M-LSA-Net: A Hybrid Morphology-Aware Deep Learning Framework for Lumpy Skin Disease Detection

Abstract

Lumpy Skin Disease (LSD) is a highly contagious viral disease that significantly impacts livestock productivity and agricultural economies worldwide. Traditional diagnostic approaches, including manual inspection and laboratory testing, are often time-consuming, resource-intensive, and unsuitable for rapid field-level deployment. To address these limitations, this study proposes M-LSA-Net, a hybrid morphology-aware deep learning framework for accurate and efficient detection of LSD from cattle skin images. The proposed model integrates MixUp-based data augmentation for improved generalization, LUMPNet-based region-focused feature extraction for capturing lesion-specific patterns, an Attention-Based Sparse Autoencoder (AB-SAE) for adaptive feature selection, and a Morphology-Aware Deep Classifier (MADC) for capturing both local texture and global structural patterns. The model was evaluated on a publicly available dataset using an 80:20 train–test split and assessed using standard performance metrics. Experimental results demonstrate that M-LSA-Net achieves an accuracy of 98.71%, precision of 98.68%, recall of 98.71%, F1-score of 98.69%, and ROC-AUC of 0.9978, outperforming baseline models such as CoAtNet, SwinT V2, and MaxViT. The findings highlight the effectiveness of combining attention-based feature refinement and morphology-aware learning for robust disease detection, making the proposed framework suitable for real-world veterinary diagnostic applications.