ConRad uses segmentation uncertainty for efficient radiomic conformal prediction

Radiomic features derived from medical images and segmentation masks are increasingly used in clinical imaging pipelines, but they often rely on predicted masks from segmentation models that can be overconfident or poorly calibrated. Standard conformal prediction offers distribution-free coverage guarantees, but its black-box intervals for segmentation-derived radiomics ignore test-time information about image appearance, mask geometry, and segmentation uncertainty, leading to inefficiency.

ConRad addresses this by constructing adaptive prediction intervals using covariates from the predicted mask, input image, predicted radiomics, and boundary uncertainty. In experiments across five 2D medical imaging datasets and 171 retained radiomic targets, ConRad consistently improved efficiency compared to baseline conformal methods while preserving near-nominal empirical coverage. Ablation studies revealed that segmentation boundary uncertainty features contributed most to the gains, highlighting the value of incorporating model confidence into radiomic uncertainty quantification. The code is publicly available.