Hybrid Hard-Negative Learning for Breast Cancer Risk Classification in Mammography with Explainable AI

Abstract

Whole-image mammography models must identify small abnormalities within predominantly normal tissue, while their predictions remain difficult to audit. This paper evaluates a Hybrid Hard-Negative learning pipeline for binary mammographic risk classification with dual-level
Grad-CAM explanations. The VinDr-Mammo dataset was mapped operationally as BI-RADS 1-2 (negative/benign assessment) versus BI-RADS 4-5 (suspicious/high-risk assessment); BIRADS 3 was excluded. The balanced subset contained 1,976 images, including a held-out test
set of 296 images. Training used lesion, lesion-context, full-image, ordinary negative-tissue, and model-mined hard-negative patches. At inference, 512 × 512 windows were scanned at stride 128 and combined using 0.75 times the mean of the three highest patch probabilities plus 0.25 times the full-image probability. Seven CNN backbones were compared, and an attention based Scanner-MIL  configuration served as a weakly supervised baseline. MobileNetV3Large achieved the strongest observed scan-level performance (ROC-AUC 0.7877, sensitivity 0.7055, F1-score 0.7128; 43 false negatives among 146 BI-RADS 4-5 images) with 3.18 million parameters. The matched Scanner-MIL baseline achieved sensitivity 0.5616. When lesion locations were supplied, pretrained backbones reached ROC-AUC 0.969-0.973, whereas automatic scan-level performance was 0.7321-0.7877 (0.6414 for the non-pretrained custom CNN). The complete Hybrid pipeline produced 12 fewer false negatives than the whole-image MobileNetV3Large configuration. Qualitative Grad-CAM examples were consistent with localization-related failure but were not a quantitative localization test. The results support candidate localization as a major performance constraint and position the system as a research prototype rather than an autonomous diagnostic tool.