MC-GenRef: Annotation-free mammography microcalcification segmentation with generative posterior refinement

A research team from Yonsei University has introduced MC-GenRef, a novel AI framework designed to segment microcalcifications (MCs) in mammograms without requiring expensive, manually annotated medical data. Microcalcifications—tiny calcium deposits that can be early indicators of breast cancer—are notoriously difficult for AI to detect due to their small size, sparse distribution, and the high cost and ambiguity of pixel-level labeling by radiologists. MC-GenRef tackles this by using a two-stage approach: first, it trains a base segmentation model using only synthetic data. This data is created by taking real mammogram patches without MCs and injecting physically plausible MC patterns using a lightweight image formation model, complete with local contrast modulation and blur. This process generates perfect image-mask pairs for training, completely bypassing the need for real, dense annotations.

During the inference stage, the system employs a novel technique called Test-Time Generative Posterior Refinement (TT-GPR). This treats the segmentation task as an approximate posterior inference problem. The process starts with a sparse seed from an initial prediction, uses a seed-conditioned rectified-flow generator to create refined projections, and then iteratively updates the segmentation logits with overlap-consistent and edge-aware regularization. The results are significant: on the public INbreast dataset, the synthetically-trained initial model alone achieved the best Dice score compared to other methods that didn't use real dense labels. More importantly, the TT-GPR refinement stage boosted performance on critical, miss-sensitive metrics like Recall and False Negative Rate (FNR), demonstrating strong class-balanced behavior measured by the G-Mean metric.

The system's robustness was further validated on an external, private cohort of 50 patients from Yonsei University Hospital, where it had to contend with cross-site distribution shift—a common failure point for medical AI. In this challenging real-world test, TT-GPR consistently improved upon the synthetic-only model, increasing Dice and Recall while reducing the FNR. This demonstrates that the generative refinement process effectively adapts the model to new data distributions without requiring any additional labeled examples from the target site. The research, detailed in a paper submitted to arXiv, presents a practical pathway for developing more robust and accessible AI diagnostic tools that are not bottlenecked by the scarcity and expense of expert medical annotations.