Evaluating Explicit Anatomical Priors Beyond a Mature CT Detector for Six Acute Abdominal Emergencies

Abstract To test whether anatomical priors from a general-purpose organ segmentation model (TotalSegmentator, TS), applied at inference or during training, improve an externally validated YOLOv11 detector for six acute abdominal emergencies. The frozen detector was applied to 827 external Stanford Merlin CT studies with language-model report labels; a nested 280-study radiologist-adjudicated subset was the reference standard. TS-derived slice selection (Arm A), organ-state integration (Arm B), and their combination (Arm C) were compared with the detector alone under frozen, recalibrated, and fine-tuned conditions. The detector was also retrained from scratch (identical 10-epoch budget) with image channels only, with the patient’s TS organ maps as added input channels, or with another patient’s maps. AUROC was primary; comparisons used a paired study-level bootstrap. The detector generalized externally (macro AUROC 0.816 cohort; 0.879 adjudicated). Slice selection was non-inferior but not beneficial: nearly all confident detections already lay within the anchored organ’s axial range; organ-state fusion improved AUROC only for cholecystitis (hypothesis-generating). In single runs, training with organ maps raised internal test [email protected] (0.517 versus 0.482), and the model depended on the maps, but external macro AUROC did not improve (0.822 versus 0.829; $$\Delta $$ Δ −0.007, 95% CI −0.019 to +0.007) and kidney-stone AUROC fell (0.678 versus 0.768; exploratory), unlike with another patient’s maps. Per-site recalibration was the most effective operating-point adjustment (macro F1 +0.021); annotation-free fine-tuning reduced discrimination. Explicit anatomy, applied after detection or during training, did not improve this detector’s overall external discrimination. Per-site recalibration gave the largest operating-point gain.