Phase-Mixed Versus Phase-Specific Training for Pancreatic Lesion Segmentation on Contrast-Enhanced CT: A Non-Inferiority Study on PanTS

A model that segments pancreatic lesions from either arterial or portal venous CT could simplify deployment, provided that its performance is not inferior to that of phase-specific models. We evaluated nnU-Net ResEnc-M models trained on arterial scans (A), portal venous scans (P), or the combined dataset (M-full). Quality control, reconstruction of patient groups, and automatic phase assignment yielded 3,085 development scans and 438 held-out PanTS test scans. Five-fold cross-validation was repeated with three training seeds for A, P, and M-full; five reduced, phase-balanced training subsets were evaluated using seed 0. Non-inferiority required the lower confidence bound for the paired difference in lesion Dice to exceed −0.03. M-full met this criterion in both phases during cross-validation with all three seeds. On 85 lesion-positive portal test scans, its Dice advantage over P was 0.0387–0.0473, with all corresponding 95% confidence intervals above zero. M-full had fewer complete misses (17–19 versus 21–23 of 85), but more false-positive scans (34–39 versus 23–27 of 243 lesion-negative scans). Reducing the portal training subset lowered portal cross-validation performance in all five balanced variants. The arterial test contained only four positive scans and was therefore descriptive. Evaluation on 281 annotated Task07 scans was also descriptive because of known and unresolved dataset overlap. Mixed-phase training met the specified non-inferiority criterion while requiring only one CT volume at inference, but small-lesion misses and the false-positive trade-off remained.