Anatomy-prior-guided and scan-aware 3D segmentation of the pediatric pancreas in CT images

Abstract

Objective. Pediatric pancreas CT segmentation is challenging because developmental anatomy and scan-protocol variability jointly affect organ size, image appearance, and boundary visibility. This study evaluated whether explicitly incorporating anatomy-related and scan-related case-level descriptors could improve pediatric pancreas segmentation relative to an image-only baseline. Approach. We developed APSC-Net, an anatomy-prior-guided and scan-aware 3D segmentation framework based on SegResNet. Structured anatomy-related and scan-related descriptors were encoded as separate context vectors and injected through bottleneck anatomy-prior guidance and decoder-stage low-rank FiLM modulation. The model was trained and internally tested on GE pediatric CT cases, then evaluated without retraining on a held-out Siemens subset in a cross-vendor evaluation within the same cohort and institution. Main results. On the full held-out GE test set, APSC-Net improved the mean dice similarity coefficient (DSC) from 0.665 to 0.770 and reduced the 95% Hausdorff distance (HD95) from 22.2 to 11.6 mm compared with an image-only SegResNet baseline. Stratified analysis showed DSC gains across pediatric age groups, including the youngest subgroup (0.718–0.751 for 0–5 years), and across exposure-stage strata. In the GE stream ablation, the full and scan-only configurations performed similarly. In secondary testing on held-out Siemens cases, the full model achieved lower average symmetric surface distance (ASSD) than either single-stream variant and achieved the lowest HD95 and ASSD among the compared methods, while age-group domain-regularized baselines remained competitive in DSC. Significance. On the evaluated cohorts, APSC-Net achieved lower boundary-distance errors than the image-only SegResNet baseline. The findings support further evaluation of case-level metadata conditioning for pediatric imaging tasks affected by developmental and acquisition-related variability.