Faculty Bibliography

Machine-learning techniques have led to remarkable advances in data extraction and analysis of medical imaging. Applications of machine learning to breast MRI continue to expand rapidly as increasingly accurate 3D breast and lesion segmentation allows the combination of radiologist-level interpretation (eg, BI-RADS lexicon), data from advanced multiparametric imaging techniques, and patient-level data such as genetic risk markers. Advances in breast MRI feature extraction have led to rapid dataset analysis, which offers promise in large pooled multiinstitutional data analysis. The object of this review is to provide an overview of machine-learning and deep-learning techniques for breast MRI, including supervised and unsupervised methods, anatomic breast segmentation, and lesion segmentation. Finally, it explores the role of machine learning, current limitations, and future applications to texture analysis, radiomics, and radiogenomics. Level of Evidence: 3 Technical Efficacy Stage: 2 J. Magn. Reson. Imaging 2019.

Breast MRI offers high sensitivity for breast cancer detection, with preferential detection of high-grade invasive cancers when compared to mammography and ultrasound. Despite the clear benefits of breast MRI in cancer screening, its cost, patient tolerance, and low utilization remain key issues. Abbreviated breast MRI, in which only a select number of sequences and postcontrast imaging are acquired, exploits the high sensitivity of breast MRI while reducing table time and reading time to maximize availability, patient tolerance, and accessibility. Worldwide studies of varying patient populations have demonstrated that the comparable diagnostic accuracy of abbreviated breast MRI is comparable to a full diagnostic protocol, highlighting the emerging role of abbreviated MRI screening in patients with an intermediate and high lifetime risk of breast cancer. The purpose of this review is to summarize the background and current literature relating to abbreviated MRI, highlight various protocols utilized in current multicenter clinical trials, describe workflow and clinical implementation issues, and discuss the future of abbreviated protocols, including advanced MRI techniques.

RATIONALE AND OBJECTIVES/OBJECTIVE:Pathologic complete response (pCR) in patients with human epidermal growth factor receptor 2 (HER2)-positive breast cancer after HER2-targeted therapy correlates increased disease-free survival and decreased mastectomy rates. The aim of this study was to explore tumor shrinkage patterns and initial tumor enhancement with pCR in HER2-positive breast cancer. MATERIALS AND METHODS/METHODS:This was an institutional review board-approved retrospective analysis of 51 HER2 positive breast cancer patients with breast MRI both pre- and post-HER2-targeted therapy. Initial enhancement ratio (IER, initial enhancement percentage over baseline at first postcontrast imaging), pattern of tumor shrinkage, and Dynamic contrast enhanced (DCE)-MRI imaging features were assessed. Wilcoxon rank, Spearman correlation, Fisher's exact, and Mann-Whitney tests were used to correlate MRI imaging features with pCR. IER reader agreement was evaluated by intraclass correlation. Binary logistic regression was used to evaluate multivariate associations with pCR. RESULTS:56.9% (29/51) of patients had pCR at surgery. Concentric tumor shrinkage pattern was associated with pCR (p = 0.001, Area under the curve (AUC) 0.778): accuracy 80.4%, specificity 96.6%, and sensitivity of 59.1%. There was no association with pCR and imaging response as defined by RECIST criteria (p = 0.169), pretreatment IER (Reader 1 (R1) p = 0.665, Reader 2 (R2) p = 0.766), or lesion size (p = 0.69). IER was associated with axillary metastases (R1 p = 0.016, R2 < 0.001) and ki-67 (R1 r = 0.52, p = 0.008, R2 r = -0.44, p = 0.028). CONCLUSION/CONCLUSIONS:The shrinkage pattern of HER2-positive tumors after targeted therapy may be associated with pCR. There was no association between IER and pCR. Future studies evaluating the correlation of shrinkage patterns to texture radiomics are of interest.

We present a deep convolutional neural network for breast cancer screening exam classification, trained and evaluated on over 200,000 exams (over 1,000,000 images). Our network achieves an AUC of 0.895 in predicting the presence of cancer in the breast, when tested on the screening population. We attribute the high accuracy to a few technical advances. (i) Our network's novel two-stage architecture and training procedure, which allows us to use a high-capacity patch-level network to learn from pixel-level labels alongside a network learning from macroscopic breast-level labels. (ii) A custom ResNet-based network used as a building block of our model, whose balance of depth and width is optimized for high-resolution medical images. (iii) Pretraining the network on screening BI-RADS classification, a related task with more noisy labels. (iv) Combining multiple input views in an optimal way among a number of possible choices. To validate our model, we conducted a reader study with 14 readers, each reading 720 screening mammogram exams, and show that our model is as accurate as experienced radiologists when presented with the same data. We also show that a hybrid model, averaging the probability of malignancy predicted by a radiologist with a prediction of our neural network, is more accurate than either of the two separately. To further understand our results, we conduct a thorough analysis of our network's performance on different subpopulations of the screening population, the model's design, training procedure, errors, and properties of its internal representations. Our best models are publicly available at https://github.com/nyukat/breastcancerclassifier.

RATIONALE AND OBJECTIVES/OBJECTIVE:The automated segmentation of organs and tissues throughout the body using computed tomography and magnetic resonance imaging has been rapidly increasing. Research into many medical conditions has benefited greatly from these approaches by allowing the development of more rapid and reproducible quantitative imaging markers. These markers have been used to help diagnose disease, determine prognosis, select patients for therapy, and follow responses to therapy. Because some of these tools are now transitioning from research environments to clinical practice, it is important for radiologists to become familiar with various methods used for automated segmentation. MATERIALS AND METHODS/METHODS:The Radiology Research Alliance of the Association of University Radiologists convened an Automated Segmentation Task Force to conduct a systematic review of the peer-reviewed literature on this topic. RESULTS:The systematic review presented here includes 408 studies and discusses various approaches to automated segmentation using computed tomography and magnetic resonance imaging for neurologic, thoracic, abdominal, musculoskeletal, and breast imaging applications. CONCLUSION/CONCLUSIONS:These insights should help prepare radiologists to better evaluate automated segmentation tools and apply them not only to research, but eventually to clinical practice.

Purpose Our aim was to evaluate the accuracy of a new algorithm to automatically delineate the breast region from the chest on T1-weighted, non-fat-suppressed MR images. This process is also referred to as the chest wall detection. There is a general agreement that this step is very difficult to automate. At the same time it is crucially needed for clinically important processing workflows [1]. These workflows include 3D measurement of breast density and of the breast parenchymal enhancement. Both measures reveal patients at risk of breast cancer [2]. Manually traced chest wall was used as the ground truth when estimating the segmentation errors. Segmentation accuracy was evaluated using the Hausdorff distance and the volumetric error. We also estimated the inter-observer agreement in defining the chest wall surface. Methods The program starts by generating the mid-sagittal 2D section by averaging the signal across 20 mm thick mid-sagittal slab. We determine the chest wall boundary on this image by modeling the signal profiles along the antero-posterior direction as a sequence of three tissues: background air, skin and fat layer, muscle. Non-uniformity correction is then applied to the entire 3D volume. The mid-sagittal boundary, represented as a polyline P, is then propagated in two opposite (left and right) directions. At each sagittal section the algorithm adjusts the control points of the polyline received from an adjacent slice. The adjustment is estimated from the weighted sum of six measures that combine specific local and global signal statistics. These include: the local gradient, the signal uniformity, the gradient similarity, the contour-gradient consistency, the global contour uniformity and the normal vector consistency. At each iteration we form a candidate shift vector, we apply it to shift P to its new position, and then we smooth the resulting polyline. The process terminates when the magnitude of the shift becomes negligible or when the specified number of iterations is exceeded. Two metrics were used to estimate accuracy. The conventional volumetric error was obtained by dividing the volume DV of misclassified breast voxels over the true breast volume V. The Hausdorff distance, HD, is the distance between each voxel on the true breast/ chest wall border and the closest boundary voxel produced by the algorithm. HD is averaged over the entire chest wall surface. From a clinical database of screening breast MRIs acquired at our medical center we have randomly selected 16 test exams. The selection was constrained to enforce that there were four exams in each of the four breast density categories [3]. Bilateral breasts were imaged on Siemens 3T Magnetom Trio equipped with a 7-element surface breast coil. The parameters of the T1-weighted non-fat-suppressed sequence were: TR = 4.74 ms, TE = 1.79 ms, FOV = 320 mm2, matrix = 448 9 358 9*150, 0.7 9 0.7 9 1.1 mm voxels, TA = 2-3 min. Three experts in breast and chest anatomy drew contours to separate the chest wall from the breast (Fig. 1). The pectoralis fascia and pectoralis muscles were used as reference points for the anterolateral borders. The medial border of the axilla was the posterolateral boundary. The axillary tail was considered as the breast tissue. The ground truth references were constructed by a software designed to perform voxel-based ROI averaging [4]. (Figure Presented) Results The border distance error HD was 0.84 +/- 0.8 mm (average +/- standard deviation) and ranged from 0.57 to 2.45 mm. The volume error DV/V was 6.43 +/- 6.82%. There was no correlation between the HD and DV/V (R2 = 0.23, p = 0.12). The test cases covered a wide range 411-3439 ml of breast volumes. There was a significant positive correlation (R2 = 0.40, p = 0.02) between volumetric error and the true breast volume V, but there was no correlation between HD and V (R2 = 0.08, p = 0.44). The average execution time was under 1.5 min per case on a standard 8-core workstation. The inter-observer agreement measured in term of HD was 0.56 +/- 0.15 mm (average +/- standard deviation). The agreement expressed in terms of volumetric discrepancy (relative to breast volume) was 1.61% +/- 0.71%. Conclusion Breast density, defined as fraction of fibroglandular tissue, and postcontrast enhancement, are considered significant risk factors for breast cancer. These MRI measures are recommended for radiologic reports and are promising cancer biomarkers. Radiologists currently visually estimate these measure. Unfortunately, readers agreement for qualitative evaluation is only fair, requiring better standardization and reproducibility. Computer-assisted quantitative assessment is needed, but the task is challenging due to image nonuniformity (breast coils cause loss of MR signal in remote regions) and to the anatomical complexity of chest wall boundary (Fig. 2). (Figure Presented) Given its accuracy and speed, our breast segmentation method appears to be ready for clinical use as a part of larger workflow to generate routine diagnostic reports

BACKGROUND: Screening breast MRI has been shown to preferentially detect high-grade ductal carcinoma in situ (DCIS) and invasive carcinoma, likely due to increased angiogenesis resulting in early initial uptake of contrast. As interest grows in abbreviated screening breast MRI (AB-MRI), markers of early contrast washin that can predict tumor grade and potential aggressiveness are of clinical interest. PURPOSE: To evaluate the feasibility of using the initial enhancement ratio (IER) as a surrogate marker for tumor grade, hormone receptor status, and prognostic markers, as an initial step to being incorporated into AB-MRI. STUDY TYPE: Retrospective. SUBJECTS: In all, 162 women (mean 55.0 years, range 32.8-87.7 years) with 187 malignancies imaged January 2012-November 2015. FIELD STRENGTH/SEQUENCE: Images were acquired at 3.0T with a T1 -weighted gradient echo fat-suppressed-volume interpolated breath-hold sequence. ASSESSMENT: Subjects underwent dynamic contrast-enhanced breast MRI with a 7-channel breast coil. IER (% signal increase over baseline at the first postcontrast acquisition) was assessed and correlated with background parenchymal enhancement, washout curves, stage, and final pathology. STATISTICAL TESTS: Chi-square test, Spearman rank correlation, Mann-Whitney U-tests, Bland-Altman analysis, and receiver operating characteristic curve analysis. RESULTS: IER was higher for invasive cancer than for DCIS (R1/R2, P < 0.001). IER increased with tumor grade (R1: r = 0.56, P < 0.001, R2: r = 0.50, P < 0.001), as ki-67 increased (R1: r = 0.35, P < 0.001; R2 r = 0.35, P < 0.001), and for node-positive disease (R1/R2, P = 0.001). IER was higher for human epidermal growth factor receptor two-positive and triple negative cancers than for estrogen receptor-positive / progesterone receptor-positive tumors (R1 P < 0.001-0.002; R2 P = 0.0.001-0.011). IER had higher sensitivity (80.6% vs. 75.5%) and specificity (55.8% vs. 48.1%) than washout curves for positive nodes, higher specificity (48.1% vs. 36.5%) and positive predictive value (70.2% vs. 66.7%) for high ki-67, and excellent interobserver agreement (intraclass correlation coefficient = 0.82). DATA CONCLUSION: IER, a measurement of early contrast washin, is associated with higher-grade malignancies and tumor aggressiveness and might be potentially incorporated into an AB-MRI protocol. LEVEL OF EVIDENCE: 3 Technical Efficacy Stage 2 J. Magn. Reson. Imaging 2017.

This article reviews new developments in breast imaging. There is growing interest in creating a shorter, less expensive MR protocol with broader applicability. There is an increasing focus on and consideration for the additive impact that functional analysis of breast pathology have on identifying and characterizing lesions. These developments apply to MR imaging and molecular imaging. This article reviews evolving breast imaging techniques with attention to strengths, weaknesses, and applications of these approaches. We aim to give the reader familiarity with the state of current developments in the field and to increase awareness of what to expect in breast imaging.

RATIONALE AND OBJECTIVES: Here we review the current state of multicenter radiology research (MRR), and utilize a survey of experienced researchers to identify common advantages, barriers, and resources to guide future investigators. MATERIALS AND METHODS: The Association of University Radiologists established a Radiology Research Alliance task force, Multi-center Research Studies in Radiology, composed of 12 society members to review MRR. A REDCap survey was designed to gain more insight from experienced researchers. Recipients were authors identified from a PubMed database search, utilizing search terms "multicenter" or "multisite" and "radiology." The survey included investigator background information, reasons why, barriers to, and resources that investigators found helpful in conducting or participating in MRR. RESULTS: The survey was completed by 23 of 80 recipients (29%), the majority (76%) of whom served as a primary investigator on at least one MRR project. Respondents reported meeting collaborators at national or international (74%) and society (39%) meetings. The most common perceived advantages of MRR were increased sample size (100%) and improved generalizability (91%). External funding was considered the most significant barrier to MRR, reported by 26% of respondents. Institutional funding, setting up a central picture archiving and communication system, and setting up a central database were considered a significant barrier by 30%, 22%, and 22% of respondents, respectively. Resources for overcoming barriers included motivated staff (74%), strong leadership (70%), regular conference calls (57%), and at least one face-to-face meeting (57%). CONCLUSIONS: Barriers to MRR include funding and establishing a central database and a picture archiving and communication system. Upon embarking on an MRR project, forming a motivated team who meets and speaks regularly is essential.