Signal, Image and Video Processing Signal Processing
HYDRA-YOLOv10: An automated approach for detection and classification of Monkeypox skin lesion stages using YOLOv10 Adapted with Circular-Responsive Attention (CRA) and Normalization-Based Attention Mechanism (NAM)
Abstract
Mpox is a zoonotic viral disease whose clinical manifestations often resemble other dermatological conditions, making accurate and timely diagnosis a challenge. Traditional diagnostic methods, such as PCR, are accurate but costly, time-consuming, and often inaccessible in low-resource environments. This study proposes HYDRA-YOLOv10, a deep learning-based architecture designed to automatically detect and classify Mpox skin lesions across five clinical stages: macules, papules, vesicles, pustules, and scabs. The model incorporates key architectural innovations, including hybrid backbone connections, extended use of the C2f module for multi-scale feature fusion, and the replacement of the SPPF module with the Circular-Responsive Attention (CRA) mechanism-designed to better capture the geometry of rounded lesions using polar-aware convolutions. Additionally, the Normalization-based Attention Mechanism (NAM) replaces conventional spatial attention, enhancing feature integration and robustness to visual variations. Experimental results demonstrate the model’s superiority over state-of-the-art approaches, achieving 95.4% precision, 94.8% recall, 95.1% F1 score, 95.5% mAP@0.5, and 83.4% mAP@0.5:0.95. These results validate HYDRA-YOLOv10 as a reliable and scalable solution for automated Mpox diagnosis, offering valuable support for clinical decision-making and public health response.