WSR betting bound re-certifies CT segmentation AI with 25 local cases

Deploying AI segmentation models across hospitals is risky: performance degrades under clinical domain shift. A new arXiv paper by Souraj Adhikary, Negar Chabi, and Andre Mastmeyer tackles this with distribution-free risk control that adds per-organ recall guarantees to frozen segmentation models. Using an AMOS-trained nnU-Net as a case study, the team audited transfer to the RAOS dataset and discovered that 7 out of 12 organs exceeded an α=0.10 recall threshold—meaning over a tenth of diseased tissue would be missed without adjustment.

The paper systematically compares Risk-Controlling Prediction Sets (RCPS), Conformal Risk Control (CRC), and a novel Waudby-Smith-Ramdas (WSR) betting bound for local re-certification. Classic fixed or global thresholds offer no per-organ guarantees, and smaller calibration sets can mask exceedances with conservative or vacuous thresholds. WSR re-certifies six Tier-1 organs with only 25 local cases versus 30-40 for Hoeffding-Bentkus (HB). CRC needs even fewer (10-15), but shows a heavier individual-case tail—a critical safety concern in medical imaging. No Tier-2 organ met the illustrative precision criterion at 25 cases, highlighting the trade-off between recall and precision in low-data settings.