical spine can be accompanied by neck pain, which can result from narrowing of the intervertebral disc space and growth of osteophytes. In a lateral x-ray image of the cervical spine, degenerative changes are characterized by vertebral bodies that have indistinct boundaries and limited spacing between vertebrae. In this paper, we present a machine learning approach to detect and localize degenerative changes in lateral x-ray images of the cervical spine. Starting from a user-supplied set of points in the center of each vertebral body, we fit a central spline, from which a region of interest is extracted and image features are computed. A Random Forest classifier labels regions as degenerative change or normal. Leave-one-out cross-validation studies performed on a dataset of 103 patients demonstrates performance of above 95% accuracy.

The aim of this study was to investigate imaging statistical approaches for classifying 3D osteoarthritic morphological variations among 169 Temporomandibular Joint (TMJ) condyles. Cone beam Computed Tomography (CBCT) scans were acquired from 69 patients with long-term TMJ Osteoarthritis (OA) (39.1 ± 15.7 years), 15 patients at initial diagnosis of OA (44.9 ± 14.8 years) and 7 healthy controls (43 ± 12.4 years). 3D surface models of the condyles were constructed and Shape Correspondence was used to establish correspondent points on each model. The statistical framework included a multivariate analysis of covariance (MANCOVA) and Direction-Projection- Permutation (DiProPerm) for testing statistical significance of the differences between healthy control and the OA group determined by clinical and radiographic diagnoses. Unsupervised classification using hierarchical agglomerative clustering (HAC) was then conducted. Condylar morphology in OA and healthy subjects varied widely. Compared with healthy controls, OA average condyle was statistically significantly smaller in all dimensions except its anterior surface. Significant flattening of the lateral pole was noticed at initial diagnosis (p < 0.05). It was observed areas of 3.88 mm bone resorption at the superior surface and 3.10 mm bone apposition at the anterior aspect of the long-term OA average model. 1000 permutation statistics of DiProPerm supported a significant difference between the healthy control group and OA group (t = 6.7, empirical p-value = 0.001). Clinically meaningful unsupervised classification of TMJ condylar morphology determined a preliminary diagnostic index of 3D osteoarthritic changes, which may be the first step towards a more targeted diagnosis of this condition.

During image segmentation, 3D Statistical Shape Models (SSM) usually conduct a limited search for target landmarks within one-dimensional search profiles perpendicular to the model surface. In addition, landmark appearance is modeled only locally based on linear profiles and weak learners, altogether leading to segmentation errors from landmark ambiguities and limited search coverage. We present a new method for 3D SSM segmentation based on 3D Random Forest Regression Voting. For each surface landmark, a Random Regression Forest is trained that learns a 3D spatial displacement function between the according reference landmark and a set of surrounding sample points, based on an infinite set of non-local randomized 3D Haar-like features. Landmark search is then conducted omni-directionally within 3D search spaces, where voxelwise forest predictions on landmark position contribute to a common voting map which reflects the overall position estimate. Segmentation experiments were conducted on a set of 45 CT volumes of the human liver, of which 40 images were randomly chosen for training and 5 for testing. Without parameter optimization, using a simple candidate selection and a single resolution approach, excellent results were achieved, while faster convergence and better concavity segmentation were observed, altogether underlining the potential of our approach in terms of increased robustness from distinct landmark detection and from better search coverage.

Pulmonary embolism (PE) affects up to 600,000 patients and contributes to at least 100,000 deaths every year in the United States alone. Diagnosis of PE can be difficult as most symptoms are unspecific and early diagnosis is essential for successful treatment. Computed Tomography (CT) images can show morphological anomalies that suggest the existence of PE. Various image-based procedures have been proposed for improving computer-aided diagnosis of PE. We propose a novel method for detecting PE based on localized vessel-based features computed in Dual Energy CT (DECT) images. DECT provides 4D data indexed by the three spatial coordinates and the energy level. The proposed features encode the variation of the Hounsfield Units across the different levels and the CT attenuation related to the amount of iodine contrast in each vessel. A local classification of the vessels is obtained through the classification of these features. Moreover, the localization of the vessel in the lung provides better comparison between patients. Results show that the simple features designed are able to classify pulmonary embolism patients with an AUC (area under the receiver operating curve) of 0.71 on a lobe basis. Prior segmentation of the lung lobes is not necessary because an automatic atlas-based segmentation obtains similar AUC levels (0.65) for the same dataset. The automatic atlas reaches 0.80 AUC in a larger dataset with more control cases.

We have developed a computer-aided detection (CAD) system for assisting radiologists in detection of pulmonary embolism (PE) in computed tomographic pulmonary angiographic (CTPA) images. The CAD system includes stages of pulmonary vessel segmentation, prescreening of PE candidates and false positive (FP) reduction to identify suspicious PEs. The system was trained with 59 CTPA PE cases collected retrospectively from our patient files (UM set) with IRB approval. Five feature groups containing 139 features that characterized the intensity texture, gradient, intensity homogeneity, shape, and topology of PE candidates were initially extracted. Stepwise feature selection guided by simplex optimization was used to select effective features for FP reduction. A linear discriminant analysis (LDA) classifier was formulated to differentiate true PEs from FPs. The purpose of this study is to evaluate the performance of our CAD system using an independent test set of CTPA cases. The test set consists of 50 PE cases from the PIOPED II data set collected by multiple institutions with access permission. A total of 537 PEs were manually marked by experienced thoracic radiologists as reference standard for the test set. The detection performance was evaluated by freeresponse receiver operating characteristic (FROC) analysis. The FP classifier obtained a test Az value of 0.847 and the FROC analysis indicated that the CAD system achieved an overall sensitivity of 80% at 8.6 FPs/case for the PIOPED test set.

We present a fully automatic method for the assessment of spiculation of pulmonary nodules in low-dose Computed Tomography (CT) images. Spiculation is considered as one of the indicators of nodule malignancy and an important feature to assess in order to decide on a patient-tailored follow-up procedure. For this reason, lung cancer screening scenario would benefit from the presence of a fully automatic system for the assessment of spiculation. The presented framework relies on the fact that spiculated nodules mainly differ from non-spiculated ones in their morphology. In order to discriminate the two categories, information on morphology is captured by sampling intensity profiles along circular patterns on spherical surfaces centered on the nodule, in a multi-scale fashion. Each intensity profile is interpreted as a periodic signal, where the Fourier transform is applied, obtaining a spectrum. A library of spectra is created by clustering data via unsupervised learning. The centroids of the clusters are used to label back each spectrum in the sampling pattern. A compact descriptor encoding the nodule morphology is obtained as the histogram of labels along all the spherical surfaces and used to classify spiculated nodules via supervised learning. We tested our approach on a set of nodules from the Danish Lung Cancer Screening Trial (DLCST) dataset. Our results show that the proposed method outperforms other 3-D descriptors of morphology in the automatic assessment of spiculation.

The availability of large medical image datasets is critical in training and testing of computer aided diagnosis (CAD) systems. However, collection of data and establishment of ground truth for medical images are both costly and difficult. To address this problem, we have developed an image composition tool that allows users to modify or supplement existing datasets by seamlessly inserting a clinical lesion extracted from a source image into a different location on a target image. In this study we focus on the application of this tool to the training of a CAD system designed to detect pulmonary nodules in chest CT. To compare the performance of a CAD system without and with the use of our image composition tool, we trained the system on two sets of data. The first training set was obtained from original CT cases, while the second set consisted of the first set plus nodules in the first set inserted into new locations. We then compared the performance of the two CAD systems in differentiating nodules from normal areas by testing each trained system against a fixed dataset containing natural nodules, and using the area under the ROC curve (AUC) as the figure of merit. The performance of the system trained with the augmented dataset was found to be significantly better than that trained with the original dataset under several training scenarios.

The recently published Vancouver model for lung nodule malignancy prediction holds great promise as a practically feasible tool to mitigate the clinical decision problem of how to act on a lung nodule detected at baseline screening. It provides a formula to compute a probability of malignancy from only nine clinical and radiologic features. The feature values are provided by user interaction but in principle could also be automatically pre-filled by appropriate image processing algorithms and RIS requests. Nodule diameter is a feature with crucial influence on the predicted malignancy, and leads to uncertainty caused by inter-reader variability. The purpose of this paper is to analyze how strongly the malignancy prediction of a lung nodule found with CT screening is affected by the inter-reader variation of the nodule diameter estimation. To this aim we have estimated the magnitude of the malignancy variability by applying the Vancouver malignancy model to the LIDC-IDRI database which contains independent delineations from several readers. It can be shown that using fully automatic nodule segmentation can significantly lower the variability of the estimated malignancy, while demonstrating excellent agreement with the expert readers.

This work aimed to compare two different types of volume estimation methods (a model-based and a segmentationbased method) in terms of identifying factors affecting measurement uncertainty. Twenty-nine synthetic nodules with varying size, radiodensity, and shape were placed in an anthropomorphic thoracic phantom and scanned with a 16- detector row CT scanner. Ten repeat scans were acquired using three exposures and two slice collimations, and were reconstructed with varying slice thicknesses. Nodule volumes were estimated from the reconstructed data using a matched-filter and a segmentation approach. Log transformed volumes were used to obtain measurement error with truth obtained through micro-CT. ANOVA and multiple linear regression were applied to measurement error to identify significant factors affecting volume estimation for each method. Root mean square of measurement errors (RMSE) for meaningful subgroups, repeatability coefficients (RC) for different imaging protocols, and reproducibility coefficients (RDC) for thin and thick collimation conditions were evaluated. Results showed that for both methods, nodule size, shape and slice thickness were significant factors. Collimation was significant for the matched-filter method. RMSEs for matched-filter measurements were in general smaller than segmentation. To achieve RMSE on the order of 15% or less for {5, 8, 9, 10mm} nodules, the corresponding maximum allowable slice thicknesses were {3, 5, 5, 5mm} for the matched-filter and {0.8, 3, 3, 3mm} for the segmentation method. RCs showed similar patterns for both methods, increasing with slice thickness. For 8-10mm nodules, the measurements were highly repeatable provided the slice thickness was ≤3mm, regardless of method and across varying acquisition conditions. RDCs were lower for thin collimation than thick collimation protocols. While RDC of matched filter volume estimation results was always lower than segmentation results, for 8-10mm nodules with thin collimation protocols, the measurements for both approaches were highly reproducible (RDC on the order of 15% or less). These findings are valuable for validating lung nodule volume as a quantitative imaging biomarker.

Calcium burden determined in CT images acquired in lung cancer screening is a strong predictor of cardiovascular events (CVEs). This study investigated whether subjects undergoing such screening who are at risk of a CVE can be identified using automatic image analysis and subject characteristics. Moreover, the study examined whether these individuals can be identified using solely image information, or if a combination of image and subject data is needed. A set of 3559 male subjects undergoing Dutch-Belgian lung cancer screening trial was included. Low-dose non-ECG synchronized chest CT images acquired at baseline were analyzed (1834 scanned in the University Medical Center Groningen, 1725 in the University Medical Center Utrecht). Aortic and coronary calcifications were identified using previously developed automatic algorithms. A set of features describing number, volume and size distribution of the detected calcifications was computed. Age of the participants was extracted from image headers. Features describing participants' smoking status, smoking history and past CVEs were obtained. CVEs that occurred within three years after the imaging were used as outcome. Support vector machine classification was performed employing different feature sets using sets of only image features, or a combination of image and subject related characteristics. Classification based solely on the image features resulted in the area under the ROC curve (A_z) of 0.69. A combination of image and subject features resulted in an A_z of 0.71. The results demonstrate that subjects undergoing lung cancer screening who are at risk of CVE can be identified using automatic image analysis. Adding subject information slightly improved the performance.

We are developing an automated method to select the best-quality vessels from coronary arterial trees in multiplephase cCTA and build a best-quality tree to facilitate the detection of stenotic plaques. Using our previously developed vessel registration method, the vessels from different phases were automatically registered. Branching points on the centerline are projected onto the registered trees. The centerlines are split into branches based on the projected branching points. Each tree branch is then straightened. The registered trees and centerline branches are used to determine the correspondence of branches between phases so that each branch can be compared to its corresponding branches in the other phases. A vessel quality measure (VQM) is calculated as the average radial gradients at the vessel wall over the entire vessel branch. The quality of the corresponding branches from all phases is automatically compared using the VQM. An observer preference study was conducted with two radiologists to visually compare the quality of the vessels. For each branch, the pair that was automatically determined to be the best and worst quality by the VQM was used for the radiologists’ visual assessment. Each radiologist, blinded to the VQM, evaluated pairs of corresponding branches and provided their preference. The performance of the automatic selection using VQM was evaluated as the percentage of the total number of vessel pairs for which the automatic selection agreed with the radiologist’s selection of the higher quality branch in the pair. The agreement between the first radiologist and the automated selection was 80% and that between the second radiologist and the automated selection was 82%. In comparison, the agreement between the two radiologists was 90%. This preliminary study demonstrates the feasibility of using an automated method to select the best-quality vessels from multiple cCTA phases.

Swelling of the optic nerve head (ONH) is subjectively assessed by clinicians using the Frisén scale. It is believed that a direct measurement of the ONH volume would serve as a better representation of the swelling. However, a direct measurement requires optic nerve imaging with spectral domain optical coherence tomography (SD-OCT) and 3D segmentation of the resulting images, which is not always available during clinical evaluation. Furthermore, telemedical imaging of the eye at remote locations is more feasible with non-mydriatic fundus cameras which are less costly than OCT imagers. Therefore, there is a critical need to develop a more quantitative analysis of optic nerve swelling on a continuous scale, similar to SD-OCT. Here, we select features from more commonly available 2D fundus images and use them to predict ONH volume. Twenty-six features were extracted from each of 48 color fundus images. The features include attributes of the blood vessels, optic nerve head, and peripapillary retina areas. These features were used in a regression analysis to predict ONH volume, as computed by a segmentation of the SD-OCT image. The results of the regression analysis yielded a mean square error of 2.43 mm³ and a correlation coefficient between computed and predicted volumes of R = 0:771, which suggests that ONH volume may be predicted from fundus features alone.

Cardiac visceral fat was segmented from low-dose non-contrast chest CT images using a fully automated method. Cardiac visceral fat is defined as the fatty tissues surrounding the heart region, enclosed by the lungs and posterior to the sternum. It is measured by constraining the heart region with an Anatomy Label Map that contains robust segmentations of the lungs and other major organs and estimating the fatty tissue within this region. The algorithm was evaluated on 124 low-dose and 223 standard-dose non-contrast chest CT scans from two public datasets. Based on visual inspection, 343 cases had good cardiac visceral fat segmentation. For quantitative evaluation, manual markings of cardiac visceral fat regions were made in 3 image slices for 45 low-dose scans and the Dice similarity coefficient (DSC) was computed. The automated algorithm achieved an average DSC of 0.93. Cardiac visceral fat volume (CVFV), heart region volume (HRV) and their ratio were computed for each case. The correlation between cardiac visceral fat measurement and coronary artery and aortic calcification was also evaluated. Results indicated the automated algorithm for measuring cardiac visceral fat volume may be an alternative method to the traditional manual assessment of thoracic region fat content in the assessment of cardiovascular disease risk.

Diabetic retinopathy is a condition of the eye of diabetic patients where the retina is damaged because of long-term diabetes. The condition deteriorates towards irreversible blindness in extreme cases of diabetic retinopathy. Hence, early detection of diabetic retinopathy is important to prevent blindness. Regular screening of fundus images of diabetic patients could be helpful in preventing blindness caused by diabetic retinopathy. In this paper, we propose techniques for staging of diabetic retinopathy in fundus images using several shape and texture features computed from detected microaneurysms, exudates, and hemorrhages. The classification accuracy is reported in terms of the area (A_z) under the receiver operating characteristic curve using 200 fundus images from the MESSIDOR database. The value of A_z for classifying normal images versus mild, moderate, and severe nonproliferative diabetic retinopathy (NPDR) is 0:9106. The value of A_z for classification of mild NPDR versus moderate and severe NPDR is 0:8372. The A_z value for classification of moderate NPDR and severe NPDR is 0:9750.

The segmentation of whole breast serves as the first step towards automated breast lesion detection. It is also necessary for automatically assessing the breast density, which is considered to be an important risk factor for breast cancer. In this paper we present a fully automated algorithm to segment the whole breast in low-dose chest CT images (LDCT), which has been recommended as an annual lung cancer screening test. The automated whole breast segmentation and potential breast density readings as well as lesion detection in LDCT will provide useful information for women who have received LDCT screening, especially the ones who have not undergone mammographic screening, by providing them additional risk indicators for breast cancer with no additional radiation exposure. The two main challenges to be addressed are significant range of variations in terms of the shape and location of the breast in LDCT and the separation of pectoral muscles from the glandular tissues. The presented algorithm achieves robust whole breast segmentation using an anatomy directed rule-based method. The evaluation is performed on 20 LDCT scans by comparing the segmentation with ground truth manually annotated by a radiologist on one axial slice and two sagittal slices for each scan. The resulting average Dice coefficient is 0.880 with a standard deviation of 0.058, demonstrating that the automated segmentation algorithm achieves results consistent with manual annotations of a radiologist.

Blood vessels are a major cause of false positives in computer aided detection systems for the detection of breast cancer. Therefore, the purpose of this study is to construct a framework for the segmentation of blood vessels in screening mammograms. The proposed framework is based on supervised learning using a cascade classifier. This cascade classifier consists of several stages where in each stage a GentleBoost classifier is trained on Haar-like features. A total of 30 cases were included in this study. In each image, vessel pixels were annotated by selecting pixels on the centerline of the vessel, control samples were taken by annotating a region without any visible vascular structures. This resulted in a total of 31,000 pixels marked as vascular and over 4 million control pixels. After training, the classifier assigns a vesselness likelihood to the pixels. The proposed framework was compared to three other vessel enhancing methods, i) a vesselness filter, ii) a gaussian derivative filter, and iii) a tubeness filter. The methods were compared in terms of area under the receiver operating characteristics curves, the Az values. The Az value of the cascade approach is 0:85. This is superior to the vesselness, Gaussian, and tubeness methods, with Az values of 0:77, 0:81, and 0:78, respectively. From these results, it can be concluded that our proposed framework is a promising method for the detection of vessels in screening mammograms.

Digital breast tomosynthesis (DBT) has the potential to replace digital mammography (DM) for breast cancer screening. An effective computer-aided detection (CAD) system for microcalcification clusters (MCs) on DBT will facilitate the transition. In this study, we collected a data set with corresponding DBT and DM for the same breasts. DBT was acquired with IRB approval and informed consent using a GE GEN2 DBT prototype system. The DM acquired with a GE Essential system for the patient’s clinical care was collected retrospectively from patient files. DM-based CAD (CAD_DM) and DBT-based CAD (CAD_DBT) were previously developed by our group. The major differences between the CAD systems include: (a) CAD_DBT uses two parallel processes whereas CAD_DM uses a single process for enhancing MCs and removing the structured background, (b) CAD_DBT has additional processing steps to reduce the false positives (FPs), including ranking of candidates of cluster seeds and cluster members and the use of adaptive CNR and size thresholds at clustering and FP reduction, (c) CAD_DM uses convolution neural network (CNN) and linear discriminant analysis (LDA) to differentiate true microcalcifications from FPs based on their morphological and CNN features. The performance difference is assessed by FROC analysis using test set (100 views with MCs and 74 views without MCs) independent of their respective training sets. At sensitivities of 70% and 80%, CAD_DBT achieved FP rates of 0.78 and 1.57 per view compared to 0.66 and 2.10 per image for the CAD_DM. JAFROC showed no significant difference between MC detection on DM and DBT by the two CAD systems.

Digital breast tomosynthesis (DBT) addresses limitations of 2-D projection imaging for detection of masses. Microcalcification clusters may be more difficult to appreciate in DBT as individual calcifications within clusters may appear on different slices. This research aims to evaluate the performance of ImageChecker 3D Calc CAD v1.0. Women were recruited as part of the TOMMY trial. From the trial, 169 were included in this study. The DBT images were processed with the computer aided detection (CAD) algorithm. Three consultant radiologists reviewed the images and recorded whether CAD prompts were on or off target. 79/80 (98.8%) malignant cases had a prompt on the area of microcalcification. In these cases, there were 1-15 marks (median 5) with the majority of false prompts (n=326/431) due to benign (68%) and vascular (24%) calcifications. Of 89 normal/benign cases, there were 1-13 prompts (median 3), 27 (30%) had no prompts and the majority of false prompts (n=238) were benign (77%) calcifications. CAD is effective in prompting malignant microcalcification clusters and may overcome the difficulty of detecting clusters in slice images. Although there was a high rate of false prompts, further advances in the software may improve specificity.

Breast magnetic resonance imaging (MRI) is recommended as an adjunct to mammography for women who are considered at elevated risk of developing breast cancer. As a key component of breast MRI, dynamic contrast-enhanced MRI (DCE-MRI) uses a contrast agent to provide high intensity contrast between breast tissues, making it sensitive to tissue composition and vascularity. Breast DCE-MRI characterizes certain physiologic properties of breast tissue that are potentially related to breast cancer risk. Studies have shown that increased background parenchymal enhancement (BPE), which is the contrast enhancement occurring in normal cancer-unaffected breast tissues in post-contrast sequences, predicts increased breast cancer risk. Signal enhancement ratio (SER) computed from pre-contrast and post-contrast sequences in DCE-MRI measures change in signal intensity due to contrast uptake over time and is a measure of contrast enhancement kinetics. SER quantified in breast tumor has been shown potential as a biomarker for characterizing tumor response to treatments. In this work we investigated the relationship between quantitative measures of SER and breast cancer risk. A pilot retrospective case-control study was performed using a cohort of 102 women, consisting of 51 women who had diagnosed with unilateral breast cancer and 51 matched controls (by age and MRI date) with a unilateral biopsy-proven benign lesion. SER was quantified using fully-automated computerized algorithms and three SER-derived quantitative volume measures were compared between the cancer cases and controls using logistic regression analysis. Our preliminary results showed that SER is associated with breast cancer risk, after adjustment for the Breast Imaging Reporting and Data System (BI-RADS)-based mammographic breast density measures. This pilot study indicated that SER has potential for use as a risk factor for breast cancer risk assessment in women at elevated risk of developing breast cancer.

Many medical centers offer digital breast tomosynthesis (DBT) and 2D digital mammography acquired under the same compression (i.e., “Combo” examination) for screening. This paper compares a conventional 2D CAD algorithm (Hologic® ImageChecker® CAD v9.4) for calcification detection against a prototype 3D algorithm (Hologic® ImageChecker® 3D Calc CAD v1.0). Due to the newness of DBT, the development of this 3D CAD algorithm is ongoing, and it is currently not FDA-approved in the United States. For this study, DBT screening cases with suspicious calcifications were identified retrospectively at the University of Pennsylvania. An expert radiologist (E.F.C.) reviewed images with both 2D and DBT CAD marks, and compared the marks to biopsy results. Control cases with one-year negative follow-up were also studied; these cases either possess clearly benign calcifications or lacked calcifications. To allow the user to alter the sensitivity for cancer detection, an operating point is assigned to each CAD mark. As expected from conventional 2D CAD, increasing the operating point in 3D CAD increases sensitivity and reduces specificity. Additionally, we showed that some cancers are occult to 2D CAD at all operating points. By contrast, 3D CAD allows for detection of some cancers that are missed on 2D CAD. We also demonstrated that some non-cancerous CAD marks in 3D are not present at analogous locations in the 2D image. Hence, there are additional marks when using both 2D and 3D CAD in combination, leading to lower specificity than with conventional 2D CAD alone.

Accurate and efficient automated tumor segmentation in breast dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) is highly desirable for computer-aided tumor diagnosis. We propose a novel automatic segmentation framework which incorporates mean-shift smoothing, superpixel-wise classification, pixel-wise graph-cuts partitioning, and morphological refinement. A set of 15 breast DCE-MR images, obtained from the American College of Radiology Imaging Network (ACRIN) 6657 I-SPY trial, were manually segmented to generate tumor masks (as ground truth) and breast masks (as regions of interest). Four state-of-the-art segmentation approaches based on diverse models were also utilized for comparison. Based on five standard evaluation metrics for segmentation, the proposed framework consistently outperformed all other approaches. The performance of the proposed framework was: 1) 0.83 for Dice similarity coefficient, 2) 0.96 for pixel-wise accuracy, 3) 0.72 for VOC score, 4) 0.79 mm for mean absolute difference, and 5) 11.71 mm for maximum Hausdorff distance, which surpassed the second best method (i.e., adaptive geodesic transformation), a semi-automatic algorithm depending on precise initialization. Our results suggest promising potential applications of our segmentation framework in assisting analysis of breast carcinomas.

This paper proposes a connection method of separated luminal regions of the intestine for Crohn's disease diagnosis. Crohn's disease is an inflammatory disease of the digestive tract. Capsule or conventional endoscopic diagnosis is performed for Crohn's disease diagnosis. However, parts of the intestines may not be observed in the endoscopic diagnosis if intestinal stenosis occurs. Endoscopes cannot pass through the stenosed parts. CT image-based diagnosis is developed as an alternative choice of the Crohn's disease. CT image-based diagnosis enables physicians to observe the entire intestines even if stenosed parts exist. CAD systems for Crohn's disease using CT volumes are recently developed. Such CAD systems need to reconstruct separated luminal regions of the intestines to analyze intestines. We propose a connection method of separated luminal regions of the intestines segmented from CT volumes. The luminal regions of the intestines are segmented from a CT volume. The centerlines of the luminal regions are calculated by using a thinning process. We enumerate all the possible sequences of the centerline segments. In this work, we newly introduce a condition using distance between connected ends points of the centerline segments. This condition eliminates unnatural connections of the centerline segments. Also, this condition reduces processing time. After generating a sequence list of the centerline segments, the correct sequence is obtained by using an evaluation function. We connect the luminal regions based on the correct sequence. Our experiments using four CT volumes showed that our method connected 6.5 out of 8.0 centerline segments per case. Processing times of the proposed method were reduced from the previous method.

CT colonography (CTC) uses orally administered fecal-tagging agents to enhance retained fluid and feces that would otherwise obscure or imitate polyps on CTC images. To visualize the complete region of colon without residual materials, electronic cleansing (EC) can be used to perform virtual subtraction of the tagged materials from CTC images. However, current EC methods produce subtraction artifacts and they can fail to subtract unclearly tagged feces. We developed a novel multi-material EC (MUMA-EC) method that uses dual-energy CTC (DE-CTC) and machine-learning methods to improve the performance of EC. In our method, material decomposition is performed to calculate wateriodine decomposition images and virtual monochromatic (VIM) images. Using the images, a random forest classifier is used to label the regions of lumen air, soft tissue, fecal tagging, and their partial-volume boundaries. The electronically cleansed images are synthesized from the multi-material and VIM image volumes. For pilot evaluation, we acquired the clinical DE-CTC data of 7 patients. Preliminary results suggest that the proposed MUMA-EC method is effective and that it minimizes the three types of image artifacts that were present in previous EC methods.

Feature classification plays an important role in computer-aided diagnosis (CADx) of suspicious lesions or polyps in this concerned study. As one of the simplest machine learning algorithms, the k-nearest neighbor (k-NN) classifier has been widely used in many classification problems. However, the k-NN classifier has a drawback that the majority classes will dominate the prediction of a new sample. To mitigate this drawback, efforts have been devoted to set weight on each neighbor to avoid the influence of the “majority” classes. As a result, various weighted or wk-NN strategies have been explored. In this paper, we explored an alternative strategy, called “distance weighted inside disc” (DWID) classifier, which is different from the k-NN and wk-NN by such a way that it classifies the test point by assigning a corresponding label (instead a weight) with consideration of only those points inside the disc whose center is the test point instead of the k-nearest points. We evaluated this new DWID classifier with comparison to the k-NN, wk-NN, support vector machine (SVM) and random forest (RF) classifiers by experiments on a database of 153 polyps, including 116 neoplastic (malignance) polyps and 37 hyperplastic (benign) polyps, in terms of CADx or differentiation of benign from malignancy. The evaluation outcomes were documented quantitatively by the Receiver Operating Characteristics (ROC) analysis and the merit of area under the ROC curve (AUC), which is a well-established evaluation criterion to various classifiers. The results showed noticeable gain on the polyp differentiation by this new classifier according to the AUC values, as compared to the k-NN and wk-NN, as well as the SVM and RF. In the meantime, this new classifier also showed a noticeable reduction of computing time.

In computer-aided diagnosis (CAD) systems for prostate cancer, dynamic contrast enhanced (DCE) magnetic resonance imaging is useful for distinguishing cancerous and benign tissue. The Tofts physiological model is a commonly used representation of the DCE image data, but the parameters require extensive computation. Hence, we developed an alternative representation based on the Hilbert transform of the DCE images. The time maximum of the Hilbert transform, a binary metric of early enhancement, and a pre-DCE value was assigned to each voxel and appended to a standard feature set derived from T2-weighted images and apparent diffusion coefficient maps. A cohort of 40 patients was used for training the classifier, and 20 patients were used for testing. The AUC was calculated by pooling the voxel-wise prediction values and comparing with the ground truth. The resulting AUC of 0.92 (95% CI [0.87 0.97]) is not significantly different from an AUC calculated using Tofts physiological models of 0.92 (95% CI [0.87 0.97]), as validated by a Wilcoxon signed rank test on each patient’s AUC (p = 0.19). The time required for calculation and feature extraction is 11.39 seconds (95% CI [10.95 11.82]) per patient using the Hilbert-based feature set, two orders of magnitude faster than the 1319 seconds (95% CI [1233 1404]) required for the Tofts parameter-based feature set (p<0.001). Hence, the features proposed herein appear useful for CAD systems integrated into clinical workflows where efficiency is important.

In prostate cancer, the detection of metastatic lymph nodes indicates progression from localized disease to metastasized cancer. The detection of positive lymph nodes is, however, a complex and time consuming task for experienced radiologists. Assistance of a two-stage Computer-Aided Detection (CAD) system in MR Lymphography (MRL) is not yet feasible due to the large number of false positives in the first stage of the system. By introducing a multi-structure, multi-atlas segmentation, using an affine transformation followed by a B-spline transformation for registration, the organ location is given by a mean density probability map. The atlas segmentation is semi-automatically drawn with ITK-SNAP, using Active Contour Segmentation. Each anatomic structure is identified by a label number. Registration is performed using Elastix, using Mutual Information and an Adaptive Stochastic Gradient optimization. The dataset consists of the MRL scans of ten patients, with lymph nodes manually annotated in consensus by two expert readers. The feature map of the CAD system consists of the Multi-Atlas and various other features (e.g. Normalized Intensity and multi-scale Blobness). The voxel-based Gentleboost classifier is evaluated using ROC analysis with cross validation. We show in a set of 10 studies that adding multi-structure, multi-atlas anatomical structure likelihood features improves the quality of the lymph node voxel likelihood map. Multiple structure anatomy maps may thus make MRL CAD more feasible.

In this work, we examine the strength of deep learning approaches for pathology detection in chest radiograph data. Convolutional neural networks (CNN) deep architecture classification approaches have gained popularity due to their ability to learn mid and high level image representations. We explore the ability of a CNN to identify different types of pathologies in chest x-ray images. Moreover, since very large training sets are generally not available in the medical domain, we explore the feasibility of using a deep learning approach based on non-medical learning. We tested our algorithm on a dataset of 93 images. We use a CNN that was trained with ImageNet, a well-known large scale nonmedical image database. The best performance was achieved using a combination of features extracted from the CNN and a set of low-level features. We obtained an area under curve (AUC) of 0.93 for Right Pleural Effusion detection, 0.89 for Enlarged heart detection and 0.79 for classification between healthy and abnormal chest x-ray, where all pathologies are combined into one large class. This is a first-of-its-kind experiment that shows that deep learning with large scale non-medical image databases may be sufficient for general medical image recognition tasks.

Medical images contain a large amount of visual information about structures and anomalies in the human body. To make sense of this information, human interpretation is often essential. On the other hand, computer-based approaches can exploit information contained in the images by numerically measuring and quantifying specific visual features. Annotation of organs and other anatomical regions is an important step before computing numerical features on medical images. In this paper, a texture-based organ classification algorithm is presented, which can be used to reduce the time required for annotating medical images. The texture of organs is analyzed using a combination of state-of-the-art techniques: the Riesz transform and a bag of meaningful visual words. The effect of a meaningfulness transformation in the visual word space yields two important advantages that can be seen in the results. The number of descriptors is enormously reduced down to 10% of the original size, whereas classification accuracy is improved by up to 25% with respect to the baseline approach.

Cervical cancer is a leading most common type of cancer for women worldwide. Existing screening programs for cervical cancer suffer from low sensitivity. Using images of the cervix (cervigrams) as an aid in detecting pre-cancerous changes to the cervix has good potential to improve sensitivity and help reduce the number of cervical cancer cases. In this paper, we present a method that utilizes multi-modality information extracted from multiple tests of a patient’s visit to classify the patient visit to be either low-risk or high-risk. Our algorithm integrates image features and text features to make a diagnosis. We also present two strategies to estimate the missing values in text features: Image Classifier Supervised Mean Imputation (ICSMI) and Image Classifier Supervised Linear Interpolation (ICSLI). We evaluate our method on a large medical dataset and compare it with several alternative approaches. The results show that the proposed method with ICSLI strategy achieves the best result of 83.03% specificity and 76.36% sensitivity. When higher specificity is desired, our method can achieve 90% specificity with 62.12% sensitivity.

We present a method for the automatic delineation of two neuromelanin rich brainstem structures –substantia nigra pars compacta (SN) and locus coeruleus (LC)- in neuromelanin sensitive magnetic resonance images of the brain. The segmentation method uses a dynamic multi-image reference atlas and a pre-registration atlas selection strategy. To create the atlas, a pool of 35 images of healthy subjects was pair-wise pre-registered and clustered in groups using an affinity propagation approach. Each group of the atlas is represented by a single exemplar image. Each new target image to be segmented is registered to the exemplars of each cluster. Then all the images of the highest performing clusters are enrolled into the final atlas, and the results of the registration with the target image are propagated using a majority voting approach. All registration processes used combined one two-stage affine and one elastic B-spline algorithm, to account for global positioning, region selection and local anatomic differences. In this paper, we present the algorithm, with emphasis in the atlas selection method and the registration scheme. We evaluate the performance of the atlas selection strategy using 35 healthy subjects and 5 Parkinson’s disease patients. Then, we quantified the volume and contrast ratio of neuromelanin signal of these structures in 47 normal subjects and 40 Parkinson’s disease patients to confirm that this method can detect neuromelanin-containing neurons loss in Parkinson’s disease patients and could eventually be used for the early detection of SN and LC damage.

Automatic Anatomy Recognition (AAR) is a recently developed approach for the automatic whole body wide organ segmentation. We previously tested that methodology on image cases with some pathology where the organs were not distorted significantly. In this paper, we present an advancement of AAR to handle organs which may have been modified or resected by surgical intervention. We focus on MRI of the neck in pediatric Obstructive Sleep Apnea Syndrome (OSAS). The proposed method consists of an AAR step followed by support vector machine techniques to detect the presence/absence of organs. The AAR step employs a hierarchical organization of the organs for model building. For each organ, a fuzzy model over a population is built. The model of the body region is then described in terms of the fuzzy models and a host of other descriptors which include parent to offspring relationship estimated over the population. Organs are recognized following the organ hierarchy by using an optimal threshold based search. The SVM step subsequently checks for evidence of the presence of organs. Experimental results show that AAR techniques can be combined with machine learning strategies within the AAR recognition framework for good performance in recognizing missing organs, in our case missing tonsils in post-tonsillectomy images as well as in simulating tonsillectomy images. The previous recognition performance is maintained achieving an organ localization accuracy of within 1 voxel when the organ is actually not removed. To our knowledge, no methods have been reported to date for handling significantly deformed or missing organs, especially in neck MRI.

We propose a novel non-invasive brain tumor type classification using Multi-fractal Detrended Fluctuation Analysis (MFDFA) [1] in structural magnetic resonance (MR) images. This preliminary work investigates the efficacy of the MFDFA features along with our novel texture feature known as multifractional Brownian motion (mBm) [2] in classifying (grading) brain tumors as High Grade (HG) and Low Grade (LG). Based on prior performance, Random Forest (RF) [3] is employed for tumor grading using two different datasets such as BRATS-2013 [4] and BRATS-2014 [5]. Quantitative scores such as precision, recall, accuracy are obtained using the confusion matrix. On an average 90% precision and 85% recall from the inter-dataset cross-validation confirm the efficacy of the proposed method.

Cerebral small vessel disease (SVD) is a common finding on magnetic resonance images of elderly people. White matter lesions (WML) are important markers for not only the small vessel disease, but also neuro-degenerative diseases including multiple sclerosis, Alzheimer’s disease and vascular dementia. Volumetric measurements such as the “total lesion load”, have been studied and related to these diseases. With respect to SVD we conjecture that small lesions are important, as they have been observed to grow over time and they form the majority of lesions in number. To study these small lesions they need to be annotated, which is a complex and time-consuming task. Existing (semi) automatic methods have been aimed at volumetric measurements and large lesions, and are not suitable for the detection of small lesions. In this research we established a supervised voxel classification CAD system, optimized and trained to exclusively detect small WMLs. To achieve this, several preprocessing steps were taken, which included a robust standardization of subject intensities to reduce inter-subject intensity variability as much as possible. A number of features that were found to be well identifying small lesions were calculated including multimodal intensities, tissue probabilities, several features for accurate location description, a number of second order derivative features as well as multi-scale annular filter for blobness detection. Only small lesions were used to learn the target concept via Adaboost using random forests as its basic classifiers. Finally the results were evaluated using Free-response receiver operating characteristic.

Assessment of the extent of cerebral damage on admission in patients with acute ischemic stroke could play an important role in treatment decision making. Computed tomography perfusion (CTP) imaging can be used to determine the extent of damage. However, clinical application is hindered by differences among vendors and used methodology. As a result, threshold based methods and visual assessment of CTP images has not yet shown to be useful in treatment decision making and predicting clinical outcome. Preliminary results in MR studies have shown the benefit of using supervised classifiers for predicting tissue outcome, but this has not been demonstrated for CTP. We present a novel method for the automatic prediction of tissue outcome by combining multi-parametric CTP images into a tissue outcome probability map. A supervised classification scheme was developed to extract absolute and relative perfusion values from processed CTP images that are summarized by a trained classifier into a likelihood of infarction. Training was performed using follow-up CT scans of 20 acute stroke patients with complete recanalization of the vessel that was occluded on admission. Infarcted regions were annotated by expert neuroradiologists. Multiple classifiers were evaluated in a leave-one-patient-out strategy for their discriminating performance using receiver operating characteristic (ROC) statistics. Results showed that a RandomForest classifier performed optimally with an area under the ROC of 0.90 for discriminating infarct tissue. The obtained results are an improvement over existing thresholding methods and are in line with results found in literature where MR perfusion was used.

The thyroid gland plays an important role in clinical practice, especially for radiation therapy treatment planning. For patients with head and neck cancer, radiation therapy requires a precise delineation of the thyroid gland to be spared on the pre-treatment planning CT images to avoid thyroid dysfunction. In the current clinical workflow, the thyroid gland is normally manually delineated by radiologists or radiation oncologists, which is time consuming and error prone. Therefore, a system for automated segmentation of the thyroid is desirable. However, automated segmentation of the thyroid is challenging because the thyroid is inhomogeneous and surrounded by structures that have similar intensities. In this work, the thyroid gland segmentation is initially estimated by multi-atlas label fusion algorithm. The segmentation is refined by supervised statistical learning based voxel labeling with a random forest algorithm. Multiatlas label fusion (MALF) transfers expert-labeled thyroids from atlases to a target image using deformable registration. Errors produced by label transfer are reduced by label fusion that combines the results produced by all atlases into a consensus solution. Then, random forest (RF) employs an ensemble of decision trees that are trained on labeled thyroids to recognize features. The trained forest classifier is then applied to the thyroid estimated from the MALF by voxel scanning to assign the class-conditional probability. Voxels from the expert-labeled thyroids in CT volumes are treated as positive classes; background non-thyroid voxels as negatives. We applied this automated thyroid segmentation system to CT scans of 20 patients. The results showed that the MALF achieved an overall 0.75 Dice Similarity Coefficient (DSC) and the RF classification further improved the DSC to 0.81.

Alzheimer's disease (AD) is one of the most frequent forms of dementia and an increasing challenging public health problem. In the last two decades, structural magnetic resonance imaging (MRI) has shown potential in distinguishing patients with Alzheimer's disease and elderly controls (CN). To obtain AD-specific biomarkers, previous research used either statistical testing to find statistically significant different regions between the two clinical groups, or ℓ₁ sparse learning to select isolated features in the image domain. In this paper, we propose a new framework that uses structural MRI to simultaneously distinguish the two clinical groups and find the bio-markers of AD, using a group lasso support vector machine (SVM). The group lasso term (mixed ℓ₁- ℓ₂ norm) introduces anatomical information from the image domain into the feature domain, such that the resulting set of selected voxels are more meaningful than the ℓ₁ sparse SVM. Because of large inter-structure size variation, we introduce a group specific normalization factor to deal with the structure size bias. Experiments have been performed on a well-designed AD vs. CN dataset¹ to validate our method. Comparing to the ℓ₁ sparse SVM approach, our method achieved better classification performance and a more meaningful biomarker selection. When we vary the training set, the selected regions by our method were more stable than the ℓ₁ sparse SVM. Classification experiments showed that our group normalization lead to higher classification accuracy with fewer selected regions than the non-normalized method. Comparing to the state-of-art AD vs. CN classification methods, our approach not only obtains a high accuracy with the same dataset, but more importantly, we simultaneously find the brain anatomies that are closely related to the disease.

The ability to identify aggressive tumors from indolent tumors using quantitative analysis on dynamic contrast enhanced magnetic resonance imaging (DCE-MRI) would dramatically change the breast cancer treatment paradigm. With this prognostic information, patients with aggressive tumors that have the ability to spread to distant sites outside of the breast could be selected for more aggressive treatment and surveillance regimens. Conversely, patients with tumors that do not have the propensity to metastasize could be treated less aggressively, avoiding some of the morbidity associated with surgery, radiation and chemotherapy. We propose a computer aided detection framework to determine which breast cancers will metastasize to the loco-regional lymph nodes as well as which tumors will eventually go on to develop distant metastses using quantitative image analysis and radiomics. We defined a new contrast based tumor habitat and analyzed textural kinetic features from this habitat for classification purposes. The proposed tumor habitat, which we call combined-habitat, is derived from the intersection of two individual tumor sub-regions: one that exhibits rapid initial contrast uptake and the other that exhibits rapid delayed contrast washout. Hence the combined-habitat represents the tumor sub-region within which the pixels undergo both rapid initial uptake and rapid delayed washout. We analyzed a dataset of twenty-seven representative two dimensional (2D) images from volumetric DCE-MRI of breast tumors, for classification of tumors with no lymph nodes from tumors with positive number of axillary lymph nodes. For this classification an accuracy of 88.9% was achieved. Twenty of the twenty-seven patients were analyzed for classification of distant metastatic tumors from indolent cancers (tumors with no lymph nodes), for which the accuracy was 84.3%.

We recently investigated a new mammographic image feature based risk factor to predict near-term breast cancer risk after a woman has a negative mammographic screening. We hypothesized that unlike the conventional epidemiology-based long-term (or lifetime) risk factors, the mammographic image feature based risk factor value will increase as the time lag between the negative and positive mammography screening decreases. The purpose of this study is to test this hypothesis. From a large and diverse full-field digital mammography (FFDM) image database with 1278 cases, we collected all available sequential FFDM examinations for each case including the “current” and 1 to 3 most recently “prior” examinations. All “prior” examinations were interpreted negative, and “current” ones were either malignant or recalled negative/benign. We computed 92 global mammographic texture and density based features, and included three clinical risk factors (woman’s age, family history and subjective breast density BIRADS ratings). On this initial feature set, we applied a fast and accurate Sequential Forward Floating Selection (SFFS) feature selection algorithm to reduce feature dimensionality. The features computed on both mammographic views were individually/ separately trained using two artificial neural network (ANN) classifiers. The classification scores of the two ANNs were then merged with a sequential ANN. The results show that the maximum adjusted odds ratios were 5.59, 7.98, and 15.77 for using the 3rd, 2nd, and 1st “prior” FFDM examinations, respectively, which demonstrates a higher association of mammographic image feature change and an increasing risk trend of developing breast cancer in the near-term after a negative screening.

High overall mammographic density is associated with both an increased risk of developing breast cancer and the risk of cancer being masked. We compared local density at cancer sites in diagnostic images with corresponding previous screening mammograms (priors), and matched controls. Volpara^TM density maps were obtained for 54 mammograms showing unilateral breast cancer and their priors which had been previously read as normal. These were each matched to 3 controls on age, menopausal status, hormone replacement therapy usage, body mass index and year of prior. Local percent density was computed in 15mm square regions at lesion sites and similar locations in the corresponding images. Conditional logistic regression was used to predict case-control status. In diagnostic and prior images, local density was increased at the lesion site compared with the opposite breast (medians 21.58%, 9.18%, p<0.001 diagnostic; 18.82%, 9.45%, p <0.001 prior). Women in the highest tertile of local density in priors were more likely to develop cancer than those in the lowest tertile (OR 42.09, 95% CI 5.37-329.94). Those in the highest tertile of Volpara^TM gland volume were also more likely to develop cancer (OR 2.89, 95% CI 1.30-6.42). Local density is increased where cancer will develop compared with corresponding regions in the opposite breast and matched controls, and its measurement could enhance computer-aided mammography.

Computer-aided detection (CAD) has the potential to aid radiologists in detection of microcalcification clusters (MCs). CAD for digital breast tomosynthesis (DBT) can be developed by using the reconstructed volume, the projection views or other derivatives as input. We have developed a novel method of generating a single planar projection (PPJ) image from a regularized DBT volume to emphasize the high contrast objects such as microcalcifications while removing the anatomical background and noise. In this work, we adapted a CAD system developed for digital mammography (CAD_DM) to the PPJ image and compared its performance with our CAD system developed for DBT volumes (CAD_DBT) in the same set of cases. For microcalcification detection in the PPJ image using the CAD_DM system, the background removal preprocessing step designed for DM was not needed. The other methods and processing steps in the CAD_DM system were kept without modification while the parameters were optimized with a training set. The linear discriminant analysis classifier using cluster based features was retrained to generate a discriminant score to be used as decision variable. For view-based FROC analysis, at 80% sensitivity, an FP rate of 1.95/volume and 1.54/image were achieved, respectively, for CAD_DBT and CAD_DM in an independent test set. At a threshold of 1.2 FPs per image or per DBT volume, the nonparametric analysis of the area under the FROC curve shows that the optimized CADDM for PPJ is significantly better than CADDBT. However, the performance of CADDM drops at higher sensitivity or FP rate, resulting in similar overall performance between the two CAD systems. The higher sensitivity of the CAD_DM in the low FP rate region and vice versa for the CAD_DBT indicate that a joint CAD system combining detection in the DBT volume and the PPJ image has the potential to increase the sensitivity and reduce the FP rate.

In this study, a novel mass detection framework that utilizes the information from synthetic mammograms has been developed for detecting masses in digital breast tomosynthesis (DBT). In clinical study, it is demonstrated that the combination of DBT and full field digital mammography (FFDM) increases the reader performance. To reduce the radiation dose in this approach, synthetic mammogram has been developed in previous researches and it is demonstrated that synthetic mammogram can alternate the FFDM when it is used with DBT. In this study, we investigate the feasibility of the combined approach of DBT and synthetic mammogram in point of computer-aided detection (CAD). As a synthetic mammogram, two-dimensional image was generated by adopting conspicuous voxels of three-dimensional DBT volume in our study. The mass likelihood scores estimated for each mass candidates in synthetic mammogram and DBT are merged to differentiate masses and false positives (FPs) in combined approach. We compared the performance of detecting masses in the proposed combined approach and DBT alone. A clinical data set of 196 DBT volumes was used to evaluate the different detection schemes. The combined approach achieved sensitivity of 80% and 89% with 1.16 and 2.37 FPs per DBT volume. The DBT alone approach achieved same sensitivities with 1.61 and 3.46 FPs per DBT volume. Experimental results show that statistically significant improvement (p = 0.002) is achieved in combined approach compared to DBT alone. These results imply that the information fusion of synthetic mammogram and DBT is a promising approach to detect masses in DBT.

Prostate cancer is the most prevalently diagnosed and the second cause of cancer-related death in North American men. Several approaches have been proposed to augment detection of prostate cancer using different imaging modalities. Due to advantages of ultrasound imaging, these approaches have been the subject of several recent studies. This paper presents the results of a feasibility study on differentiating between lower and higher grade prostate cancer using ultrasound RF time series data. We also propose new spectral features of RF time series to highlight aggressive prostate cancer in small ROIs of size 1 mm × 1 mm in a cohort of 19 ex vivo specimens of human prostate tissue. In leave-one-patient-out cross-validation strategy, an area under accumulated ROC curve of 0.8 has been achieved with overall sensitivity and specificity of 81% and 80%, respectively. The current method shows promising results on differentiating between lower and higher grade of prostate cancer using ultrasound RF time series.

The common practice for biopsy guidance is through transrectal ultrasound, with the fusion of ultrasound and MRI-based targets when available. However, ultrasound is only used as a guidance modality in MR-targeted ultrasound-guided biopsy, even though previous work has shown the potential utility of ultrasound, particularly ultrasound vibro-elastography, as a tissue typing approach. We argue that multiparametric ultrasound, which includes B-mode and vibro-elastography images, could contain information that is not captured using multiparametric MRI (mpMRI) and therefore play a role in refining the biopsy and treatment strategies. In this work, we combine mpMRI with multiparametric ultrasound features from registered tissue areas to examine the potential improvement in cancer detection. All the images were acquired prior to radical prostatectomy and cancer detection was validated based on 36 whole mount histology slides. We calculated a set of 24 texture features from vibro-elastography and B-mode images, and five features from mpMRI. Then we used recursive feature elimination (RFE) and sparse regression through LASSO to find an optimal set of features to be used for tissue classification. We show that the set of these selected features increases the area under ROC curve from 0.87 with mpMRI alone to 0.94 with the selected mpMRI and multiparametric ultrasound features, when used with support vector machine classification on features extracted from peripheral zone. For features extracted from the whole-gland, the area under the curve was 0.75 and 0.82 for mpMRI and mpMRI along with ultrasound, respectively. These preliminary results provide evidence that ultrasound and ultrasound vibro-elastography could be used as modalities for improved cancer detection in combination with MRI.

Prostate cancer (PCa) has a high 10-year recurrence rate, making PCa the second leading cause of cancer-specific mortality among men in the USA. PCa recurrences are often predicted by assessing the status of surgical margins (SM) with positive surgical margins (PSM) increasing the chances of biochemical recurrence by 2-4 times. To this end, an SM assessment system using Electrical Impedance Spectroscopy (EIS) was developed with a microendoscopic probe. This system measures the tissue bioimpedance over a range of frequencies (1 kHz to 1MHz), and computes a Composite Impedance Metric (CIM). CIM can be used to classify tissue as benign or cancerous. The system was used to collect the impedance spectra from excised prostates, which were obtained from men undergoing radical prostatectomy. The data revealed statistically significant (p<0.05) differences in the impedance properties of the benign and tumorous tissues, and between different tissue morphologies. To visualize the results of SM-assessment, a visualization tool using da Vinci stereo laparoscope is being developed. Together with the visualization tool, the EIS-based SM assessment system can be potentially used to intraoperatively classify tissues and display the results on the surgical console with a video feed of the surgical site, thereby augmenting a surgeon’s view of the site and providing a potential solution to the intraoperative SM assessment needs.

The use of magnetic resonance enterography (MRE) has become a mainstay in the evaluation, assessment and follow up of inflammatory bowel diseases, such as Crohn’s disease (CD), thanks to its high image quality and its non-ionizing nature. In particular, the advent of faster MRE sequences less sensitive to image-motion artifacts offers the possibility to obtain visual, structural and functional information of the patient’s small bowel. However, the inherent subjectivity of the mere visual inspection of these images often hinders the accurate identification and monitoring of the pathological areas. In this paper, we present a framework that provides quantitative and objective motility information of the small bowel from free-breathing MRE dynamic sequences. After compensating for the breathing motion of the patient, we create personalized peristaltic activity maps via optical flow analysis. The result is the creation of a new set of images providing objective and precise functional information of the small bowel. The accuracy of the new method was also evaluated from two different perspectives: objective accuracy (1.1 ± 0.6 mm/s of error), i.e., the ability of the system to provide quantitative and accurate information about the motility of moving bowel landmarks, and subjective accuracy (avg. difference of 0.7 ± 0.7 in a range of 1 to 5), i.e., the degree of agreement with the subjective evaluation of an expert. Finally, the practical utility of the new method was successfully evaluated in a preliminary study with 32 studies of healthy and CD cases, showing its potential for the fast and accurate assessment and follow up of CD in the small bowel.

This paper proposes a content-based image retrieval method for optical colonoscopy images that can find images similar to ones being diagnosed. Optical colonoscopy is a method of direct observation for colons and rectums to diagnose bowel diseases. It is the most common procedure for screening, surveillance and treatment. However, diagnostic accuracy for intractable inflammatory bowel diseases, such as ulcerative colitis (UC), is highly dependent on the experience and knowledge of the medical doctor, because there is considerable variety in the appearances of colonic mucosa within inflammations with UC. In order to solve this issue, this paper proposes a content-based image retrieval method based on image recognition techniques. The proposed retrieval method can find similar images from a database of images diagnosed as UC, and can potentially furnish the medical records associated with the retrieved images to assist the UC diagnosis. Within the proposed method, color histogram features and higher order local auto-correlation (HLAC) features are adopted to represent the color information and geometrical information of optical colonoscopy images, respectively. Moreover, considering various characteristics of UC colonoscopy images, such as vascular patterns and the roughness of the colonic mucosa, we also propose an image enhancement method to highlight the appearances of colonic mucosa in UC. In an experiment using 161 UC images from 32 patients, we demonstrate that our method improves the accuracy of retrieving similar UC images.

Further improvement of computer-aided detection (CADe) of colonic polyps is vital to advance computed tomographic colonography (CTC) toward a screening modality, where the detection of flat polyps is especially challenging because limited image features can be extracted from flat polyps, and the traditional geometric features-based CADe methods usually fail to detect such polyps. In this paper, we present a novel pipeline to automatically detect initial polyp candidates (IPCs), especially flat polyps, from CTC images. First, the colon wall mucosa was extracted via a partial volume segmentation approach as a volumetric layer, where the inner border of colon wall can be obtained by shrinking the volumetric layer using level set based adaptive convolution. Then the outer border of colon wall (or the colon wall serosa) was segmented via a combined implementation of geodesic active contour and Mumford-Shah functional in a coarse-to-fine manner. Finally, the wall thickness was estimated along a unique path between the segmented inner and outer borders with consideration of the volumetric layers and was mapped onto a patient-specific three-dimensional (3D) colon wall model. The IPC detection results can usually be better visualized in a 2D image flattened from the 3D model, where abnormalities were detected by Z-score transformation of the thickness values. The proposed IPC detection approach was validated on 11 patients with 22 CTC scans, and each scan has at least one flat poly annotation. The above presented novel pipeline was effective to detect some flat polyps that were missed by our CADe system while keeping false detections in a relative low level. This preliminary study indicates that the presented pipeline can be incorporated into an existing CADe system to enhance the polyp detection power, especially for flat polyps.

A new measurement of the pulmonary artery diameter is obtained where the artery may be robustly segmented between the heart and the artery bifurcation. An automated algorithm is presented that can make this pulmonary artery measurement in low-dose non-contrast chest CT images. The algorithm uses a cylinder matching method following geometric constraints obtained from other adjacent organs that have been previously segmented. This new measurement and the related ratio of pulmonary artery to aortic artery measurement are compared to traditional manual approaches for pulmonary artery characterization. The algorithm was qualitatively evaluated on 124 low-dose and 223 standard-dose non-contrast chest CT scans from two public datasets; 324 out of the 347 cases had good segmentations and in the other 23 cases there was significant boundary inaccuracy. For quantitative evaluation, the comparison was to manually marked pulmonary artery boundary in an axial slice in 45 cases; the resulting average Dice Similarity Coefficient was 0.88 (max 0.95, min 0.74). For the 45 cases with manual markings, the correlation between the automated pulmonary artery to ascending aorta diameter ratio and manual ratio at pulmonary artery bifurcation level was 0.81. Using Bland-Altman analysis, the mean difference of the two ratios was 0.03 and the limits of agreement was (-0.12, 0.18). This automated measurement may have utility as an alternative to the conventional manual measurement of pulmonary artery diameter at the bifurcation level especially in the context of noisy low-dose CT images.

We are developing a computer-aided detection system to assist radiologists in detection of non-calcified plaques (NCPs) in coronary CT angiograms (cCTA). In this study, we performed quantitative analysis of arterial flow properties in each vessel branch and extracted flow information to differentiate the presence and absence of stenosis in a vessel segment. Under rest conditions, blood flow in a single vessel branch was assumed to follow Poiseuille’s law. For a uniform pressure distribution, two quantitative flow features, the normalized arterial compliance per unit length (C_u) and the normalized volumetric flow (Q) along the vessel centerline, were calculated based on the parabolic Poiseuille solution. The flow features were evaluated for a two-class classification task to differentiate NCP candidates obtained by prescreening as true NCPs and false positives (FPs) in cCTA. For evaluation, a data set of 83 cCTA scans was retrospectively collected from 83 patient files with IRB approval. A total of 118 NCPs were identified by experienced cardiothoracic radiologists. The correlation between the two flow features was 0.32. The discriminatory ability of the flow features evaluated as the area under the ROC curve (AUC) was 0.65 for C_u and 0.63 for Q in comparison with AUCs of 0.56-0.69 from our previous luminal features. With stepwise LDA feature selection, volumetric flow (Q) was selected in addition to three other luminal features. With FROC analysis, the test results indicated a reduction of the FP rates to 3.14, 1.98, and 1.32 FPs/scan at sensitivities of 90%, 80%, and 70%, respectively. The study indicated that quantitative blood flow analysis has the potential to provide useful features for the detection of NCPs in cCTA.

Age-related Macular Degeneration (AMD) is a common eye disorder with high prevalence in elderly people. The disease mainly affects the central part of the retina, and could ultimately lead to permanent vision loss. Optical Coherence Tomography (OCT) is becoming the standard imaging modality in diagnosis of AMD and the assessment of its progression. However, the evaluation of the obtained volumetric scan is time consuming, expensive and the signs of early AMD are easy to miss. In this paper we propose a classification method to automatically distinguish AMD patients from healthy subjects with high accuracy. The method is based on an unsupervised feature learning approach, and processes the complete image without the need for an accurate pre-segmentation of the retina. The method can be divided in two steps: an unsupervised clustering stage that extracts a set of small descriptive image patches from the training data, and a supervised training stage that uses these patches to create a patch occurrence histogram for every image on which a random forest classifier is trained. Experiments using 384 volume scans show that the proposed method is capable of identifying AMD patients with high accuracy, obtaining an area under the Receiver Operating Curve of 0:984. Our method allows for a quick and reliable assessment of the presence of AMD pathology in OCT volume scans without the need for accurate layer segmentation algorithms.

Diabetic retinopathy (DR) and age related macular degeneration (ARMD) are among the major causes of visual impairment all over the world. DR is mainly characterized by small red spots, namely microaneurysms and bright lesions, specifically exudates. However, ARMD is mainly identified by tiny yellow or white deposits called drusen. Since exudates might be the only visible signs of the early diabetic retinopathy, there is an increase demand for automatic diagnosis of retinopathy. Exudates and drusen may share similar appearances; as a result discriminating between them plays a key role in improving screening performance. In this research, we investigative the role of bag of words approach in the automatic diagnosis of retinopathy diabetes. Initially, the color retinal images are preprocessed in order to reduce the intra and inter patient variability. Subsequently, SURF (Speeded up Robust Features), HOG (Histogram of Oriented Gradients), and LBP (Local Binary Patterns) descriptors are extracted. We proposed to use single-based and multiple-based methods to construct the visual dictionary by combining the histogram of word occurrences from each dictionary and building a single histogram. Finally, this histogram representation is fed into a support vector machine with linear kernel for classification. The introduced approach is evaluated for automatic diagnosis of normal and abnormal color retinal images with bright lesions such as drusen and exudates. This approach has been implemented on 430 color retinal images, including six publicly available datasets, in addition to one local dataset. The mean accuracies achieved are 97.2% and 99.77% for single-based and multiple-based dictionaries respectively.

Medical imaging technology has always provided radiologists with the opportunity to view and keep records of anatomy of the patient. With the development of machine learning and intelligent computing, these images can be used to create Computer-Aided Diagnosis (CAD) systems, which can assist radiologists in analyzing image data in various ways to provide better health care to patients. This paper looks at increasing accuracy and reducing cost in creating CAD systems, specifically in predicting the malignancy of lung nodules in the Lung Image Database Consortium (LIDC). Much of the cost in creating an accurate CAD system stems from the need for multiple radiologist diagnoses or annotations of each image, since there is rarely a ground truth diagnosis and even different radiologists' diagnoses of the same nodule often disagree. To resolve this issue, this paper outlines an method of selective iterative classification that predicts lung nodule malignancy by using multiple radiologist diagnoses only for cases that can benefit from them. Our method achieved 81% accuracy while costing only 46% of the method that indiscriminately used all annotations, which achieved a lower accuracy of 70%, while costing more.

Radiologists often compare sequential radiographs side-by-side in order to identify regions of change and evaluate the clinical significance of such regions. Some changes in pathology, however, may be overlooked or misinterpreted. For this reason, temporal subtraction (TS) images provide an important tool for enhanced visualization. Not all areas of “change” demonstrated on a TS image, however, are caused by pathology. The purpose of this study was to develop an automated computer-aided diagnosis (CAD) system to locate regions of change in TS images as well as to classify such regions as being true regions of pathologic change or false regions of change caused by misregistration artifacts. The dataset used in this study contained 120 images, on which an experienced radiologist outlined 74 regions of true pathologic change that were used as the gold standard. Through gray-level thresholding and initial false-positive reduction, an initial set of candidates was extracted and inputted into a classifier. A five-fold cross-validation method was employed to create training and testing groups. Both false-candidate regions as well as the gold-standard regions were used as training data. Of the three classifiers tested (support vector machine, logistic regression, and linear discriminant analysis), the logistic regression classifier performed the best with a sensitivity of 96% and specificity of 84%; receiver operating characteristic (ROC) analysis resulted in an area under the ROC curve of 0.94. These results show promise in the performance of the CAD system to detect regions of pathologic changes in TS images of the chest.

Due to the promotion of lung cancer screening, more Stage I non-small-cell lung cancers (NSCLC) are currently detected, which usually have favorable prognosis. However, a high percentage of the patients have cancer recurrence after surgery, which reduces overall survival rate. To achieve optimal efficacy of treating and managing Stage I NSCLC patients, it is important to develop more accurate and reliable biomarkers or tools to predict cancer prognosis. The purpose of this study is to investigate a new quantitative image analysis method to predict the risk of lung cancer recurrence of Stage I NSCLC patients after the lung cancer surgery using the conventional chest computed tomography (CT) images and compare the prediction result with a popular genetic biomarker namely, protein expression of the excision repair cross-complementing 1 (ERCC1) genes. In this study, we developed and tested a new computer-aided detection (CAD) scheme to segment lung tumors and initially compute 35 tumor-related morphologic and texture features from CT images. By applying a machine learning based feature selection method, we identified a set of 8 effective and non-redundant image features. Using these features we trained a naïve Bayesian network based classifier to predict the risk of cancer recurrence. When applying to a test dataset with 79 Stage I NSCLC cases, the computed areas under ROC curves were 0.77±0.06 and 0.63±0.07 when using the quantitative image based classifier and ERCC1, respectively. The study results demonstrated the feasibility of improving accuracy of predicting cancer prognosis or recurrence risk using a CAD-based quantitative image analysis method.

Advantages of CT scanners with high resolution have allowed the improved detection of lung cancers. In the recent release of positive results from the National Lung Screening Trial (NLST) in the US showing that CT screening does in fact have a positive impact on the reduction of lung cancer related mortality. While this study does show the efficacy of CT based screening, physicians often face the problems of deciding appropriate management strategies for maximizing patient survival and for preserving lung function. Several key manifold-learning approaches efficiently reveal intrinsic low-dimensional structures latent in high-dimensional data spaces. This study was performed to investigate whether the dimensionality reduction can identify embedded structures from the CT histogram feature of non-small-cell lung cancer (NSCLC) space to improve the performance in predicting the likelihood of RFS for patients with NSCLC.

Computer-Aided Detection (CAD) has been shown to be a promising tool for automatic detection of pulmonary nodules from computed tomography (CT) images. However, the vast majority of detected nodules are benign and do not require any treatment. For effective implementation of lung cancer screening programs, accurate identification of malignant nodules is the key. We investigate strategies to improve the performance of a CAD system in detecting nodules with a high probability of being cancers. Two strategies were proposed: (1) combining CAD detections with a recently published lung cancer risk prediction model and (2) the combination of multiple CAD systems. First, CAD systems were used to detect the nodules. Each CAD system produces markers with a certain degree of suspicion. Next, the malignancy probability was automatically computed for each marker, given nodule characteristics measured by the CAD system. Last, CAD degree of suspicion and malignancy probability were combined using the product rule. We evaluated the method using 62 nodules which were proven to be malignant cancers, from 180 scans of the Danish Lung Cancer Screening Trial. The malignant nodules were considered as positive samples, while all other findings were considered negative. Using a product rule, the best proposed system achieved an improvement in sensitivity, compared to the best individual CAD system, from 41.9% to 72.6% at 2 false positives (FPs)/scan and from 56.5% to 88.7% at 8 FPs/scan. Our experiment shows that combining a nodule malignancy probability with multiple CAD systems can increase the performance of computerized detection of lung cancer.

A novel framework for automatic detection of pneumothorax abnormality in chest radiographs is presented. The suggested method is based on a texture analysis approach combined with supervised learning techniques. The proposed framework consists of two main steps: at first, a texture analysis process is performed for detection of local abnormalities. Labeled image patches are extracted in the texture analysis procedure following which local analysis values are incorporated into a novel global image representation. The global representation is used for training and detection of the abnormality at the image level. The presented global representation is designed based on the distinctive shape of the lung, taking into account the characteristics of typical pneumothorax abnormalities. A supervised learning process was performed on both the local and global data, leading to trained detection system. The system was tested on a dataset of 108 upright chest radiographs. Several state of the art texture feature sets were experimented with (Local Binary Patterns, Maximum Response filters). The optimal configuration yielded sensitivity of 81% with specificity of 87%. The results of the evaluation are promising, establishing the current framework as a basis for additional improvements and extensions.

We are developing a computer-aided detection system for bladder cancer on CTU. The bladder was automatically segmented with our Conjoint Level set Analysis and Segmentation System (CLASS). In this preliminary study, we developed a system for detecting mass within the contrast-enhanced (C) region of the bladder. The C region was delineated from the segmented bladders using a method based on maximum intensity projection. The bladder wall of the C region was extracted using thresholding to remove the contrast material. The wall on each slice was transformed into a wall profile. Morphology and voxel intensity along the profile were analyzed and suspicious locations were labeled as lesion candidates. The candidates were segmented and 20 morphological features were extracted from each candidate. A data set of 35 patients with 45 biopsy-proven bladder lesions within the C region was used for system evaluation. Stepwise feature selection with simplex optimization and leave-one-case-out method was used for training and validation. For each partition in the leave-one-case-out method, features were selected from the training cases and a linear discriminant (LDA) classifier was designed to merge the selected features into a single score for classification of the lesion candidates into bladder lesions and normal findings in the left-out case. A single score was generated for each lesion candidate. The performance of the CAD system was evaluated by FROC analysis. At an FP rate of 2.5 FPs/case, the system achieved a sensitivity of 82%, while at 1.7 FPs/case, a sensitivity of 71%.

Station-labeling of mediastinal lymph nodes is typically performed to identify the location of enlarged nodes for cancer staging. Stations are usually assigned in clinical radiology practice manually by qualitative visual assessment on CT scans, which is time consuming and highly variable. In this paper, we developed a method that automatically recognizes the lymph node stations in thoracic CT scans based on the anatomical organs in the mediastinum. First, the trachea, lungs, and spines are automatically segmented to locate the mediastinum region. Then, eight more anatomical organs are simultaneously identified by multi-atlas segmentation. Finally, with the segmentation of those anatomical organs, we convert the text definitions of the International Association for the Study of Lung Cancer (IASLC) lymph node map into patient-specific color-coded CT image maps. Thus, a lymph node station is automatically assigned to each lymph node. We applied this system to CT scans of 86 patients with 336 mediastinal lymph nodes measuring equal or greater than 10 mm. 84.8% of mediastinal lymph nodes were correctly mapped to their stations.

We are developing an automated method for the segmentation of ureters in CTU, referred to as COmbined Modelguided Path-finding Analysis and Segmentation System (COMPASS). Ureter segmentation is a critical component for computer-aided detection of ureter cancer. A challenge for ureter segmentation is inconsistent opacification of the region of interest, which makes it difficult to be distinguished from other tissue in the surrounding area. COMPASS consists of four stages: (1) region finding and adaptive thresholding, (2) segmentation accuracy analysis, (3) potential backtracking and branching, and (4) edge profile extraction and feature analysis. In this study, we evaluated a new method in which CTU images were pre-processed with 3D multiscale Hessian filtering that enhances tubular structures. Our goal is to compare the performance of COMPASS with and without multiscale Hessian enhancement. With IRB approval, 79 cases with 124 ureters and 10 cases with 18 ureters were collected retrospectively from patient files as training and test sets, respectively. On average, the ureters spanned 289 CT slices (range: 115-405, median: 302). The segmentation performance was quantitatively assessed as the percentage of length of each ureter that was successfully tracked relative to manually tracking. COMPASS alone segmented, on average, 99.16% and 98.74% of each ureter in the training and test sets, respectively. COMPASS with Hessian enhancement segmented, on average, 97.89% and 99.63% of each ureter in the training and test sets, respectively. Although the difference did not reach statistical significance in this small test set, Hessian-enhanced tracking shows promise for overcoming certain types of difficult cases.

This paper presents a method for generating branching pattern reports of abdominal blood vessels for laparoscopic gastrectomy. In gastrectomy, it is very important to understand branching structure of abdominal arteries and veins, which feed and drain specific abdominal organs including the stomach, the liver and the pancreas. In the real clinical stage, a surgeon creates a diagnostic report of the patient anatomy. This report summarizes the branching patterns of the blood vessels related to the stomach. The surgeon decides actual operative procedure. This paper shows an automated method to generate a branching pattern report for abdominal blood vessels based on automated anatomical labeling. The report contains 3D rendering showing important blood vessels and descriptions of branching patterns of each vessel. We have applied this method for fifty cases of 3D abdominal CT scans and confirmed the proposed method can automatically generate branching pattern reports of abdominal arteries.

Soft tissue sarcomas are malignant tumors which develop from tissues like fat, muscle, nerves, fibrous tissue or blood vessels. They are challenging to physicians because of their relative infrequency and diverse outcomes, which have hindered development of new therapeutic agents. Additionally, assessing imaging response of these tumors to therapy is also difficult because of their heterogeneous appearance on magnetic resonance imaging (MRI). In this paper, we assessed standard of care MRI sequences performed before and after treatment using 36 patients with soft tissue sarcoma. Tumor tissue was identified by manually drawing a mask on contrast enhanced images. The Otsu segmentation method was applied to segment tumor tissue into low and high signal intensity regions on both T1 post-contrast and T2 without contrast images. This resulted in four distinctive subregions or “habitats.” The features used to predict metastatic tumors and necrosis included the ratio of habitat size to whole tumor size and components of 2D intensity histograms. Individual cases were correctly classified as metastatic or non-metastatic disease with 80.55% accuracy and for necrosis ≥ 90 or necrosis <90 with 75.75% accuracy by using meta-classifiers which contained feature selectors and classifiers.

Research studies have investigated a number of factors that may impact the performance assessment of computer aided detection (CAD) effectiveness, such as the inherent design of the CAD, the image and reader samples, and the assessment methods. In this study, we focused on the effect of prevalence on cue validity (co-occurrence of cue and signal) and learning as potentially important factors in CAD assessment. For example, the prevalence of cases with breast cancer is around 50% in laboratory CAD studies, which is 100 times higher than that in breast cancer screening. Although ROC is prevalence-independent, an observer’s use of CAD involves tasks that are more complicated than binary classification, including: search, detection, classification, cueing and learning. We developed models to investigate the potential impact of prevalence on cue validity and the learning of cue validity tasks. We hope this work motivates new studies that investigate previously under-explored factors involved in image interpretation with a new modality in its assessment.

The right and left mammograms of a patient are assumed to be bilaterally symmetric for image readings. The detection of asymmetry in bilateral mammograms is a reliable indicator for detecting possible breast abnormalities. The purpose of this study was to examine the potential usefulness of a new method in terms of spatial frequency components for exploration of similarity and abnormality between the right and left mammograms. A total of 98 normal and 119 abnormal cases with calcifications were used for this study. Each case included two mediolateral oblique views. The spatial frequency components were determined from the symmetric regions in the right and left mammograms by Fourier transform. The degrees of conformity between the two spatial frequency components in the right and left mammograms were calculated for the same and different patients. The degrees of conformity were also examined for cases with and without calcifications for the same patient to show if the proposed method was useful for indicating the existence of calcifications or not. The average degrees of conformity and the standard deviations for the same and different patients were 0.911 ± 0.0165 and 0.857 ± 0.0328, respectively. The degrees of conformity calculated from abnormal cases (0.836 ± 0.0906) showed statistically lower values compared with those measured from normal cases (0.911 ± 0.0165). Our results indicated that histogram analysis of spatial frequency components could be useful as a similarity measure between bilateral mammograms for the same patient and abnormal signs in a mammogram.

We propose an automated method in detecting lesions to assist radiologists in interpreting dynamic contrast enhanced magnetic resonance imaging (DCE-MRI) of breast. The aim is to highlight the suspicious regions of interest to reduce the searching time of the lesions and the possibility of radiologists overlooking small regions. In our method, we locate the suspicious regions by applying a threshold on essential features. The features are normalized to reduce the variation between patients. Support vector machine classifier is then applied to exclude normal tissues from these regions, using both kinetic and morphological features extracted in the lesions. In the evaluation of the system on 21 patients with 50 lesions, all lesions were successfully detected with 5.02 false positive regions per breast.

Amount of fibroglandular tissue (FGT) and background parenchymal enhancement (BPE) in dynamic contrast enhanced magnetic resonance (DCE-MR) images are two important indices for breast cancer risk assessment in the clinical practice. The purpose of this study is to develop and evaluate a fully automated scheme for quantitative analysis of FGT and BPE in DCE-MR images. Our fully automated method consists of three steps, i.e., segmentation of whole breast, fibroglandular tissues, and enhanced fibroglandular tissues. Based on the volume of interest extracted automatically, dynamic programming method was applied in each 2-D slice of a 3-D MR scan to delineate the chest wall and breast skin line for segmenting the whole breast. This step took advantages of the continuity of chest wall and breast skin line across adjacent slices. We then further used fuzzy c-means clustering method with automatic selection of cluster number for segmenting the fibroglandular tissues within the segmented whole breast area. Finally, a statistical method was used to set a threshold based on the estimated noise level for segmenting the enhanced fibroglandular tissues in the subtraction images of pre- and post-contrast MR scans. Based on the segmented whole breast, fibroglandular tissues, and enhanced fibroglandular tissues, FGT and BPE were automatically computed. Preliminary results of technical evaluation and clinical validation showed that our fully automated scheme could obtain good segmentation of the whole breast, fibroglandular tissues, and enhanced fibroglandular tissues to achieve accurate assessment of FGT and BPE for quantitative analysis of breast cancer risk.

Imaging plays a central role in the evaluation of breast tumor response to neoadjuvant chemotherapy. Image-based assessment of tumor change via deformable registration is a powerful, quantitative method potentially to explore novel information of tumor heterogeneity, structure, function, and treatment response. In this study, we continued a previous pilot study to further validate the feasibility of an open source deformable registration algorithm DRAMMS developed within our group as a means to analyze spatio-temporal tumor changes for a set of 14 patients with DCE-MR imaging. Two experienced breast imaging radiologists marked landmarks according to their anatomical meaning on image sets acquired before and during chemotherapy. Yet, chemotherapy remarkably changed the anatomical structure of both tumor and normal breast tissue, leading to significant discrepancies between both raters for landmarks in certain areas. Therefore, we proposed a novel method to grade the manually denoted landmarks into different challenge levels based on the inter-rater agreement, where a high level indicates significant discrepancies and considerable amounts of anatomical structure changes, which would indeed impose giant problem for the following registration algorithm. It is interesting to observe that DRAMMS performed in a similar manner as the human raters: landmark errors increased as inter-rater differences rose. Among all selected six deformable registration algorithms, DRAMMS achieves the highest overall accuracy, which is around 5.5 mm, while the average difference between human raters is 3 mm. Moreover, DRAMMS performed consistently well within both tumor and normal tissue regions. Lastly, we comprehensively tuned the fundamental parameters of DRAMMS to better understand DRAMMS to guide similar works in the future. Overall, we further validated that DRAMMS is a powerful registration tool to accurately quantify tumor changes and potentially predict early tumor response to chemotherapy. Therefore, future studies that aim at examining if DRAMMS can generate valuable biomarkers for tumor response prediction during chemotherapy become feasible.

The purpose of this study is to develop and test a new content-based image retrieval (CBIR) scheme that enables to achieve higher reproducibility when it is implemented in an interactive computer-aided diagnosis (CAD) system without significantly reducing lesion classification performance. This is a new Fourier transform based CBIR algorithm that determines image similarity of two regions of interest (ROI) based on the difference of average regional image pixel value distribution in two Fourier transform mapped images under comparison. A reference image database involving 227 ROIs depicting the verified soft-tissue breast lesions was used. For each testing ROI, the queried lesion center was systematically shifted from 10 to 50 pixels to simulate inter-user variation of querying suspicious lesion center when using an interactive CAD system. The lesion classification performance and reproducibility as the queried lesion center shift were assessed and compared among the three CBIR schemes based on Fourier transform, mutual information and Pearson correlation. Each CBIR scheme retrieved 10 most similar reference ROIs and computed a likelihood score of the queried ROI depicting a malignant lesion. The experimental results shown that three CBIR schemes yielded very comparable lesion classification performance as measured by the areas under ROC curves with the p-value greater than 0.498. However, the CBIR scheme using Fourier transform yielded the highest invariance to both queried lesion center shift and lesion size change. This study demonstrated the feasibility of improving robustness of the interactive CAD systems by adding a new Fourier transform based image feature to CBIR schemes.

This paper presents a novel method for enhancing the contrast of microcalcifications in digital breast tomosynthesis projection views for detection purposes. The proposed method relies on the correlation between the projection views in order to reduce the effect of noise, due to the low-dose exposure, and increase the contrast of the microcalcification particles for microcalcification cluster detection purposes. The method performs a series of multi-shift operations to capture the microcalcification particle movement information and compensate it in order to enhance microcalcification particles contrast. Furthermore, the proposed approach utilizes the projection view correlation in order to reduce the falsely detected regions of interest, and improve the classification of the detected regions into false positives or actual microcalcification clusters. Comparative experiments have been performed to quantitatively measure the contrast enhancement of microcalcification particles and its effect on the MC cluster detection. To that end, the contrast to noise ratio have been calculated and compared with some with previous methods. Furthermore, the free response receiver operating characteristic (FROC) curve have been used to measure the effect of the proposed enhancement on the microcalcification cluster detectability.

A novel breast cancer risk analysis approach is proposed for enhancing performance of computerized breast cancer risk analysis using bilateral mammograms. Based on the intensity of breast area, five different sub-regions were acquired from one mammogram, and bilateral features were extracted from every sub-region. Our dataset includes 180 bilateral mammograms from 180 women who underwent routine screening examinations, all interpreted as negative and not recalled by the radiologists during the original screening procedures. A computerized breast cancer risk analysis scheme using four image processing modules, including sub-region segmentation, bilateral feature extraction, feature selection, and classification was designed to detect and compute image feature asymmetry between the left and right breasts imaged on the mammograms. The highest computed area under the curve (AUC) is 0.763 ± 0.021 when applying the multiple sub-region features to our testing dataset. The positive predictive value and the negative predictive value were 0.60 and 0.73, respectively. The study demonstrates that (1) features extracted from multiple sub-regions can improve the performance of our scheme compared to using features from whole breast area only; (2) a classifier using asymmetry bilateral features can effectively predict breast cancer risk; (3) incorporating texture and morphological features with density features can boost the classification accuracy.

We propose an automatic segmentation method of chest wall in 3D ABUS images using rib shadow enhancement and multi-planar cumulative probability enhanced map. For the identification of individual dark rib shadows, each rib shadow is enhanced using intensity transfer function and 3D sheet-like enhancement filtering. Then, wrongly enhanced intercostal regions and small fatty tissues are removed using coronal and sagittal cumulative probability enhanced maps. The large fatty tissues with globular and sheet-like shapes at the top of rib shadow are removed using shape and orientation analysis based on moment matrix. Detected chest walls are connected with cubic B-spline interpolation. Experimental results show that the Dice similarity coefficient of proposed method as comparison with two manually outlining results provides over 90% in average.

Since performance and clinical utility of current computer-aided detection (CAD) schemes of detecting and classifying soft tissue lesions (e.g., breast masses and lung nodules) is not satisfactory, many researchers in CAD field call for new CAD research ideas and approaches. The purpose of presenting this opinion paper is to share our vision and stimulate more discussions of how to overcome or compensate the limitation of current lesion-detection based CAD schemes in the CAD research community. Since based on our observation that analyzing global image information plays an important role in radiologists’ decision making, we hypothesized that using the targeted quantitative image features computed from global images could also provide highly discriminatory power, which are supplementary to the lesion-based information. To test our hypothesis, we recently performed a number of independent studies. Based on our published preliminary study results, we demonstrated that global mammographic image features and background parenchymal enhancement of breast MR images carried useful information to (1) predict near-term breast cancer risk based on negative screening mammograms, (2) distinguish between true- and false-positive recalls in mammography screening examinations, and (3) classify between malignant and benign breast MR examinations. The global case-based CAD scheme only warns a risk level of the cases without cueing a large number of false-positive lesions. It can also be applied to guide lesion-based CAD cueing to reduce false-positives but enhance clinically relevant true-positive cueing. However, before such a new CAD approach is clinically acceptable, more work is needed to optimize not only the scheme performance but also how to integrate with lesion-based CAD schemes in the clinical practice.

In this paper, new mass features based on inter-view similarity in DBT projection views are proposed for classifying masses, aiming to effectively reducing false-positives (FPs). The proposed features are focused on utilizing inter-view information in projection views. The FPs induced by overlapping tissues of different depth could be observed differently between projection views, while masses could appear similar. To utilize the observation, the inter-view similarity measure is developed by utilizing the normalized cross-correlation between ROIs in projection views. In the analysis of inter-view similarities of masses and FPs, it is showed that inter-view similarities of FPs are lower than those of masses. To that end, new features are proposed to encode aforementioned difference of inter-view similarities between FPs and masses. Experimental results show that the proposed features can improve the mass classification performance in projection views in terms of the area under the ROC curve.

A computer-aided detection (CADe) system for clustered microcalcifications (MCs) in reconstructed digital breast tomosynthesis (DBT) volumes was suggested. The system consisted of prescreening, MC detecting, clustering, and falsepositive reduction steps. In the prescreening stage, the MC-like objects were enhanced by a multiscale-based 3D calcification response function. A connected component segmentation method was used to detect cluster seed objects, which were considered as potential clustering centers of MCs. Starting with each cluster seed object as the initial cluster center, a cluster candidate was formed by including nearby MC candidates within a 3D neighborhood of the cluster seed object satisfying the clustering criteria during the clustering step. The size and number of the clustered MCs in a cluster seed candidate were used to reduce the number of FPs. A bounding cube for each MCC was generated for each accepted seed candidates. Then, the overlapping cubes were combined and examined according to the FP reduction criteria. After FP reduction step, we obtained the average number of FPs of 2.47 per DBT volume with sensitivity of 83.3%. Our study indicates the simplified false-positive reduction approach applied to the detection of clustered MCs in DBT is promising as an efficient CADe system.

Breast parenchymal density is considered a strong indicator of breast cancer risk and therefore useful for preventive tasks. Measurement of breast density is often qualitative and requires the subjective judgment of radiologists. Here we explore an automatic breast composition classification workflow based on convolutional neural networks for feature extraction in combination with a support vector machines classifier. This is compared to the assessments of seven experienced radiologists. The experiments yielded an average kappa value of 0.58 when using the mode of the radiologists’ classifications as ground truth. Individual radiologist performance against this ground truth yielded kappa values between 0.56 and 0.79.

Dynamic contrast-enhanced breast magnetic resonance imaging (DCE-MRI) is used increasingly in diagnosis of breast cancer and assessment of treatment efficacy in current clinical practice. The purpose of this preliminary study is to develop and test a new quantitative kinetic image feature analysis method and biomarker to predict response of breast cancer patients to neoadjuvant chemotherapy using breast MR images acquired before the chemotherapy. For this purpose, we developed a computer-aided detection scheme to automatically segment breast areas and tumors depicting on the sequentially scanned breast MR images. From a contrast-enhancement map generated by subtraction of two image sets scanned pre- and post-injection of contrast agent, our scheme computed 38 morphological and kinetic image features from both tumor and background parenchymal regions. We applied a number of statistical data analysis methods to identify effective image features in predicting response of the patients to the chemotherapy. Based on the performance assessment of individual features and their correlations, we applied a fusion method to generate a final image biomarker. A breast MR image dataset involving 68 patients was used in this study. Among them, 25 had complete response and 43 had partially response to the chemotherapy based on the RECIST guideline. Using this image feature fusion based biomarker, the area under a receiver operating characteristic curve is AUC = 0.850±0.047. This study demonstrated that a biomarker developed from the fusion of kinetic image features computed from breast MR images acquired pre-chemotherapy has potentially higher discriminatory power in predicting response of the patients to the chemotherapy.

Breast cancer incidence tends to rise globally and the mortality rate for breast cancer is increasing in Japan. There are various screening modalities for breast cancer, and MRI examinations with high detection rate are used for high-risk groups, which are genetically prone to develop breast cancer. In the breast MRI examination, unenhanced T1 and T2 weighted images shows no significant difference in signal value between tumor and normal tissue. Therefore, tumors are identified with use of contrast enhanced kinetic curve obtained by dynamic scan using contrast agent. Some computer aided diagnosis methods using dynamic contrast enhanced MR images also have been proposed. However, contrast agent produces the allergic reaction in rare case; it should not be used for screening examinees. Here, MRI provides the anatomical and functional information by using various sequences without contrast agents. According to the reports, this information can discriminate between tumor and normal tissue. In this study, we analyzed unenhanced MR images by using plural sequences and developed an automated method for the detection of tumors. First, we extracted the breast region from the T1-weighted image semi-automatically. Next, using the threshold determined by considering the signal intensities of tumor and normal tissue, a thresholding method was applied for diffusion-weighted image to extract the first candidate regions. After labeling processing, the breast region removes outside candidates from Initial candidates. Then false positives are reduced by the rule-based classifier. Finally, we examined the remaining candidates as possible tumor regions. We applied the proposed method to 54 cases of MR images and evaluated its usefulness. As a result, the detection sensitivity was 71.9% and the abnormal regions were clearly detected. These results indicate that the proposed method may be useful for tumor detection in unenhanced breast MR images.

Mammography is a standard tool for breast cancer diagnosis. In current clinical practice, typically two mammograms of each breast are taken from different angles. A fundamental step when using ipsilateral mammograms for the diagnosis of breast cancer, is the identification of corresponding locations/structures in both views, which is a very challenging task due to the projective nature of the images and the different compression parameters used for each view. In this contribution, four different approaches for the estimation of corresponding locations in ipsilateral mammograms are systematically compared using 46 mammogram pairs (50 point-to-point correspondences). The evaluation includes simple heuristic methods (annular bands and straight strips) as well as methods based on geometric and physically motivated breast compression models, which aim to simulate the mammogram acquisition process. The evaluation results show that on average no significant differences exist between the estimation accuracies obtained using the simple heuristic methods and the more involved compression models. However, the results of this study indicate the potential of a method that optimally combines the different approaches.

The use of biomechanics-based numerical simulations has attracted growing interest in recent years for computer-aided diagnosis and treatment planning. With this in mind, a method for automatic mesh generation of brain structures of interest, using statistical models of shape (SSM) and appearance (SAM), for personalised computational modelling is presented. SSMs are constructed as point distribution models (PDMs) while SAMs are trained using intensity profiles sampled from a training set of T1-weighted magnetic resonance images. The brain structures of interest are, the cortical surface (cerebrum, cerebellum & brainstem), lateral ventricles and falx-cerebri membrane. Two methods for establishing correspondences across the training set of shapes are investigated and compared (based on SSM quality): the Coherent Point Drift (CPD) point-set registration method and B-spline mesh-to-mesh registration method. The MNI-305 (Montreal Neurological Institute) average brain atlas is used to generate the template mesh, which is deformed and registered to each training case, to establish correspondence over the training set of shapes. 18 healthy patients’ T1-weightedMRimages form the training set used to generate the SSM and SAM. Both model-training and model-fitting are performed over multiple brain structures simultaneously. Compactness and generalisation errors of the BSpline-SSM and CPD-SSM are evaluated and used to quantitatively compare the SSMs. Leave-one-out cross validation is used to evaluate SSM quality in terms of these measures. The mesh-based SSM is found to generalise better and is more compact, relative to the CPD-based SSM. Quality of the best-fit model instance from the trained SSMs, to test cases are evaluated using the Hausdorff distance (HD) and mean absolute surface distance (MASD) metrics.

Dental implant recognition in patients without available records is a time-consuming and not straightforward task. The traditional method is a complete user-dependent process, where the expert compares a 2D X-ray image of the dental implant with a generic database. Due to the high number of implants available and the similarity between them, automatic/semi-automatic frameworks to aide implant model detection are essential. In this study, a novel computer-aided framework for dental implant recognition is suggested. The proposed method relies on image processing concepts, namely: (i) a segmentation strategy for semi-automatic implant delineation; and (ii) a machine learning approach for implant model recognition. Although the segmentation technique is the main focus of the current study, preliminary details of the machine learning approach are also reported. Two different scenarios are used to validate the framework: (1) comparison of the semi-automatic contours against implant’s manual contours of 125 X-ray images; and (2) classification of 11 known implants using a large reference database of 601 implants. Regarding experiment 1, 0.97±0.01, 2.24±0.85 pixels and 11.12±6 pixels of dice metric, mean absolute distance and Hausdorff distance were obtained, respectively. In experiment 2, 91% of the implants were successfully recognized while reducing the reference database to 5% of its original size. Overall, the segmentation technique achieved accurate implant contours. Although the preliminary classification results prove the concept of the current work, more features and an extended database should be used in a future work.

Brain micro-bleeds (BMBs) are used as surrogate markers for detecting diffuse axonal injury in traumatic brain injury (TBI) patients. The location and number of BMBs have been shown to influence the long-term outcome of TBI. To further study the importance of BMBs for prognosis, accurate localization and quantification are required. The task of annotating BMBs is laborious, complex and prone to error, resulting in a high inter- and intra-reader variability. In this paper we propose a computer-aided detection (CAD) system to automatically detect BMBs in MRI scans of moderate to severe neuro-trauma patients. Our method consists of four steps. Step one: preprocessing of the data. Both susceptibility (SWI) and T1 weighted MRI scans are used. The images are co-registered, a brain-mask is generated, the bias field is corrected, and the image intensities are normalized. Step two: initial candidates for BMBs are selected as local minima in the processed SWI scans. Step three: feature extraction. BMBs appear as round or ovoid signal hypo-intensities on SWI. Twelve features are computed to capture these properties of a BMB. Step four: Classification. To identify BMBs from the set of local minima using their features, different classifiers are trained on a database of 33 expert annotated scans and 18 healthy subjects with no BMBs. Our system uses a leave-one-out strategy to analyze its performance. With a sensitivity of 90% and 1.3 false positives per BMB, our CAD system shows superior results compared to state-of-the-art BMB detection algorithms (developed for non-trauma patients).

Abnormal accumulation of brain iron has been detected in various neurodegenerative diseases. Quantitative susceptibility mapping (QSM) is a novel contrast mechanism in magnetic resonance (MR) imaging and enables the quantitative analysis of local tissue susceptibility property. Therefore, automatic segmentation tools of brain regions on QSM images would be helpful for radiologists’ quantitative analysis in various neurodegenerative diseases. The purpose of this study was to develop an automatic segmentation and classification method of striatum regions on QSM images. Our image database consisted of 22 QSM images obtained from healthy volunteers. These images were acquired on a 3.0 T MR scanner. The voxel size was 0.9×0.9×2 mm. The matrix size of each slice image was 256×256 pixels. In our computerized method, a template mating technique was first used for the detection of a slice image containing striatum regions. An image registration technique was subsequently employed for the classification of striatum regions in consideration of the anatomical knowledge. After the image registration, the voxels in the target image which correspond with striatum regions in the reference image were classified into three striatum regions, i.e., head of the caudate nucleus, putamen, and globus pallidus. The experimental results indicated that 100% (21/21) of the slice images containing striatum regions were detected accurately. The subjective evaluation of the classification results indicated that 20 (95.2%) of 21 showed good or adequate quality. Our computerized method would be useful for the quantitative analysis of Parkinson diseases in QSM images.

Brain tumor heterogeneity remains a challenge for probing brain cancer evolutionary dynamics. In light of evolution, it is a priority to inspect the cancer system from a time-domain perspective since it explicitly tracks the dynamics of cancer variations. In this paper, we study the problem of exploring brain tumor heterogeneity from temporal clinical magnetic resonance imaging (MRI) data. Our goal is to discover evidence-based knowledge from such temporal imaging data, where multiple clinical MRI scans from Glioblastoma multiforme (GBM) patients are generated during therapy. In particular, we propose a quantitative histogram-based approach that builds a prediction model to measure the difference in histograms obtained from pre- and post-treatment. The study could significantly assist radiologists by providing a metric to identify distinctive patterns within each tumor, which is crucial for the goal of providing patient-specific treatments. We examine the proposed approach for a practical application - clinical survival group prediction. Experimental results show that our approach achieved 90.91% accuracy.

Leukoaraiosis (LA) describes diffuse white matter abnormalities on CT or MR brain scans, often seen in the normal elderly and in association with vascular risk factors such as hypertension, or in the context of cognitive impairment. The mechanism of cognitive dysfunction is still unclear. The recent clinical studies have revealed that the severity of LA was not corresponding to the cognitive level, and functional connectivity analysis is an appropriate method to detect the relation between LA and cognitive decline. However, existing functional connectivity analyses of LA have been mostly limited to linear associations. In this investigation, a novel measure utilizing the extended maximal information coefficient (eMIC) was applied to construct non-linear functional connectivity in 44 LA subjects (9 dementia, 25 mild cognitive impairment (MCI) and 10 cognitively normal (CN)). The strength of non-linear functional connections for the first 1% of discriminative power increased in MCI compared with CN and dementia, which was opposed to its linear counterpart. Further functional network analysis revealed that the changes of the non-linear and linear connectivity have similar but not completely the same spatial distribution in human brain. In the multivariate pattern analysis with multiple classifiers, the non-linear functional connectivity mostly identified dementia, MCI and CN from LA with a relatively higher accuracy rate than the linear measure. Our findings revealed the non-linear functional connectivity provided useful discriminative power in classification of LA, and the spatial distributed changes between the non-linear and linear measure may indicate the underlying mechanism of cognitive dysfunction in LA.

Findings on dental panoramic radiographs (DPRs) have shown that mandibular cortical index (MCI) based on the morphology of mandibular inferior cortex was significantly correlated with osteoporosis. MCI on DPRs can be categorized into one of three groups and has the high potential for identifying patients with osteoporosis. However, most DPRs are used only for diagnosing dental conditions by dentists in their routine clinical work. Moreover, MCI is not generally quantified but assessed subjectively. In this study, we investigated a computer-aided diagnosis (CAD) system that automatically classifies mandibular cortical bone for detection of osteoporotic patients at early stage. First, an inferior border of mandibular bone was detected by use of an active contour method. Second, regions of interest including the cortical bone are extracted and analyzed for its thickness and roughness. Finally, support vector machine (SVM) differentiate cases into three MCI categories by features including the thickness and roughness. Ninety eight DPRs were used to evaluate our proposed scheme. The number of cases classified to Class I, II, and III by a dental radiologist are 56, 25 and 17 cases, respectively. Experimental result based on the leave-one-out cross-validation evaluation showed that the sensitivities for the classes I, II, and III were 94.6%, 57.7% and 94.1%, respectively. Distribution of the groups in the feature space indicates a possibility of MCI quantification by the proposed method. Therefore, our scheme has a potential in identifying osteoporotic patients at an early stage.

Accurate computer-aided prediction of survival time for brain tumor patients requires a thorough understanding of clinical data, since it provides useful prior knowledge for learning models. However, to simplify the learning process, traditional settings often assume datasets with equally distributed classes, which clearly does not reflect a typical distribution. In this paper, we investigate the problem of mining knowledge from an imbalanced dataset (i.e., a skewed distribution) to predict survival time. In particular, we propose an algorithmic framework to predict survival groups of brain tumor patients using multi-modality MRI data. Both an imbalanced distribution and classifier design are jointly considered: 1) We used the Synthetic Minority Over-sampling Technique to compensate for the imbalanced distribution; 2) A predictive linear regression model was adopted to learn a pair of class-specific dictionaries, which were derived from reformulated balanced data. We tested the proposed framework using a dataset of 42 patients with Glioblastoma Multiforme (GBM) tumors whose scans were obtained from the cancer genome atlas (TCGA). Experimental results showed that the proposed method achieved 95.24% accuracy.

Up to 25% of children who undergo brain tumour resection surgery in the posterior fossa develop posterior fossa syndrome (PFS). This syndrome is characterised by mutism and disturbance in speech. Our hypothesis is that there is a correlation between PFS and the occurrence of hypertrophic olivary degeneration (HOD) in lobes within the posterior fossa, known as the inferior olivary nuclei (ION). HOD is exhibited as an increase in size and intensity of the ION on an MR image. Intra-operative MRI (IoMRI) is used during surgical procedures at the Alder Hey Children’s Hospital, Liver- pool, England, in the treatment of Posterior Fossa tumours and allows visualisation of the brain during surgery. The final MR scan on the IoMRI allows early assessment of the ION immediately after the surgical procedure. The longitudinal MRI data of 28 patients was analysed in a collaborative study with Alder Hey Children’s Hospital, in order to identify the most relevant imaging features that relate to the development of PFS, specifically related to HOD. A semi-automated segmentation process was carried out to delineate the ION on each MRI. Feature selection techniques were used to identify the most relevant features amongst the MRI data, demographics and clinical data provided by the hospital. A support vector machine (SVM) was used to analyse the discriminative ability of the selected features. The results indicate the presence of HOD as the most efficient feature that correlates with the development of PFS, followed by the change in intensity and size of the ION and whether HOD occurred bilaterally or unilaterally.

Recurrent nerve paralysis (RP) is one of the most frequent complications of thyroid surgery. It reduces vocal fold mobility. Nasal endoscopy, a mini-invasive procedure, is the conventional way to detect RP. We suggest a new approach based on laryngeal ultrasound and a specific data analysis was designed to help with the automated detection of RP. Ten subjects were enrolled for this feasibility study: four controls, three patients with RP and three patients without RP according to nasal endoscopy. The ultrasound protocol was based on a ten seconds B-mode acquisition in a coronal plane during normal breathing. Image processing included three steps: 1) automated detection of two consecutive closing and opening images, corresponding to extreme positions of vocal folds in the sequence of B-mode images, using principal component analysis of the image sequence; 2) positioning of three landmarks and robust tracking of these points using a multi-pyramidal refined optical flow approach; 3) estimation of quantitative parameters indicating left and right fractions of mobility, and motion symmetry. Results provided by automated image processing were compared to those obtained by an expert. Detection of extreme images was accurate; tracking of landmarks was reliable in 80% of cases. Motion symmetry indices showed similar values for controls and patients without RP. Fraction of mobility was reduced in cases of RP. Thus, our CAD system helped in the detection of RP. Laryngeal ultrasound combined with appropriate image processing helped in the diagnosis of recurrent nerve paralysis and could be proposed as a first–line method.

Detection of larynx cancer from medical imaging is important for the quantification and for the definition of target volumes in radiotherapy treatment planning (RTP). Magnetic resonance imaging (MRI) is being increasingly used in RTP due to its high resolution and excellent soft tissue contrast. Manually detecting larynx cancer from sequential MRI is time consuming and subjective. The large diversity of cancer in terms of geometry, non-distinct boundaries combined with the presence of normal anatomical regions close to the cancer regions necessitates the development of automatic and robust algorithms for this task. A new automatic algorithm for the detection of larynx cancer from 2D gadoliniumenhanced T1-weighted (T1+Gd) MRI to assist clinicians in RTP is presented. The algorithm employs edge detection using spatial neighborhood information of pixels and incorporates this information in a fuzzy c-means clustering process to robustly separate different tissues types. Furthermore, it utilizes the information of the expected cancerous location for cancer regions labeling. Comparison of this automatic detection system with manual clinical detection on real T1+Gd axial MRI slices of 2 patients (24 MRI slices) with visible larynx cancer yields an average dice similarity coefficient of 0.78±0.04 and average root mean square error of 1.82±0.28 mm. Preliminary results show that this fully automatic system can assist clinicians in RTP by obtaining quantifiable and non-subjective repeatable detection results in a particular time-efficient and unbiased fashion.

Computed tomography has been used for assessing structural abnormalities associated with emphysema. It is important to develop a robust CT based imaging biomarker that would allow quantification of emphysema progression in early stage. This paper presents effect of smoking on emphysema progression using annual changes of low attenuation volume (LAV) by each lung lobe acquired from low-dose CT images in longitudinal screening for lung cancer. The percentage of LAV (LAV%) was measured after applying CT value threshold method and small noise reduction. Progression of emphysema was assessed by statistical analysis of the annual changes represented by linear regression of LAV%. This method was applied to 215 participants in lung cancer CT screening for five years (18 nonsmokers, 85 past smokers, and 112 current smokers). The results showed that LAV% is useful to classify current smokers with rapid progression of emphysema (0.2%/year, p<0.05). This paper demonstrates effectiveness of the proposed method in diagnosis and prognosis of early emphysema in CT screening for lung cancer.

Early detection of pulmonary nodules is crucial for improving prognosis of patients with lung cancer. Computer-aided detection of lung nodules in thoracic computed tomography (CT) scans has a great potential to enhance the performance of the radiologist in detecting nodules. In this paper we present a computer-aided lung nodule detection system for computed tomography (CT) scans that works in three steps. The system first segments the lung using thresholding and hole filling. From this segmentation the system extracts candidate nodules using Laplacian of Gaussian. To reject false positives among the detected candidate nodules, multiple established features are calculated. We propose a novel feature based on a spherical shell filter, which is specifically designed to distinguish between vascular structures and nodular structures. The performance of the proposed CAD system was evaluated by partaking in the ANODE09 challenge, which presents a platform for comparing automatic nodule detection programs. The results from the challenge show that our CAD system ranks third among the submitted works, demonstrating the efficacy of our proposed CAD system. The results also show that our proposed spherical shell filter in combination with conventional features can significantly reduce the number of false positives from the detected candidate nodules.

Analysis of medical images by radiologists is often a time consuming and costly process. Computer Aided diagnosis (CAD) systems can be used to provide second opinion and assist radiologists in diagnosis. Crowdsourcing has been a tool used by many domains in order to generate fast and cheap labels. However, exposing medical images to a large crowd is problematic. Our system takes a machine sourcing approach to CADs – that is, the idea that multiple weak segmentations can be used to create predictions with the same quality as an expert. We propose that the use of a large amount of weak segmentations can simulate crowdsourcing to create accurate predictions of semantic characteristics. In addition, we investigated the idea that outliers filtering will better performance of this CAD. Three segmentation algorithms were employed to create 20 weak segmentations. Low quality segmentations were then removed using outlier filtering method. Segmentations were then classified using an ensemble classifier to create final predictions. It was found that this CAD preformed just as well or better than a radiologist. In contrast, the removal of outliers from the set of segmentations does not improve the result further.

Automated segmentation of pathological bearing region is the first step towards the development of lung CAD. Most of the work reported in the literature related to automated analysis of lung tissue aims towards classification of fixed sized block into one of the classes. This block level classification of lung tissues in the image never results in accurate or smooth boundaries between different regions. In this work, effort is taken to investigate the performance of three automated image segmentation algorithms those results in smooth boundaries among lung tissue patterns commonly encountered in HRCT images of the thorax. A public database that consists of HRCT images taken from patients affected with Interstitial Lung Diseases (ILDs) is used for the evaluation. The algorithms considered are Markov Random Field (MRF), Gaussian Mixture Model (GMM) and Mean Shift (MS). 2-fold cross validation approach is followed for the selection of the best parameter value for individual algorithm as well as to evaluate the performance of all the algorithms. Mean shift algorithm is observed as the best performer in terms of Jaccard Index, Modified Hausdorff Distance, accuracy, Dice Similarity Coefficient and execution speed.

Reflecting global interest in lung cancer screening, considerable attention has been paid to automatic segmentation and volumetric measurement of lung nodules on CT. Ground glass opacity (GGO) nodules deserve special consideration in this context, since it has been reported that they are more likely to be malignant than solid nodules. However, due to relatively low contrast and indistinct boundaries of GGO nodules, segmentation is more difficult for GGO nodules compared with solid nodules. To overcome this difficulty, we propose a method for accurately segmenting not only solid nodules but also GGO nodules without prior information about nodule types. First, the histogram of CT values in pre-extracted lung regions is modeled by a Gaussian mixture model and a threshold value for including high-attenuation regions is computed. Second, after setting up a region of interest around the nodule seed point, foreground regions are extracted by using the threshold and quick-shift-based mode seeking. Finally, for separating vessels from the nodule, a vessel-likelihood map derived from elongatedness of foreground regions is computed, and a region growing scheme starting from the seed point is applied to the map with the aid of fast marching method. Experimental results using an anthropomorphic chest phantom showed that our method yielded generally lower volumetric measurement errors for both solid and GGO nodules compared with other methods reported in preceding studies conducted using similar technical settings. Also, our method allowed reasonable segmentation of GGO nodules in low-dose images and could be applied to clinical CT images including part-solid nodules.

Current computer-aided diagnosis (CAD) models, developed to determine the malignancy of pulmonary nodules, characterize the nodule’s shape, density, and border. Analyzing the lung parenchyma surrounding the nodule is an area that has been minimally explored. We hypothesize that improved classification of nodules can be achieved through the inclusion of features quantified from the surrounding lung tissue. From computed tomography (CT) data, feature extraction techniques were developed to quantify the parenchymal and nodule textures, including a three-dimensional application of Laws’ Texture Energy Measures. Border irregularity was investigated using ray-casting and rubber-band straightening techniques, while histogram features characterized the densities of the nodule and parenchyma. The feature set was reduced by stepwise feature selection to a few independent features that best summarized the dataset. Using leave-one-out cross-validation, a neural network was used for classification. The CAD tool was applied to 50 nodules (22 malignant, 28 benign) from high-resolution CT scans. 47 features, including 39 parenchymal features, were statistically significant, with both nodule and parenchyma features selected for classification, yielding an area under the ROC curve (AUC) of 0.935. This was compared to classification solely based on the nodule yielding an AUC of 0.917. These preliminary results show an increase in performance when the surrounding parenchyma is included in analysis. While modest, the improvement and large number of significant parenchyma features supports our hypothesis that the parenchyma contains meaningful data that can assist in CAD development.

Accurate quantification of vessel diameter in low-dose Computer Tomography (CT) images is important to study pulmonary diseases, in particular for the diagnosis of vascular diseases and the characterization of morphological vascular remodeling in Chronic Obstructive Pulmonary Disease (COPD). In this study, we objectively compare several vessel diameter estimation methods using a physical phantom. Five solid tubes of differing diameters (from 0.898 to 3.980 mm) were embedded in foam, simulating vessels in the lungs. To measure the diameters, we first extracted the vessels using either of two approaches: vessel enhancement using multi-scale Hessian matrix computation, or explicitly segmenting them using intensity threshold. We implemented six methods to quantify the diameter: three estimating diameter as a function of scale used to calculate the Hessian matrix; two calculating equivalent diameter from the crosssection area obtained by thresholding the intensity and vesselness response, respectively; and finally, estimating the diameter of the object using the Full Width Half Maximum (FWHM). We find that the accuracy of frequently used methods estimating vessel diameter from the multi-scale vesselness filter depends on the range and the number of scales used. Moreover, these methods still yield a significant error margin on the challenging estimation of the smallest diameter (on the order or below the size of the CT point spread function). Obviously, the performance of the thresholding-based methods depends on the value of the threshold. Finally, we observe that a simple adaptive thresholding approach can achieve a robust and accurate estimation of the smallest vessels diameter.