Search

Advances in both imaging and computers have led to the rise in the potential use of artificial intelligence (AI) in various tasks in breast imaging, going beyond the current use in computer-aided detection to include diagnosis, prognosis, response to therapy, and risk assessment. The automated capabilities of AI offer the potential to enhance the diagnostic expertise of clinicians, including accurate demarcation of tumor volume, extraction of characteristic cancer phenotypes, translation of tumoral phenotype features to clinical genotype implications, and risk prediction. The combination of image-specific findings with the underlying genomic, pathologic, and clinical features is becoming of increasing value in breast cancer. The concurrent emergence of newer imaging techniques has provided radiologists with greater diagnostic tools and image datasets to analyze and interpret. Integrating an AI-based workflow within breast imaging enables the integration of multiple data streams into powerful multidisciplinary applications that may lead the path to personalized patient-specific medicine. In this article we describe the goals of AI in breast cancer imaging, in particular MRI, and review the literature as it relates to the current application, potential, and limitations in breast cancer. Level of Evidence: 3 Technical Efficacy: Stage 3 J. Magn. Reson. Imaging 2020;51:1310-1324.

PURPOSE Radiomic texture analysis is typically performed on images acquired under specific, homogeneous imaging conditions. These controlled conditions may not be representative of the range of imaging conditions implemented clinically. We aim to develop a two-stage method of radiomic texture analysis that incorporates the reproducibility of individual texture features across imaging conditions to guide the development of texture signatures which are robust across mammography unit vendors. METHODS Full-field digital mammograms were retrospectively collected for women who underwent screening mammography on both a Hologic Lorad Selenia and GE Senographe 2000D system. Radiomic features were calculated on manually placed regions of interest in each image. In stage one (robustness assessment), we identified a set of nonredundant features that were reproducible across the two different vendors. This was achieved through hierarchical clustering and application of robustness metrics. In stage two (classification evaluation), we performed stepwise feature selection and leave-one-out quadratic discriminant analysis (QDA) to construct radiomic signatures. We refer to this two-state method as robustness assessment, classification evaluation (RACE). These radiomic signatures were used to classify the risk of breast cancer through receiver operator characteristic (ROC) analysis, using the area under the ROC curve as a figure of merit in the task of distinguishing between women with and without high-risk factors present. Generalizability was investigated by comparing the classification performance of a feature set on the images from which they were selected (intravendor) to the classification performance on images from the vendor on which it was not selected (intervendor). Intervendor and intravendor performances were also compared to the performance obtained by implementing ComBat, a feature-level harmonization method and to the performance by implementing ComBat followed by RACE. RESULTS Generalizability, defined as the difference between intervendor and intravendor classification performance, was shown to monotonically decrease as the number of clusters used in stage one increased (Mann-Kendall P

The goals of this review paper on deep learning (DL) in medical imaging and radiation therapy are to (a) summarize what has been achieved to date; (b) identify common and unique challenges, and strategies that researchers have taken to address these challenges; and (c) identify some of the promising avenues for the future both in terms of applications as well as technical innovations. We introduce the general principles of DL and convolutional neural networks, survey five major areas of application of DL in medical imaging and radiation therapy, identify common themes, discuss methods for dataset expansion, and conclude by summarizing lessons learned, remaining challenges, and future directions.

Background This article is a summary of the quantitative imaging subgroup of the 2017 AAPM Practical Big Data Workshop (PBDW-2017) on progress and challenges in big data applied to cancer treatment and research supplemented by a draft white paper following an American Association of Physicists in Medicine FOREM meeting on Imaging Genomics in 2014. Aims The goal of PBDW-2017 was to close the gap between theoretical vision and practical experience with encountering and solving challenges in curating and analyzing data. Conclusions Recommendations based on the meetings are summarized.

Background Deep learning methods for radiomics/computer-aided diagnosis (CADx) are often prohibited by small datasets, long computation time, and the need for extensive image preprocessing. Aims We aim to develop a breast CADx methodology that addresses the aforementioned issues by exploiting the efficiency of pre-trained convolutional neural networks (CNNs) and using pre-existing handcrafted CADx features. Materials & Methods We present a methodology that extracts and pools low- to mid-level features using a pretrained CNN and fuses them with handcrafted radiomic features computed using conventional CADx methods. Our methodology is tested on three different clinical imaging modalities (dynamic contrast enhanced-MRI [690 cases], full-field digital mammography [245 cases], and ultrasound [1125 cases]). Results From ROC analysis, our fusion-based method demonstrates, on all three imaging modalities, statistically significant improvements in terms of AUC as compared to previous breast cancer CADx methods in the task of distinguishing between malignant and benign lesions. (DCE-MRI [AUC = 0.89 (se = 0.01)], FFDM [AUC = 0.86 (se = 0.01)], and ultrasound [AUC = 0.90 (se = 0.01)]). Discussion/Conclusion We proposed a novel breast CADx methodology that can be used to more effectively characterize breast lesions in comparison to existing methods. Furthermore, our proposed methodology is computationally efficient and circumvents the need for image preprocessing.

Purpose To develop and assess a full-coverage, sensitivity encoding (SENSE)-accelerated breast high spatial and spectral resolution (HiSS) magnetic resonance imaging (MRI) within clinically reasonable times as a potential nonenhanced MRI protocol for breast density measurement or breast cancer screening. Materials and Methods Sixteen women with biopsy-proven cancer or suspicious lesions, and 13 women who were healthy volunteers or were screened for breast cancer, received 3T breast MRI exams, including SENSE-accelerated HiSS MRI, which was implemented as a submillimeter spatial resolution echo-planar spectroscopic imaging (EPSI) sequence. In postprocessing, fat and water resonance peak height and integral images were generated from EPSI data. The postprocessing software was custom-designed, and new algorithms were developed to enable processing of whole-coverage axial HiSS datasets. Water peak height HiSS images were compared to pre- and postcontrast T1-weighted images. Fat suppression was quantified as parenchymal-to-suppressed-fat signal ratio in HiSS water peak height and nonenhanced T1-weighted images, and artifact levels were scored. Results Approximately a 4-fold decrease in acquisition speed, with a concurrent 2.5-fold decrease in voxel size, was achieved, with low artifact levels, and with spectral signal-to-noise ratio (SNR) of 45:1. Fat suppression was 1.9 times more effective (P < 0.001) in HiSS images than in T1-weighted images (SPAIR), and HiSS images showed higher SNR in the axilla. HiSS MRI visualized 10 of 13 malignant lesions identified on dynamic contrast-enhanced (DCE)-MRI, and did not require skin removal in postprocessing to generate maximum intensity projection images. Conclusion We demonstrate full-coverage, SENSE-accelerated breast HiSS MRI within clinically reasonable times, as a potential protocol for breast density measurement or breast cancer screening. Level of Evidence: 2 J. Magn. Reson. Imaging 2017.

Objective In lupus nephritis, tubulointerstitial inflammation (TII) is associated with in situ adaptive immune cell networks that amplify local tissue damage. Since conventional therapy appears ineffective for severe TII, and these patients often progress to renal failure, understanding in situ mechanisms might reveal new therapeutic targets. This study was undertaken to assess whether dysregulated apoptotic regulators maintain local adaptive immunity and drive inflammation in TII. Methods This study utilized novel computational approaches that, when applied to multicolor confocal images, quantified apoptotic regulator protein expression in selected lymphocyte subsets. This approach was validated using laser-capture microdissection (LCM) coupled to quantitative polymerase chain reaction (qPCR). Furthermore, the consequences of dysregulated apoptotic mediator expression were explored in a murine model of lupus nephritis. Results Analyses of renal biopsy tissue from patients with lupus nephritis and those with mixed cellular renal allograft rejection revealed that the B cell lymphoma 2 protein (Bcl-2) was frequently expressed in infiltrating lymphocytes, whereas expression of myeloid cell leukemia 1 was low. In contrast, the reciprocal pattern of expression was observed in tonsil germinal centers. These results were consistent with RNA expression data obtained using LCM and qPCR. Bcl-2 was also highly expressed in tubulointerstitial infiltrates in (NZB × NZW)F1 (NZB/NZW) mice. Furthermore, treatment of NZB/NZW mice with ABT-199, a selective oral inhibitor of Bcl-2, prolonged survival and prevented proteinuria and development of TII in a lupus prevention model. Interestingly, glomerular immune complexes were partially ameliorated by ABT-199 treatment, and serum anti-double-stranded DNA antibody titers were unaffected. Conclusion These data demonstrate that Bcl-2 is an attractive therapeutic target in patients with lupus nephritis who manifest TII.

Purpose To use manganese-enhanced MRI (MEMRI) to detect changes in calcium handling associated with cardiac hypertrophy in a mouse model, and to determine whether the impact of creatine kinase ablation is detectable using this method.

Purpose To compare the performance of computer-aided diagnosis (CADx) analysis of precontrast high spectral and spatial resolution (HiSS) MRI to that of clinical dynamic contrast-enhanced MRI (DCE-MRI) in the diagnostic classification of breast lesions. Materials and Methods Thirty-four malignant and seven benign lesions were scanned using two-dimensional (2D) HiSS and clinical 4D DCE-MRI protocols. Lesions were automatically segmented. Morphological features were calculated for HiSS, whereas both morphological and kinetic features were calculated for DCE-MRI. After stepwise feature selection, Bayesian artificial neural networks merged selected features, and receiver operating characteristic (ROC) analysis evaluated the performance with leave-one-lesion-out validation. Results AUC (area under the ROC curve) values of 0.92 ± 0.06 and 0.90 ± 0.05 were obtained using CADx on HiSS and DCE-MRI, respectively, in the task of classifying benign and malignant lesions. While we failed to show that the higher HiSS performance was significantly better than DCE-MRI, noninferiority testing confirmed that HiSS was not worse than DCE-MRI. Conclusion CADx of HiSS (without contrast) performed similarly to CADx on clinical DCE-MRI; thus, computerized analysis of HiSS may provide sufficient information for diagnostic classification. The results are clinically important for patients in whom contrast agent is contra-indicated. Even in the limited acquisition mode of 2D single slice HiSS, by using quantitative image analysis to extract characteristics from the HiSS images, similar performance levels were obtained as compared with those from current clinical 4D DCE-MRI. As HiSS acquisitions become possible in 3D, CADx methods can also be applied. Because HiSS and DCE-MRI are based on different contrast mechanisms, the use of the two protocols in combination may increase diagnostic accuracy. J. Magn. Reson. Imaging 2014;39:59–67. © 2013 Wiley Periodicals, Inc.

Purpose: To develop an automated method to detect breast masses on dedicated breast CT (BCT) volumes and to conduct a preliminary evaluation of its performance. This method can be used in a computer-aided detection (CADe) system for noncontrast enhanced BCT. Methods: The database included patient images, which were acquired under an IRB-approved protocol. The database in this study consisted of 132 cases. 50 cases contained 58 malignant masses, and 23 cases contained 24 benign masses. 59 cases did not contain any biopsy-proven lesions. Each case consisted of an unenhanced CT volume of a single breast. First, each breast was segmented into adipose and glandular tissues using a fuzzy c-means clustering algorithm. The glandular breast regions were then sampled at a resolution of 2 mm. At each sampling step, a 3.5-cm3 volume-of-interest was subjected to constrained region segmentation and 17 characteristic features were extracted, yielding 17 corresponding feature volumes. Four features were selected using step-wise feature selection and merged with linear discriminant analysis trained in the task of distinguishing between normal breast glandular regions and masses. Detection performance was measured using free-response receiver operating characteristic analysis (FROC) with leave-one-case-out evaluation. Results: The feature selection stage selected features that characterized the shape and margin strength of the segmented region. CADe sensitivity per case was 84% (std = 4.2%) at 2.6 (std = 0.06) false positives per volume, or 6 × 10−3 per slice (at an average of 424 slices per volume in this data set). Conclusions: This preliminary study demonstrates the feasibility of our approach for CADe for BCT.

Purpose: In cardiac computed tomography (CT), important clinical indices, such as the coronary calcium score and the percentage of coronary artery stenosis, are often adversely affected by motion artifacts. As a result, the expert observer must decide whether or not to use these indices during image interpretation. Computerized methods potentially can be used to assist in these decisions. In a previous study, an artificial neural network (ANN) regression model provided assessability (image quality) indices of calcified plaque images from the software NCAT phantom that were highly agreeable with those provided by expert observers. The method predicted assessability indices based on computer-extracted features of the plaque. In the current study, the ANN-predicted assessability indices were used to identify calcified plaque images with diagnostic calcium scores (based on mass) from a physical dynamic cardiac phantom. The basic assumption was that better quality images were associated with more accurate calcium scores. Methods: A 64-channel CT scanner was used to obtain 500 calcified plaque images from a physical dynamic cardiac phantom at different heart rates, cardiac phases, and plaque locations. Two expert observers independently provided separate sets of assessability indices for each of these images. Separate sets of ANN-predicted assessability indices tailored to each observer were then generated within the framework of a bootstrap resampling scheme. For each resampling iteration, the absolute calcium score error between the calcium scores of the motion-contaminated plaque image and its corresponding stationary image served as the ground truth in terms of indicating images with diagnostic calcium scores. The performances of the ANN-predicted and observer-assigned indices in identifying images with diagnostic calcium scores were then evaluated using ROC analysis. Results: Assessability indices provided by the first observer and the corresponding ANN performed similarly (AUCOBS1=0.80 [0.73,0.86] vs AUCANN1=0.88 [0.82,0.92]) as that of the second observer and the corresponding ANN (AUCOBS2=0.87 [0.83,0.91] vs AUCANN2=0.90 [0.85,0.94]). Moreover, the ANN-predicted indices were generated in a fraction of the time required to obtain the observer-assigned indices. Conclusions: ANN-predicted assessability indices performed similar to observer-assigned assessability indices in identifying images with diagnostic calcium scores from the physical dynamic cardiac phantom. The results of this study demonstrate the potential of using computerized methods for identifying images with diagnostic clinical indices in cardiac CT images.

Purpose: The aim of this study was to investigate the concept of repeatability in a case-based performance evaluation of two classifiers commonly used in computer-aided diagnosis in the task of distinguishing benign from malignant lesions. Methods: The authors performed .632 + bootstrap analyses using a data set of 1251 sonographic lesions of which 212 were malignant. Several analyses were performed investigating the impact of sample size and number of bootstrap iterations. The classifiers investigated were a Bayesian neural net (BNN) with five hidden units and linear discriminant analysis(LDA). Both used the same four input lesion features. While the authors did evaluate classifier performance using receiver operating characteristic (ROC) analysis, the main focus was to investigate case-based performance based on the classifier output for individual cases, i.e., the classifier outputs for each test case measured over the bootstrap iterations. In this case-based analysis, the authors examined the classifier output variability and linked it to the concept of repeatability. Repeatability was assessed on the level of individual cases, overall for all cases in the data set, and regarding its dependence on the case-based classifier output. The impact of repeatability was studied when aiming to operate at a constant sensitivity or specificity and when aiming to operate at a constant threshold value for the classifier output. Results: The BNN slightly outperformed the LDA with an area under the ROC curve of 0.88 versus 0.85 ( p 0.05 ) . In the repeatability analysis on an individual case basis, it was evident that different cases posed different degrees of difficulty to each classifier as measured by the by-case output variability. When considering the entire data set, however, the overall repeatability of the BNN classifier was lower than for the LDA classifier, i.e., the by-case variability for the BNN was higher. The dependence of the by-case variability on the average by-case classifier output was markedly different for the classifiers. The BNN achieved the lowest variability (best repeatability) when operating at high sensitivity ( > 90 % ) and low specificity ( 66 % ) , while the LDA achieved this at moderate sensitivity ( ∼ 74 % ) and specificity ( ∼ 84 % ) . When operating at constant 90% sensitivity or constant 90% specificity, the width of the 95% confidence intervals for the corresponding classifier output was considerable for both classifiers and increased for smaller sample sizes. When operating at a constant threshold value for the classifier output, the width of the 95% confidence intervals for the corresponding sensitivity and specificity ranged from 9 percentage points (pp) to 30 pp. Conclusions: The repeatability of the classifier output can have a substantial effect on the obtained sensitivity and specificity. Knowledge of classifier repeatability, in addition to overall performance level, is important for successful translation and implementation of computer-aided diagnosis in clinical decision making.

The roles of physicists in medical imaging have expanded over the years, from the study of imaging systems (sources and detectors) and dose to the assessment of image quality and perception, the development of image processing techniques, and the development of image analysis methods to assist in detection and diagnosis. The latter is a natural extension of medical physicists' goals in developing imaging techniques to help physicians acquire diagnostic information and improve clinical decisions. Studies indicate that radiologists do not detect all abnormalities on images that are visible on retrospective review, and they do not always correctly characterize abnormalities that are found. Since the 1950s, the potential use of computers had been considered for analysis of radiographic abnormalities. In the mid-1980s, however, medical physicists and radiologists began major research efforts for computer-aided detection or computer-aided diagnosis (CAD), that is, using the computer output as an aid to radiologists-as opposed to a completely automatic computer interpretation-focusing initially on methods for the detection of lesions on chest radiographs and mammograms. Since then, extensive investigations of computerized image analysis for detection or diagnosis of abnormalities in a variety of 2D and 3D medical images have been conducted. The growth of CAD over the past 20 years has been tremendous-from the early days of time-consuming film digitization and CPU-intensive computations on a limited number of cases to its current status in which developed CAD approaches are evaluated rigorously on large clinically relevant databases. CAD research by medical physicists includes many aspects-collecting relevant normal and pathological cases; developing computer algorithms appropriate for the medical interpretation task including those for segmentation, feature extraction, and classifier design; developing methodology for assessing CAD performance; validating the algorithms using appropriate cases to measure performance and robustness; conducting observer studies with which to evaluate radiologists in the diagnostic task without and with the use of the computer aid; and ultimately assessing performance with a clinical trial. Medical physicists also have an important role in quantitative imaging, by validating the quantitative integrity of scanners and developing imaging techniques, and image analysis tools that extract quantitative data in a more accurate and automated fashion. As imaging systems become more complex and the need for better quantitative information from images grows, the future includes the combined research efforts from physicists working in CAD with those working on quantitative imaging systems to readily yield information on morphology, function, molecular structure, and more-from animal imaging research to clinical patient care. A historical review of CAD and a discussion of challenges for the future are presented here, along with the extension to quantitative image analysis.

Identifying the corresponding images of a lesion in different views is an essential step in improving the diagnostic ability of both radiologists and computer-aided diagnosis (CAD) systems. Because of the nonrigidity of the breasts and the 2D projective property of mammograms, this task is not trivial. In this pilot study, we present a computerized framework that differentiates between corresponding images of the same lesion in different views and noncorresponding images, i.e., images of different lesions. A dual-stage segmentation method, which employs an initial radial gradient index (RGI) based segmentation and an active contour model, is applied to extract mass lesions from the surrounding parenchyma. Then various lesion features are automatically extracted from each of the two views of each lesion to quantify the characteristics of density, size, texture and the neighborhood of the lesion, as well as its distance to the nipple. A two-step scheme is employed to estimate the probability that the two lesion images from different mammographic views are of the same physical lesion. In the first step, a correspondence metric for each pairwise feature is estimated by a Bayesian artificial neural network (BANN). Then, these pairwise correspondence metrics are combined using another BANN to yield an overall probability of correspondence. Receiver operating characteristic (ROC) analysis was used to evaluate the performance of the individual features and the selected feature subset in the task of distinguishing corresponding pairs from noncorresponding pairs. Using a FFDM database with 123 corresponding image pairs and 82 noncorresponding pairs, the distance feature yielded an area under the ROC curve (AUC) of 0.81±0.02 with leave-one-out (by physical lesion) evaluation, and the feature metric subset, which included distance, gradient texture, and ROI-based correlation, yielded an AUC of 0.87±0.02. The improvement by using multiple feature metrics was statistically significant compared to single feature performance.

To perform a pilot study investigating whether the sensitivity and specificity of kinetic parameters can be improved by considering mass and nonmass breast lesions separately. The contrast media uptake and washout kinetics in benign and malignant breast lesions were analyzed using an empirical mathematical model (EMM), and model parameters were compared in lesions with mass-like and nonmass-like enhancement characteristics. 34 benign and 78 malignant breast lesions were selected for review. Dynamic MR protocol: 1 pre and 5 postcontrast images acquired in the coronal plane using a 3D T1-weighted SPGR with 68 s timing resolution. An experienced radiologist classified the type of enhancement as mass, nonmass, or focus, according to the BI-RADS lexicon. The kinetic curve obtained from a radiologist-drawn region within the lesion was analyzed quantitatively using a three parameter EMM. Several kinetic parameters were then derived from the EMM parameters: the initial slope (Slope(ini)), curvature at the peak (kappa(peak)), time to peak (T(peak)), initial area under the curve at 30 s (iAUC30), and the signal enhancement ratio (SER). The BI-RADS classification of the lesions yielded: 70 mass lesions, 38 nonmass, 4 focus. For mass lesions, the contrast uptake rate (alpha), contrast washout rate (beta), iAUC30, SER, Slope(ini), T(peak) and kappa(peak) differed substantially between benign and malignant lesions, and after correcting for multiple tests of significance SER and T(peak) demonstrated significance (p 0.5). Kinetic parameters could distinguish benign and malignant mass lesions effectively, but were not quite as useful in discriminating benign from malignant nonmass lesions. If the results of this pilot study are validated in a larger trial, we expect that to maximize diagnostic utility, it will be better to classify lesion morphology as mass or nonmass-like enhancement prior to kinetic analysis.

Periprosthetic osteolysis is one of the most serious long-term problems in total hip arthroplasty. It has been primarily attributed to the body's inflammatory response to submicron polyethylene particles worn from the hip implant, and it leads to bone loss and structural deterioration in the surrounding bone. It was previously demonstrated that radiographic texture analysis (RTA) has the ability to distinguish between osteolysis and normal cases at the time of clinical detection of the disease; however, that analysis did not take into account the changes in texture over time. The goal of this preliminary analysis, however, is to assess the ability of temporal radiographic texture analysis (tRTA) to distinguish between patients who develop osteolysis and normal cases. Two tRTA methods were used in the study: the RTA feature change from baseline at various follow-up intervals and the slope of the best-fit line to the RTA data series. These tRTA methods included Fourier-based and fractal-based features calculated from digitized images of 202 total hip replacement cases, including 70 that developed osteolysis. Results show that separation between the osteolysis and normal groups increased over time for the feature difference method, as the disease progressed, with area under the curve (AUC) values from receiver operating characteristic analysis of 0.65 to 0.72 at 15 years postsurgery. Separation for the slope method was also evident, with AUC values ranging from 0.65 to 0.76 for the task of distinguishing between osteolysis and normal cases. The results suggest that tRTA methods have the ability to measure changes in trabecular structure, and may be useful in the early detection of periprosthetic osteolysis.

Automated image analysis aims to extract relevant information from contrast-enhanced magnetic resonance images (CE-MRI) of the breast and improve the accuracy and consistency of image interpretation. In this work, we extend the traditional 2D gray-level co-occurrence matrix (GLCM) method to investigate a volumetric texture analysis approach and apply it for the characterization of breast MR lesions. Our database of breast MR images was obtained using a T1-weighted 3D spoiled gradient echo sequence and consists of 121 biopsy-proven lesions (77 malignant and 44 benign). A fuzzy c-means clustering (FCM) based method is employed to automatically segment 3D breast lesions on CE-MR images. For each 3D lesion, a nondirectional GLCM is then computed on the first postcontrast frame by summing 13 directional GLCMs. Texture features are extracted from the nondirectional GLCMs and the performance of each texture feature in the task of distinguishing between malignant and benign breast lesions is assessed by receiver operating characteristics (ROC) analysis. Our results show that the classification performance of volumetric texture features is significantly better than that based on 2D analysis. Our investigations of the effects of various of parameters on the diagnostic accuracy provided means for the optimal use of the approach.

The recently developed weighted backprojection filtration (WBPF) algorithm using data redundancy has capabilities that make this algorithm an attractive candidate for reconstructing images from motion-contaminated projection data. First, the WBPF algorithm is capable of reconstructing region-of-interest (ROI) images from reduced-scan fan-beam data, which have less data than the short-scan data required to reconstruct the entire field of view (FOV). Second, this algorithm can reconstruct ROI images from truncated data. Using phantom simulation studies, we demonstrate how these unique capabilities can be exploited to reduce the amount of motion-contaminated data used for reconstruction. In particular, we use examples from cardiac imaging to illustrate how off-center phantom positioning combined with phase-interval ROI reconstruction can result in the suppression of motion artifacts. In terms of temporal resolution, reduced-scan reconstruction with 45% of a full-scan dataset can be used to improve the temporal resolution of a short-scan reconstruction by 25.8% if ungated data are used. For data gated at 66 beats per minute, reduced-scan reconstruction with 45% of a full-scan dataset can be used to improve the temporal resolution of a short-scan reconstruction by 7.9%. As a result of our studies, we believe that the WBPF algorithm demonstrates the potential for reconstructing quality ROImore » images from motion-contaminated fan-beam data.« less

Digital breast tomosynthesis (DBT) has recently emerged as a new and promising three-dimensional modality in breast imaging. In DBT, the breast volume is reconstructed from 11 projection images, taken at source angles equally spaced over an arc of 50 degrees . Reconstruction algorithms for this modality are not fully optimized yet. Because computerized lesion detection in the reconstructed breast volume will be affected by the reconstruction technique, we are developing a novel mass detection algorithm that operates instead on the set of raw projection images. Mass detection is done in three stages. First, lesion candidates are obtained for each projection image separately, using a mass detection algorithm that was initially developed for screen-film mammography. Second, the locations of a lesion candidate are backprojected into the breast volume. In this feature volume, voxel intensities are a combined measure of detection frequency (e.g., the number of projections in which a given lesion candidate was detected), and a measure of the angular range over which a given lesion was detected. Third, features are extracted after reprojecting the three-dimensional (3-D) locations of lesion candidates into projection images. Features are combined using linear discriminant analysis. The database used to test the algorithm consisted of 21 mass cases (13 malignant, 8 benign) and 15 cases without mass lesions. Based on this database, the algorithm yielded a sensitivity of 90% at 1.5 false positives per breast volume. Algorithm performance is positively biased because this dataset was used for development, training, and testing, and because the number of algorithm parameters was approximately the same as the number of patient cases. Our results indicate that computerized mass detection in the sequence of projection images for DBT may be effective despite the higher noise level in those images.

The advantages of breast MRI using contrast agent Gd-DTPA in the diagnosis of breast cancer have been well established. The variation of interpretation criteria and absence of interpretation guidelines, however, is a major obstacle for applications of MRI in the routine clinical practice of breast imaging. Our study aims to increase the objectivity and reproducibility of breast MRI interpretation by developing an automated interpretation approach for ultimate use in computer-aided diagnosis. The database in this study contains 121 cases: 77 malignant and 44 benign masses as revealed by biopsy. Images were obtained using a T1-weighted 3D spoiled gradient echo sequence. After the acquisition of the precontrast series, Gd-DTPA contrast agent was injected intravenously by power injection with a dose of 0.2 mmol/kg. Five postcontrast series were then taken with a time interval of 60 s. Each series contained 64 coronal slices with a matrix of 128 x 256 pixels and an in-plane resolution of 1.25 x 1.25 mm2. Slice thickness ranged from 2 to 3 mm depending on breast size. The lesions were delineated by an experienced radiologist as well as independently by computer using an automatic volume-growing algorithm. Fourteen features that were extracted automatically from the lesions could be grouped into three categories based on (I) morphology, (II) enhancement kinetics, and (III) time course of enhancement-variation over the lesion. A stepwise feature selection procedure was employed to select an effective subset of features, which were then combined by linear discriminant analysis (LDA) into a discriminant score, related to the likelihood of malignancy. The classification performances of individual features and the combined discriminant score were evaluated with receiver operating characteristic (ROC) analysis. With the radiologist-delineated lesion contours, stepwise feature selection yielded four features and an Az value of 0.80 for the LDA in leave-one-out cross-validation testing. With the computer-segmented lesion volumes, it yielded six features and an Az value of 0.86 for the LDA in the leave-one-out testing.

The long-term goal of our research is to develop computerized radiographic markers for assessing breast density and parenchymal patterns that may be used together with clinical measures for determining the risk of breast cancer and assessing the response to preventive treatment. In our earlier studies, we found that women at high risk tended to have dense breasts with mammographic patterns that were coarse and low in contrast. With our method, computerized texture analysis is performed on a region of interest (ROI) within the mammographic image. In our current study, we investigate the effect of ROI size and ROI location on the computerized texture features obtained from 90 subjects (30 BRCA1/BRCA2 gene-mutation carriers and 60 age-matched women deemed to be at low risk for breast cancer). Mammograms were digitized at 0.1 mm pixel size and various ROI sizes were extracted from different breast regions in the craniocaudal (CC) view. Seventeen features, which characterize the density and texture of the parenchymal patterns, were extracted from the ROIs on these digitized mammograms. Stepwise feature selection and linear discriminant analysis were applied to identify features that differentiate between the low-risk women and the BRCA1/BRCA2 gene-mutation carriers. ROC analysis was used to assess the performance of the features in the task of distinguishing between these two groups. Our results show that there was a statistically significant decrease in the performance of the computerized texture features, as the ROI location was varied from the central region behind the nipple. However, we failed to show a statistically significant decrease in the performance of the computerized texture features with decreasing ROI size for the range studied.

In recent years heavy focus has been placed on clinical careers in medical physics, and the pathways to such careers are well-known and well-described. For the future advancement of the field of medical physics, it is necessary the focus also include research-oriented careers. It is important for the younger generation of medical physicists to know that there are many career opportunities in research, and that it is quite possible to build a successful career in which research is a major component. We will hear from those who have already blazed a trail of success in academic and industry research careers in order to understand some of the most important decisions to make and skills to develop in order to be successful. Topics such as how to build a research program, how to develop your own lab, and how to initiate good collaborations are not generally taught explicitly but are integral to succeeding in a research career. Skills such as these are often best learned from the experiences of others. The challenges encountered in transitioning from graduate student to faculty member, securing the first grant, and moving into an industry research position will be highlighted and discussed. Learning Objectives: 1.Identify career paths built on interest in research 2.Understand basic steps to starting and maintaining a research career 3.Identify differences between research careers in academia and industry M. Giger, support by NIH and COI will be listed. J. Dempsey, ViewRay, Inc. has provided financial support for my research. R. Mohan, support from NIH.

Purpose: We investigate deep learning in the task of distinguishing between malignant and benign breast lesions on dynamic contrast-enhanced MR images (DCE-MRIs), eliminating the need for lesion segmentation and extraction of tumor features. We evaluate convolutional neural network (CNN) after transfer learning with ImageNet, a database of thousands of non-medical images. Methods: Under a HIPAA-compliant IRB protocol, a database of 551 (357 malignant and 194 benign) breast MRI cases was collected. ROIs around each lesion were extracted from the DCE-MRI slices at the second post-contrast time point. Depending on the size of the tumor, the dimensions of ROIs varied between 1 and 1.5 times the maximum diameter of the lesion, which were further upsampled or downsampled to yield 256×256-pixel ROIs. Next, these ROIs were directly input to the convolutional neural network ConvNet, which had been pre-trained by AlexNet on the ImageNet database. The internal layer of ConvNet output 4,096 features, which were subsequently used as the input to a support vector machine (SVM) to classify the lesions as malignant or benign. Area under the receiver operating characteristic curve (AUC) was used as a figure of merit for the classification task. We performed 10-fold cross-validation with training and testing sets consisting of 90% and 10% of the database, respectively. Results: The CNN with transfer learning and subsequent SVM yielded an an AUC value of 0.85, demonstrating the predictive value of the CNN. In future work, we will compare the results obtained with ConvNet with the results obtained using conventional tumor radiomics features. Our approach returns the prediction within minutes, due to incorporation of transfer learning. Conclusion: A CNN pre-trained on non-medical images can be used to extract image characteristics from breast DCE-MR images relevant to diagnostic decision-making. Our work shows that transfer learning can aid in prediction of breast cancer malignancy. M. Giger is a stockholder in R2/Hologic, co-founder and equity holder in Quantitative Insights, and receives royalties from Hologic, GE Medical Systems, MEDIAN Technologies, Riverain Medical, Mitsubishi, and Toshiba

Peer-reviewed journals and conference proceedings publish hundreds of papers that describe new medical imaging algorithms, including, for example, techniques for computer-aided diagnosis or characterization, segmentation, image registration, image reconstruction, and radiomics. It is difficult, if not impossible, to fairly compare the performance of these algorithms as investigators must either use different data sets, or if using the same data, use different implementations of competing algorithms. Grand Challenges facilitate the fair comparison of algorithms by providing a common data set to all participants and by having each participant be responsible for implementation of their own algorithm. The dissemination of findings from Grand Challenges provides important information to the scientific community and helps to determine which approaches have the greatest promise for successful translation to clinical practice. In this session we will review the outcomes and lessons learned from the 2015 SPIE-AAPM-NCI Lung Nodule Classification Challenge. We will then turn to the 2016 NIH-AAPM-Mayo Clinic Low Dose CT Grand Challenge, providing an overview of denoising and iterative reconstruction approaches and a description of the Challenge. The top 3 performing participants will be announced, and each will give a short presentation on their technique. Learning Objectives: 1. Understand the role of Grand Challenges in the field of medical imaging 2. Be able to summarize the outcomes of the 2015 lung nodule classification challenge 3. Be able to review the primary types of noise reduction techniques used in CT 4. Be familiar with a library of patient CT projection data available to researchers 5. Learn which techniques performed best in the Low Dose CT Grand Challenge Pelc: GE Healthcare, Philips Healthcare; McCollough: Research grant, Siemens Healthcare; Low Dose CT Grand Challenge supported by the AAPM Science Council and NIH (grant EB 017185), and hosted by the Mayo Clinic; Giger: stockholder R2 technology/Hologic, royalties from Hologic, GE Medical Systems, MEDIAN Technologies, Riverain Medical, Mitsubishi/Toshiba. Cofounder/stockholder Quantitative Insights.

We are using Bayesian artificial neural networks (BANNs) to classify mammographic masses in schemes for computer-aided diagnosis, and we are extending this methodology to a three-class classification task. We investigated whether a BANN can estimate ideal observer decision variables to distinguish malignant, benign, and false-positive computer detections. Five features were calculated for 63 malignant and 29 benign computer-detected mass lesions, and for 1049 false-positive computer detections, in 440 mammograms randomly divided into a training and testing set. A BANN was trained on the training set features and applied to the testing set features. We then used a known relation between three-class ideal observer decision variables and that used by a two-class ideal observer when two of three classes are grouped into one class, giving one decision variable for distinguishing malignant from nonmalignant detections, and a second for distinguishing true-positive from false-positive computer detections. For comparison, we grouped the training data into two classes in the same two ways and trained two-class BANNs for these two tasks. The three-class BANN decision variables were essentially identical in performance to the specifically trained two-class BANNs, with the average difference in area under the ROC curves being less than 0.0035 and no differences in area being statistically significant. Thus, the BANN outputs obey the same theoretical relationship as do the three-class and two-class ideal observer decision variables, which is consistent with the claim that the three-class BANN output can provide good estimates of the decision variables used by a three-class ideal observer.

Our purpose in this study was to identify computer-extracted, mammographic parenchymal patterns that are associated with breast cancer risk. We extracted 14 features from the central breast region on digitized mammograms to characterize the mammographic parenchymal patterns of women at different risk levels. Two different approaches were employed to relate these mammographic features to breast cancer risk. In one approach, the features were used to distinguish mammographic patterns seen in low-risk women from those who inherited a mutated form of the BRCA1/BRCA2 gene, which confers a very high risk of developing breast cancer. In another approach, the features were related to risk as determined from existing clinical models (Gail and Claus models), which use well-known epidemiological factors such as a woman's age, her family history of breast cancer, reproductive history, etc. Stepwise linear discriminant analysis was employed to identify features that were useful in differentiating between "low-risk" women and BRCA1/BRCA2-mutation carriers. Stepwise linear regression analysis was employed to identify useful features in predicting the risk, as estimated from the Gail and Claus models. Similar computer-extracted mammographic features were identified in the two approaches. Results show that women at high risk tend to have dense breasts and their mammographic patterns tend to be coarse and low in contrast.

Computer-aided diagnosis has the potential of increasing diagnostic accuracy by providing a second reading to radiologists. In many computerized schemes, numerous features can be extracted to describe suspect image regions. A subset of these features is then employed in a data classifier to determine whether the suspect region is abnormal or normal. Different subsets of features will, in general, result in different classification performances. A feature selection method is often used to determine an "optimal" subset of features to use with a particular classifier. A classifier performance measure (such as the area under the receiver operating characteristic curve) must be incorporated into this feature selection process. With limited datasets, however, there is a distribution in the classifier performance measure for a given classifier and subset of features. In this paper, we investigate the variation in the selected subset of "optimal" features as compared with the true optimal subset of features caused by this distribution of classifier performance. We consider examples in which the probability that the optimal subset of features is selected can be analytically computed. We show the dependence of this probability on the dataset sample size, the total number of features from which to select, the number of features selected, and the performance of the true optimal subset. Once a subset of features has been selected, the parameters of the data classifier must be determined. We show that, with limited datasets and/or a large number of features from which to choose, bias is introduced if the classifier parameters are determined using the same data that were employed to select the "optimal" subset of features.

Purpose: To determine the capability of machine learning and computer-extracted image phenotypes to distinguish between ductal carcinoma in situ and invasive breast cancers, which may facilitate timely assessment of disease prognosis and treatment. Methods: This study used a HIPAA compliant data set of 248 breast dynamic contrast-enhanced MR images of breast cancers collected under IRB-approved protocols (58 in situ and 190 invasive cancers). After automated 3D tumor segmentation, 38 image phenotypes were computer-extracted, describing tumor size, shape, morphology, enhancement texture, variance kinetics, and kinetic curve characteristics. A random forest classifier was used in a leave-one-case-out training/testing paradigm for the distinction between in situ and invasive breast cancer. Performance for this task was assessed using the area under the receiver operating characteristic curve (AUC). The random forest classifier also automatically determined the relevance of all 38 phenotypes to the task at hand. Results: The random forest classifier obtained an AUC of 0.90 (95% confidence interval [0.86;0.95]) in the task of distinguishing between in situ and invasive breast cancers. Phenotypes most important here were shape (sphericity), enhancement texture (variance, difference variance, heterogeneity, contrast), and kinetic curve characteristics (normalized total rate variance, washout rate, time to peak). Phenotypes least effective were size (maximum diameter, surface area), morphology (margin gradient, margin sharpness), variance kinetics, and a kinetic curve characteristic (maximum enhancement). Conclusion: We obtained promising results in automated MR image-based assessment of breast tumor invasiveness. In the current era of personalized medicine such analysis may positively impact patient care by identifying which breast tumors require timely work-up because of invasive components, and which findings could potentially be monitored without aggressive treatment. The image-based phenotypes that emerged in this study are markedly different from those relevant to the distinction between benign and malignant breast tumors, where mainly size and morphology influenced diagnostic decision making. Funding: NIH S10 RR021039, University of Chicago Dean Bridge Fund. and Carole Segal. COI: M.L.G. is a stockholder in R2 Technology/Hologic and receives royalties from Hologic, GE Medical Systems, MEDIAN Technologies, Riverain Medical, Mitsubishi and Toshiba. MLG is a cofounder of and stockholder in Quantitative Insights.

Purpose: To determine associations between radiologist-annotated MRI features and genomic measurements in breast invasive carcinoma (BRCA) from the Cancer Genome Atlas (TCGA). Methods: 98 TCGA patients with BRCA were assessed by a panel of radiologists (TCGA Breast Phenotype Research Group) based on a variety of mass and non-mass features according to the Breast Imaging Reporting and Data System (BI-RADS). Batch corrected gene expression data was obtained from the TCGA Data Portal. The Kruskal-Wallis test was used to assess correlations between categorical image features and tumor-derived genomic features (such as gene pathway activity, copy number and mutation characteristics). Image-derived features were also correlated with estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2/neu) status. Multiple hypothesis correction was done using Benjamini-Hochberg FDR. Associations at an FDR of 0.1 were selected for interpretation. Results: ER status was associated with rim enhancement and peritumoral edema. PR status was associated with internal enhancement. Several components of the PI3K/Akt pathway were associated with rim enhancement as well as heterogeneity. In addition, several components of cell cycle regulation and cell division were associated with imaging characteristics.TP53 and GATA3 mutations were associated with lesion size. MRI features associated with TP53 mutation status were rim enhancement and peritumoral edema. Rim enhancement was associated with activity of RB1, PIK3R1, MAP3K1, AKT1,PI3K, and PIK3CA. Margin status was associated with HIF1A/ARNT, Ras/ GTP/PI3K, KRAS, and GADD45A. Axillary lymphadenopathy was associated with RB1 and BCL2L1. Peritumoral edema was associated with Aurora A/GADD45A, BCL2L1, CCNE1, and FOXA1. Heterogeneous internal nonmass enhancement was associated with EGFR, PI3K, AKT1, HF/MET, and EGFR/Erbb4/neuregulin 1. Diffuse nonmass enhancement was associated with HGF/MET/MUC20/SHIP, and HGF/MET/RANBP9. Linear nonmass enhancement was associated with PIK3R1 and AKT activity. Conclusion: MRI-genomic association analysis revealed that several BRCA-associated gene features were associated with radiologist-annotated image features.

A shift-invariant artificial neutral network (SIANN) has been applied to eliminate the false-positive detections reported by a rule-based computer aided-diagnosis (CAD) scheme developed in our laboratory. Regions of interest (ROIs) were selected around the centers of the rule-based CAD detections and analyzed by the SIANN. In our previous study, background-trend correction and pixel-value normalization were used as the preprocessing of the ROIs prior to the SIANN. A ROI is classified as a positive ROI, if the total number of microcalcifications detected in the ROI is greater than a certain number. In this study, modifications were made to improve the performance of the SIANN. First, the preprocessing is removed because the result of the background-trend correction is affected by the size of ROIs. Second, image-feature analysis is employed to the output of the SIANN in an effort to eliminate some of the false detections by the SIANN. In order to train the SIANN to detect microcalcifications and also to extract image features of microcalcifications, the zero-mean-weight constraint and training-free-zone techniques have been developed. A cross-validation training method was also applied to avoid the overtraining problem. The performance of the SIANN was evaluated by means of ROC analysis using a database of 39 mammograms for training and 50 different mammograms for testing. The analysis yielded an average area under the ROC curve (A(z)) of 0.90 for the testing set. Approximately 62% of false-positive clusters detected by the rule-based scheme were eliminated without any loss of the true-positive clusters by using the improved SIANN with image feature analysis techniques.

Purpose: To assess the performance of using transferred features from pre-trained deep convolutional networks (CNNs) in the task of classifying cancer in breast ultrasound images, and to compare this method of transfer learning with previous methods involving human-designed features. Methods: A breast ultrasound dataset consisting of 1125 cases and 2393 regions of interest (ROIs) was used. Each ROI was labeled as cystic, benign, or malignant. Features were extracted from each ROI using pre-trained CNNs and used to train support vector machine (SVM) classifiers in the tasks of distinguishing non-malignant (benign+cystic) vs malignant lesions and benign vs malignant lesions. For a baseline comparison, classifiers were also trained on prior analytically-extracted tumor features. Five-fold cross-validation (by case) was conducted with the area under the receiver operating characteristic curve (AUC) as the performance metric. Results: Classifiers trained on CNN-extracted features were comparable to classifiers trained on human-designed features. In the non-malignant vs malignant task, both the SVM trained on CNN-extracted features and the SVM trained on human-designed features obtained an AUC of 0.90. In the task of determining benign vs malignant, the SVM trained on CNN-extracted features obtained an AUC of 0.88, compared to the AUC of 0.85 obtained by the SVM trained on human-designed features. Conclusion: We obtained strong results using transfer learning to characterize ultrasound breast cancer images. This method allows us to directly classify a small dataset of lesions in a computationally inexpensive fashion without any manual input. Modern deep learning methods in computer vision are contingent on large datasets and vast computational resources, which are often inaccessible for clinical applications. Consequently, we believe transfer learning methods will be important for computer-aided diagnosis schemes in order to utilize advancements in deep learning and computer vision without the associated costs. This work was partially funded by NIH grant U01 CA195564 and the University of Chicago Metcalf program. M.L.G. is a stockholder in R2/Hologic, co-founder and equity holder in Quantitative Insights, and receives royalties from Hologic, GE Medical Systems, MEDIAN Technologies, Riverain Medical, Mitsubishi, and Toshiba. K.D. received royalties from Hologic.

State-of-the-Art in Radiomics in Radiology and Radiation Oncology Radiomics is the science of converting medical images into mineable data, data that are descriptive of “phenotypes,” which may provide diagnostic, prognostic, or therapeutic information. Genomics is the science of sequencing and analyzing the function and structure of genomes; the complete set of DNA in a single cell of an organism. In turn, imaging genomics (or radiogenomics) is concerned with the correlation between image-based features, as determined by radiomics, and gene expression, as determined by genomics. Imaging genomics arose from decades of work in at least three key areas: 1) sequencing of the human genome, 2) quantitative imaging, computer-aided diagnosis, therapy prognostics, and assessment of therapy response in preclinical and clinical research and practice; and 3) data science, including the burgeoning areas of genomics, preclinical and population-based disease modeling, individualized medicine, and big data. Imaging genomics may answer important questions in medicine by correlating validated, quantitative image phenotypes (through validated imaging biomarkers) with clinical data, histopathologic data, molecular classifications, genomic assays, and treatment outcomes. This approach could address some of the greatest health burdens, including cancer, cardiac disease, and arthritis. Participants will discuss the state-of-the-art of radiomics across multiple disease sites and modalities. Aspects of the presentations will include how to improve the quality of image-based phenotypes of normal and diseased tissue, how to better determine the relationships between these phenotypes and the underlying biology associated with the images, and how to create predictive models using the image-based phenotypes. Topics will also include: 1) creating, archiving, curating and sharing ultra-large datasets (“big data”); 2) standardizing image acquisition and processing methods; 3) standardizing and validating phenotype extraction methods and classifier designs; and 4) using high-throughput, robust, and validated phenotyping systems. H. Aerts, NIH; NCIH. Li, This research was funded in part by the University of Chicago Dean Bridge Fund,and by NCI U24-CA143848-05, P50- CA58223 Breast SPORE program. Hui Li received royalties from Hologic.W. Lu, This work was supported in part by the National Cancer Institute Grants R01CA172638.

Purpose: For this study, we investigated quantitative radiomics of breast tumors on diffusion weighted imaging and dynamic contrast-enhanced MRIs in the task of assessing the prognostic status of breast cancers. Methods: Our IRB-approved, retrospectively-collected dataset included 316 breast cancers with 235 ER+ and 81 ER- cases. All images were acquired during clinical breast MRI incorporating dynamic-contrast MRI and diffusion-weighted MRI. Phenotypic categories extracted quantitatively from DCE-MRI included tumor size, shape, margin sharpness, enhancement texture, kinetics, and variance kinetics, and from DWI-DCE ADC features (average, range, variation) for DWI. Phenotypes, as well as merged tumor signatures from round robin evaluation, were assessed for the prognostic tasks using area under the ROC curve (AUC) as the index of performance. Results: In the task of distinguishing between ER+ and ER- cancers, computer-extracted phenotypes from DCE and DWI yielded comparable performance levels, however, we found that the phenotypes, as well as the modality-specific tumor signatures, showed only slight correlation (r=−0.44), thus indicating the promise of multi-modality signatures. In the tasks of ER+ vs. ER-. PR+ vs. PR-, lymph node positive vs negative, we obtained AUC values of 0.66 (0.03), 0.64 (0.03), and 0.64 (0.03) for DCE-MRI, and AUC values of 0.64 (0.03), 0.61 (0.03), and 0.61 (0.03) for DWI-MRI, respectively. The combination of the modalities yielded AUC values of 0.67 (0.03), 0.64 (0.03), and 0.62 (0.03), respectively. Conclusion The correlation and performance results obtained from merging radiomic features from DCE-MRI and DWI-MRI indicate that the additional benefit of multimodality breast MRI in assessing prognosis is promising. Funded by an NIH (PREP) (R25) Grant and the University of Chicago Dean Bridge Fund. COI: M.L. Giger is a stockholder in R2 technology/Hologic and receives royalties from Hologic, GE Medical Systems, MEDIAN Technologies, Riverrain Medical, Mitsubushi, and Toshiba. MLG is a co-founder and stockholder in Quantitative Insights.

A computer‐aided diagnosis(CAD) scheme has been developed in our laboratory for the detection of clustered microcalcifications in digital mammograms. In this study, we apply a shift‐invariant neural network to eliminate false‐positive detections reported by the CAD scheme. The shift‐invariant neural network is a multilayer back‐propagation neural network with local, shift‐invariant interconnections. The advantage of the shift‐invariant neural network is that the result of the network is not dependent on the locations of the clustered microcalcifications in the input layer. The neural network is trained to detect each individual microcalcification in a given region of interest (ROI) reported by the CAD scheme. A ROI is classified as a positive ROI if the total number of microcalcifications detected in the ROI is greater than a certain number. The performance of the shift‐invariant neural network was evaluated by means of a jackknife (or holdout) method and ROC analysis using a database of 168 ROIs, as reported by the CAD scheme when applied to 34 mammograms. The analysis yielded an average area under the ROC curve (A z ) of 0.91. Approximately 55% of false‐positive ROIs were eliminated without any loss of the true‐positive ROIs. The result is considerably better than that obtained in our previous study using a conventional three‐layer, feed‐forward neural network. The effect of the network structure on the performance of the shift‐invariant neural network is also studied.

The choice of clinical cases used to train and test a computer-aided diagnosis (CAD) scheme can affect the test results (i.e., error rate). In this study, we deliberately modified the components of our testing database to study the effects of this modification on measured performance. Using a computerized scheme for the automated detection of breast masses from mammograms, it was found that the sensitivity of the scheme ranged between 26% and 100% (at a false positive rate of 1.0 per image) depending on the cases used to test the scheme. Even a 20% change in the cases comprising the database can reduce the measured sensitivity by 15%-25%. Because of the strong dependence of measured performance on the testing database, it is difficult to estimate reliably the accuracy of a CAD scheme. Furthermore, it is questionable to compare different CAD schemes when different cases are used for testing. Sharing databases, creating a common database, or using a quantitative measure to characterize databases are possible solutions to this problem. However, none of these solutions exists or is practiced at present. Therefore, as a short-term solution, it is recommended that the method used for selecting cases, and histograms or mean and standard deviations of relevant image features be reported whenever performance data are presented.

The segmentation and three-dimensional representation of the liver from a computed tomography (CT) scan is an important step in many medical applications, such as in the surgical planning for a living-donor liver transplant and in the automatic detection and documentation of pathological states. A method is being developed to automatically extract liver structure from abdominal CT scans using a priori information about liver morphology and digital image-processing techniques. Segmentation is performed sequentially image-by-image (slice-by-slice), starting with a reference image in which the liver occupies almost the entire right half of the abdomen cross section. Image processing techniques include gray-level thresholding, Gaussian smoothing, and eight-point connectivity tracking. For each case, the shape, size, and pixel density distribution of the liver are recorded for each CT image and used in the processing of other CT images. Extracted boundaries of the liver are smoothed using mathematical morphology techniques and B-splines. Computer-determined boundaries were compared with those drawn by a radiologist. The boundary descriptions from the two methods were in agreement, and the calculated areas were within 10%.

A high quality film‐duplication system was developed in order to improve the image quality of duplicated radiographs and to recover improperly exposed films. The system consists of a laser film digitizer, a laser film printer, a workstation, and a magneto‐optical disk. Radiographs are digitized by the laser digitizer, processed by the computer for image enhancement, and then printed on a film by the laser printer. A nonlinear density‐correction technique is employed in recovering improperly exposed radiographs using the HD however, for high spatial frequencies, the MTF of the duplication system is superior. The noise in the duplication system is about half of that in the CR system.

Purpose: Diffusion-weighted imaging (DWI) for breast MRI is a non-invasive technique which maps the diffusion process of water molecules, with the use of Brownian motion, in biological tissues, yielding apparent diffusion coefficients (ADC). Quantitative image analyses of the resulting images can yield information about the surrounding tissues in vivo. In this study, we investigated computerized image analyses of DWI images, including automated lesion segmentation, calculation of ADC values, and assessment of tumor heterogeneity, on a dataset of malignant and benign breast lesions. Method and Database: The IRB-approved, retrospectively collected dataset included 46 breast lesions -- 36 malignant and 10 benign. The DW images had been acquired during clinical breast MRI on a high-field 1.5 T echo-planar system using five b values [b = 0, 500, 1000, 1500, and 2000 s/mm2]. Seed-point-initiated 3D Gaussian-based lesion segmentation was conducted to yield the lesion margins, and within the margin and after b-value fitting, ADC values were calculated. Average ADC values and various measures of heterogeneity were calculated for benign and malignant lesions. Results: Our findings showed that malignant lesions tended, as expected, to have lower ADC values relative to benign lesions with average ADC values for benign and malignant lesions being 1.70 +/− 0.50 × 10(−3) and 1.07 +/− 0.37 × 10(−3), respectively. In addition, malignant lesions showed higher heterogeneity. In the task of distinguishing between benign and malignant lesions, ROC analysis, from histogram analysis, yielded AUC values of 0.86 +/− 0.07 and 0.88 +/− 0.06 for average ADC values and standard deviation of ADC values, respectively. Conclusion: DWI exhibits potential for differentiating between benign and malignant lesions, and inclusion of computerized quantitative image analysis methods allow for more objectives and efficient calculations of the ADC values and their variations, for potential use in decision making. M. L. Giger is a stockholder in R2 Technology/Hologic, is a shareholder/investor in Quantitative Insights, Qview Medical,receives royalties from Hologic, GE Medical Systems, MEDIAN Technologies, Riverain Medical, Mitsubishi and Toshiba.

The expectation maximization (EM) algorithm has received considerable attention in the area of positron emitted tomography (PET) as a restoration and reconstruction technique. In this paper, the restoration capabilities of the EM algorithm when applied to radiographic images is investigated. This application does not involve reconstruction. The performance of the EM algorithm is quantitatively evaluated using a "perceived" signal-to-noise ratio (SNR) as the image quality metric. This perceived SNR is based on statistical decision theory and includes both the observer's visual response function and a noise component internal to the eye-brain system. For a variety of processing parameters, the relative SNR (ratio of the processed SNR to the original SNR) is calculated and used as a metric to compare quantitatively the effects of the EM algorithm with two other image enhancement techniques: global contrast enhancement (windowing) and unsharp mask filtering. The results suggest that the EM algorithm's performance is superior when compared to unsharp mask filtering and global contrast enhancement for radiographic images which contain objects smaller than 4 mm.

A computerized scheme is being developed for the detection of masses in digital mammograms. Based on the deviation from the normal architectural symmetry of the right and left breasts, a bilateral subtraction technique is used to enhance the conspicuity of possible masses. The scheme employs two pairs of conventional screen-film mammograms (the right and left mediolateral oblique views and craniocaudal views), which are digitized by a TV camera/Gould digitizer. The right and left breast images in each pair are aligned manually during digitization. A nonlinear bilateral subtraction technique that involves linking multiple subtracted images has been investigated and compared to a simple linear subtraction method. Various feature-extraction techniques are used to reduce false-positive detections resulting from the bilateral subtraction. The scheme has been evaluated using 46 pairs of clinical mammograms and was found to yield a 95% true-positive rate at an average of three false-positive detections per image. This preliminary study indicates that the scheme is potentially useful as an aid to radiologists in the interpretation of screening mammograms.

To compare the diagnostic quality of images obtained with a computed radiography(CR) system based on storage phosphor technology with that obtained with conventional screen–film systems, a dual‐image recording technique was devised. With this technique, a CRimaging plate is placed behind a screen–film system in a conventional cassette. This makes it possible to obtain two images simultaneously, one from each system, in a clinical examination with the same patient positioning, the same degree of patient motion, the same geometric unsharpness, and no additional exposure. The modulation transfer functions (MTFs) of the CR system with and without the dual‐image recording technique were greater at low frequencies, but lower at high frequencies, that the MTFs of the screen–film systems used. The noise Wiener spectra of the CRimages at the plane of the imaging plate were greater than those of the screen–film systems, but were comparable to those of the screen–film systems at the plane of the printed film due to the reduction in image size. Clinical chest images obtained with the dual‐image recording technique appeared comparable, probably because of the image size reduction and the use of mild unsharp mask processing.

A curve fitting technique combined with an angulated slit image has been developed for the measurement of the presampling modulation transfer function(MTF) of film digitizers. The noisy line spread functions (LSFs) acquired from an angulated slit image are fitted using a combination of two functions by means of a nonlinear least‐square fitting technique. The parameters in the model function for each LSF are obtained by minimizing the residual root mean square (RMS), and then averaged over all the LSF fittings. The resulting analytical function is representative of the continuous presampled LSF. We have found that a combination of Gaussian and exponential functions provides a good fit to the LSFs obtained with film digitizers. The corresponding analytical Fourier transformation of the model function yields the presampling MTF, without Nyquist frequency limitation. Measurements of spatial resolutionproperties using this method were performed for two laser scanners and an optical drum scanner.

Currently, radiologists can fail to detect lung nodules in up to 30% of actually positive cases. If a computerized scheme could alert the radiologist to locations of suspected nodules, then potentially the number of missed nodules could be reduced. We are developing such a computerized scheme that involves a difference-image approach and various feature-extraction techniques. In this paper, we describe our use of digital morphological processing in the reduction of computer-identified false-positive detections. A feature-extraction technique, which includes the sequential application of nonlinear filters of erosion and dilation, is employed to reduce the camouflaging effect of ribs and vessels on nodule detection. This additional feature-extraction technique reduced the true-positive rate of the computerized scheme by 13% and the false-positive rate by 50%. In a comparison of the scheme with and without the additional feature-extraction technique, inclusion of the additional technique increased the detection sensitivity by about half at the level of three to four false-positive detections per chest image.

Purpose: To investigate the heterogeneity of breast tumors in terms of variation in contrast enhancement as observed in computer-extracted MRI-based tumor phenotypes. Methods: Analysis was conducted on a retrospective dataset of 84 de-identified, multi-institutional breast magnetic-resonance images (MRIs) from the National Cancer Institute repository, The Cancer Imaging Archive, along with clinical, histopathological, and genomic data from The Cancer Genome Atlas and gene assay data. The 84 cases were classified into Normal-like, Luminal A, Luminal B, HER2-enriched, and Basal-like molecular subtypes based on gene expression classifications. For each case, analysis of the dynamic-contrast enhanced MRIs included computerized 3D lesion segmentation and phenotype extraction, which characterized tumor size, shape, morphology, enhancement texture, kinetic curve assessment, and enhancement variance kinetics of the breast tumors. Enhancement texture over the imaging sequence was calculated at the first, second and third post-contrast time frames. Associations between computer-extracted tumor phenotypes and molecular subtypes were assessed through inferences on the Kendall τ rank correlation coefficient. Results: By examining the relationship between image-based phenotype and the molecular subtypes, there is a positive trend for enhancement texture for the various molecular subtypes. The Kendall τ rank correlation coefficient of 0.2342 with a p-value of 0.0055 was obtained between enhancement texture of entropy which was calculated at the first post-contrast time frame of DCE-MR image and the molecular subtypes. This enhancement texture quantitatively characterizes the heterogeneous nature of contrast uptake within the breast tumor. Conclusion: Computer-extracted image-based phenotypes show promise as a means for high-throughput discrimination of breast cancer molecular subtypes. Funding: University of Chicago Dean Bridge Fund, NCI U24-CA143848-05, and Breast Cancer Research Foundation. COI: MLG is a stockholder in R2 technology/Hologic and receives royalties from Hologic, GE Medical Systems, MEDIAN Technologies, Riverain Medical, Mitsubishi, and Toshiba. MLG is a co-founder and stockholder in Quantitative Insights.

Purpose: To gain insight into the role of parenchyma stroma in the characterization of breast tumors by incorporating computerized mammographic parenchyma assessment into breast CADx in the task of distinguishing between malignant and benign lesions. Methods: This study was performed on 182 biopsy-proven breast mass lesions, including 76 benign and 106 malignant lesions. For each full-field digital mammogram (FFDM) case, our quantitative imaging analysis was performed on both the tumor and a region-of-interest (ROI) from the normal contralateral breast. The lesion characterization includes automatic lesion segmentation and feature extraction. Radiographic texture analysis (RTA) was applied on the normal ROIs to assess the mammographic parenchymal patterns of these contralateral normal breasts. Classification performance of both individual computer extracted features and the output from a Bayesian artificial neural network (BANN) were evaluated with a leave-one-lesion-out method using receiver operating characteristic (ROC) analysis with area under the curve (AUC) as the figure of merit. Results: Lesion characterization included computer-extracted phenotypes of spiculation, size, shape, and margin. For parenchymal pattern characterization, five texture features were selected, including power law beta, contrast, and edge gradient. Merging of these computer-selected features using BANN classifiers yielded AUC values of 0.79 (SE=0.03) and 0.67 (SE=0.04) in the taskmore » of distinguishing between malignant and benign lesions using only tumor phenotypes and texture features from the contralateral breasts, respectively. Incorporation of tumor phenotypes with parenchyma texture features into the BANN yielded improved classification performance with an AUC value of 0.83 (SE=0.03) in the task of differentiating malignant from benign lesions. Conclusion: Combining computerized tumor and parenchyma phenotyping was found to significantly improve breast cancer diagnostic accuracy highlighting the need to consider both tumor and stroma in decision making. Funding: University of Chicago Dean Bridge Fund, NCI U24-CA143848-05, P50-CA58223 Breast SPORE program, and Breast Cancer Research Foundation. COI: MLG is a stockholder in R2 technology/Hologic and receives royalties from Hologic, GE Medical Systems, MEDIAN Technologies, Riverain Medical, Mitsubishi, and Toshiba. MLG is a cofounder and stockholder in Quantitative Insights.« less

Purpose: For this study, we investigated the computer-extracted tumor phenotypes from diffusion weighted imaging, dynamic contrast-enhanced, and T2-weighted magnetic resonance imaging modalities on a dataset of malignant and benign breast lesions. Methods: The IRB-approved, retrospectively-collected dataset included 118 breast lesions with 105 malignant and 13 benign. All images were acquired during clinical breast MRI at both 1.5T and 3.0T magnet strength. Phenotypic categories extracted with each modality included tumor size, shape, margin sharpness, enhancement texture, kinetics, and variance kinetics for DCE, size, shape, margin sharpness, texture for T2w, and ADC features for DWI. Results: In the task of distinguishing between benign and malignant lesions, each modality’s performance was analyzed by Round Robin evaluation using Receiver Operating Characteristic (ROC) analysis. DCE alone outperformed DWI and T2w with an AUC value of 0.89 +/−0.06. DWI and T2w yielded AUC values of 0.86 +/−0.05 and 0.84 +/−0.06 respectively. The combination of all three modalities yielded an AUC value of 0.88 +/−0.04 under single-loop Round Robin evaluation. The contrast phenotype from T2w and the standard deviation phenotype from DWI were found to be statistically different between the malignant and benign multimodality lesion groups. Conclusion: The results obtained from merging radiomic features from multimodality breast MRI (DCE, T2w, and DWI) indicate that the additional benefit of multimodality breast MRI in cancer diagnosis could be significant. This method also has potential to determine the most discriminatory radiomic phenotype from each modality. APPM DREAM Fellowship and the University of Chicago Dean Bridge Fund. M. L. Giger is a stockholder in R2 technology/Hologic and receives royalties from Hologic, GE Medical Systems, MEDIAN Technologies, Riverain Medical, Mitsubishi, and Toshiba. MLG is a co-founder and stockholder in Quantitative Insights.

Purpose: Magnetic Resonance Imaging(MRI) has been routinely used for diagnosis and assessment of breast cancer. Despite its wide applications in clinical practice, the relationship between the observed tumor MRI phenotypes and the genomic mechanism of tumorigenesis remains under-explored, largely due to lack of data on both imaging and genomics for the same tumors. Methods: We combined data from The Cancer Genome Atlas(TCGA) and The Cancer Image Archive(TCIA), that included quantitatively extracted MRI phenotypes of 91 breast invasive carcinomas and their multi-layer genomic data. Gene set enrichment analysis and regression analysis were performed to identify associations between tumor MRI radiomic phenotypes and various genomic and molecular subtypes of tumors. Patient groups defined by radiomic phenotypes and genomic platforms were also associated with tumor pathological stages and molecular receptor status using Fisher’s exact test. Results: Significant associations (adjusted p-values ≤ 0.1) were identified between radiomic phenotypes (characterizing tumor size, shape, margin, enhancement texture, and blood flow kinetics) and genomic features involved in multiple molecular regulation layers (including pathway gene expressions, pathway copy number variations, gene somatic mutations, miRNA expressions, and protein expressions). Transcriptional activities of various genetic pathways were dominantly positively associated with tumor size and blurred tumor margin. miRNA activity significantly associated with tumor size and enhancement textures, but not with phenotypes describing tumor shape, margin, and blood flow kinetics. Patient groups defined by radiomic phenotypes were associated with tumor T stage and overall stage (p-values ≤ 0.072). Genomic platforms defined patient groups associated with the status of progesterone and estrogen receptors (p-values ≤ 0.0000427) and pathological stages (p-values ≤ 0.056). Conclusion: We present these findings as a resource shedding insight on the connection between underlying genetic mechanisms and observed tumor radiomic phenotypes, which forms a basis for future studies using non-invasive MRI techniques for accurate cancer diagnosis and prognosis.

Purpose: Computer-extracted image phenotypes (CEIPs) are being investigated as complimentary attributes in the characterization of breast cancer in radiomics/ radiogenomics research. To be useful, CIEPs need to be robust across data obtained with different manufacturers’ MRI scanners and imaging protocols. Methods: Our research involved two HIPAA-compliant retrospectively-collected MRI datasets: Database 1 included 91 imaged breast cancers from the National Cancer Institute repository (imaged using General Electric equipment) and Database 2 included 117 breast cancers (imaged at our site using Phillips equipment). For each case, information on clinical lymph node status and histopathology on ER, PR, and Her2 receptor status was available. Each lesion underwent quantitative radiomics analysis yielding CEIPs characterizing tumor size, shape, morphology, enhancement texture, kinetic curve assessment, and enhancement variance kinetics. The robustness of CEIPs was assessed through statistical comparisons across the two datasets in terms of average CEIP values, t-test results on the subgroups of interest, and non-inferiority testing of performance in the prognostic tasks of distinguishing ER, PR, and Her2 receptor status and lymph node status using area under the receiver operating characteristic curve (AUC). Results: We failed to find any statistically significant differences in the average value of the CEIP distributions across the 2 scanners for subgroups possessing enough cases. We found greater variation in average feature values for the clinical subgroups having less than 20 cases. Non-inferiority analysis demonstrated varying degrees of robustness for different MRI phenotypes. The most enhancing volume and total rate variation showed the best agreement with absolute value of the lower bound of the 90% confidence for delta AUC