This PDF file contains the front matter associated with SPIE Proceedings Volume 6915, including the Title Page, Copyright information, Table of Contents, Introduction, and the Conference Committee listing.

The ability to acquire and store radiological images digitally has made this data available to mathematical and scientific methods. With the step from subjective interpretation to reproducible measurements and knowledge, it is also possible to develop and apply models that give additional information which is not directly visible in the data. In this context, it is important to know the characteristics and limitations of each model. Four characteristics assure the clinical relevance of models for computer-assisted diagnosis and therapy: ability of patient individual adaptation, treatment of errors and uncertainty, dynamic behavior, and in-depth evaluation. We demonstrate the development and clinical application of a model in the context of liver surgery. Here, a model for intrahepatic vascular structures is combined with individual, but in the degree of vascular details limited anatomical information from radiological images. As a result, the model allows for a dedicated risk analysis and preoperative planning of oncologic resections as well as for living donor liver transplantations. The clinical relevance of the method was approved in several evaluation studies of our medical partners and more than 2900 complex surgical cases have been analyzed since 2002.

In this study we investigated different feature selection methods for use in computer-aided mass detection. The data set we used (1357 malignant mass regions and 58444 normal regions) was much larger than used in previous research where feature selection did not directly improve the performance compared to using the entire feature set. We introduced a new performance measure to be used during feature selection, defined as the mean sensitivity in an interval of the free response operating characteristic (FROC) curve computed on a logarithmic scale. This measure is similar to the final validation performance measure we were optimizing. Therefore it was expected to give better results than more general feature selection criteria. We compared the performance of feature sets selected using the mean sensitivity of the FROC curve to sets selected using the Wilks' lambda statistic and investigated the effect of reducing the skewness in the distribution of the feature values before performing feature selection. In the case of Wilks' lambda, we found that reducing skewness had a clear positive effect, yielding performances similar or exceeding performances obtained when the entire feature set was used. Our results indicate that a general measure like Wilks' lambda selects better performing feature sets than the mean sensitivity of the FROC curve.

Classifier design for a given classification task needs to take into consideration both the complexity of the classifier and the size of the data set that is available for training the classifier. With limited training data, as often is the situation in computer-aided diagnosis of medical images, a classifier with simple structure (e.g., a linear classifier) is more robust and therefore preferred. We consider the two-class classification problem in which the feature data arise from two multivariate normal distributions. A linear function is used to combine the multi-dimensional feature vector onto a scalar variable. This scalar variable, however, is generally not an ideal decision variable unless the covariance matrices of the two classes are equal. We propose using the likelihood ratio of this scalar variable as a decision variable and, thus, generalizing the traditional classification paradigm to a hybrid two-stage procedure: a linear combination of the feature vector elements to form a scalar variable followed by a nonlinear, nonmonotic transformation that maps the scalar variable onto its likelihood ratio (i.e., the ideal decision variable, given the scalar variable). We show that the traditional Fisher's linear discriminant function is generally not the optimal linear function for the first stage in this two-stage paradigm. We further show that the optimal linear function can be obtained with a numerical optimization procedure using the area under the "proper" ROC curve as the objective function.

The purpose of this project is to study two Computer Aided Detection (CADe) systems for breast masses for digital tomosynthesis using reconstructed slices. This study used eighty human subject cases collected as part of on-going clinical trials at Duke University. Raw projections images were used to identify suspicious regions in the algorithm's high sensitivity, low specificity stage using a Difference of Gaussian filter. The filtered images were thresholded to yield initial CADe hits that were then shifted and added to yield a 3D distribution of suspicious regions. The initial system performance was 95% sensitivity at 10 false positives per breast volume. Two CADe systems were developed. In system A, the central slice located at the centroid depth was used to extract a 256 X 256 Regions of Interest (ROI) database centered at the lesion coordinates. For system B, 5 slices centered at the lesion coordinates were summed before the extraction of 256 × 256 ROIs. To avoid issues associated with feature extraction, selection, and merging, information theory principles were used to reduce false positives for both the systems resulting in a classifier performance of 0.81 and 0.865 Area Under Curve (AUC) with leave-one-case-out sampling. This resulted in an overall system performance of 87% sensitivity with 6.1 FPs/ volume and 85% sensitivity with 3.8 FPs/ volume for systems A and B respectively. This system therefore has the potential to detect breast masses in tomosynthesis data sets.

We are developing computer-aided diagnosis (CADx) methods for classification of masses on digital breast tomosynthesis mammograms (DBTs). A DBT data set containing 107 masses (56 malignant and 51 benign) collected at the Massachusetts General Hospital was used. The DBTs were obtained with a GE prototype system which acquired 11 projection views (PVs) over a 50-degree arc. We reconstructed the DBTs at 1-mm slice interval using a simultaneous algebraic reconstruction technique. The regions of interest (ROIs) containing the masses in the DBT volume and the corresponding ROIs on the PVs were identified. The mass on each slice or each PV was segmented by an active contour model. Spiculation measures, texture features, and morphological features were extracted from the segmented mass. Four feature spaces were formed: (1) features from the central DBT slice, (2) average features from 5 DBT slices centered at the central slice, (3) features from the central PV, and (4) average features from all 11 PVs. In each feature space, a linear discriminant analysis classifier with stepwise feature selection was trained and tested using a two loop leave-one-case-out procedure. The test Az of 0.91±0.03 from the 5-DBT-slice feature space was significantly (p=0.003) higher than that of 0.84±0.04 from the 1-DBT-slice feature space. The test Az of 0.83±0.04 from the 11-PV feature space was not significantly different (p=0.18) from that of 0.79±0.04 from the 1-PV feature space. The classification accuracy in the 5-DBT-slice feature space was significantly better (p=0.006) than that in the 11-PV feature space. The results demonstrate that the features of breast lesions extracted from the DBT slices may provide higher classification accuracy than those from the PV images.

Digital breast tomosynthesis (DBT) has emerged as a promising imaging modality for screening mammography. However, visually detecting micro-calcification clusters depicted on DBT images is a difficult task. Computer-aided detection (CAD) schemes for detecting micro-calcification clusters depicted on mammograms can achieve high performance and the use of CAD results can assist radiologists in detecting subtle micro-calcification clusters. In this study, we compared the performance of an available 2D based CAD scheme with one that includes a new grouping and scoring method when applied to both projection and reconstructed DBT images. We selected a dataset involving 96 DBT examinations acquired on 45 women. Each DBT image set included 11 low dose projection images and a varying number of reconstructed image slices ranging from 18 to 87. In this dataset 20 true-positive micro-calcification clusters were visually detected on the projection images and 40 were visually detected on the reconstructed images, respectively. We first applied the CAD scheme that was previously developed in our laboratory to the DBT dataset. We then tested a new grouping method that defines an independent cluster by grouping the same cluster detected on different projection or reconstructed images. We then compared four scoring methods to assess the CAD performance. The maximum sensitivity level observed for the different grouping and scoring methods were 70% and 88% for the projection and reconstructed images with a maximum false-positive rate of 4.0 and 15.9 per examination, respectively. This preliminary study demonstrates that (1) among the maximum, the minimum or the average CAD generated scores, using the maximum score of the grouped cluster regions achieved the highest performance level, (2) the histogram based scoring method is reasonably effective in reducing false-positive detections on the projection images but the overall CAD sensitivity is lower due to lower signal-to-noise ratio, and (3) CAD achieved higher sensitivity and higher false-positive rate (per examination) on the reconstructed images. We concluded that without changing the detection threshold or performing pre-filtering to possibly increase detection sensitivity, current CAD schemes developed and optimized for 2D mammograms perform relatively poorly and need to be re-optimized using DBT datasets and new grouping and scoring methods need to be incorporated into the schemes if these are to be used on the DBT examinations.

Digital tomosynthesis mammography (DTM) can provide quasi-3D structural information of the breast by reconstructing the breast volume from projection views (PV) acquired in a limited angular range. Our purpose is to design an effective classifier to distinguish breast masses from normal tissues in DTMs. A data set of 100 DTM cases collected with a GE first generation prototype DTM system at the Massachusetts General Hospital was used. We reconstructed the DTMs using a simultaneous algebraic reconstruction technique (SART). Mass candidates were identified by 3D gradient field analysis. Three approaches to distinguish breast masses from normal tissues were evaluated. In the 3D approach, we extracted morphological and run-length statistics texture features from DTM slices as input to a linear discriminant analysis (LDA) classifier. In the 2D approach, the raw input PVs were first preprocessed with a Laplacian pyramid multi-resolution enhancement scheme. A mass candidate was then forward-projected to the preprocessed PVs in order to determine the corresponding regions of interest (ROIs). Spatial gray-level dependence (SGLD) texture features were extracted from each ROI and averaged over 11 PVs. An LDA classifier was designed to distinguish the masses from normal tissues. In the combined approach, the LDA scores from the 3D and 2D approaches were averaged to generate a mass likelihood score for each candidate. The A_z values were 0.87±0.02, 0.86±0.02, and 0.91±0.02 for the 3D, 2D, and combined approaches, respectively. The difference between the A_z values of the 3D and 2D approaches did not achieve statistical significance. The performance of the combined approach was significantly (p<0.05) better than either the 3D or 2D approach alone. The combined classifier will be useful for false-positive reduction in computerized mass detection in DTM.

In this paper we propose a method to classify masses in digital breast tomosynthesis (DBT) datasets. First, markers of potential lesions are extracted and matched over the different projections. Then two level-set models are applied on each finding corresponding to spiculated and circumscribed mass assumptions respectively. The formulation of the active contours within this framework leads to several candidate contours for each finding. In addition, a membership value to the class contour is derived from the energy of the segmentation model, and allows associating several fuzzy contours from different projections to each set of markers corresponding to a lesion. Fuzzy attributes are computed for each fuzzy contour. Then the attributes corresponding to fuzzy contours associated to each set of markers are aggregated. Finally, these cumulated fuzzy attributes are processed by two distinct fuzzy decision trees in order to validate/invalidate the spiculated or circumscribed mass assumptions. The classification has been validated on a database of 23 real lesions using the leave-one-out method. An error classification rate of 9% was obtained with these data, which confirms the interest of the proposed approach.

Studies have demonstrated a relationship between mammographic texture and breast cancer risk. To date, texture analysis has been limited by tissue superimposition in mammography. Digital Breast Tomosynthesis (DBT) is a novel x-ray imaging modality in which 3D images of the breast are reconstructed from a limited number of source projections. Tomosynthesis alleviates the effect of tissue superimposition and offers the ability to perform tomographic texture analysis; having the potential to ultimately yield more accurate measures of risk. In this study, we analyzed texture in DBT and digital mammography (DM). Our goal was to compare tomographic versus mammographic texture characterization and evaluate the robustness of texture descriptors in reflecting characteristic parenchymal properties. We analyzed DBT and DM images from 40 women with recently detected abnormalities and/or previously diagnosed breast cancer. Texture features, previously shown to correlate with risk, were computed from the retroareolar region. We computed the texture correlation between (i) the DBT and DM, and (ii) between contralateral and ipsilateral breasts. The effect of the gray-level quantization on the observed correlations was investigated. Low correlation was detected between DBT and DM features. The correlation between contralateral and ipsilateral breasts was significant for both modalities, and overall stronger for DBT. We observed that the selection of the gray-level quantization algorithm affects the detected correlations. The strong correlation between contralateral and ipsilateral breasts supports the hypothesis that parenchymal properties appear to be inherent in an individual woman; the texture of the unaffected breast could potentially be used as a marker of risk.

Computed tomographic colonography (CTC) is a feasible and minimally invasive method for the detection of colorectal polyps and cancer screening. In current practice, a patient will be scanned twice during the CTC examination - once supine and once prone. In order to assist the radiologists in evaluating colon polyp candidates in both scans, we expect the computer aided detection (CAD) system can provide not only the locations of suspicious polyps, but also the possible matched pairs of polyps in two scans. In this paper, we propose a new automated matching method based on the extracted features of polyps by using principal component analysis (PCA) and Support Vector Machines (SVMs). Our dataset comes from the 104 CT scans of 52 patients with supine and prone positions collected from three medical centers. From it we constructed two groups of matched polyp candidates according to the size of true polyps: group A contains 12 true polyp pairs (> 9 mm) and 454 false pairs; group B contains 24 true polyp pairs (6-9 mm) and 514 false pairs. By using PCA, we reduced the dimensions of original data (with 157 attributes) to 30 dimensions. We did leave-one-patient-out test on the two groups of data. ROC analysis shows that it is easier to match bigger polyps than that of smaller polyps. On group A data, when false alarm probability is 0.18, the sensitivity of SVM achieves 0.83 which shows that automated matching of polyp candidates is practicable for clinical applications.

Computer-aided diagnosis systems have been shown to be feasible for polyp detection on computed tomography (CT) scans. After 3-D image segmentation and feature extraction, the dataset of colonic polyp candidates has large-scale and high dimension characteristics. In this paper, we propose a large-scale dimensionality reduction method based on Diffusion Map and Locally Linear Embedding for detection of polyps in CT colonography. By selecting partial data as landmarks, we first map the landmarks into a low dimensional embedding space using Diffusion Map. Then by using Locally Linear Embedding algorithm, non-landmark samples are embedded into the same low dimensional space according to their nearest landmark samples. The local geometry of samples is preserved in both the original space and the embedding space. We applied the proposed method called DMLLE to a colonic polyp dataset which contains 58336 candidates (including 85 6-9mm true polyps) with 155 features. Visual inspection shows that true polyps with similar shapes are mapped to close vicinity in the low dimensional space. FROC analysis shows that SVM with DMLLE achieves higher sensitivity with lower false positives per patient than that of SVM using all features. At the false positives of 8 per patient, SVM with DMLLE improves the average sensitivity from 64% to 75% for polyps whose sizes are in the range from 6 mm to 9 mm (p < 0.05). This higher sensitivity is comparable to unaided readings by trained radiologists.

Electronic cleansing (EC) is a method that segments fecal material tagged by an X-ray-opaque oral contrast agent in CT colonography (CTC) images, and effectively removes the material for digitally cleansing the colon. In this study, we developed a novel EC method, called mosaic decomposition, for reduction of the artifacts due to incomplete cleansing of heterogeneously tagged fecal material in reduced- or non-cathartic fecal-tagging CTC examinations. In our approach, a segmented colonic lumen, including the tagged regions, was first partitioned into a set of local homogeneous regions by application of a watershed transform to the gradient of the CTC images. Then, each of the local homogeneous regions was classified into five different material classes, including air, soft tissue, tagged feces, air bubbles, and foodstuff, based on texture features of the tile. A single index, called a soft-tissue index, is formulated for differentiation of these materials from the submerged solid soft-tissue structures such as polyps and folds. Here, a larger value of the index indicates a higher likelihood of soft tissue. Then, EC is performed by first initializing the level-set front with the classified tagged regions, and the front is evolved by use of a speed function that was designed, based on the soft-tissue index, to reserve the submerged soft-tissue structures while suppressing the air bubbles and foodstuff. Visual assessment and application of our computer-aided detection (CAD) of polyps showed that the use of our new EC method substantially improved the detection performance of CAD, indicating the effectiveness of our EC method in reducing incomplete cleansing artifacts.

Selecting a set of relevant features is a crucial step in the process of building robust classifiers. Searching all possible subsets of features is computationally impractical for large number of features. Generally, classifiers are used for the evaluation of the separability of a certain feature subset. The performance of these classifiers depends on some predefined parameters. However, the choice of these parameters for a given classifier is influenced by the given feature subset and vice versa. The computational cost for feature selection would be largely increased by including the selection of optimal parameters for the classifier (for each subset). This paper attempts to tackle the problem by introducing genetic algorithms (GAs) to combine the processes. The proposed approach can choose the most relevant features from a feature set whilst simultaneously optimising the parameters of the classifier. Its performance was tested on a colon polyp database from a cohort study using a weighted support vector machine (SVM) classifier. As a general approach, other classifiers such as artificial neural networks (ANN) and decision trees could be used. This approach could also be applied to other classification problems such as other computer aided detection/diagnosis applications.

A major challenge in computer-aided detection (CAD) of polyps in CT colonography (CTC) is the detection of "difficult" polyps which radiologists are likely to miss. Our purpose was to develop a CAD scheme incorporating massive-training artificial neural networks (MTANNs) and to evaluate its performance on false-negative (FN) cases in a large multicenter clinical trial. We developed an initial polyp-detection scheme consisting of colon segmentation based on CT value-based analysis, detection of polyp candidates based on morphologic analysis, and quadratic discriminant analysis based on 3D pattern features for classification. For reduction of false-positive (FP) detections, we developed multiple expert 3D MTANNs designed to differentiate between polyps and seven types of non-polyps. Our independent database was obtained from CTC scans of 155 patients with polyps from a multicenter trial in which 15 medical institutions participated nationwide. Among them, about 45% patients received FN interpretations in CTC. For testing our CAD, 14 cases with 14 polyps/masses were randomly selected from the FN cases. Lesion sizes ranged from 6-35 mm, with an average of 10 mm. The initial CAD scheme detected 71.4% (10/14) of "missed" polyps, including sessile polyps and polyps on folds, with 18.9 (264/14) FPs per case. The MTANNs removed 75% (197/264) of the FPs without loss of any true positives; thus, the performance of our CAD scheme was improved to 4.8 (67/14) FPs per case. With our CAD scheme incorporating MTANNs, 71.4% of polyps "missed" by radiologists in the trial were detected correctly, with a reasonable number of FPs.

We developed a self-assessment method in which the CADx system provided a confidence level for its lesion segmentations. The self-assessment was performed by a fuzzy-inference system based on 4 computer-extracted features of the computer-segmented lesions in a leave-one-case-out evaluation protocol. In instances where the initial segmentation received a low assessment rating, lesions were re-segmented using the same segmentation method but based on a user-defined region-of-interest. A total of 542 cases with 1133 lesions were collected in this study, and we focused here on the 97 normal lymph nodes in this dataset since these pose challenges for automated segmentation due to their inhomogeneous appearance. The percentage of all lesions with satisfactory segmentation (i.e., normalized overlap with the radiologist-delineated lesion >=0.3) was 85%. For normal lymph nodes, however, this percentage was only 36%. Of the lymph nodes, 53 received a low confidence rating (<0.3) for their initial segmentation. When those lymph nodes were re-segmented, the percentage with a satisfactory segmentation improved to 80.0%. Computerassessed confidence levels demonstrated potential to 1) help radiologists decide whether to use or disregard CADx output, and 2) provide a guide for improvement of lesion segmentation.

Segmentation of masses is the first step in most computer-aided diagnosis (CAD) systems for characterization of breast masses as malignant or benign. In this study, we designed an automated method for segmentation of masses on ultrasound (US) images. The method automatically estimated an initial contour based on a manually-identified point approximately at the mass center. A two-stage active contour (AC) method iteratively refined the initial contour and performed self-examination and correction on the segmentation result. To evaluate our method, we compared it with manual segmentation by an experienced radiologists (R1) on a data set of 226 US images containing biopsy-proven masses from 121 patients (44 malignant and 77 benign). Four performance measures were used to evaluate the segmentation accuracy; two measures were related to the overlap between the computer and radiologist segmentation, and two were related to the area difference between the two segmentation results. To compare the difference between the segmentation results by the computer and R1 to inter-observer variation, a second radiologist (R2) also manually segmented all masses. The two overlap measures between the segmentation results by the computer and R1 were 0.87+ 0.16 and 0.73+ 0.17 respectively, indicating a high agreement. However, the segmentation results between two radiologists were more consistent. To evaluate the effect of the segmentation method on classification accuracy, three feature spaces were formed by extracting texture, width-to-height, and posterior shadowing features using the computer segmentation, R1's manual segmentation, and R2's manual segmentation. A linear discriminant analysis classifier using stepwise feature selection was tested and trained by a leave-one-case-out method to characterize the masses as malignant or benign. For case-based classification, the area A_z under the test receiver operating characteristic (ROC) curve was 0.90±0.03, 0.87±0.03 and 0.87±0.03 for the feature sets based on computer segmentation, R1's manual segmentation, and R2's manual segmentation, respectively.

Ultrasound has been an important imaging technique for detecting breast tumors. As opposed to the conventional B-mode image, the ultrasound elastography is a new technique for imaging the elasticity and applied to detect the stiffness of tissues. The red region of color elastography indicates the soft tissue and the blue one indicates the hard tissue, and the harder tissue usually is classified to malignancy. In this paper, we proposed a CAD system on elastography to measure whether this system is effective and accurate to classify the tumor into benign and malignant. According to the features of elasticity, the color elastography was transferred to HSV color space and extracted meaningful features from hue images. Then the neural network was utilized in multiple features to distinguish tumors. In this experiment, there are 180 pathology-proven cases including 113 benign and 67 malignant cases used to examine the classification. The results of the proposed system showed an accuracy of 83.89%, a sensitivity of 85.07% and a specificity of 83.19%. Compared with the physician's diagnosis, an accuracy of 78.33%, a sensitivity of 53.73% and a specificity of 92.92%, the proposed CAD system had better performance. Moreover, the agreement of the proposed CAD system and the physician's diagnosis was calculated by kappa statistics, the kappa 0.54 indicated there is a fair agreement of observers.

The kinetic characteristics of tissue in dynamic contrast-enhanced magnetic resonance imaging data are an important source of information for the differentiation of benign and malignant lesions. Kinetic curves measured for each lesion voxel allow to infer information about the state of the local tissue. As a whole, they reflect the heterogeneity of the vascular structure within a lesion, an important criterion for the preoperative classification of lesions. Current clinical practice in analysis of tissue kinetics however is mainly based on the evaluation of the "most-suspect curve", which is only related to a small, manually or semi-automatically selected region-of-interest within a lesion and does not reflect any information about tissue heterogeneity. We propose a new method which exploits the full range of kinetic information for the automatic classification of lesions. Instead of breaking down the large amount of kinetic information to a single curve, each lesion is considered as a probability distribution in a space of kinetic features, efficiently represented by its kinetic signature obtained by adaptive vector quantization of the corresponding kinetic curves. Dissimilarity of two signatures can be objectively measured using the Mallows distance, which is a metric defined on probability distributions. The embedding of this metric in a suitable kernel function enables us to employ modern kernel-based machine learning techniques for the classification of signatures. In a study considering 81 breast lesions, the proposed method yielded an A_z value of 0.89±0.01 for the discrimination of benign and malignant lesions in a nested leave-one-lesion-out evaluation setting.

We propose a new model for pharmacokinetic analysis based on the one proposed by Tofts. Our model both eliminates the need for estimating the Arterial Input Function (AIF) and normalizes analysis so that comparisons across patients can be performed. Previous methods have attempted to circumvent the AIF estimation by using the pharmacokinetic parameters of multiple reference regions (RR). Viewing anatomical structures as filters, pharmacokinetic analysis tells us that 'similar' structures will be similar filters. By cascading the inverse filter at a RR with the filter at the voxel being analyzed, we obtain a transfer function relating the concentration of a voxel to that of the RR. We show that this transfer function simplifies into a five-parameter nonlinear model with no reference to the AIF. These five parameters are combinations of the three parameters of the original model at the RR and the region of interest. Contrary to existing methods, ours does not require explicit estimation of the pharmacokinetic parameters of the RR. Also, cascading filters in the frequency domain allows us to manipulate more complex models, such as accounting for the vascular tracer component. We believe that our model can improve analysis across MR parameters because the analyzed and reference enhancement series are from the same image. Initial results are promising with the proposed model parameters exhibiting values that are more consistent across lesions in multiple patients. Additionally, our model can be applied to multiple voxels to estimate the original pharmacokinetic parameters as well as the AIF.

Today, mammography is recognized as the most effective technique for breast cancer screening. Unfortunately, the low positive predictive value of breast biopsy examinations resulting from mammogram interpretation leads to many unnecessary biopsies performed on benign lesions. In the last years, several computer assisted diagnosis (CADx) systems have been proposed with the goal to assist the radiologist in the discrimination of benign and malignant breast lesions and thus to reduce the high number of unnecessary biopsies. In this paper we present a novel, knowledge-based approach to the computer aided discrimination of mammographic mass lesions that uses computer-extracted attributes of mammographic masses and clinical data as input attributes to a case-based reasoning system. Our approach emphasizes a transparent reasoning process which is important for the acceptance of a CADx system in clinical practice. We evaluate the performance of the proposed system on a large publicly available mammography database using receiver operating characteristic curve analysis. Our results indicate that the proposed CADx system has the potential to significantly reduce the number of unnecessary breast biopsies in clinical practice.

Motion artifacts in cardiac CT are an obstacle to obtaining diagnostically usable images. Although phase-specific reconstruction can produce images with improved assessability (image quality), this requires that the radiologist spend time and effort evaluating multiple image sets from reconstructions at different phases. In this study, ordinal logistic regression (OLR) and artificial neural network (ANN) models were used to automatically assign assessability to images of coronary calcified plaques obtained using a physical, dynamic cardiac phantom. 350 plaque images of 7 plaques from five data sets (heart rates 60, 60, 70, 80, 90) and ten phases of reconstruction were obtained using standard cardiac CT scanning parameters on a Phillips Brilliance 64-channel clinical CT scanner. Six features of the plaques (velocity, acceleration, edge-based volume, threshold-based volume, sphericity, and standard deviation of intensity) as well as mean feature values and heart rate were used for training the OLR and ANN in a round-robin re-sampling scheme based on training and testing groups with independent plaques. For each image, an ordinal assessability index rating on a 1-5 scale was assigned by a cardiac radiologist (D.B.) for use as a "truth" in training the OLR and ANN. The mean difference between the assessability index truth and model-predicted assessability index values was +0.111 with SD=0.942 for the OLR and +0.143 with SD=0.916 for the ANN. Comparing images from the repeat 60 bpm scans gave concordance correlation coefficients (CCCs) of 0.794 [0.743, 0.837] (value, 95% CI) for the radiologist assigned values, 0.894 [0.856, 0.922] for the OLR, and 0.861 [0.818, 0.895] for the ANN. Thus, the variability of the OLR and ANN assessability index values appear to lie within the variability of the radiologist assigned values.

We have developed a new single photon emission computerized tomography (SPECT) hotspot quantification method incorporating extra cardiac activity correction and hotspot normal limit estimation. The method was validated for estimation accuracy of myocardial tracer focal uptake in a chronic canine model of myocardial infarction (MI). Dogs (n = 4) at 2 weeks post MI were injected with Tl-201 and a Tc-99m-labeled hotspot tracer targeted at matrix metalloproteinases (MMPs). An external point source filled with Tc-99m was used for a reference of absolute radioactivity. Dual-isotope (Tc-99m/Tl-201) SPECT images were acquired simultaneously followed by an X-ray CT acquisition. Dogs were sacrificed after imaging for myocardial gamma well counting. Images were reconstructed with CT-based attenuation correction (AC) and without AC (NAC) and were quantified using our quantification method. Normal limits for myocardial hotspot uptake were estimated based on 3 different schemes: maximum entropy, meansquared-error minimization (MSEM) and global minimization. Absolute myocardial hotspot uptake was quantified from SPECT images using the normal limits and compared with well-counted radioactivity on a segment-by-segment basis (n = 12 segments/dog). Radioactivity was expressed as % injected dose (%ID). There was an excellent correlation (r = 0.78-0.92) between the estimated activity (%ID) derived using the SPECT quantitative approach and well-counting, independent of AC. However, SPECT quantification without AC resulted in the significant underestimation of radioactivity. Quantification using SPECT with AC and the MSEM normal limit yielded the best results compared with well-counting. In conclusion, focal myocardial "hotspot" uptake of a targeted radiotracer can be accurately quantified in vivo using a method that incorporates SPECT imaging with AC, an external reference, background scatter compensation, and a suitable normal limit. This hybrid SPECT/CT approach allows for the serial non-invasive quantitative evaluation of molecular targeted tracers in the heart.

Cardiac CT has been reported to be an effective means for clinical diagnosis of coronary artery plaque disease. We are investigating the feasibility of developing a computer-assisted image analysis (CAA) system to assist radiologist in detection of coronary artery plaque disease in ECG-gated cardiac CT scans. The heart region was first extracted using morphological operations and an adaptive EM thresholding method. Vascular structures in the heart volume were enhanced by 3D multi-scale filtering and analysis of the eigenvalues of Hessian matrices using a vessel enhancement response function specially designed for coronary arteries. The enhanced vascular structures were then segmented by an EM estimation method. Finally, our newly developed 3D rolling balloon vessel tracking method (RBVT) was used to track the segmented coronary arteries. Starting at two manually identified points located at the origins of left and right coronary artery (LCA and RCA), the RBVT method moved a sphere of adaptive diameter along the vessels, tracking the vessels and identifying its branches automatically to generate the left and right coronary arterial trees. Ten cardiac CT scans that contained various degrees of coronary artery diseases were used as test data set for our vessel segmentation and tracking method. Two experienced thoracic radiologists visually examined the computer tracked coronary arteries on a graphical interface to count untracked false-negative (FN) branches (segments). A total of 27 artery segments were identified to be FNs in the 10 cases, ranging from 0 to 6 FN segments in each case. No FN artery segment was found in 2 cases.

Neuroblastoma is a cancer of the nervous system and one of the most common tumors in children. In clinical practice, pathologists examine the haematoxylin and eosin (H&E) stained tissue slides under the microscope for the diagnosis. According to the International Neuroblastoma Classification System, neuroblastoma tumors are categorized into favorable and unfavorable histologies. The subsequent treatment planning is based on this classification. However, this qualitative evaluation is time consuming, prone to error and subject to inter- and intra-reader variations and sampling bias. To overcome these shortcomings, we are developing a computerized system for the quantitative analysis of neuroblastoma slides. In this study, we present a novel image analysis system to determine the degree of stromal development from digitized whole-slide neuroblastoma samples. The developed method uses a multi-resolution approach that works similar to how pathologists examine slides. Due to their very large resolutions, the whole-slide images are divided into non-overlapping image tiles and the proposed image analysis steps are applied to each image tile using a parallel computation infrastructure developed earlier by our group. The computerized system classifies image tiles as stroma-poor or stroma-rich subtypes using texture characteristics. The developed method has been independently tested on 20 whole-slide neuroblastoma slides and it has achieved 95% classification accuracy.

We proposed a novel approach to automatic classification of Diabetic Retinopathy (DR) images and retrieval of clinically-relevant DR images from a database. Given a query image, our approach first classifies the image into one of the three categories: microaneurysm (MA), neovascularization (NV) and normal, and then it retrieves DR images that are clinically-relevant to the query image from an archival image database. In the classification stage, the query DR images are classified by the Multi-class Multiple-Instance Learning (McMIL) approach, where images are viewed as bags, each of which contains a number of instances corresponding to non-overlapping blocks, and each block is characterized by low-level features including color, texture, histogram of edge directions, and shape. McMIL first learns a collection of instance prototypes for each class that maximizes the Diverse Density function using Expectation- Maximization algorithm. A nonlinear mapping is then defined using the instance prototypes and maps every bag to a point in a new multi-class bag feature space. Finally a multi-class Support Vector Machine is trained in the multi-class bag feature space. In the retrieval stage, we retrieve images from the archival database who bear the same label with the query image, and who are the top K nearest neighbors of the query image in terms of similarity in the multi-class bag feature space. The classification approach achieves high classification accuracy, and the retrieval of clinically-relevant images not only facilitates utilization of the vast amount of hidden diagnostic knowledge in the database, but also improves the efficiency and accuracy of DR lesion diagnosis and assessment.

Classification plays an important role in the reduction of false positives in many computer aided detection and diagnosis methods. The difficulty of classifying polyps lies in the variation of possible polyp shapes and sizes and the imbalance between the number of polyp and non-polyp regions available in the training data. CAD schemes for medical applications demand high levels of sensitivity even at the expense of keeping a certain number of false positives. In this paper, we investigate some state-of-the-art solutions to the imbalanced data problem: Synthetic Minority Over-sampling Technique (SMOTE) and weighted Support Vector Machines (SVM). We tested these methods using a diverse database of CT colonography, which included a wide spectrum of dificult cases to detect polyps. We performed several experiments with different combinations of over-sampling techniques on training data. The results demonstrated that SVMs have achieved much better performance over C4.5 with different over-sampling techniques. Also, the results show that weighted SVM without over-sampling can achieve comparable performance in terms of sensitivity and specificity to conventional SVM combined with the over-sampling approach.

This paper presents an adaptive neural network based Fuzzy Inference System (ANFIS) to reduce the false positive (FP) rate of detected colonic polyps in Computed Tomography Colonography (CTC) scans. Extracted fuzzy rules establish linguistically interpretable relationships in the data that are easy to understand, validate, and extend. The system takes several features identified from regions extracted by a segmentation algorithm and decides whether the regions are true polyps. In the training phase, subtractive clustering is used to down-sample the negative regions in order to get balanced data. The rule extraction method is based on estimating clusters in the data using the subtractive clustering algorithm; each cluster obtained corresponds to a fuzzy rule that maps a region in the input space to an output class. After the number of rules and initial rule parameters are obtained by cluster estimation, the rule parameters are optimized using a hybrid learning algorithm which is a combination of least-squares estimation with back propagation. The evolved Sugeno-type FIS has been tested on a total of 129 scans with 99 polyps of sizes 5-15 mm by experienced radiologists. The results indicate that for 93% detection sensitivity (on polyps), the evolved FIS method is able to remove 88% of FPs generated by the segmentation algorithm leaving 7.5 FP per scan. The high sensitivity rate of our results show the promise of neuro-fuzzy classifiers as an aid for interpreting CTC examinations.

One of the key factors which may lead to a greater patient compliance in virtual colonoscopy is a well tolerated bowel preparation; there is evidence that this may be obtained by faecal tagging techniques. In so-called "sameday faecal tagging" (SDFT) preparations, iodine contrast material is administered on the day of the exam in a hospital ward. The administration of oral contrast on the day of the procedure, however, often results in a less homogenous marking of faecal residues and computer aided-detection (CAD) systems must be able to treat these kinds of preparations. The aim of this work is to present a CAD scheme capable of achieving good performances on CT datasets obtained using SDFT, both in terms of sensitivity and specificity. The electronically cleansed datasets is processed by a scheme composed of three steps: colon surface extraction, polyp candidate segmentation through curvature-based features and discrimination between true polyps and false alarms. The system was evaluated on a dataset including 102 patients from three different centers. A specificity of 8.2 false positives per scan was obtained at a 100% sensitivity for polyps larger than 10 mm. In conclusion CAD schemes for SDFT may be designed to obtain high performances.

Current techniques for localization of prostatic adenocarcinoma (CaP) via blinded trans-rectal ultrasound biopsy are associated with a high false negative detection rate. While high resolution endorectal in vivo Magnetic Resonance (MR) prostate imaging has been shown to have improved contrast and resolution for CaP detection over ultrasound, similarity in intensity characteristics between benign and cancerous regions on MR images contribute to a high false positive detection rate. In this paper, we present a novel unsupervised segmentation method that employs manifold learning via consensus schemes for detection of cancerous regions from high resolution 1.5 Tesla (T) endorectal in vivo prostate MRI. A significant contribution of this paper is a method to combine multiple weak, lower-dimensional representations of high dimensional feature data in a way analogous to classifier ensemble schemes, and hence create a stable and accurate reduced dimensional representation. After correcting for MR image intensity artifacts, such as bias field inhomogeneity and intensity non-standardness, our algorithm extracts over 350 3D texture features at every spatial location in the MR scene at multiple scales and orientations. Non-linear dimensionality reduction schemes such as Locally Linear Embedding (LLE) and Graph Embedding (GE) are employed to create multiple low dimensional data representations of this high dimensional texture feature space. Our novel consensus embedding method is used to average object adjacencies from within the multiple low dimensional projections so that class relationships are preserved. Unsupervised consensus clustering is then used to partition the objects in this consensus embedding space into distinct classes. Quantitative evaluation on 18 1.5 T prostate MR data against corresponding histology obtained from the multi-site ACRIN trials show a sensitivity of 92.65% and a specificity of 82.06%, which suggests that our method is successfully able to detect suspicious regions in the prostate.

Computer-aided diagnosis (CAD) systems for the detection of cancer in medical images require precise labeling of training data. For magnetic resonance (MR) imaging (MRI) of the prostate, training labels define the spatial extent of prostate cancer (CaP); the most common source for these labels is expert segmentations. When ancillary data such as whole mount histology (WMH) sections, which provide the gold standard for cancer ground truth, are available, the manual labeling of CaP can be improved by referencing WMH. However, manual segmentation is error prone, time consuming and not reproducible. Therefore, we present the use of multimodal image registration to automatically and accurately transcribe CaP from histology onto MRI following alignment of the two modalities, in order to improve the quality of training data and hence classifier performance. We quantitatively demonstrate the superiority of this registration-based methodology by comparing its results to the manual CaP annotation of expert radiologists. Five supervised CAD classifiers were trained using the labels for CaP extent on MRI obtained by the expert and 4 different registration techniques. Two of the registration methods were affi;ne schemes; one based on maximization of mutual information (MI) and the other method that we previously developed, Combined Feature Ensemble Mutual Information (COFEMI), which incorporates high-order statistical features for robust multimodal registration. Two non-rigid schemes were obtained by succeeding the two affine registration methods with an elastic deformation step using thin-plate splines (TPS). In the absence of definitive ground truth for CaP extent on MRI, classifier accuracy was evaluated against 7 ground truth surrogates obtained by different combinations of the expert and registration segmentations. For 26 multimodal MRI-WMH image pairs, all four registration methods produced a higher area under the receiver operating characteristic curve compared to that obtained from expert annotation. These results suggest that in the presence of additional multimodal image information one can obtain more accurate object annotations than achievable via expert delineation despite vast differences between modalities that hinder image registration.

In this study, we investigate the diagnostic performance of our CAD system when discriminating prostate cancer from benign lesions and normal peripheral zone using registered multi-modal images. We have developed a method that automatically extracts quantitative T2 values out of acquired T2-w images and evaluated its additional value to the discriminating performance of our CAD system. This study addresses 2 issues when using both T2-w and dynamic MR images for the characterization of prostate lesions. Firstly, T2-w images do not provide quantitative values, and secondly, images can be misaligned due to patient movements. To compensate, a mutual information registration strategy is performed after which T2 values are estimated using the acquired proton density images. From the resulted quantitative T2 maps as well as the dynamic images relevant features were extracted for training a support vector machine as classfier. The output of the classifier was used as a measure of likelihood of malignancy. General performance of the scheme was evaluated using the area under the ROC curve. We conclude that it is feasible to automatically extract diagnostic T2 values out of acquired T2-w images. Furthermore, a discriminating performance of 0.75 (0.66-0.85) was obtained when only using T2-values as feature. Combining the T2 values with pharmacokinetic parameters did not increase diagnostic performance in a pilot study.

This paper presents a new computer-aided detection scheme for lung nodules attached to the pleural or mediastinal surface in low dose CT scans. First the lungs are automatically segmented and smoothed. Any connected set of voxels attached to the wall - with each voxel above minus 500 HU and the total object within a specified volume range - was considered a candidate finding. For each candidate, a refined segmentation was computed using morphological operators to remove attached structures. For each candidate, 35 features were defined, based on their position in the lung and relative to other structures, and the shape and density within and around each candidate. In a training procedure an optimal set of 15 features was determined with a k-nearest-neighbor classifier and sequential floating forward feature selection. The algorithm was trained with a data set of 708 scans from a lung cancer screening study containing 224 pleural nodules and tested on an independent test set of 226 scans from the same program with 58 pleural nodules. The algorithm achieved a sensitivity of 52% with an average of 0.76 false positives per scan. At 2.5 false positive marks per scan, the sensitivity increased to 80%.

We developed a computer-aided diagnostic (CAD) scheme for detection of lung nodules in CT, and investigated its performance levels for nodules in different size and pattern groups. Our database consisted of 117 thin-slice CT scans with 153 nodules. There were 68 (44.4%) small, 52 (34.0%) medium-sized, and 33 (21.6%) large nodules; 101 (66.0%) solid and 52 (34.0%) nodules with ground glass opacity (GGO) in the database. Our CAD scheme consisted of lung segmentation, selective nodule enhancement, initial nodule detection, accurate nodule segmentation, and feature extraction and analysis techniques. We employed a case-based four-fold cross-validation method to evaluate the performance levels of our CAD scheme. We detected 87% of nodules (small: 74%, medium-sized: 98%, large: 94%; solid: 85%, GGO: 90%) with 6.5 false positives per scan; 82% of nodules (small: 68%, medium-sized: 94%, large: 91%; solid: 78%, GGO: 89%) with 2.8 false positives per scan; and 77% of nodules (small: 63%, medium-sized: 90%, large: 89%; solid: 71%, GGO: 89%) with 1.5 false positives per scan. Our CAD scheme achieved a higher sensitivity for GGO nodules than for solid nodules, because most of small nodules were solid. In conclusion, our CAD scheme achieved a low false positive rate and a relatively high detection rate for nodules with a large variation in size and pattern.

Chest X-ray (CXR) data is a 2D projection image. The main drawback of such an image is that each pixel of it represents a volumetric integration. This poses a challenge in detection and estimation of nodules and their characteristics. Due to human anatomy there are a lot of lung structures which can be falsely identified as nodules in a projection data. Detection of nodules with a large number of false positives (FP) adds more work for the radiologists. With the help of CAD algorithms we aim to identify regions which cause higher FP readings or provide additional information for nodule detection based on the human anatomy. Different lung regions have different image characteristics we take advantage of this and propose an automatic lung partitioning into vessel, apical, basal and exterior pulmonary regions. Anatomical landmarks like aortic arch and end of cardiac-notch along-with inter intra-rib width and their shape characteristics were used for this partitioning. Likelihood of FPs is more in vessel, apical and exterior pulmonary regions due to rib-crossing, overlap of vessel with rib and vessel branching. For each of these three cases, special features were designed based on histogram of rib slope and the structural properties of rib segments information. These features were assigned different weights based on the partitioning. An experiment was carried out using a prototype CAD system 150 routine CXR studies were acquired from three institutions (24 negatives, rest with one or more nodules). Our algorithm provided a sensitivity of 70.4% with 5 FP/image for cross-validation without partition. Inclusion of the proposed techniques increases the sensitivity to 78.1% with 4.1 FP/image.

Accurate nodule volume estimation is necessary in order to estimate the clinically relevant growth rate or change in size over time. An automated nodule volume-measuring algorithm was applied to a set of pulmonary nodules that were documented by the Lung Image Database Consortium (LIDC). The LIDC process model specifies that each scan is assessed by four experienced thoracic radiologists and that boundaries are to be marked around the visible extent of the nodules for nodules 3 mm and larger. Nodules were selected from the LIDC database with the following inclusion criteria: (a) they must have a solid component on a minimum of three CT image slices and (b) they must be marked by all four LIDC radiologists. A total of 113 nodules met the selection criterion with diameters ranging from 3.59 mm to 32.68 mm (mean 9.37 mm, median 7.67 mm). The centroid of each marked nodule was used as the seed point for the automated algorithm. 95 nodules (84.1%) were correctly segmented, but one was considered not meeting the first selection criterion by the automated method; for the remaining ones, eight (7.1%) were structurally too complex or extensively attached and 10 (8.8%) were considered not properly segmented after a simple visual inspection by a radiologist. Since the LIDC specifications, as aforementioned, instruct radiologists to include both solid and sub-solid parts, the automated method core capability of segmenting solid tissues was augmented to take into account also the nodule sub-solid parts. We ranked the distances of the automated method estimates and the radiologist-based estimates from the median of the radiologist-based values. The automated method was in 76.6% of the cases closer to the median than at least one of the values derived from the manual markings, which is a sign of a very good agreement with the radiologists' markings.

Computer aided characterization aims to support the differential diagnosis of indeterminate pulmonary nodules. A number of published studies have correlated automatically computed features from image processing with clinical diagnoses of malignancy vs. benignity. Often, however, a high number of features was trained on a relatively small number of diagnosed nodules, raising a certain skepticism as to how salient and numerically robust the various features really are. On the way towards computer aided diagnosis which is trusted in clinical practice, the credibility of the individual numerical features has to be carefully established. Nodule volume is the most crucial parameter for nodule characterization, and a number of studies are testing its repeatability. Apart from functional parameters (such as dynamic CT enhancement and PET uptake values), the next most widely used parameter is the surface characteristic (vascularization, spicularity, lobulation, smoothness). In this study, we test the repeatability of two simple surface smoothness features which can discriminate between smoothly delineated nodules and those with a high degree of surface irregularity. Robustness of the completely automatically computed features was tested with respect to the following aspects: (a) repeated CT scan of the same patient with equal dose, (b) repeated CT scan with much lower dose and much higher noise, (c) repeated automatic segmentation of the nodules using varying segmentation parameters, resulting in differing nodule surfaces. The tested nodules (81) were all solid or partially solid and included a high number of sub- and juxtapleural nodules. We found that both tested surface characterization features correlated reasonably well with each other (80%), and that in particular the mean-surface-shape-index showed an excellent repeatability: 98% correlation between equal dose CT scans, 93% between standard-dose and low-dose scan (without systematic shift), and 97% between varying HU-threshold of the automatic segmentation, which makes it a reliable feature to be used in computer aided diagnosis.

A computerized system for segmenting lesions in head and neck CT scans was developed to assist radiologists in estimation of the response to treatment of malignant lesions. The system performs 3D segmentations based on a level set model and uses as input an approximate bounding box for the lesion of interest. In this preliminary study, CT scans from a pre-treatment exam and a post one-cycle chemotherapy exam of 13 patients containing head and neck neoplasms were used. A radiologist marked 35 temporal pairs of lesions. 13 pairs were primary site cancers and 22 pairs were metastatic lymph nodes. For all lesions, a radiologist outlined a contour on the best slice on both the pre- and post treatment scans. For the 13 primary lesion pairs, full 3D contours were also extracted by a radiologist. The average pre- and post-treatment areas on the best slices for all lesions were 4.5 and 2.1 cm², respectively. For the 13 primary site pairs the average pre- and post-treatment primary lesions volumes were 15.4 and 6.7 cm³ respectively. The correlation between the automatic and manual estimates for the pre-to-post-treatment change in area for all 35 pairs was r=0.97, while the correlation for the percent change in area was r=0.80. The correlation for the change in volume for the 13 primary site pairs was r=0.89, while the correlation for the percent change in volume was r=0.79. The average signed percent error between the automatic and manual areas for all 70 lesions was 11.0±20.6%. The average signed percent error between the automatic and manual volumes for all 26 primary lesions was 37.8±42.1%. The preliminary results indicate that the automated segmentation system can reliably estimate tumor size change in response to treatment relative to radiologist's hand segmentation.

Response assessment of cancer therapy is a crucial component towards a more effective and patient individualized cancer therapy. Integrated PET/CT systems provide the opportunity to combine morphologic with functional information. However, dealing simultaneously with several PET/CT scans poses a serious workflow problem. It can be a difficult and tedious task to extract response criteria based upon an integrated analysis of PET and CT images and to track these criteria over time. In order to improve the workflow for serial analysis of PET/CT scans we introduce in this paper a fast lesion tracking algorithm. We combine a global multi-resolution rigid registration algorithm with a local block matching and a local region growing algorithm. Whenever the user clicks on a lesion in the base-line PET scan the course of standardized uptake values (SUV) is automatically identified and shown to the user as a graph plot. We have validated our method by a data collection from 7 patients. Each patient underwent two or three PET/CT scans during the course of a cancer therapy. An experienced nuclear medicine physician manually measured the courses of the maximum SUVs for altogether 18 lesions. As a result we obtained that the automatic detection of the corresponding lesions resulted in SUV measurements which are nearly identical to the manually measured SUVs. Between 38 measured maximum SUVs derived from manual and automatic detected lesions we observed a correlation of 0.9994 and a average error of 0.4 SUV units.

Cirrhosis of the liver is a chronic disease. It is characterized by the presence of widespread nodules and fibrosis in the liver which results in characteristic texture patterns. Computerized analysis of hepatic texture patterns is usually based on regions-of-interest (ROIs). However, not all ROIs are typical representatives of the disease stage of the liver from which the ROIs originated. This leads to uncertainties in the ROI labels (diseased or non-diseased). On the other hand, supervised classifiers are commonly used in determining the assignment rule. This presents a problem as the training of a supervised classifier requires the correct labels of the ROIs. The main purpose of this paper is to investigate the use of an unsupervised classifier, the k-means clustering, in classifying ROI based data. In addition, a procedure for generating a receiver operating characteristic (ROC) curve depicting the classification performance of k-means clustering is also reported. Hepatic MRI images of 44 patients (16 cirrhotic; 28 non-cirrhotic) are used in this study. The MRI data are derived from gadolinium-enhanced equilibrium phase images. For each patient, 10 ROIs selected by an experienced radiologist and 7 texture features measured on each ROI are included in the MRI data. Results of the k-means classifier are depicted using an ROC curve. The area under the curve (AUC) has a value of 0.704. This is slightly lower than but comparable to that of LDA and ANN classifiers which have values 0.781 and 0.801, respectively. Methods in constructing ROC curve in relation to k-means clustering have not been previously reported in the literature.

Emerging whole-body imaging technologies push computer aided detection/diagnosis (CAD) to scale up to a whole-body level, which involves multiple organs or anatomical structure. To be exploited in this paper is the fact that the various tasks in whole-body CAD are often highly dependent (e.g., the localization of the femur heads strongly predicts the position of the iliac bifurcation of the aorta). One way to effectively employ task dependency is to schedule the tasks such that outputs of some tasks are used to guide the others. In this sense, optimal task scheduling is key to improve overall performance of a whole-body CAD system. In this paper, we propose a method for task scheduling that is optimal in an information-theoretic sense. The central idea is to schedule tasks in such an order that each operation achieves maximum expected information gain over all the tasks. The formulation embeds two intuitive principles: (1) a task with higher confidence tends to be scheduled earlier; (2) a task with higher predictive power for other tasks tends to be scheduled earlier. More specifically, task dependency is modeled by conditional probability; the outcome of each task is assumed to be probabilistic as well; and the objective function is based on the reduction of the summed conditional entropy over all tasks. The validation is carried out on a challenging CAD problem, multi-organ localization in whole-body CT. Compared to unscheduled and ad hoc scheduled organ detection/localization, our scheduled execution achieves higher accuracy with much less computation time.

Tissue abnormality characterization is a generalized segmentation problem which aims at determining a continuous score that can be assigned to the tissue which characterizes the extent of tissue deterioration, with completely healthy tissue being one end of the spectrum and fully abnormal tissue such as lesions, being on the other end. Our method is based on the assumptions that there is some tissue that is neither fully healthy or nor completely abnormal but lies in between the two in terms of abnormality; and that the voxel-wise score of tissue abnormality lies on a spatially and temporally smooth manifold of abnormality. Unlike in a pure classification problem which associates an independent label with each voxel without considering correlation with neighbors, or an absolute clustering problem which does not consider a priori knowledge of tissue type, we assume that diseased and healthy tissue lie on a manifold that encompasses the healthy tissue and diseased tissue, stretching from one to the other. We propose a semi-supervised method for determining such as abnormality manifold, using multi-parametric features incorporated into a support vector machine framework in combination with manifold regularization. We apply the framework towards the characterization of tissue abnormality to brains of multiple sclerosis patients.

Breast vascular calcifications (BVCs) are calcifications that line the blood vessel walls in the breast and appear as parallel or tubular tracks on mammograms. BVC is one of the major causes of the false positive (FP) marks from computer-aided detection (CADe) systems for screening mammography. With the detection of BVCs and the calcified vessels identified, these FP clusters can be excluded. Moreover, recent studies reported the increasing interests in the correlation between mammographically visible BVCs and the risk of coronary artery diseases. In this study, we developed an automated BVC detection method based on microcalcification prescreening and a new k-segments clustering algorithm. The mammogram is first processed with a difference-image filtering technique designed to enhance calcifications. The calcification candidates are selected by an iterative process that combines global thresholding and local thresholding. A new k-segments clustering algorithm is then used to find a set of line segments that may be caused by the presence of calcified vessels. A linear discriminant analysis (LDA) classifier was designed to reduce false segments that are not associated with BVCs. Four features for each segment selected with stepwise feature selection were used for this LDA classification. Finally, the neighboring segments were linked and dilated with morphological dilation to cover the regions of calcified vessels. A data set of 16 FFDM cases with vascular calcifications was collected for this preliminary study. Our preliminary result demonstrated that breast vascular calcifications can be accurately detected and the calcified vessels identified. It was found that the automated method can achieve a detection sensitivity of 65%, 70%, and 75% at 6.1 mm, 8.4 mm, and 12.6mm FP segments/image, respectively, without any true clustered microcalcifications being falsely marked. Further work is underway to improve this method and to incorporate it into our FFDM CADe system.

A wide range of pulmonary diseases, including common ones such as COPD, affect the airways. If the dimensions of airway can be measured with high confidence, the clinicians will be able to better diagnose diseases as well as monitor progression and response to treatment. In this paper, we introduce a method to assess the airway dimensions from CT scans, including the airway segments that are not oriented axially. First, the airway lumen is segmented and skeletonized, and subsequently each airway segment is identified. We then represent each airway segment using a segment-centric generalized cylinder model and assess airway lumen diameter (LD) and wall thickness (WT) for each segment by determining inner and outer wall boundaries. The method was evaluated on 14 healthy patients from a Weill Cornell database who had two scans within a 2 month interval. The corresponding airway segments were located in two scans and measured using the automated method. The total number of segments identified in both scans was 131. When 131 segments were considered altogether, the average absolute change over two scans was 0.31 mm for LD and 0.12 mm for WT, with 95% limits of agreement of [-0.85, 0.83] for LD and [-0.32, 0.26] for WT. The results were also analyzed on per-patient basis, and the average absolute change was 0.19 mm for LD and 0.05 mm for WT. 95% limits of agreement for per-patient changes were [-0.57, 0.47] for LD and [-0.16, 0.10] for WT.

The ability to automatically detect and monitor implanted devices may serve an important role in patient care and the evaluation of device and treatment efficacy. The purpose of this research was to develop a system for the automated detection of one-way endobronchial valves implanted as part of a clinical trial for less invasive lung volume reduction. Volumetric thin section CT data was obtained for 275 subjects; 95 subjects implanted with 246 devices were used for system development and 180 subjects implanted with 354 devices were reserved for testing. The detection process consisted of pre-processing, pattern-recognition based detection, and a final device selection. Following the pre-processing, a set of classifiers were trained using AdaBoost to discriminate true devices from false positives (such as calcium deposits). The classifiers in the cascade used simple features (the mean or max attenuation) computed near control points relative to a template model of the valve. Visual confirmation of the system output served as the gold standard. FROC analysis was performed for the evaluation; the system could be set so the mean sensitivity was 96.5% with a mean of 0.18 false positives per subject. These generic device modeling and detection techniques may be applicable to other devices and useful for monitoring the placement and function of implanted devices.

We have developed a new computer-aided diagnostic (CAD) scheme for detection of diffuse lung disease in computed radiographic (CR) chest images. One hundred ninety-four chest images (56 normals and 138 abnormals with diffuse lung diseases) were used. The 138 abnormal cases were classified into three levels of severity (34 mild, 60 moderate, and 44 severe) by an experienced chest radiologist with use of five different patterns, i.e., reticular, reticulonodular, nodular, air-space opacity, and emphysema. In our computerized scheme, the first moment of the power spectrum, the root-mean-square variation, and the average pixel value were determined for each region of interest (ROI), which was selected automatically in the lung fields. The average pixel value and its dependence on the location of the ROI were employed for identifying abnormal patterns due to air-space opacity or emphysema. A rule-based method was used for determining three levels of abnormality for each ROI (0: normal, 1: mild, 2: moderate, and 3: severe). The distinction between normal lungs and abnormal lungs with diffuse lung disease was determined based on the fractional number of abnormal ROIs by taking into account the severity of abnormalities. Preliminary results indicated that the area under the ROC curve was 0.889 for the 44 severe cases, 0.825 for the 104 severe and moderate cases, and 0.794 for all cases. We have identified a number of problems and reasons causing false positives on normal cases, and also false negatives on abnormal cases. In addition, we have discussed potential approaches for improvement of our CAD scheme. In conclusion, the CAD scheme for detection of diffuse lung diseases based on texture features extracted from CR chest images has the potential to assist radiologists in their interpretation of diffuse lung diseases.

Lung cancer remains the leading cause of cancer death in the United States and is expected to account for nearly 30% of all cancer deaths in 2007. Central to the lung-cancer diagnosis and staging process is the assessment of the central chest lymph nodes. This assessment typically requires two major stages: (1) location of the lymph nodes in a three-dimensional (3D) high-resolution volumetric multi-detector computed-tomography (MDCT) image of the chest; (2) subsequent nodal sampling using transbronchial needle aspiration (TBNA). We describe a computer-based system for automatically locating the central chest lymph-node stations in a 3D MDCT image. Automated analysis methods are first run that extract the airway tree, airway-tree centerlines, aorta, pulmonary artery, lungs, key skeletal structures, and major-airway labels. This information provides geometrical and anatomical cues for localizing the major nodal stations. Our system demarcates these stations, conforming to criteria outlined for the Mountain and Wang standard classification systems. Visualization tools within the system then enable the user to interact with these stations to locate visible lymph nodes. Results derived from a set of human 3D MDCT chest images illustrate the usage and efficacy of the system.

Pulmonary embolism (PE) is a serious medical condition, characterized by the partial/complete blockage of an artery within the lungs. We have previously developed a fast yet effective approach for computer aided detection of PE in computed topographic pulmonary angiography (CTPA),¹ which is capable of detecting both acute and chronic PEs, achieving a benchmark performance of 78% sensitivity at 4 false positives (FPs) per volume. By reviewing the FPs generated by this system, we found the most dominant type of FP, roughly one third of all FPs, to be lymph/connective tissue. In this paper, we propose a novel approach that specifically aims at reducing this FP type. Our idea is to explicitly exploit the anatomical context configuration of PE and lymph tissue in the lungs: a lymph FP connects to the airway and is located outside the artery, while a true PE should not connect to the airway and must be inside the artery. To realize this idea, given a detected candidate (i.e. a cluster of suspicious voxels), we compute a set of contextual features, including its distance to the airway based on local distance transform and its relative position to the artery based on fast tensor voting and Hessian "vesselness" scores. Our tests on unseen cases show that these features can reduce the lymph FPs by 59%, while improving the overall sensitivity by 3.4%.

Differences in the size distribution of malignant and benign pulmonary nodules in databases used for training and testing characterization systems have a significant impact on the measured performance. The magnitude of this effect and methods to provide more relevant performance results are explored in this paper. Two- and three-dimensional features, both including and excluding size, and two classifiers, logistic regression and distance-weighted nearest-neighbors (dwNN), were evaluated on a database of 178 pulmonary nodules. For the full database, the area under the ROC curve (AUC) of the logistic regression classifier for 2D features with and without size was 0.721 and 0.614 respectively, and for 3D features with and without size, 0.773 and 0.737 respectively. In comparison, the performance using a simple size-threshold classifier was 0.675. In the second part of the study, the performance was measured on a subset of 46 nodules from the entire subset selected to have a similar size-distribution of malignant and benign nodules. For this subset, performance of the size-threshold was 0.504. For logistic regression, the performance for 2D, with and without size, were 0.578 and 0.478, and for 3D, with and without size, 0.671 and 0.767. Over all the databases, logistic regression exhibited better performance using 3D features than 2D features. This study suggests that in systems for nodule classification, size is responsible for a large part of the reported performance. To address this, system performance should be reported with respect to the performance of a size-threshold classifier.

Chest X-ray (CXR) data is a projection image where each pixel of it represents a volumetric integration. Consequently identification of nodules and their characteristics is a difficult task in such images. Using a novel application of random process-based fractal image processing technique we extract features for nodule characterization. The uniqueness of the proposed technique lies in the fact that instead of relying on apriori information from user as in other random process inspired measures, we translate the random walk process into a feature which is based on its realization values. The Normalized Fractional Brownian Motion (NFBM) Model is derived from the random walk process. Using neighborhood region information in an incremental manner we can characterize the smoothness or roughness of a surface. The NFBM system gives a measure of roughness of a surface which in our case is a suspicious region (probable nodule). A classification procedure uses this measure to categorize nodule and non-nodule structures in the lung. The NFBM feature set is integrated in a prototype CAD system for nodule detection in CXR. Our algorithm provided a sensitivity of 75.9% with 3.1 FP/image on an independent test set of 50 CXR studies.

The performance of automated pulmonary nodule detection systems is typically qualified with respect to some minimum size of nodule to be detected. Also, an evaluation dataset is typically constructed by expert radiologists with all nodules larger than the minimum size being designated as true positives while all other smaller detected "nodules" are considered to be false positives. In this paper, we consider the negative impact that size estimation error, either in the establishment of ground truth for the evaluation dataset or by the automated detection method for the size estimate of nodule candidates, has on the measured performance of the detection system. Furthermore, we propose a modified evaluation procedure that addresses the size estimation error issue. The impact of the size measurement error was estimated for a documented research image database consisting of whole-lung CT scans for 509 cases in which 690 nodules have been documented. We compute FROC curves both with and without size error compensation and we found that for a minimum size limit of 4 mm the performance of the system is underestimated by a sensitivity reduction of 5% and a false positive rate increase of 0.25 per case. Therefore, error in nodule size estimation should be considered in the evaluation of automated detection systems.

Lung nodule segmentation in computed tomography (CT) plays an important role in computer-aided detection, diagnosis, and quantification systems for lung cancer. In this study, we developed a simple but accurate nodule segmentation method in three-dimensional (3D) CT. First, a volume of interest (VOI) was determined at the location of a nodule. We then transformed the VOI into a two-dimensional (2D) image by use of a "spiral-scanning" technique, in which a radial line originating from the center of the VOI spirally scanned the VOI. The voxels scanned by the radial line were arranged sequentially to form a transformed 2D image. Because the surface of a nodule in 3D image became a curve in the transformed 2D image, the spiral-scanning technique considerably simplified our segmentation method and enabled us to obtain accurate segmentation results. We employed a dynamic programming technique to delineate the "optimal" outline of a nodule in the 2D image, which was transformed back into the 3D image space to provide the interior of the nodule. The proposed segmentation method was trained on the first and was tested on the second Lung Image Database Consortium (LIDC) datasets. An overlap between nodule regions provided by computer and by the radiologists was employed as a performance metric. The experimental results on the LIDC database demonstrated that our segmentation method provided relatively robust and accurate segmentation results with mean overlap values of 66% and 64% for the nodules in the first and second LIDC datasets, respectively, and would be useful for the quantification, detection, and diagnosis of lung cancer.

One challenge facing radiologists is the characterization of whether a pulmonary nodule detected in a CT scan is likely to be benign or malignant. We have developed an image processing and machine learning based computer-aided diagnosis (CADx) method to support such decisions by estimating the likelihood of malignancy of pulmonary nodules. The system computes 192 image features which are combined with patient age to comprise the feature pool. We constructed an ensemble of 1000 linear discriminant classifiers using 1000 feature subsets selected from the feature pool using a random subspace method. The classifiers were trained on a dataset of 125 pulmonary nodules. The individual classifier results were combined using a majority voting method to form an ensemble estimate of the likelihood of malignancy. Validation was performed on nodules in the Lung Imaging Database Consortium (LIDC) dataset for which radiologist interpretations were available. We performed calibration to reduce the differences in the internal operating points and spacing between radiologist rating and the CADx algorithm. Comparing radiologists with the CADx in assigning nodules into four malignancy categories, fair agreement was observed (κ=0.381) while binary rating yielded an agreement of (κ=0.475), suggesting that CADx can be a promising second reader in a clinical setting.

The high-frequency of suspicious non-malignant image features limit the use of CT lung-cancer screening of asymptomatic individuals. A reference database of 6 radiologists' localizations of suspicious image features was created. The frequency, sizes, shapes, margins and degree of calcification of the images features were determined. The radiologist exam report identified 50% findings reported on 5 or 6 occasions, while the CAD system identified 40%. The radiologist exam report missed 80% to 50% of suspicious, retrospectively identified in CT lung cancer images. Many were less than 4 mm. Radiologists can use a lenient criterion in experiments of previously read screening cases.

Although ensemble techniques have been investigated in supervised machine learning, their potential with knowledge-based systems is unexplored. The purpose of this study is to investigate the ensemble approach with a knowledge-based (KB) CAD system for the detection of masses in screening mammograms. The system is designed to determine the presence of a mass in a query mammographic region of interest (ROI) based on its similarity with previously acquired examples of mass and normal cases. Similarity between images is assessed using normalized mutual information. Two different approaches of knowledge database decomposition were investigated to create the ensemble. The first approach was random division of the knowledge database into a pre-specified number of equal size, separate groups. The second approach was based on k-means clustering of the knowledge cases according to common texture features extracted from the ROIs. The ensemble components were fused using a linear classifier. Based on a database of 1820 ROIs (901 masses and 919 and the leave-one-out crossvalidation scheme, the ensemble techniques improved the performance of the original KB-CAD system (A_z = 0.86±0.01). Specifically, random division resulted in ROC area index of A_z = 0.90 ± 0.01 while k-means clustering provided further improvement (A_z = 0.91 ± 0.01). Although marginally better, the improvement was statistically significant. The superiority of the k-means clustering scheme was robust regardless of the number of clusters. This study supports the idea of incorporation of ensemble techniques with knowledge-based systems in mammography.