A Comparative Analysis of Deep Learning Models for Detection of Lumpy Skin Disease with emphasis on Shifted Window Transformers

Abstract

Lumpy Skin Disease (LSD) in cattle is an increasingly prevalent viral infection with significant economic impact. Traditional detection methods are often labor-intensive and delayed. In this study, five state-of-the-art deep learning (DL) architectures—ResNet50, EfficientNetB0, MobileNetV2, Vision Transformer (ViT-B16), and Swin Transformer Tiny (Swin-T)—were evaluated and compared for image-based LSD classification. Publicly available Kaggle datasets of infected and healthy cattle were used. All models were fine-tuned using transfer learning and tested for classification accuracy, F1-score, inference time, explainability (via Grad-CAM), and real-world deployability. Results show that Swin-T achieved the highest classification accuracy of 95.3%, while MobileNetV2 emerged as the most deployment-friendly model. Grad-CAM visualizations confirmed that transformer-based models captured relevant lesion features with greater spatial sensitivity than CNNs. The study highlights the promise of hybrid transformer-CNN models for practical livestock diagnostics, especially in resource-constrained environments.