<![CDATA[Accurate classification of feline skin diseases can support early identification and improve access to preliminary veterinary information. This study evaluates three pretrained convolutional neural network (CNN) architectures, ConvNeXtV2 (Atto), MobileNetV4 (ConvSmall), and RegNet (Y-008), for classifying five categories of cat images: feline acne, flea allergy, scabies, healthy cats, and ear mites. We utilize a balanced dataset of 3,000 images, with 600 images per class, and apply image preprocessing and online data augmentation to improve model training. We evaluate the models using three-fold cross-validation and a separate test set of 300 images. The results show that ConvNeXtV2 achieves the highest mean cross-validation accuracy of 96.48%, followed by RegNet at 96.33% and MobileNetV4 at 92.30%. However, RegNet achieves the best test performance, with 89.67% accuracy, 92.61% macro precision, 89.67% macro recall, and an 89.81% macro F1-score. RegNet also achieves 70.00% recall for scabies, outperforming ConvNeXtV2 and MobileNetV4. All three models achieve 100% recall for healthy cats, while scabies remains the most challenging class due to frequent misclassification as feline acne and ear mites. The application rejects predictions with confidence below 50% and clarifies that its output does not replace a veterinary diagnosis. These findings demonstrate the potential of pretrained CNN architectures for feline skin disease classification and identify RegNet as the most effective model among the evaluated architectures.]]>
Copyrights © 2026
