CONReg: Uncertainty-Aware Medical Image Registration Using Conformal Prediction.

Deep learning (DL) has advanced medical image registration, but most models produce only point estimates of dense displacement fields (DDFs) without providing information on model uncertainty, which limits clinical reliability. Uncertainty quantification (UQ) addresses this gap by identifying regions and cases where predictions may be less trustworthy. We present CONReg, a framework that integrates quantile regression with conformal prediction (CP) to provide voxelwise and case-level UQ in registration. A 3D U-Net was trained on publicly available brain and lung datasets to predict DDFs along with lower and upper quantile bounds, yielding predictive intervals for voxel displacements. Uncertainty was also estimated at anatomically defined keypoints (KPs), and CP calibrated these intervals to produce 95% statistically valid confidence bounds, represented as uncertainty bounding boxes (UBBxs). Clustering based on UBBx volumes stratified both KPs and entire cases into certain and uncertain groups. Across all test datasets, the achieved empirical coverage ranged from 0.92 to 0.98, indicating reliable UQ. Certain KPs and cases consistently exhibited significantly lower target registration error than uncertain ones (p < 0.05). For brain datasets, a pretrained VoxelMorph model was also applied to subgroup cases, and mean squared error (MSE) comparisons showed significantly lower MSE in certain cases (p < 0.05). These findings demonstrate that CONReg provides a statistically principled approach to registration uncertainty, improving both the reliability and interpretability of DL-based medical image registration.