A Comparative Analysis of Deep Convolutional Networks for Automated Diagnosis of Retinal Detachment in Dogs

OBJECTIVE: To compare ImageNet-pretrained deep convolutional neural networks for automated detection of retinal detachment (RD) in canine fundus photographs. ANIMALS STUDIED: Archived fundus images from 275 dogs. PROCEDURES: In this multicenter retrospective study, 2000 color fundus photographs (793 RD; 1207 normal) acquired between 2020 and 2025 were included after quality filtering. Data were split at the patient level into training (80%) and an independent validation set (20%). Transfer learning was applied to three pretrained architectures (ResNet50V2, VGG16, EfficientNetB0) using standardized preprocessing and real-time augmentation. Performance on the validation set was assessed using accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC). Ninety-five percent confidence intervals were estimated by bootstrapping. RESULTS: ResNet50V2 achieved the best overall discrimination (accuracy 0.8909; AUC 0.9194), followed by EfficientNetB0 (accuracy 0.8182; AUC 0.8831). VGG16 showed limited reliability (accuracy 0.6182; AUC 0.6868) due to a high false-positive rate. Gradient-weighted class activation mapping indicated that the best-performing model consistently attended to regions consistent with retinal detachment. CONCLUSIONS: ResNet50V2-based analysis of canine fundus photographs shows strong potential as a scalable screening support tool for RD. Prospective external validation across additional devices and practice settings is warranted before routine clinical implementation.