Scholarlink

<p>Melanocytic Tumors (MTs), which arise from melanocytes in the skin, demand accurate and early classification to support effective treatment. Conventional approaches primarily emphasize lesion regions, often neglecting the surrounding skin context and multi-scale analysis that are essential for precise boundary detection and robust diagnosis. To overcome these limitations, this paper proposes an efficient joint context-aware framework that integrates clinical metadata and dermoscopic images for MT analysis. The process begins with pre-processing to enhance image quality and clinical data reliability. Lesion and surrounding skin structures are segmented using an Exponential Cosine Decay U-Net (ECD-U-Net), and the Melanocytic Contextual Feature Index (MCFI) is introduced to quantify contextual scoring of non-tumor structures. Local and global features are extracted through a Deep Residual Learning–Based Deformable Attention Hierarchical Vision Transformer (DRLBDaHViT), while complementary features from clinical and contextual sources are fused using Symmetrized Kullback–Leibler Autoencoders (SKLA). For classification, a Deep CutMix Multi-Layer Hat Perceptron (Deep CMLHP) is employed, offering improved generalization through CutMix regularization and novel activation. To enhance trust and interpretability, Tsallis Entropy Shekel SHAP (TESSHAP) is incorporated for explainability of model predictions. Experimental validation using the ISIC 2017 dataset demonstrates the effectiveness of the proposed framework, achieving <b>98.86% accuracy</b> along with superior precision, recall, and segmentation performance. Comparative analysis confirms that the framework significantly outperforms existing CNN-, Transformer-, and GAN-based methods in classification robustness, boundary detection, and explainability, making it a promising tool for clinical melanoma diagnosis.</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

An Efficient Framework for Joint Context-Aware Based Melanocytic Tumor Analysis Using DRLBDaHViT and Deep CMLHP

  • 期刊论文

  • 发表日期：

    2025年9月10日

查看全文
  • Supraja Eduru,
  • Walter Priesnitz Filho,
  • Maria Emilia Camargo,
  • T. Y. Satheesha,
  • Mithileysh Sathiyanarayanan

摘要

Melanocytic Tumors (MTs), which arise from melanocytes in the skin, demand accurate and early classification to support effective treatment. Conventional approaches primarily emphasize lesion regions, often neglecting the surrounding skin context and multi-scale analysis that are essential for precise boundary detection and robust diagnosis. To overcome these limitations, this paper proposes an efficient joint context-aware framework that integrates clinical metadata and dermoscopic images for MT analysis. The process begins with pre-processing to enhance image quality and clinical data reliability. Lesion and surrounding skin structures are segmented using an Exponential Cosine Decay U-Net (ECD-U-Net), and the Melanocytic Contextual Feature Index (MCFI) is introduced to quantify contextual scoring of non-tumor structures. Local and global features are extracted through a Deep Residual Learning–Based Deformable Attention Hierarchical Vision Transformer (DRLBDaHViT), while complementary features from clinical and contextual sources are fused using Symmetrized Kullback–Leibler Autoencoders (SKLA). For classification, a Deep CutMix Multi-Layer Hat Perceptron (Deep CMLHP) is employed, offering improved generalization through CutMix regularization and novel activation. To enhance trust and interpretability, Tsallis Entropy Shekel SHAP (TESSHAP) is incorporated for explainability of model predictions. Experimental validation using the ISIC 2017 dataset demonstrates the effectiveness of the proposed framework, achieving 98.86% accuracy along with superior precision, recall, and segmentation performance. Comparative analysis confirms that the framework significantly outperforms existing CNN-, Transformer-, and GAN-based methods in classification robustness, boundary detection, and explainability, making it a promising tool for clinical melanoma diagnosis.