Attention-Guided Erasing for Enhanced Transfer Learning in Breast Abnormality Classification

Scientific summary

AGE uses self-supervised attention maps to guide stochastic erasing, preserving task-relevant mammography regions during transfer learning.

Abstract

This journal article evaluates Attention-Guided Erasing as a self-supervised, task-specific augmentation strategy across image-level and patch-level mammography classification tasks.

Why it matters

The work targets a practical transfer-learning limitation in medical imaging: improving task-specific representation learning when annotated data are limited.

Contribution

The journal article tests whether Attention-Guided Erasing generalizes beyond a single task across digital mammography, contrast-enhanced mammography, and patch-level abnormality classification.

Method overview

A DINO-pretrained Vision Transformer produces attention-head maps. Selected maps are converted to binary masks, used for stochastic erasing, and evaluated during downstream transfer learning.

Key findings

• Reported statistically significant mean F1-score gains across four of five evaluated classification tasks.
• Image-level gains were reported for breast density in digital mammography and malignancy in contrast-enhanced mammography.
• The mass classification task showed only marginal improvement, marking a useful boundary case.

Citation

TODO: Add BibTeX.