{QUALITY} {CONTROL} {FOR} {MEDICAL} {IMAGE} {SEGMENTATION} {UNDER} {DOMAIN} {SHIFT} {WITH} {HETEROSCEDASTIC} {REGRESSION}

Trustworthy medical image segmentation systems require robust quality control mechanisms so that inadequately segmented cases can be excluded from downstream processing. This need becomes particularly acute when inputs deviate from the training distribution, such as images from unseen pathologies or novel scanner types. We propose heteroscedastic regression as a new approach for this, explicitly modeling both the expected accuracy and the uncertainty in our estimate. This framework provides flexibility to perform failure detection with varying thresholds for segmentation quality and confidence. Moreover, it can be combined with a recently proposed adversarial training strategy to achieve OOD generalization. Our approach permits computationally efficient inference and more reliably filters out low-quality segmentations than score agreement, while retaining substantially more high-quality results than predictive entropy or anomaly detection based approaches.