Attention-Guided Erasing: A Novel Augmentation Method for Enhancing Downstream Breast Density Classification

Scientific summary

Attention-Guided Erasing improves breast-density transfer learning by preserving attention-derived tissue regions while stochastically suppressing less informative background.

Abstract

This paper introduces Attention-Guided Erasing for breast density classification, using DINO attention maps to guide erasing during transfer learning.

Why it matters

Breast density directly affects screening sensitivity, making robust automated density assessment relevant for risk stratification and workflow support.

Contribution

The work presents AGE as a targeted augmentation strategy for BI-RADS breast density classification in mammography.

Method overview

DINO attention maps from a Vision Transformer are used to identify relevant mammography regions; background regions are erased with random probabilities during transfer learning.

Key findings

• On VinDr-Mammo, the reported mean F1-score was 0.5910.
• The method outperformed no-AGE and random-erasing baselines reported at 0.5594 and 0.5691 mean F1.
• The improvement was reported as statistically significant with p < 0.0001.

Citation

TODO: Add BibTeX.