Scholarlink

<p>Melanoma remains one of the most aggressive forms of skin cancer, and early diagnosis is critical to improving patient survival. This study presents an Adaptive Hybrid AI Framework (AHA-Net) designed for accurate and interpretable skin lesion segmentation and melanoma classification. The proposed architecture enhances a modified UNet + + backbone with an Adaptive Scaled Dot Attention Mechanism (A-SDAM) that dynamically regulates attention sharpness across multiple lesion scales. A Residual Cross-Attention Bridge (RCAB) enables effective contextual fusion between segmentation and classification pathways, while Vision Transformer (ViT)–based multi-scale attention layers capture both local and global dependencies. A hybrid CNN–ViT classifier, refined with dynamically weighted SVM post-classification, improves decision boundary precision under class imbalance. Furthermore, self-distillation between ViT layers enhances feature coherence and cross-dataset generalization. The framework was rigorously evaluated across four benchmark datasets i.e., HAM10000, ISIC 2019; ISIC 2020, and PH2, representing diverse lesion types and imaging conditions. Experimental results demonstrate that Proposed AHA-Net consistently outperforms existing state-of-the-art architectures, including U-Net, ResNet, UNet++, and TransUNet, in both segmentation and classification tasks. Statistical analysis confirms the significance and reproducibility of performance gains (<i>p</i> &lt; 0.05). Quantitative explainability assessment using Grad-CAM shows that the model’s attention maps align closely with clinically relevant melanoma features such as irregular borders, color heterogeneity, and asymmetric structures. These results establish Proposed AHA-Net as a robust, generalizable, and explainable AI framework with strong potential for integration into real-world dermatological diagnostic workflows.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Adaptive Hybrid AI Framework for Robust and Explainable Skin Lesion Segmentation and Melanoma Detection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月4日

查看全文
  • Umesh Kumar Lilhore,
  • D. Anitha,
  • R. Priya,
  • Wasudeo P. Rahane,
  • Sunil L. Bangare,
  • R. Kavitha,
  • Arvind Panwar,
  • Lidia Gosy Tekeste,
  • Ehab Ghith,
  • Hanaa A. Abdallah,
  • Sarita Simaiya

摘要

Melanoma remains one of the most aggressive forms of skin cancer, and early diagnosis is critical to improving patient survival. This study presents an Adaptive Hybrid AI Framework (AHA-Net) designed for accurate and interpretable skin lesion segmentation and melanoma classification. The proposed architecture enhances a modified UNet + + backbone with an Adaptive Scaled Dot Attention Mechanism (A-SDAM) that dynamically regulates attention sharpness across multiple lesion scales. A Residual Cross-Attention Bridge (RCAB) enables effective contextual fusion between segmentation and classification pathways, while Vision Transformer (ViT)–based multi-scale attention layers capture both local and global dependencies. A hybrid CNN–ViT classifier, refined with dynamically weighted SVM post-classification, improves decision boundary precision under class imbalance. Furthermore, self-distillation between ViT layers enhances feature coherence and cross-dataset generalization. The framework was rigorously evaluated across four benchmark datasets i.e., HAM10000, ISIC 2019; ISIC 2020, and PH2, representing diverse lesion types and imaging conditions. Experimental results demonstrate that Proposed AHA-Net consistently outperforms existing state-of-the-art architectures, including U-Net, ResNet, UNet++, and TransUNet, in both segmentation and classification tasks. Statistical analysis confirms the significance and reproducibility of performance gains (p < 0.05). Quantitative explainability assessment using Grad-CAM shows that the model’s attention maps align closely with clinically relevant melanoma features such as irregular borders, color heterogeneity, and asymmetric structures. These results establish Proposed AHA-Net as a robust, generalizable, and explainable AI framework with strong potential for integration into real-world dermatological diagnostic workflows.