This PDF file contains the front matter associated with SPIE Proceedings Volume 10718, including the Title Page, Copyright information, Table of Contents, and Conference Committee listing.

Background

Breast compression is used in mammography to improve image quality and reduce radiation dose. However, the compression may lead to discomfort or pain for the women. Breast compression time is longer with digital breast tomosynthesis (DBT) than with digital mammography (DM). We aimed to explore breast compression parameters with DM and DBT.

Material and Methods

We used information from 16,832 women participating in the Tomosynthesis Trial in Bergen between January 2016 and April 2017. We compared mean values of applied compression force (N), compression pressure (kPa) and compressed breast thickness (mm) for DM and DBT, by view (craniocaudal, CC, and mediolateral-oblique, MLO). Two-sample ttests were used to test statistical significance.

Results

Number of women screened with DM or DBT were similar (DM: n= 8354 and DBT: n= 8478). Mean compression force was statistically significantly higher for DM compared to DBT (CC: 108.6 N versus 102.7 N; MLO: 122.4 N versus 120.8 N, p <0.01). Mean compression pressure was higher for DM compared to DBT for CC view (13.9 kPa versus 13.0 kPa, p<0.01), however, not for MLO view (DM and DBT: 9.7 kPa, p= 0.55). Mean compressed breast thickness did not differ statistically significantly for DM compared to DBT (CC: 58.7 mm N versus 58.6 mm, p= 0.72; MLO: 60.1 mm versus 59.9 mm, p= 0.23).

Conclusion

Radiographers applied less breast compression with DBT compared to DM. However, the observed differences were negligible. Further research should investigate the clinical implications of the differences, such as image quality.

For more than a decade, radiologists have used traditional computer aided detection systems to read mammograms, but mainly because of a low computer specificity may not improve their screening performance, according to several studies. The breakthrough in deep learning techniques has boosted the performance of machine learning algorithms, also for breast cancer detection in mammography. The objective of this study was to determine whether radiologists improve their breast cancer detection performance when they concurrently use a deep learningbased computer system for decision support, compared to when they read mammography unaided. A retrospective, fully-crossed, multi-reader multi-case (MRMC) study was designed to compare this. The employed decision support system was Transpara™ (Screenpoint Medical, Nijmegen, the Netherlands). Radiologists interact by clicking an area on the mammogram, for which the computer system displays its cancer likelihood score (1-100). In total, 240 cases (100 cancers, 40 false positive recalls, 100 normals) acquired with two different mammography systems were retrospectively collected. Seven radiologists scored each case once with, and once without the use of decision support, providing a forced BI-RADS® score and a level of suspiciousness (1-100). MRMC analysis of variance of the area under the receiver operating characteristic curves (AUC), and specificity and sensitivity were computed. When using decision support, the AUC increased from 0.87 to 0.89 (P=0.043) and specificity increased from 73% to 78% (P=0.030), while sensitivity did not significantly increment (84% to 87%, P=0.180). In conclusion, radiologists significantly improved their performance when using a deep learningbased computer system as decision support.

Can radiologists distinguish prior mammograms with no overt signs of cancer from women who were later diagnosed with breast cancer from the prior mammograms of women reported as normal and subsequently confirmed to be cancerfree? Twenty-three radiologists and breast physicians viewed 200 craniocaudial mammograms for a half-second and rated whether the woman would be recalled on a scale of 0 (clearly normal) to 100 (clearly abnormal). The dataset included five categories of mammograms, with each category containing 40 cases. The categories were Cancer (current cancer-containing mammograms), Prior-Vis (prior mammograms with visible cancer signs), Contra (current ‘normal’ mammograms contralateral to the cancer), Prior-Invis (priors without visible cancer signs), and Normal (priors of normal cases). For each radiologist, four pairs of analyses were performed to evaluate whether the radiologists could distinguish mammograms in each category from the normal mammograms: Cancer vs Normal, Prior-Vis vs Normal, Contra vs Normal, and Prior-Invis vs Normal. The Area under Receiver Operating Characteristic curves (AUC) was calculated for each paired grouping and each radiologist. Wilcoxon Signed Rank test showed the AUC values were above-chance for all comparisons: Cancer (z=4.20, P<0.001); Prior-Vis (z=4.11, P<0.001); Contra (z=4.17, P<0.001); Prior-Invis (z=3.71, P<0.001). The results suggest that radiologists can distinguish patients who were diagnosed with cancer from individuals without breast cancer at an above-chance level based on a half-second glimpse of mammogram even before the lesion becomes apparently visible (Prior-Invis). Apparently, something about the breast parenchyma can look abnormal before the appearance of a localized lesion

Computer-aided detection or decision support systems aim to improve breast cancer screening programs by helping radiologists to evaluate digital mammography (DM) exams. Commonly such methods proceed in two steps: selection of candidate regions for malignancy, and later classification as either malignant or not. In this study, we present a candidate detection method based on deep learning to automatically detect and additionally segment soft tissue lesions in DM. A database of DM exams (mostly bilateral and two views) was collected from our institutional archive. In total, 7196 DM exams (28294 DM images) acquired with systems from three different vendors (General Electric, Siemens, Hologic) were collected, of which 2883 contained malignant lesions verified with histopathology. Data was randomly split on an exam level into training (50%), validation (10%) and testing (40%) of deep neural network with u-net architecture. The u-net classifies the image but also provides lesion segmentation. Free receiver operating characteristic (FROC) analysis was used to evaluate the model, on an image and on an exam level. On an image level, a maximum sensitivity of 0.94 at 7.93 false positives (FP) per image was achieved. Similarly, per exam a maximum sensitivity of 0.98 at 7.81 FP per image was achieved. In conclusion, the method could be used as a candidate selection model with high accuracy and with the additional information of lesion segmentation.

We investigated the additive role of breast parenchyma stroma in the computer-aided diagnosis (CADx) of tumors on full-field digital mammograms (FFDM) by combining images of the tumor and contralateral normal parenchyma information via deep learning. The study included 182 breast lesions in which 106 were malignant and 76 were benign. All FFDM images were acquired using a GE 2000D Senographe system and retrospectively collected under an Institution Review Board (IRB) approved, Health Insurance Portability and Accountability Act (HIPAA) compliant protocol. Convolutional neutral networks (CNNs) with transfer learning were used to extract image-based characteristics of lesions and of parenchymal patterns (on the contralateral breast) directly from the FFDM images. Classification performance was evaluated and compared between analysis of only tumors and that of combined tumor and parenchymal patterns in the task of distinguishing between malignant and benign cases with the area under the Receiver Operating Characteristic (ROC) curve (AUC) used as the figure of merit. Using only lesion image data, the transfer learning method yielded an AUC value of 0.871 (SE=0.025) and using combined information from both lesion and parenchyma analyses, an AUC value of 0.911 (SE=0.021) was observed. This improvement was statistically significant (p-value=0.0362). Thus, we conclude that using CNNs with transfer learning to combine extracted image information of both tumor and parenchyma may improve breast cancer diagnosis.

Breast cancer is considered to have a high incidence among women worldwide. Recent development in biomedical image analysis using deep learning based neural networks have motivated researches to enhance the performance of Computer Aided Diagnosis (CAD) systems. In this paper, the performance of four different deep neural networks was compared for malignant/benign classification of mammographic mass abnormalities. For this aim, different annotated mammography repositories were introduced and the classification performance of four deep Convolutional Neural Networks (CNNs) on each dataset and on their combination was investigated. The robustness to over-fitting regarding the size of data and the approach of transfer learning were compared. Our quantitative results indicate the importance of training samples regardless of acquisition methods when training with various deep CNN models. We achieved an average accuracy of 85% and an average AUC of 0.83 in our best result on the combination of all datasets. However, we conclude that several runs with different samples are needed to understand the variation in the results, especially with smaller datasets.

Recently, both Deep Cascade classifiers and Convolutional Neural Networks (CNNs) have achieved state-ofthe-art microcalcification (MC) detection performance in digital mammography. Deep Cascades consist in long sequences of weak classifiers designed to effectively learn from heavily unbalanced data as in the case of MCs (∼ 1 MC every 10, 000 non-MC samples). CNNs are powerful models that achieve impressive results for image classification thanks to the ability to automatically extract general-purpose features from the data, but require balanced classes. In this work, we introduce a two-stage classification scheme that combines the benefits of both systems. Firstly, Deep Cascades are trained by requiring a very high sensitivity (99.5%) throughout the sequence of classifiers. As a result, while the number of MC samples remains practically unchanged, the number of non-MC samples is greatly reduced. The remaining data, approximately balanced, are used to train an additional stage of classification with a CNN. We evaluated the proposed approach on a database of 1, 066 digital mammograms. MC detection results of the combined classification were statistically significantly higher than Deep Cascade and CNN alone, yielding an average improvement in mean sensitivity of 3.19% and 2.45%, respectively. Remarkably, the proposed system also yielded a faster per-mammogram processing time (2.0s) compared to Deep Cascade (2.5s) and CNN (5.7s).

Presentation of reference images that are similar to a query image can be helpful in medical image diagnosis and treatment planning. The purpose of this study is to investigate a method for retrieving relevant images of breast masses on mammograms as a diagnostic reference. In our previous studies, subjective similarities for pairs of masses were obtained from experienced radiologists and used as the gold standard for retrieving visually similar images. By use of multidimensional scaling, a subjective similarity space was spanned so that masses that were placed close to a query image can be retrieved as reference images. This method, however, required manual outlines of masses for image feature determination. In this study, we modelled this similarity space using convolutional neural network. The result was evaluated using the leave-one-out cross validation method in terms of the correlation between the subjective ratings and determined similarity measures. The relevance of retrieved images was also evaluated in terms of the precision, which is the fraction of pathology matched images in the retrieved images. The correlation coefficient between the subjective ratings and the determined similarity measure was moderate with 0.735, which was slightly lower than those of the previous methods. The average precision was high with 0.852 when the most similar image was retrieved, which was higher than those in the previous studies. The results indicate the potential usefulness of the proposed method in similar image retrieval of breast masses on mammograms.

Assessment of breast density at the point of mammographic examination could lead to optimized breast cancer screening pathways. The onsite breast density information may offer guidance when to recommend supplemental imaging for women in a screening program. In this work, performance evaluation of a new software (Insight BD, Siemens Healthcare GmbH) for fast onsite quantification of volumetric breast density is presented. Accuracy of volumetric measurement is evaluated using breast tissue equivalent phantom experiments. Reproducibility of measurement results is analyzed using 8150 4-view mammography exams. Furthermore, agreement between breast density categories computed by the software with those determined visually by radiologists is examined. The results of the performance evaluation demonstrate that the software delivers accurate and reproducible measurements that agree well with the visual assessment of breast density by radiologists.

This paper presents a novel variant of Local Binary Patterns (LBP) so called Mean-Elliptical Local Binary Patterns (M-ELBP) for breast density classification. The basic idea of (E)LBP is that an image texture consists of micropatterns and the histogram of these micropatterns could give information about the local features in an image. To include intensity based information along with texture features, we improve the ELBP descriptor by including the mean intensity around the central pixel. We use multiple orientations which facilitates capturing sufficient micropattern features from the mammographic images. The M-ELBP is computed for the fibroglandular disk area instead of the whole breast region. A Bayesian network classifier was used for performing mammogram density classification and a stratified ten-fold cross-validation scheme was used for performance evaluation of the proposed method. The proposed method achieved an accuracy of 74% which is comparable with other methods in the literature using the same dataset.

Background

We analyzed the consistency in absolute and percent density using an automated method for estimating volumetric breast density among women who attended BreastScreen Norway 2007-2015.

Material and Methods

We used information from 33,711 women aged 50-69 who underwent two to five full field digital screening examinations biennially, 2007-2015. BMI and HT-use was obtained from a self-administered questionnaire, sent together with the invitation to screening. BMI was categorized into <21.0; 21.0-24.9; ≥25.0 kg/m², while use of HT was defined as ever versus never used. An automated method estimated fibroglandular volume (cm³), breast volume (cm³) and volumetric breast density (%) for each screening examination. We applied mixed-effects linear models to estimate associations between age, fibroglandular volume, breast volume and volumetric breast density over time including data on BMI and HT.

Results

The results models indicated age to be associated with decreased fibroglandular volume and volumetric breast density, and increased breast volume. BMI <21 kg/m² was associated with higher volumetric breast density, but lower fibroglandular and breast volume. Contrary, BMI ≥25 kg/m² was associated with lower volumetric breast density and higher fibroglandular and breast volumes. Variation in volumetric breast density and fibroglandular volume within women was rather subtle: this variance did not exceed 10% for either volumetric breast density, fibroglandular volume or breast volume.

Conclusion

Automated measures of volumetric breast density and fibroglandular volume software decreased with age among women screened in BreastScreen Norway. Absolute and percent density varied with a maximum of 10% over time, from first to last screening examination.

Background: Mammographic density is an important risk factor for breast cancer. Recent research demonstrated that percentage density assessed visually using Visual Analogue Scales (VAS) showed stronger risk prediction than existing automated density measures, suggesting readers may recognise relevant image features not yet captured by automated methods.

Method: We have built convolutional neural networks (CNN) to predict VAS scores from full-field digital mammograms. The CNNs are trained using whole-image mammograms, each labelled with the average VAS score of two independent readers. They learn a mapping between mammographic appearance and VAS score so that at test time, they can predict VAS score for an unseen image. Networks were trained using 67520 mammographic images from 16968 women, and tested on a large dataset of 73128 images and case-control sets of contralateral mammograms of screen detected cancers and prior images of women with cancers detected subsequently, matched to controls on age, menopausal status, parity, HRT and BMI.

Results: Pearson's correlation coefficient between readers' and predicted VAS in the large dataset was 0.79 per mammogram and 0.83 per woman (averaging over all views). In the case-control sets, odds ratios of cancer in the highest vs lowest quintile of percentage density were 3.07 (95%CI: 1.97 - 4.77) for the screen detected cancers and 3.52 (2.22 - 5.58) for the priors, with matched concordance indices of 0.59 (0.55 - 0.64) and 0.61 (0.58 - 0.65) respectively.

Conclusion: Our fully automated method demonstrated encouraging results which compare well with existing methods, including VAS.

High mammographic density reduces the diagnostic accuracy of mammography by masking tumors, leading to interval cancers and late stage diagnosis. In this study, various models to predict masking risk are computed on a cohort of 90 interval or undiagnosed (“masked”) invasive cancers and 186 screen-detected invasive cancers, based on biometric (age and BMI) and image-based parameters (BI-RADS density, volumetric breast density (VBD) and detectability). Univariate logistic regressions were computed to predict masked cancers, and the accuracy of the regressions was evaluated using the area under receiver operator characteristic curve (AUC). The univariate AUC for BMI, age, BIRADS density, VBD and mean detectability were 0.61 [0.54-0.68], 0.65 [0.58-0.73], 0.67 [0.61–0.73], 0.72 [0.65-0.78] and 0.75 [0.68-0.81] respectively (95% confidence intervals are noted in the brackets). The models were applied to a set of 248 mammography exams from cancer-free women of the same population. A stratified screening model was tested by computing the fraction of disease-free women identified as masked (the recruitment rate) as a function of the fraction of masked cancers that were correctly identified. For BI-RADS densities 3 or 4 (4th edition), up to 60% of masked cancers could potentially be detected by supplemental tests, requiring 43% of women to be recruited for extra screening. Selecting by mean detectability would require a 29% recruitment rate for the same potential capture. Future work to develop multivariate masking risk predictors could yield more efficient stratified screening approaches for breast cancer detection.

A simulation of sequential breast pathology images is proposed, as a prerequisite for the development of virtual clinical trials (VCTs) with radiology-pathology (rad-path) correlation. The rad-path correlation of breast cancer findings is performed clinically to confirm concordance and increase confidence in diagnoses. VCTs have been used for optimization of breast imaging systems, based upon computer simulation of breast anatomy, imaging modalities, and image interpretation. Today, VCTs are used to optimize breast imaging at the “radiology” spatial scale, by simulating tissue structures seen in radiological images, namely, skin, adipose or dense tissue compartments, fibrous ligaments, and major ducts and blood vessels. We have extended this simulation to the microscopic (i.e., “pathology”) spatial scale, to allow for virtual rad-path correlation. Previously, we developed a manual simulation of adipose and dense tissue regions in pathology images, corresponding to a small region selected within a breast phantom simulated at the radiological scale. In this paper, we describe an automated simulation of adipocytes, epithelial and myoepithelial cells, collagen fibers, and fibroblasts. Adipocytes are simulated by recursive partitioning. Epithelial and myoepithelial cells are simulated radially around ductal or acinar lumen. Fibers and fibroblasts are simulated by an analogy with the electrostatic field. Our approach models the volumetric distributions of cells and various breast tissues, which allows the simulation of sequential pathology images at clinical inter-slice distances. The proposed simulation method has been evaluated by a clinical pathologist and medical physicists. The effect of the simulation approaches on the visual appearance of simulated pathology images has been evaluated.

We have developed imaging biomarkers for mammographically-occult (MO) cancer in women with dense breasts using a radiologist’s progress ratings on cancer development. MO cancer is a cancer that is occluded by dense breast tissue, or especially visually subtle that radiologists fail to recognize the cancer. We used screening mammograms of 242 normal women (half training) with dense breasts and 116 women (66 training and 50 testing) with dense breasts who had unilateral MO cancer, i.e., negative consecutive mammograms followed by a diagnosis of cancer. With full cancer diagnosis information of the cancer-diagnosed in the current year’s mammogram, an experienced breast radiologist reviewed 1- 3 prior consecutive mammograms in the training set and rated the progress score on cancer development over prior mammograms, using a [0 100] scale. We segmented a dense area in those mammograms and extracted 42 image features (5 histogram, 16 texture, and 21 bilateral asymmetries). We conducted a Pearson’s correlation analysis between image features and the radiologist’s cancer development ratings. We found 23 features correlated with the radiologist’s ratings (p-values < 0.05). We used the top six correlated image features (p-values < 0.01) with the radiologist’s ratings to develop a classifier to identify women with MO cancer. The features included three histogram, one texture, and two bilateral histogram asymmetry features. Using training and testing sets, we trained and tested a logistic regression classifier. The mean and 95% confidence interval of the area under the receiver-operating characteristic curve (AUC) of the classifier was 0.79 [0.7, 0.86].

Radiomic features extracted from breast magnetic resonance (MR) images have demonstrated potential in diagnosis and prognosis of breast cancer. However, presentation of lesions on MRI may be affected by a biopsy event. We investigated, relative to biopsy condition, the difference in radiomic features and performance (the area under the receiver operating curve (AUC)) for the task of distinguishing between benign lesions and luminal A cancers. Dynamic contrast-enhanced MR images were collected retrospectively under IRB/HIPAA compliance. The 361-case dataset included 92 benign and 30 luminal A lesions imaged pre-biopsy and 40 benign and 199 luminal A lesions imaged post-biopsy. Thirty-four radiomic non-size features were extracted and their values compared for each group of lesions by biopsy condition using the Kolmogorov-Smirnov test to determine if the two groups were drawn from the same patient distribution. For each feature by biopsy condition, the median of the AUC and the confidence interval for the difference in AUC were determined (2000 bootstrap iterations). In all analyses, after correction for multiple comparisons, a difference was considered significant when the p-value was less than 0.0015 (0.05/34) or, equivalently, the 99.85% confidence interval for the difference excluded zero.

If comparisons failed to meet p < 0.0015, features were considered potentially robust. While, as expected, the morphology feature of irregularity was significantly different (p = 0.0003) for benign lesions due to how biopsy events increased irregularity of benign lesions, most features were potentially robust between biopsy conditions. While features did well in distinguishing between luminal A and benign lesions, all failed to demonstrate significance differences in AUC between biopsy conditions.

Our clinical practice transitioned from digital mammography (DM) to digital breast tomosynthesis (DBT) screening in 2011. This study analyzes the radiation dose of diagnostic exams after recall from screening in two cohorts, before (I) and after (II) the transition to screening with four Hologic Selenia Dimensions DM/DBT systems. We considered four different imaging modes: DM, magnification DM, and the 2D and 3D components of DBT. Diagnostic exams were classified into four groups based on screening recalled finding: asymmetry, architectural distortion, masses, and calcifications. The study set consisted of 7,409 images from 1,857 women (mean age 55.3±10 yrs.) acquired at two time periods (2010-11, cohort I; 2012-13, cohort II). The average glandular dose (AGD) for the population was computed from the sum of all exposures and analyzed by finding type. The AGD, breast thickness, and exposure settings were obtained using an automated dose-reporting software that stores DICOM metadata in a database for real-time data exploration. The average AGD per patient was 5.82 mGy for cohort I (4.95, 5.40, 6.01 and 8.96 mGy for masses, asymmetry, architectural distortion and calcifications, respectively) and 7.15 mGy for cohort II (6.23, 7.19, 7.61 and 9.03 mGy, respectively). While the AGD for calcifications remained the same after the transition, a dose increase of 23% was found due to the addition of 3D imaging for the other findings. The implementation of DBT in the diagnostic setting results in increased radiation dose to the diagnostic population. However, due to the reduction of recalled patients from screening, we achieved a 15.4% reduction in the utilization of diagnostic imaging. This resulted in a net increase of only 4.0% in the total radiation dose to the screening population arising from diagnostic imaging at recall.

Purpose: To evaluate whether combining a polychromatic reconstruction algorithm for breast CT with projection data acquired using alternating high and low energy spectra allows a significant dose reduction while maintaining image quality. Materials and Methods: A breast phantom was scanned on a clinical breast CT scanner using the automatic exposure control selected exposure at the regular spectrum with a tube voltage of 49 kV and a 1.576 mm aluminum filter and with a second, higher energy spectrum created by adding a 0.254 mm copper filter. An acquisition with spectrum switching was simulated by interleaving projections from the standard and high energy datasets, and a previously developed polychromatic reconstruction algorithm was modified to reconstruct the breast CT images. Image quality was assessed using the signal difference-to-noise ratio (SDNR) of a high and a low contrast target present in the phantom. A Monte Carlo simulation was performed to determine the mean glandular dose (MGD) of each scan. Results: Acquisition of the simulated scan with spectrum switching would result in an MGD of 6.57 mGy, compared to the standard acquisition MGD of 10.4 mGy, a reduction of 37%. At the same time, the measured SDNR of the mixed spectrum reconstructions was slightly higher than that of the standard acquisition, with an increase in SDNR of 6.6% (p < 0.01) for the high contrast target and 5.3% (p = 0.12) for the low contrast target. Conclusion: Our approach combining a polychromatic reconstruction algorithm for breast CT with an advanced acquisition protocol using alternating high and low energy spectra can lower dose by at least a third without loss of target SDNR

The purpose is to develop and validate an automated method for detecting image unsharpness caused by patient motion blur in digital mammograms. The goal is that such a tool would facilitate immediate re-taking of blurred images, which has the potential to reduce the number of recalled examinations, and to ensure that sharp, high-quality mammograms are presented for reading. To meet this goal, an automated method was developed based on interpretation of the normalized image Wiener Spectrum. A preliminary algorithm was developed using 25 cases acquired using a single vendor system, read by two expert readers identifying the presence of blur, location, and severity. A predictive blur severity score was established using multivariate modeling, which had an adjusted coefficient of determination, R² =0.63±0.02, for linear regression against the average reader-scored blur severity. A heatmap of the relative blur magnitude showed good correspondence with reader sketches of blur location, with a Spearman rank correlation of 0.70 between the algorithmestimated area fraction with blur and the maximum of the blur area fraction categories of the two readers. Given these promising results, the algorithm-estimated blur severity score and heatmap are proposed to be used to aid observer interpretation. The use of this automated blur analysis approach, ideally with feedback during an exam, could lead to a reduction in repeat appointments for technical reasons, saving time, cost, potential anxiety, and improving image quality for accurate diagnosis.

The goal of this study is to investigate whether reduced breast compression in digital breast tomosynthesis (DBT) exams causes larger internal breast motion that would adversely affect DBT image quality. We designed an experiment to collect real-time breast motion data from patients using ultrasound under three levels of DBT compression (full, medium and half). The ultrasound RF data had a pixel size of 21.5 μm and 150 μm in the axial and lateral directions of the probe, allowing the tracking of very fine movement of internal structure. We have successfully collected data from six human subjects and continue to recruit patients. The data were analyzed using speckle-tracking techniques to extract internal tissue movement trajectories pixel by pixel at multiple locations. Initial data analysis showed that internal breast tissue movement is highly correlated with breathing. Based on the first four patient datasets we have processed, the internal motion magnitudes on average were smaller than 1 mm under the full and reduced compression levels. The statistical distributions of the motion magnitudes among the three compression levels were similar, indicating that the internal breast motion may not necessarily increase even when compression is reduced by half. However, more data will be collected and analyzed to strengthen this study for more solid conclusions.

We are currently building the PET/X scanner, which is a breast positron emission tomography (PET) scanner combined with a standard X-ray mammography system. The role for the PET/X scanner is to precisely measure changes in radiotracer uptake after an initial test dose of adjuvant or neoadjuvant therapy. The system performance target is that a measured 20% change in tracer uptake in 5 mm diameter lesions with standardized uptake value (SUV) of 5 g/ml should correspond to at least 95% specificity (<5% false positive rate for detecting a true change in uptake). Simulations indicate that this performance can be achieved with a 3-minute scan of a 370-MBq (10-mCi) injection of 18F-FDG or other radiotracers.

The PET scanner consists of 4 planar detectors, forming a rectangular system with complete azimuthal angular sampling (around the z-axis). The panels are formed of detector blocks each using a 20x20 array of LYSO scintillation crystals (2x2x10mm³) coupled to a 12x12 array of 3x3 mm² pixel silicon photomultipliers (SiPMs). The overall system size is 24cm wide (6 blocks) by 16cm deep (4 blocks axially) by 4-to-12cm adjustable height.

The PET system mounts on a mammography scanner by swapping out the standard 'bucky' unit. Patients are scanned in both mammography and PET modes without movement to allow co-registration of the two image sets. Here, we review the system design, construction, and initial performance measures.

The objective of our study is to optimize the acquisition parameters for imaging Ag2S nanoparticles using contrast-enhanced digital mammography (CE-DM) by varying parameters such as kV, mAs, and filtration. The efficacies of three different contrast materials (Ag2S nanoparticles, silver nanoparticles, and iodine) were assessed using a contrast-embedded gradient phantom. The phantom was constructed using tissue-equivalent materials and varied continuously in composition from 100% glandular tissue to 100% adipose tissue. Each contrast agent was prepared at six different concentrations (1, 2, 5, 10, 15, and 25 mg/mL). Holes were bored through the phantom in the direction of varying glandularity, and tubes of contrast agents were inserted into the holes. Phantoms were imaged at four different energies (26 kV, 32 kV, 45 kV, and 49 kV); 5 energy pairs were considered in this study. Our results demonstrate that for a given contrast agent, the contrast-to-noise ratio is linearly proportional to concentration, and its magnitude is dependent on the energy of the low-energy (LE) image. In our study, it was shown that the LE images at 26 kV are better suited for imaging silver-based nanoparticles, and the LE images at 32 kV are better suited for imaging iodine contrast. Thus, the energy of the LE image should be chosen so that it is as close as possible to the k-edge of the contrast material. Preliminary results from CE-DM imaging indicate that silver contrast has a significantly higher signal than iodine contrast when imaging at lower energies, thus demonstrating the feasibility of using silver-based nanoparticles in breast imaging.

This study examines the potential of a multisource x-ray system to reduce cone beam artifacts in a dedicated breast CT acquisition geometry. A breast CT scanner (Doheny), built at our institution, was used to demonstrate the potential of multiple x-ray sources in a single x-ray tube housing. Both 3 focal spot and 5 focal spot thermionic systems were physically simulated in this study. The x-ray tube is mounted on a vertical actuator on the breast CT system gantry, allowing the single x-ray source to be positioned at different vertical locations in the field of view. Five acquisition geometries were used to acquire raw cone beam CT data with the x-ray source locations placed at 2 cm intervals. Data was collected using a 15-cm tall Defrise phantom. The individual acquisitions of raw CT data were reconstructed using filtered back projection, aligned and summed. The reconstructed CT volume data set using three sources and five sources were compared to that produced from a single source. Both multi-source datasets demonstrated less visible cone beam artifact, and the contrast clearly improved. The resolvable field of view in the vertical direction was extended by 50% when comparing the one source to the three source geometry and extended by 120% when comparing the one source to the five source geometry. This physical simulation of a multisource x-ray CT system successfully demonstrated that a reduction in cone beam CT artifacts could be achieved using a multi-source x-ray tube on a breast CT scanner.

A method to modify mammographic images so that they appear as if acquired with different radiographic factors has been developed for use in virtual clinical trials. The methodology was validated using a custom-made 3D printed breast phantom. The radiographic output, 1st half value layer, noise coefficients, grid factor, glare and scatter-to-primary ratio were measured on a GE Senographe Essential digital mammography system. Three images of the phantom at each of four dose levels were acquired with a Mo/Mo 26 kV spectrum using the same digital mammography system. In addition, three images were acquired at each of 24 kV and 28 kV. High dose images at 26 kV were modified to simulate the ones acquired at a lower dose, and at the lower and higher tube voltage. The signal level was changed to account for different dose level, scatter, grid attenuation and change of signal across the image. Noise was then added to the image to match the expected noise for the target image. The power spectra of the target and the simulated images were calculated and compared. This showed that they match to within an average difference of 4%. Also, an average structural similarity index of 0.998 was obtained, meaning that the images are almost equivalent. Using breast phantom images, it was shown that this method can replicate images as if acquired with different radiographic factors. Images modified in this manner could be used for clinical evaluation of image quality or for other types of virtual or human observer studies.

Because breast phantoms are central for evaluating 2D and 3D breast imaging systems, it is important to develop anthropomorphic, realistic phantoms that can be used in task-based assessment. The current phantoms available for use with full field digital mammography (FFDM) and digital breast tomosynthesis (DBT) are uniform or have otherwise unrealistic background texture. However, this work presents a task-based methodology for evaluating FFDM and DBT systems using two components: an anthropomorphic breast phantom created through inkjet printing and realistic pathological inserts with two types of microcalcifications (MCs). To create the phantom, a virtual breast was first modeled, then digitally compressed to a 4 cm thickness. To realize the phantom, virtual model was printed with an inkjet printer in a slice-by-slice fashion using iohexol-doped ink. The MC inserts were created using two types of materials. One set of MCs were made using calcium hydroxyapatite (HA) powder, and the other was made from sodalime coated glass (SLG) microspheres. The composite phantom plus two types of inserts was imaged on two commercially available FFDM/DBT systems: the GE Senographe Essential and the Hologic Selenia Dimensions. For the acquisitions, the automatic exposure control was used to determine a typical mammographic beam based on the phantom thickness. In addition, a similar a similar average glandular dose was used for both systems. A four alternative forced choice (4AFC) study was conducted to demonstrate the utility of the methodology and to evaluate the systems. Results of the 4AFC study showed higher detection of calcs for DBT versus FFDM on the GE system, but the reverse on the Hologic system. In this work, a breast phantom as created from inexpensive and easily accessible materials. This methodology is promising for the objective evaluation of task performance for 2D and 3D breast imaging systems.

In this paper, texture calculations are used to validate the realism of a physical anthropomorphic phantom for digital breast tomosynthesis. The texture features were compared against clinical mammography data. Three groups of features (grey-level histogram, co-occurrence, and run-length) were considered. The features were analyzed over a broad range of technique settings (kV and mAs). These calculations were done in the central slice of the reconstruction as well as the synthetic 2D mammogram. For each feature, the clinical data were binned into strata based on the compressed breast thickness. It was demonstrated that the clinical features vary by thickness. To evaluate the realism of the phantom, each feature was compared against clinical data in the same thickness stratum. For the purpose of this paper, a feature was considered to be realistic if it was within the middle 95% of the statistical distribution of clinical values. In the reconstruction, most features were found to exhibit realism; specifically, all 12 grey-level histogram features, four out of seven co-occurrence features, and three out of seven run-length features. The realism of most features was robust to changes in the technique settings. However, in the synthetic 2D mammogram, fewer features were found to exhibit realism. In conclusion, this paper provides a validation of the textural realism of the phantom in the reconstruction, and shows that there is less realism in the synthetic 2D mammogram. We identify the features that should be considered to refine the design of the phantom in future work.

In this work we present a new methodology for creating a physical anthropomorphic breast phantom for use in dedicated breast CT (bCT) systems. We also present a method for modeling microcalcifications (MCs) in the phantom. An uncompressed digital breast was first created though analytical modeling. This virtual model is then physically realized through inkjet printing using doped ink and paper with radiographic properties similar to that of glandular and adipose tissues. In addition, MC cluster inserts were created using calcium hydroxyapatite (HA) spheres that ranged in diameter from 212-250um to 250-300um and could be inserted inside the phantom. The phantom was then assessed in terms of material realism and reproducibility using x-ray spectroscopy, a clinical full field digital mammography (FFDM) system, and a benchtop CT system. Results show that both the paper and ink used to create the physical phantom have radiographic properties which closely match reference values for glandular and adipose tissue. Further, reproducibility was confirmed through multiple prints. The phantom and MC inserts can be used to objectively assess image quality of dedicated bCT systems.

We characterize the detectability of fiber-like signals in digital breast tomosynthesis (DBT) for linear iterative image reconstruction (IIR) algorithms. The detectability is investigated as a function of signal orientation and IIR regularization strength. The detectability is computed with a region-of-interest (ROI) Hotelling observer (HO) and applied to two linear IIR algorithms. Trends in detectability are compared with conspicuity of signals reconstructed in both simulation and real data studies. A common trend is observed with both algorithms in which signals oriented parallel to the detector and the plane containing the source-trajectory have lower detectability than their orthogonal counterparts at low regularization strengths. The orientation dependence is gradually reduced with increasing regularization strength. These trends in detectability are seen to match well with trends in the conspicuity of reconstructed signals in both simulation and real data studies.

Studies have shown that there are discrepancies among pathologists in the classification of breast histopathological slides. In this study we propose a framework for categorizing hematoxylin-eosin stained breast images either as benign or malignant at four magnification factors, and then aggregating the classification results of a patient’s images from different magnification factors to make the ultimate diagnosis for each patient. We used a publicly available database, containing 7786 images from 81 patients. The images were acquired in four visual magnification factors, namely x40, x100, x200, and x400, with an effective pixel size of 0.49 μm, 0.20 μm, 0.10 μm, and 0.05 μm respectively. In order to mitigate the inconsistencies in the color of the images, stain normalization was performed. Next, for each magnification factor, a deep residual network (ResNet) with 152 layers has been trained for classifying patches from the images as benign or malignant. Then, a meta-decision tree was used to combine classification results of a patient’s images from different magnification factors to provide a patient-level diagnosis. The ResNets achieved correct classification rates (CCR) of 98.52%, 97.90%, 98.33%, and 97.66% at x40, x100, x200, and x400 magnification factors, respectively. For classification of patients either as benign or malignant, a CCR of 98.77% was obtained. In conclusion, our study showed that the proposed framework can be helpful in the categorization of breast digital slides.

Finding mamographic masses located in a dense breast tissue is a challenge even for an experienced radiologist. The difficulty comes from the similarity of intensity between the masses and the overlapped normal dense tissues. In this study, a novel method for classification of masses localized in dense background of breast is proposed. The method can identify meaningful superpixel patterns present in mammograms within mass-like regions. The topology of superpixel patterns, captured by using spatial connectivity graphs, revealed significant differences between cancerous and healthy areas of breasts. Four clinically recognizable features were extracted from the superpixel graphs and used for classification. The proposed approach was evaluated using ninety three dense ROIs selected from the publicly available Digital Database for Screening Mammography (DDSM). All 93 ROIs were localized in dense backgrounds of breasts. Among them, 41 contained malignant masses in dense backgrounds and 52 contained healthy dense breast tissues. The results indicate that the graph features generated from superpixel pattern graphs can produce very effective and efficient feature descriptors of breast masses localized in dense background. Using Fisher Linear Discriminant Analysis (LDA) classifier AUC score of 0.90 was achieved for the four dimensional feature vector.

Recently, Convolutional Neural Networks (CNNs) have been successfully used to detect microcalcifications in mammograms. An important step in CNN-based detection is image preprocessing that, in raw mammograms, is usually employed to equalize or remove the intensity-dependent quantum noise. In this work, we show how removing the noise can significantly improve the microcalcification detection performance of a CNN. To this end, we describe the quantum noise with a uniform square-root model. Under this assumption, the generalized Anscombe transformation is applied to the raw mammograms by estimating the noise characteristics from the image at hand. In the Anscombe domain, noise is filtered through an adaptive Wiener filter. The denoised images are recovered with an appropriate inverse transformation and are then used to train the CNN-based detector. Experiments were performed on 1,066 mammograms acquired with GE Senographe systems. MC detection performance of a CNN on noise-free mammograms was statistically significantly higher than on unprocessed mammograms. Results were also superior in comparison with a nonparametric noise-equalizing transformation previously proposed for digital mammograms.

The Convolutional Neural Network (CNN) is a promising technique to detect breast cancer based on mammograms. Training the CNN from scratch, however, requires a large amount of labeled data. Such a requirement usually is infeasible for some kinds of medical image data such as mammographic tumor images. Because improvement of the performance of a CNN classifier requires more training data, the creation of new training images -- image augmentation – could be one solution to this problem. In this study, we applied the Generative Adversarial Network (GAN) to generate synthetic mammographic images from the Digital Database for Screening Mammography (DDSM). From the DDSM, we cropped two sets of regions of interest (ROIs) from the images: normal and abnormal (cancer/tumor) Those ROIs were used to train the GAN, and the GAN then generated synthetic images. To compare the GAN with the affine transformation augmentation methods, such as rotation, shifting, scaling, etc., we used six groups of ROIs (three simple groups: affine augmented, GAN synthetic, real (original), and three mixture groups of each pair of the three simple groups) for each to train a CNN classifier from scratch. And, we used real ROIs that were not used in training to validate classification outcomes. Our results show that, to classify the normal ROIs and abnormal (tumor) ROIs from DDSM, adding GAN-generated ROIs to the training data can reduce overfitting of the classifier. But the affine transformations performed slightly better than GAN. Therefore, GAN could be an optional augmentation approach. The images augmented by GAN or affine transformation cannot substitute entirely for real images to train CNN classifiers because the absence of real images in the training set will cause serious over-fitting with more training.

The purpose of this study is comparing the detection performance in 2D full field digital mammography (FFDM) and digital breast tomosynthesis (DBT) using a structured phantom with inserted target objects. The phantom consists of a semi-cylindrical PMMA container, filled with water and PMMA spheres of different diameters. Microcalcifications and 3D printed masses (spiculated and non-spiculated) were inserted. The phantom was imaged ten times in both modes of five systems, using automatic exposure control (AEC) and at half and double the AEC dose. Five readers evaluated target detectability in a four-alternative forced-choice study. The percentage of correct responses (PC) was assessed based on 10 trials of each reader for each object type, size, imaging modality and dose level. Additionally, detection threshold diameters at 62.5 PC were assessed via non-linear regression fitting of the psychometric curve. Evaluation of target detection in FFDM showed that spiculated masses were better detected compared to non-spiculated masses. In DBT, detection of both mass types increased significantly (p=0.0001) compared to FFDM. Microcalcification detection thresholds ranged between 110 and 118 μm and were similar for the five systems in FFDM while larger variations (106-158 μm) were found in DBT. Mass detection was independent of dose in FFDM while weak dependence was seen for DBT. Microcalcification detection increased with increasing dose for both modalities. The phantom was able to show detectability differences between FFDM and DBT mode for five commercial systems in line with the findings from clinical trials. We suggest to use the phantom for task-based assessment methods for acceptance and commissioning testing of DBT systems.

Detection of lesions is an essential part of making a diagnosis in mammography and therefore is a main focus in the development of algorithms built for image quality assessment. We propose a hybrid approach with an accurate lesion projection model and embedding of lesions into clinical images that already contain relevant structures of anatomical noise. Using an algebraic lesion model, lesions with different sizes and contrasts are generated. The projection algorithm incorporates the modeling of blur effects due to system movement and the physical extent of the anode. Signal and background patches are extracted and used to evaluate channelized Hotelling observers with Laguerre-Gauss channels and with Gabor channels. A four-alternative forced-choice study with five medical imaging experts is performed and the inter-reader agreement with and without the model observers is determined by using Fleiss' kappa. Analyzing three different sizes for tiny, dense lesions and four density levels for larger mass-like lesions we find a good detection rate of the tiny lesions for both human as well as model observers. The inter-reader agreement using the common interpretation of Fleiss' kappa is substantial or better. Comparing full-field digital mammography and digital breast tomosynthesis w.r.t. the different mass densities we find that human readers and model observers perform well on the DBT data and the detection rate drops with lesion contrast as expected. The inter-reader agreement here is fair for the lowest contrast and substantial for the denser cases. Both human readers and model observers show difficulty in detecting the low contrast lesions in FFDM images. The inter-reader agreement is rather poor among all readers. Overall, the results indicate a good agreement between human observers and model observers and a distinctive benefit of 3-D reconstruction over FFDMs for low contrast lesions.

The 2D synthetic image (SM) generated from digital breast tomosynthesis (DBT) has the potential to replace conventional digital mammography (DM), therefore reducing patient dose without affecting the cancer detection performance. In this work, we analysed the image quality of SMs from three different manufacturers for the specific task of detecting microcalcifications (MC), in comparison to DM. A phantom with MC clusters on a uniform background was employed, thus also allowing to explore its feasibility to be used for quality control (QC). A 4-Alternative Forced Choice (4AFC) experiment was performed by four human observers, for detection of MC clusters on a region-of-interest level. We also explored the possibility to replace human observers with a virtual observer. For this, we developed a deep learning convolutional neural network (CNN) for the task of classifying the same images from the 4AFC study, and then compare the results to the human-based study. The results showed that for the four readers and all the systems, the percentage of correct answers (PC) was 100% and the visibility was 3 for the largest MC clusters. However, SM yielded worse detectability than DM for MC with sizes between 180 and 100 μm (PC was around 18% inferior in average). The CNN yielded the same relative results across modalities and systems than the 4AFC study, but in terms of the area under the receiver operating characteristic curve. This might encourage the possibility to develop QC procedures based on artificial intelligence image reading, improving reproducibility and reducing costs.

The aim of this paper is to propose a deep learning framework for micro-calcification detection in 2D mammography and in 2D synthetic mammography (C-view) from digital breast tomosynthesis (DBT). The dataset analyzed for 2D mammograms is the INbreast dataset that consists of 410 digital images and we used 360 images with annotated micro-calcifications. For the synthetic views in DBT, we used a private dataset of 245 images, where micro-calcifications were validated by an experienced radiologist. The network is trained in a patch-based fashion, where micro-calcifications are considered positive samples, while patches containing other breast tissues are considered negative. For evaluating the entire dataset, a 2-fold cross validation was performed. In addition, a sliding window method was used to classify new patches within an image with those from the trained model. Considering 5,656 positive samples and 18,000,000 of negative samples, results for the 2D mammography, on the entire dataset, showed an area under the curve (AUC) of 0.9998 and a logarithmic partial area under the curve (logPAUC), in the interval (10⁻⁶ , 1), of 0.8252. Results for the C-View, considering 3,420 positive samples and 11,395,939 of negative samples, showed an AUC, on the entire dataset, of 0.9997 and a logPAUC, in the interval (10⁻⁶ , 1), of 0.8178. In this paper, we illustrate the applied methodologies, the network architecture used for training and test, and the results obtained.

Mammography is considered the most appropriate method to defeat breast cancer in population screening. Although mammography is used for the detection of micro-calcifications, the benign/malignant classification relies on subsequent histology assessment. We describe a novel method for the automatic classification of benign and malignant micro-calcifications directly from mammographic images, which uses the morphology as well as the distribution aspects of micro-calcifications and is directly associated with the BIRADS categorization of microcalcifications. The developed approach is building on Iwanowski’s morphology work, but is adding class-extractors and multi-scale aspects. Translation to other application areas is discussed and evaluation based on the MIAS and DDSM datasets show results in line with stateof-the-art micro-calcifications classification approaches.

Literatures have showed that deep learning models can detect a breast cancer with high diagnostic accuracy in the publicly available mammography datasets. The objective of this study is to examine whether the high performance (accuracy) of a deep learning model, trained by the public mammography dataset, can be transferred into the clinic practice by applying it to a new mammography dataset obtained in an academic breast center. An end-to-end CNN architecture was trained on DDSM dataset and transferred to INbreast dataset and the in-house collected dataset. The model achieved validation AUC of 0.82 on DDSM dataset and 0.93 on INbreast dataset. However, it only achieved 0.70 when applied to the in-house dataset. Reviewing the images revealed that the in-house dataset is more challenging to classify. The mean subtlety score for DDSM dataset is 3.64 and median is 4. For in-house dataset, the mean and median scores are 2.65 and 2, respectively. In addition, the in-house dataset has more co-existing benign abnormalities as more patients with benign biopsy or prior surgery return for mammography. These observations are in line with other institutes’ finding that the relative percentage of early stage cancer cases from mammography diagnosis has more than tripled since 2002. This indicates that currently available public open datasets may be inadequate to represent the mammography seen in today’s clinical practice. It is necessary to build an updated mammography database that contains sufficient pathological heterogeneity of breast cancer and coexisting benign abnormalities that reflect the cases seen in current practice.

To reduce cumulative radiation exposure and lifetime risks for radiation-induced cancer from breast cancer screening, we developed neural network convolution (NNC) deep learning for radiation dose reduction in digital breast tomosynthesis (DBT). Our NNC deep learning employed patched-based neural network regression in a convolutional manner to convert lower-dose (LD) to higher-dose (HD) tomosynthesis images. We trained our NNC with quarter-dose (25% of the standard dose: 12 mAs at 32 kVp) raw-projection images and corresponding “teaching” higher-dose (HD) images (200% of the standard dose: 99 mAs at 32 kVp) of a breast cadaver phantom acquired with a DBT system (Selenia Dimensions, Hologic, Inc, Bedford, MA). Once trained, NNC no longer requires HD images. It converts new LD images to images that look like HD images; thus the term “virtual” HD (VHD) images. We reconstructed tomosynthesis slices on a research DBT system. To determine a dose reduction rate, we acquired 4 studies of another test phantom at 4 different radiation doses (1.35, 2.7, 4.04, and 5.39 mGy entrance dose). Structural SIMilarity (SSIM) index was used to evaluate the image quality. Our cadaver phantom experiment demonstrated up to 79% dose reduction. For further testing, we collected half-dose (50% of the standard dose: 32±14 mAs at 33±5 kVp) and full-dose (100% of the standard dose: 68±23 mAs at 33±5 kvp) images of 10 clinical cases with the DBT system at University of Iowa Hospitals and Clinics. Our NNC converted half-dose DBT images of the 10 clinical cases to VHD DBT images that were equivalent to full-dose DBT images, according our observer rating study of 10 breast radiologists. Thus, we achieved 50% dose reduction without sacrificing the image quality.

Axillary lymph node (ALN) status is a prognostic factor for patients with breast cancer. Metastasis of sentinel lymph node (SLN) indicates ALN involvement. In this study, our purpose is to develop a quantitative approach in characterizing the metastasis of ALN on spectral CT using the largest SLN (LSLN) as the surrogate. With IRB approval, a data set of 185 patients with breast cancer was retrospectively collect at Sun Yat-Sen Memorial Hospital in Guangzhou, China. Each patient underwent a preoperative spectral CT scan. A chest and axillary dual-phasic contrast media enhanced scan were acquired with a GE Discovery CT750HD CT scanner while the patient was in supine position. The LSLN was manually identified by radiologists for quantitative image analysis. We used a total of 6 sets of dual-phasic scans including 40 keV monochromatic images, 70 keV monochromatic images, and gemstone spectral images obtained at arterial and venous phases. 82 patients were positive to biopsy-proven cancer metastasis and the remaining 103 were negative. A deep convolutional neural network (DCNN) was used to extract quantitative image features as the image representation of SLN. To assess the efficacy of quantitative image features in characterization of SLN, three machine learning classifiers including KNN, SVM, and random forest were compared. Ten-fold cross validation was used for model selection. Results indicated that the AUCs on the 6 CT images for classification of LSLN metastasis ranged from 0.71-0.78 in which the best classification were observed on 70 keV monochromatic images at arterial phase. The overall classifications in arterial phase were better than those in venous phase for low (40 keV) and mixture energy setting while the findings were reversed for high (70 keV) energy setting. Future work is underway to assess our quantitative measures in axillary staging.

Microcalcifications (MC) are small deposits of calcium, which are associated with early signs of breast cancer. In this paper, a novel approach is presented to develop a computer-aided diagnosis (CADx) system for automatic differentiation between benign and malignant MC clusters based on their morphology, texture, and the distribution of individual and global features using an ensemble classifier. The images were enhanced, segmented and the feature extraction and selection phase were carried out to generate the feature space which was later fed into an ensemble classifier to classify the MC clusters. The validity of the proposed method was investigated by using two well-known digitized datasets that contain biopsy proven results for MC clusters: MIAS (24 images: 12 benign, 12 malignant) and DDSM (280 images: 148 benign and 132 malignant). A high classification accuracies (100% for MIAS and 91.39% for DDSM) and good ROC results (area under the ROC curve equal to 1 for MIAS and 0.91 for DDSM) were achieved. A full comparison with related publications is provided. The results indicate that the proposed approach is outperforming the current state-of-the-art methods.

To compensate for an insufficiency of the case images needed in the development of CAD (Computer-Aided Diagnosis), work is underway to create artificial case images by embedding tumors and other such lesions into lesion-free images. Previously, the authors have demonstrated the effectiveness of creating artificial case images for hepatic and breast tumors and utilizing them in CAD development. Thus far, however, when training data comprising 50% or more artificial cases is used in CAD development, the resulting discrimination performance on test data has tended to be somewhat inferior compared to when training data comprises only actual cases.

With the objectives of applying artificial case images to a greater range of sites and of using exclusively artificial cases to develop a high-performance discriminator, in this study, effectiveness verification was conducted that focused on breast cancer calcifications as a new target. Because the characteristics of calcification shadows differ substantially from those of the hepatic and breast cancer tumor shadows studied thus far, a new artificial image creation technique was developed. Artificial cases created using this technique were applied to CAD development. As a result, a discriminator trained with 100% artificial cases obtained detection performance equal to that of a discriminator trained with entirely actual cases.

Accurate segmentation of breast region is an essential step for quantitative analysis of breast parenchyma on mammograms. Pectoral muscle identification on mediolateral oblique (MLO) view mammograms remains a challenging problem. In this study, our purpose is to develop a supervised deep learning approach for automated identification of the pectoral muscle on MLO-view mammograms. With IRB approval, 756 MLO-view mammograms including 656 digitized film mammograms (DFM) and 100 full field digital mammograms (DM) were retrospectively collected. The film mammograms were digitized at a pixel size of 50 μm × 50 μm and the DMs were acquired with a GE Senographe system with a pixel size of 100 μm × 100 μm. All mammograms were subsampled to 800 μm × 800 μm before the pectoral muscle analysis. An experienced radiologist manually segmented the pectoral muscle boundary as the reference standard. We constructed a U-Net-like deep convolutional neural network (DCNN) to identify the boundary of the pectoral muscle. The DCNN consisted of a contracting path to capture multi-resolution image context and a symmetric expanding path for prediction of the pectoral muscle region. A total of 15 million parameters in DCNN were trained with a mini-batched gradient decent algorithm by minimizing a binary cross-entropy cost function. Ten-fold crossvalidation was used in training and evaluating the performance of our model. The DCNN-segmented pectoral muscle was compared to the reference standard with three criteria: 1) the percent overlap area (POA), 2) the Hausdorff distance (Hdist) and 3) the average Euclidean distance (AvgDist). We found that the mean POA, the mean Hdist, and the mean AvgDist were 96.0±5.3%, 2.14±1.50 mm, and 0.77± 0.97 mm, respectively. Further study is underway to evaluate its effect on quantitative analysis of mammograms.

In digital breast tomosynthesis (DBT) large number of parameters influence system performance and the requirement to achieve high quality images every day suggest the implementation of a daily quality control (DQC) procedure. In 2D digital mammography, daily QC is typically performed with homogenous plates and a minimal amount of technical inserts for assessment of NNPS, signal to noise, uniformity, defective pixels and other artefacts. This work proposes an alternative means of performing DQC in DBT with a 3D structured phantom that also includes a constancy test of reconstruction stability in the analysis. The aim of the study was to explore deep learning techniques to automatically track deviations from the normal or baseline operating point and compare the results to the standard metrics. As a first test case, changes in dose were investigated.

Feed-forward convolutional neural networks (CNN) have been successfully applied in the medical imaging domain. A 12 layer CNN model was constructed to extract features for image classification. A structured phantom was scanned on a Siemens DBT system at three dose levels: dose set by the automatic exposure control (AEC) system, half this dose and double. After training the CNN on 36 DBT acquisitions (51840 image segments), newly acquired test images were categorized by the algorithm into the dose categories with an accuracy of 99.7%. Parallel to that the standard methods as NNPS and pixel value (PV) mean and variance calculated for the projection and reconstructed planes also show ability to detect the dose level change with some limitations for the reconstructed planes. This result indicates the potential for further use of deep learning algorithms for DQC when using only the reconstructed DBT planes.

Classification of benign and malignant masses in mammograms is a challenging problem. It has wide applications in the development of Computer Aided Diagnosis (CAD) systems, however many challenges still need to be addressed. Due to the risk associated with segmenting the mass region, focus is shifting from selecting the features just from the mass area, to the whole Region of Interest (RoI) containing that mass. Bag of Visual Words (BoVW) techniques are gaining attention for classification tasks in medical imaging by considering RoI as a set of local features. In general BoVW aims to construct a global descriptor based on the extracted local features. In this work, we investigate the performance of BoVW for the classification of benign and malignant mammographic masses. Several features have been explored as the local features and different methods are applied for building the code-book. Subsequently we propose a voting-based approach to encode the features. The proposed approach is evaluated on a subset of DDSM dataset. Initial results reveal classification accuracy as high as 87% and Area Under the Curve (AUC) as 0.93, which are better than the current state-of-the-art approaches applied to the same problem.

This paper presents a novel investigation of machine learning performance by examining probability outputs in conjunction with classification accuracy (CA) and area under the curve (AUC). One of the main issues in the deployment of computer-aided detection/diagnosis (CAD) systems is lack of ‘trust’ of clinicians in the CAD system, increasing the possibility of the system not being used. Whilst most authors evaluate the performance of their breast CAD systems based on CA and AUC, we study the distribution of the classifiers’ probability outputs and use it as an additional confidence level metric to indicate the reliability of a computer system. Experimental results suggest that although most classifiers produce similar results in terms of CA and AUC (less than 2% variation), their performances are significantly different when considering confidence level (10 to 25% difference). This study may provide opportunities for refining radiologists’ interaction with CAD systems and improving the reliability of CAD systems as well as diagnostic decision making in medicine with high CA or AUC with high degree of certainty.

Previous studies found that multiple view techniques improved the accuracy of lesion detection on mammograms. One of the key components in multiple view techniques was the detection of nipple location, which is the only reliable landmark on mammograms. In this study, our purpose was to develop a novel nipple detection scheme by using geometric and radiomic information extracted on digital mammography (DM). We first extracted a region of interest (ROI) to limit the region of nipple detection by using breast area and the chest wall orientation. The geometric information along the breast boundary was used to categorize the nipples into obvious and subtle types. A top hat transform was used to identify the location of obvious nipples. For subtle type, the radiomic feature matrix was calculated on straightened ROIs along the normal direction of breast boundary. A random forest classifier was trained to combine the radiomic features and to predict the location of subtle nipples. Seven hundred and twenty one DMs were collected for evaluation of our algorithm. A radiologist manually identified the location of nipples as the reference standard. It was found that the average Euclidean distances between the computer and the reference standard were 0.93±5.0 mm for obvious nipples, and 2.74±5.0 mm for subtle nipples, respectively. Future work is underway to evaluate the automated nipples on the registration of abnormalities on multiple view mammograms.

Purpose: The purpose of this study was to apply a neural net to a dataset of women who later experienced either screening detected or interval cancers and determine if it aids in classifying risk of interval cancer compared to using BI-RADS density.

Materials and Methods: Full-field digital screening mammograms acquired in our clinics were reviewed from 2006-2015. Interval cancers were matched to screening-detected cancers based on age, race, exam date, and time since last imaging examination. A deep learning architecture (ResNet50) was trained on this dataset with the goal to classify between interval and screen detected cancers. Network weights were initialized from ImageNet training and the final fully connected layers were retrained. Prediction loss, prediction accuracy, and ROC curves were calculated using this deep learning architecture and compared to predictions from conditional logistic regression using BI-RADS density.

Results: 182 interval and 173 screening-detected cancers were found in our study group. The prediction accuracy improved from 63% using only BI-RADS density to 78% after including predictions from the deep learning model. The area under the ROC curve improved from 0.65 using only BI-RADS density to 0.84 after including the deep learning network as a predictor. Conclusions: We conclude that deep learning methods may be useful in identifying individuals at risk of interval cancer and that these methods can provide additional risk information not contained in breast density alone.

Variability in the color appearance in H and E stained histopathological images are typically observed. Color normalization has been found useful in standardizing the color appearance of H and E stained histopathological images prior to quantitative analysis with machine learning (using handcrafted features). However, its usefulness has not been previously studied when deep convolutional neural networks (CNNs) are used in classifying H and E stained breast cancer histopathological images. In this paper, we have adopted a representative CNN for classifying breast cancer histopathological images and evaluated the benefit/necessity of color normalisation using the commonly used Macenko, Khan and Reinhard color normalization methods. The representative CNN was implemented in-house and was verified. The BreaKHis dataset was used to train and test the CNN model. The preliminary results did not show significant superiority in the CNN performance when color normalization was used to standardize the color appearance of histopathological image. Furthermore, the classification performance of a magnification-independent CNN is comparable to that of magnification-specific CNNs with an additional benefit of a simpler classification scheme and training for only one CNN models (rather than multiple magnificationspecific models). It may also have an advantage in clinical practice when the magnification factor of a histopathological image is not known.

Mammography is a gold standard imaging modality and is widely used for breast cancer screening. With recent advances in the field of deep learning, the use of deep convolution neural networks (CNNs) in medical image analysis has become very encouraging. The aim of this study is to exploit CNNs for mass detection in mammograms using pre-trained networks. We use the resnet-50 CNN architecture pre-trained with the ImageNet database to perform mass detection on two publicly available image datasets: CBIS-DDSM and INbreast. We demonstrate that the CNN model pretrained using natural image database (ImageNet) can be effectively finetuned to yield better results, compared to randomly initialized models. Further, the benefit of applying transfer learning on a smaller dataset is demonstrated by using the best model obtained from CBIS-DDSM training to finetune on the INbreast database. We analyzed the adaptability of the CNN’s last fully connected (FC) layer and the all convolutional layers to detect masses. The results showed a testing accuracy of 0.92 and an area under the receiver operating characteristic curve (AUC) of 0.98 for the model finetuned on all convolutional layers, while testing accuracy of 0.86 and AUC=0.93 when the model is trained only on the last FC layer.

An accurate measurement of the breast glandular fraction, or glandularity, is important for many research and clinical applications, such as breast cancer risk assessment. We propose a method to estimate the loss of glandular tissue detail due to the limited voxel size in tomographic images of the breast. CT images of a breast tissue specimen were acquired using a CdTe single photon counting detector (nominal pixel size of 60 μm) and using a monochromatic synchrotron radiation x-ray beam. Images were reconstructed using a filtered backprojection algorithm at seven different voxel sizes (range 60-420 μm, with a 60 μm step) and twelve groups of Regions of Interest (ROIs) with different percentage and patterns of glandular tissue were extracted. All ROIs within each group contained the same portion of the image (and therefore the same glandular fraction) reconstructed at a different voxel size. The glandular tissue was segmented and the glandularity calculated for all ROIs. A machine learning algorithm was trained on the glandularity values as a function of reconstruction voxel size. After the training was completed, the algorithm could estimate, given a tomographic breast image reconstructed at a given voxel size with a certain glandularity, the increase (or decrease) of glandularity if the same image were reconstructed with a smaller (or larger) voxel dimension. The algorithm was tested on six additional groups of ROIs, resulting in an average relative standard error between the calculated and estimated glandularity of 0.02 ± 0.016.