Deep Multi-Head CNN for Unified Breast Density Segmentation and BI-RADS Estimation in 2D Mammograms
Keywords: Breast cancer, CNN, Mammography, Radiomics, Segmentation
TL;DR: Deep Unified CNN for Breast Density Analysis in 2D Mammograms
Abstract: Comprehensive breast density estimation is crucial in mammogram assessment and cancer risk stratification, yet many existing AI-based radiomic methods designed for this purpose often tackle tissue segmentation and classification as separate tasks. To address this limitation, we propose a multi-head convolutional neural network (MH-CNN) that integrates these functions into a unified end-to-end architecture. Built on a ResNet101 encoder, our approach learns high-level features for breast density segmentation while parallel network heads perform continuous density regression and BI-RADS classification overlaid in the resulting images. Evaluation on the VinDr mammogram dataset yielded a Dice coefficient of 84.57% for segmentation, a mean absolute error (MAE) of 5.92% for density regression, and 80.51% accuracy for BI-RADS classification. These results suggest that the MH-CNN can streamline clinical workflows by providing objective and reliable breast density assessments
Submission Number: 39
Loading