Dense Breast Tissue Segmentation under Limited Annotation: Comparing CNNs, Transformers, Self-Supervision, and Prompting

CNN-based dense breast tissue segmentation models achieve the best standalone performance under limited pixel-level supervision on 596 expert-annotated mammograms, balancing accuracy and computational cost.

• Why it matters: Dense tissue lowers mammographic screening sensitivity and is a breast cancer risk factor, so reliable segmentation is needed even when pixel-level expert labels are scarce.
• What they did: The study compares classical methods, CNN/U-Net encoder–decoder models, transformer models, self-supervised pre-training, fine-tuning and loss choices, plus prompt-based MedSAM variants using 596 labeled images and 20,000 in-domain unlabeled images.
• The result: Transformer-based models and the adaptive fuzzy c-means + SVM baseline perform worse; prompt quality strongly affects MedSAM, with CNN-generated dense prompts improving results, and density-calibrated hybrid loss reducing test MAE from 14.8 to 11.8 percentage points.

AI summary of the abstract · may contain errors · not medical advice · read the paper.