Comparative Heatmap Explanations for Deep Learning-Based Breast Density Classification

Abstract

Breast density is associated with a higher risk of developing breast cancer and complicates mammographic interpretation because dense tissue can obscure suspicious findings. Deep learning (DL) models are increasingly used for automated breast density classification, yet their limited interpretability remains a concern in safety-critical medical applications. In this paper, we compare several DL backbones for four-class BI-RADS density classification and analyze their predictions using three heatmap explanation methods—GradCAM, GradCAM++, and ScoreCAM. Our results show that transformer-based models achieve marginally stronger classification performance than the evaluated convolutional neural network (CNN) baselines, with the ViT model obtaining the best overall results and showing improved recognition of extremely dense cases. Qualitative heatmap analysis suggests that ScoreCAM produces the most spatially coherent and clinically plausible explanations, particularly for transformer-based models for which gradient-based maps are often diffuse or unstable. We further present an exploratory procedure that converts ScoreCAM heatmaps into coarse binary masks through thresholding, illustrating the potential of explanation maps for weak localization of dense tissue. Rather than claiming a validated segmentation framework, we position this step as a proof of concept that may support future studies on annotation-efficient dense-tissue localization.