Slice-Level Deep Learning Classification of Acute Cholecystitis on Contrast-Enhanced CT: A Single-Center Benchmark of Six CNN Architectures

Abstract

Background/Objectives: Acute cholecystitis requires timely diagnosis, yet contrast-enhanced computed tomography (CT) interpretation can be challenging under a high radiology workload. This proof-of-concept study benchmarked convolutional neural network (CNN) architectures for slice-level classification, not a clinically validated diagnostic tool. Methods: Contrast-enhanced CT from 80 patients (51 acute cholecystitis, 29 controls) at a single center (2018–2020) yielded 1060 annotated axial slices. Acute cholecystitis was defined by a Tokyo Guidelines 2018–based composite reference standard. All slices from a patient were assigned to one partition only. Six ImageNet-pretrained architectures were fine-tuned: ResNet-18/50/101, Inception-V3, MobileNet-V3-small/large. Group comparisons used Student’s t test, chi-square with Yates’ correction, or Fisher’s exact test (R version 4.0.3; Python/SciPy). Results: On the internal held-out test set (213 slices; 122 acute-cholecystitis slices, 91 control slices), ResNet-101 achieved the highest accuracy (0.915) and specificity (0.857), while MobileNet-V3-large achieved the highest area under the curve (AUC) (0.97) and tied for the highest sensitivity (0.959) with far fewer parameters; across models, accuracy ranged from 0.812 to 0.915 and AUC from 0.89 to 0.97. Conclusions: Compact architectures showed competitive slice-level performance in this single-center benchmark. Because evaluation was slice-level in a small cohort with non-inflamed controls, these findings do not represent patient-level diagnostic accuracy. Patient-level aggregation, matched sensitivity analyses, explainability assessment, and external validation are required before clinical application.