Abstract P2-06-29: Automated Breast Density Assessment on Chest CT with a Deep-Learned 3D Ordinal Regression Model
Introduction: Breast cancer is the most common malignancy and the second-leading cause of cancer mortality in women. Breast density, defined as the extent of fibroglandular tissue within the breast, is assessed semi-quantitatively on a four-grade scale, from mostly fatty (grade 1) to extremely dense (grade 4), with a severalfold increase in prospective cancer risk between the lowest and highest grades. Millions of chest CT studies are performed annually in patients eligible for routine mammographic screening, representing a significant opportunity for incidental density reporting and risk stratification. While automated density grading on mammography is robust and widely deployed, assessment on CT remains limited by comparison. To facilitate the efficient and standardized assessment of breast density on chest CT, we train and evaluate a 3D convolutional neural network (CNN) model to predict breast density grade from chest CT volumes using a retrospective dataset of patients at Columbia University Irving Medical Center.Methods: A retrospective, IRB-approved chart review identified female patients at CUIMC who underwent both mammography and chest CT acquired less than one year apart from January 2010 through March 2019. All cases were reviewed by a faculty radiologist, recording both CT breast density and whether the entire breast was present within the scanner field of view (FOV). All CT scans with or without IV contrast, with axial soft-tissue reconstruction and no breast malignancy, prosthesis or prior mastectomy were included. CT volumes were downsampled to isotropic 2.5 mm resolution and 80x80x80 voxel volumes of interest (VOIs) were extracted containing each imaged breast. Scans were randomly allocated to training, validation, and test sets in a 3:1:1 ratio with stratification by density grade.Our 3D CNN model consists of five blocks containing 3D convolutional layers with ReLU activation, convolutional block attention modules (CBAM), and batch normalization, interleaved with pooling operations. Rank-consistent ordinal regression (CORN) was applied to the final fully-connected layer to predict the density grade for each image. The model was trained to minimize ordinal cross-entropy with L2 regularization, using training batches balanced across density grades and train-time augmentation, over 60,000 training steps with early stopping conditioned on validation ordinal cross-entropy. Patient-level predictions were generated on the test cases by averaging predictions over the right- and left-breast VOIs.Results: Of 503 scans which met the inclusion criteria, 86 (17.1%) were assigned breast density grade 1, 244 (48.5%) were grade 2, 131 (26.0%) were grade 3, and 42 (8.3%) were grade 4. One or both breasts were partially outside the FOV in 134 scans (26.6%). On prediction of high (grades 3-4) versus low breast density, the optimized model achieved test-set accuracy of 90.1%, receiver-operating characteristic AUC of 0.935, and Cohen kappa of 0.787 with respect to radiologist-assessed ground truth, comparable to reported values for inter-radiologist agreement. One-vs-all accuracy for breast density grade was 67.3%, with 98.0% of predicted grades at most one level off from the ground truth. ROC AUC and Cohen kappa improved to 0.955 and 0.873, respectively, on the subset of test cases (N=80) where both breasts were contained within the FOV.Conclusions: We have introduced a 3-D CNN model for automated ordinal regression of breast density on chest CT using a retrospective clinical CT dataset, with performance relative to ground-truth comparable to published inter-reader agreement among trained radiologists. Estimating breast density grade independently of threshold-based volumetric percent density is a direct precursor to deep-learning on chest CT to improve on image-based breast cancer risk stratification. Citation Format: Artur Wysoczanski, Elsa D. Angelini, Sachin S. Jambawalikar, Andrew F. Laine, Mary M. Salvatore. Automated Breast Density Assessment on Chest CT with a Deep-Learned 3D Ordinal Regression Model [abstract]. In: Proceedings of the San Antonio Breast Cancer Symposium 2024; 2024 Dec 10-13; San Antonio, TX. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(12 Suppl):Abstract nr P2-06-29.
Read more