Scholarlink

The conventional diagnosis of skin cancer using deep learning is heavily reliant on convolutional neural networks (CNNs), which often lack generalization and interpretability. This task introduces a hybrid CNN-transformer model that integrates a visually appealing and efficient Vision Transformer (ViT). Additionally, image preprocessing techniques such as composite image formation and adaptive histogram equalization are employed to enhance the model’s robustness. The model leverages diverse data sources, including the ISIC Archives and other clinical images, to improve generalization. Interpretative techniques like Grad-CAM and SHAP are also incorporated to provide insights into the model’s decision-making process. Experimental results show a significant improvement compared to traditional CNN-based models, achieving an accuracy rate of 99.3 percent and an AUC of 0.98, which surpasses state-of-the-art methods. This paper addresses the gap between accuracy and interpretability, making the model a valuable tool for clinical practice.

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

Advancements in Skin Cancer Detection: A Hybrid CNN-Transformer Approach for Accurate Diagnosis

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • J. GnanaSoundharam,
  • P. Girija,
  • S. Fiona Branice

摘要

The conventional diagnosis of skin cancer using deep learning is heavily reliant on convolutional neural networks (CNNs), which often lack generalization and interpretability. This task introduces a hybrid CNN-transformer model that integrates a visually appealing and efficient Vision Transformer (ViT). Additionally, image preprocessing techniques such as composite image formation and adaptive histogram equalization are employed to enhance the model’s robustness. The model leverages diverse data sources, including the ISIC Archives and other clinical images, to improve generalization. Interpretative techniques like Grad-CAM and SHAP are also incorporated to provide insights into the model’s decision-making process. Experimental results show a significant improvement compared to traditional CNN-based models, achieving an accuracy rate of 99.3 percent and an AUC of 0.98, which surpasses state-of-the-art methods. This paper addresses the gap between accuracy and interpretability, making the model a valuable tool for clinical practice.