YOLOv11-HBDY: A Unified Hybrid Model Integrating NAM-R, CBH-R, and Ghost Convolutions for Multi-task Mpox Lesion Analysis

Recent mpox (monkeypox) outbreaks have intensified global public health concerns and underscored the urgent demand for accurate, scalable, and accessible diagnostic systems. Correctly distinguishing the five clinical stages of cutaneous lesions—macules, papules, vesicles, pustules, and scabs—is critical for therapeutic decision-making, monitoring disease progression, and preventing transmission. Despite its clinical relevance, automated classification of these dermatological phases remains markedly underexplored in the existing literature. To bridge this gap, we introduce YOLOv11-HBDY, a hybrid deep learning architecture derived from YOLOv11 and tailored for automated interpretation of mpox skin lesions across all clinical stages. The framework integrates the CBH-R (Circular Boundary-Enhanced Convolutional Block), which facilitates refined delineation of rounded lesion structures through Circular Convolutional Kernels (CCK) optimized via Gaussian functions. Complementing this module, the GGFE (Gaussian-Guided Feature Enhancement) mechanism mitigates noise and artifacts, thereby improving boundary fidelity and lesion contour definition. Furthermore, the model incorporates NAM-R (Normalized Attention Module for Rounded Patterns), an advanced variant of NAM designed to enhance attention toward circular morphological cues through exponential and hyperbolic cosine normalization. To reduce computational overhead while preserving representational power, conventional convolutional layers are substituted with Ghost-Conv blocks. The architecture employs MobileNetV3 as its backbone, ensuring efficient feature extraction and enabling deployment in resource-constrained clinical environments. To improve robustness and generalization under variable real-world imaging conditions, six data augmentation strategies—rotation, scaling, flipping, brightness modulation, translation, and contrast enhancement—were systematically applied. Experimental results demonstrate the superior effectiveness of YOLOv11-HBDY, achieving 87.4% precision, 85.3% recall, and an AUC of 0.84, confirming its strong potential as a reliable and clinically meaningful tool for mpox lesion stage classification.