Researchers have proposed a method for ensuring accurate segmentation of multiple organs in medical imaging scans, even when the model is trained on different datasets. The approach, called Distribution-free risk control, adds guarantees to frozen segmentation models by calibrating per-organ thresholds and auditing transfer to new datasets. However, the study found that smaller calibration sets can sometimes mask errors with overly conservative or vacuous thresholds. --- Why it matters: This matters because accurate organ segmentation is crucial in medical imaging for diagnosis and treatment planning. The proposed method's ability to provide guarantees on recall risk control can improve model reliability and trustworthiness in real-world applications. Source: https://arxiv.org/abs/2608.18193