Six CNN architectures benchmarked for slice-level acute cholecystitis detection on CT

Proof-of-concept: MobileNet-V3-large achieved AUC 0.97, sensitivity 0.959 for slice-level acute cholecystitis on contrast-enhanced CT; ResNet-101 had accuracy 0.915, specificity 0.857. Slice-level, single-center, not patient-level – external validation needed.

• ResNet-101 achieved highest accuracy (0.915) and specificity (0.857); MobileNet-V3-large had highest AUC (0.97) and sensitivity (0.959) with far fewer parameters.
• Across six CNN architectures, accuracy ranged 0.812–0.915 and AUC 0.89–0.97 on the internal held-out test set of 213 slices.
• Key limitation: slice-level evaluation in a small single-center cohort (n=80) with non-inflamed controls; patient-level aggregation and external validation are required before clinical use.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.