Unsupervised Domain Adaptation for Calcification Classification in Mammography (arXiv 2507.06826)

Forum topic · 小凯 · 2026-07-09

Summary

A paper by Xuan Liu, Derek L. Nguyen, and Emily C. Barre (arXiv 2507.06826, published 2025-07-09) proposes a framework for malignant versus benign breast disease calcification classification across multi-site mammography datasets. The framework combines (1) an unsupervised domain adaptation module based on style transfer models (AdaIN and CycleGAN) that generates vendor-specific and technique-specific training samples without additional annotations, and (2) a supervised classification module using Swin Transformer V2 as the backbone. The method was evaluated on three datasets: cross-validation on OPTIMAM (UK NHS; n=2994), with external validation on EMBED (Emory University; n=125) and the Duke Calcification Dataset v1 (n=788), covering multiple vendors including full-field digital mammography and synthetic 2D images from digital breast tomosynthesis. The framework improved cross-site performance, raising AUC on EMBED from 0.68 to 0.72 and on the Duke dataset from 0.68 to 0.73. The results indicate that unsupervised domain adaptation can reduce domain shift and improve generalization of calcification classification across multi-site mammography data.

Overview

Research area: Computer Vision (CV) Authors: Xuan Liu, Derek L. Nguyen, Emily C. Barre Published: 2025-07-09 arXiv: 2507.06826

Abstract

Deep learning-based computer-aided diagnosis (CAD) systems have shown strong performance in breast cancer diagnosis, particularly for classification tasks in mammography. However, domain shifts across multi-site datasets remain a challenge, especially when models are applied to unseen domains.

In this work, the authors propose a calcification classification framework to improve malignant versus benign breast disease classification across multi-site mammography datasets. The framework consists of two components:

1. Unsupervised domain adaptation module based on style transfer models (AdaIN and CycleGAN) to generate vendor-specific and technique-specific training samples without additional annotations. 2. Supervised classification module using Swin Transformer V2 as the backbone.

Evaluation

The method was evaluated on three datasets spanning multiple vendors, including full-field digital mammography and synthetic 2D images derived from digital breast tomosynthesis:

| Dataset | Source | Size | Role | |---|---|---|---| | OPTIMAM | UK National Health Service | n=2994 | Cross-validation | | EMBED | Emory University | n=125 | External validation | | Duke Calcification Dataset v1 | Duke University | n=788 | External validation |

Results

The proposed framework improved cross-site performance:

EMBED: AUC improved from 0.68 to 0.72
Duke Calcification Dataset: AUC improved from 0.68 to 0.73

Conclusion

These findings suggest that unsupervised domain adaptation via style transfer can reduce domain shift and improve generalization of calcification classification across multi-site mammography datasets, without requiring additional annotations.

--- *Auto-collected on 2026-07-09*