Comparative Image-Based Evaluation of Deep Learning Models for the Diagnosis of Different Ocular Diseases in Cats.

Deep learning models, especially EfficientNet-B0 and DenseNet-121, show promise supporting feline ocular disease diagnosis from images.

This study evaluated nine convolutional neural network (CNN) architectures for automated classification of feline ocular diseases using 456 clinically validated ophthalmic images spanning 13 classes (12 disease categories plus a healthy group). Architectures tested included ResNet, EfficientNet, DenseNet, MobileNet, and VGG variants, all trained via transfer learning with pre-trained ImageNet weights. Data augmentation and five-fold cross-validation were employed during training. EfficientNet-B0 achieved the highest overall classification accuracy at 78%, while DenseNet-121 led in macro-averaged F1-score (0.76) and AUC (0.9790), closely followed by EfficientNet-B0 (AUC 0.9750) and ResNet-34 (AUC 0.9724). McNemar's test confirmed no statistically significant difference among these three top-performing models (p>0.05). Disease-level performance varied considerably: cherry eye, healthy eyes, and corneal sequestration were classified with the highest precision and recall (reaching 1.00), whereas glaucoma (recall 0.25) and corneal ulcer (F1-score 0.29) showed the weakest classification results. Per-image inference times ranged from 2.67 to 5.73 milliseconds across all architectures, indicating practical clinical feasibility. Gradient-weighted Class Activation Mapping (Grad-CAM) confirmed that EfficientNet-B0 focused on clinically relevant ocular regions, supporting model interpretability. The authors conclude that deep learning-based multi-class image classification, particularly using EfficientNet-B0, DenseNet-121, and ResNet-34, shows promise as a decision-support tool for feline ophthalmic diagnostics, though challenges remain for conditions with overlapping clinical presentations.

This summary was distilled by AI and may occasionally misinterpret data. Confirm critical details with the primary literature before clinical application.