Hybrid model based on NAM-R, CBH-R, and Ghost-Conv for segmentation, detection, and classification of monkeypox stages from skin lesion images

Abstract The recent resurgence of mpox (monkeypox) has intensified the demand for automated, stage-specific lesion assessment to support screening, triage, and disease monitoring. However, most dermatological AI systems remain limited to binary recognition or coarse categorization and do not robustly discriminate the five clinically relevant stages of mpox skin lesions (macules, papules, vesicles, pustules, and scabs), where inter-stage visual overlap is substantial. To address this gap, we propose HOST-YOLOv10, a lightweight hybrid detector tailored to the rounded morphology of mpox lesions and designed for efficient deployment. HOST-YOLOv10 introduces three complementary components: (i) CBH-R, which incorporates circular convolutional kernels with Gaussian-guided feature enhancement to improve boundary delineation and suppress background noise; (ii) NAM-R, an attention mechanism that emphasizes smoothly rounded lesion patterns to enhance inter-stage discriminability; and (iii) Ghost-Conv substitutions to reduce redundancy and computational cost while preserving representational capacity. In addition, a MobileNetV3-based feature extractor is integrated to further improve efficiency under resource constraints, and robustness is strengthened through a controlled augmentation pipeline applied exclusively during training. Experimental results on held-out test data demonstrate that the complete HOST-YOLOv10 configuration achieves 89.3% precision, 87.0% recall, and an 88.1% F1-score, with mAP $$_{0.5:0.95}=58.5$$ at a computational cost of 81 GFLOPs, outperforming competing baselines under the same evaluation protocol. These findings support the clinical potential of HOST-YOLOv10 as a practical and interpretable solution for fine-grained mpox lesion staging in real-world settings.