Unsupervised Domain Adaptation for Calcification Classification in Mammography Across Multi-Site Datasets

2026-07-08 · Source: cs.CV updates on arXiv.org · Field: Health & Wellbeing — Clinical Care & Medical Practice, Medical Devices & Health Technology, Artificial Intelligence & Machine Learning · Depth: Expert, extended

Summary

This study introduces a multi-stage framework to improve malignant versus benign calcification classification in mammography across diverse multi-site datasets, addressing performance degradation due to domain shifts. The framework integrates an unsupervised domain adaptation module, utilizing style transfer models like AdaIN and CycleGAN to generate vendor- and technique-specific training samples without extra annotations, with a supervised classification module powered by Swin Transformer V2. Evaluated on OPTIMAM (n=2994), EMBED (n=125), and Duke Calcification Dataset v1 (n=788), the method significantly improved cross-site performance. Specifically, AUC increased from 0.68 to 0.72 on EMBED and from 0.68 to 0.73 on the Duke dataset, demonstrating that domain adaptation effectively reduces generalization gaps.

Key takeaway

For AI Scientists and Machine Learning Engineers developing CAD systems for mammography, you should integrate unsupervised domain adaptation techniques into your training pipelines. This approach, particularly using style transfer models like CycleGAN to generate vendor- and technique-specific training samples, can significantly improve model generalization and robustness when deploying models across diverse clinical sites and imaging modalities. Consider augmenting your training data with synthetically adapted images to reduce performance drops on unseen domains.

Key insights

Unsupervised style transfer effectively mitigates domain shifts in mammography, improving calcification classification generalization across diverse datasets.

Principles

• Deep learning model performance on natural images does not directly translate to medical imaging tasks.
• Domain shifts from hardware, techniques, and protocols significantly degrade medical AI generalization.
• Generating synthetic, style-adapted training data can reduce domain gaps without new annotations.

Method

The framework trains a style transfer model (AdaIN or CycleGAN) on source data and unlabeled vendor-specific patches to generate stylized lesion patches. These, combined with original patches, train a Swin Transformer V2 classifier.

In practice

• Incorporate AdaIN or CycleGAN for unsupervised image-to-image style translation.
• Use a 1:2 malignant to benign patch ratio in training batches to address class imbalance.

Topics

• Unsupervised Domain Adaptation
• Mammography
• Calcification Classification
• Style Transfer
• Swin Transformer V2
• Computer-Aided Diagnosis

Best for: Computer Vision Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Generative Deep Learning for Computational Destaining and Restaining of Unregistered Digital Pathology Images — Preprocessing-based domain adaptation enables cGAN models to generalize for computational pathology across institutions without retraining.
• Token-Based Dual-view Fusion and Adaptation of Large Vision Models for Breast Cancer Classification — A token-centric, multi-depth fusion framework enhances breast cancer classification by integrating complementary mammography views.
• UniRG: Scaling medical imaging report generation with multimodal reinforcement learning — Reinforcement learning with clinically meaningful rewards significantly improves medical vision-language model reliability and generality.
• When To Adapt? Adapting the Model or Data in Federated Medical Imaging — Adaptation strategy effectiveness in federated medical imaging depends on the type and magnitude of domain shift.
• Scaling Generative Foundation Models for Chest Radiography with Rectified Flow Transformers — A 1.3B-parameter generative foundation model for chest radiographs achieves expert-indistinguishable synthesis and controllable editing.
• CA-GCL: Cross-Anatomy Global-Local Contrastive Learning for Robust 3D Medical Image Understanding — CA-GCL uses global-local contrastive learning and clinical-aware text augmentation to prevent representation collapse in 3D medical FVLP.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.