Enhancing Conditional Risk Control in Image Segmentation with Adaptive Conformal Prediction

Rui Luo, Zhixin Zhou

Keywords: conformal prediction, conformal risk control, image segmentation

Abstract: Uncertainty quantification is crucial in high-stakes image segmentation, yet existing conformal risk control methods often exhibit highly variable conditional risk: some images suffer extreme false negative rates while others show minimal errors. We introduce Conformal Risk Adaptation (CRA), a framework that employs a novel score function, inspired by adaptive prediction sets, to create image-specific uncertainty regions. We formalize the risk control problem as a weighted quantile estimation task, which enables a computationally efficient, grid-search-free algorithm for threshold calculation. To ensure the reliability of our adaptive score function, we integrate a specialized non-parametric calibration method that enhances pixel-wise probability estimates. Experiments on polyp and crack segmentation demonstrate that CRA maintains valid marginal risk guarantees while delivering substantially more consistent conditional risk control across diverse images. This advancement provides practitioners with a principled approach to uncertainty quantification that adapts to individual cases while maintaining rigorous statistical guarantees, which is critical for personalized medical applications. We have provided code with implementation details in the repository below: https://anonymous.4open.science/r/conformal-risk-adaptation-3BB2.

Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)

Submission Number: 8715