Cat Skin Disease Diagnosis Using EfficientNetV2 for Lightweight Processing on Low-Resource Devices

Abstract

Skin diseases are among the most common health issues in domestic cats. However, access to veterinarians is often limited, especially in low-resource settings. Automated image-based detection offers a fast and affordable alternative for early intervention. This paper presents a lightweight approach for diagnosing feline skin diseases using EfficientNetV2 optimized for low-resource devices. A balanced custom dataset consisting of 720 images across nine classes, namely Healthy, Mild/Severe Ringworm, Mild/Severe Acne, Mild/Severe Flea, and Mild/Severe Scabies, was compiled from Kaggle, Roboflow, and Google Images, ensuring ethical use of publicly available data. The images were augmented through rotations (0°, 90°, 180°, 270°) and horizontal flips, resulting in 5,760 images, to enhance model generalization. Five CNN architectures were benchmarked: DenseNet121, MobileNetV2, MobileNetV3, EfficientNetB0, and EfficientNetV2B0. Training was conducted with grid searches over batch sizes {64, 32, 16, 8} and learning rates {1e-3, 5e-4, 2e-4, 1e-4, 5e-5} for up to 300 epochs, and with the Adam optimizer and Reduce-LR-on-Plateau (decay factor 0.5). Early stopping (patience = 10) was used to mitigate overfitting. The best model was selected based on highest validation accuracy. The experiments were conducted on an Intel Xeon 6 CPU (2.2 GHz, 2 vCPUs) in Google Colab without GPU to simulate low-resource deployment. EfficientNetV2B0 achieved the best performance with 99.62% validation accuracy and 99.79% test accuracy, with an average inference latency of 78 ms/frame. Compared to previous studies focusing on heavyweight models or conventional ML using handcrafted features, this work highlights the feasibility of deploying an accurate real-time diagnostic pipeline on edge devices.