Scholarlink

<p>Skin cancer remains one of the most life-threatening diseases worldwide, where early and accurate detection is critical for improving patient outcomes. Traditional diagnostic methods heavily rely on clinical expertise and can be subjective, time-consuming, and inaccessible in resource-limited settings. Recent advances in deep learning (DL) have shown remarkable potential in automating skin lesion classification. However, two significant challenges persist: the severe class imbalance in medical datasets and the inherent “black-box” nature of DL models, which hinders clinical trust and adoption. This paper proposes EXplainableDerm-EfficientNet, a novel DL-based framework that integrates advanced data preprocessing with explainable artificial intelligence (XAI) to address these challenges. Our approach introduces a hybrid data balancing strategy combining the Synthetic Minority Over-sampling Technique (SMOTE) with a novel trimming mechanism to simultaneously mitigate class imbalance and label noise in the widely-used HAM10000 dataset. The framework employs a transfer learning-based EfficientNet-B3 architecture as its core classifier, optimized for the seven-class classification of dermoscopic images. To enhance transparency, we integrate two XAI techniques—Local Interpretable Model-agnostic Explanations (LIME) and SHAP and Mutual Information-based Local Explanations (SMILE)—providing interpretable visual insights into the model’s decision-making process. Experimental results demonstrate that EXplainableDerm-EfficientNet outperforms six state-of-the-art DL models, including ResNet-152, VGG-16, and Inception-V3. Our model achieves superior performance metrics, with an accuracy of 87.87%, precision of 89.05%, recall of 84.87%, and an F1-score of 89.05%. The hybrid SMOTE-Trimming preprocessing contributes to a 3.97% accuracy improvement over the baseline. By delivering both high diagnostic accuracy and model interpretability, EXplainableDerm-EfficientNet presents a comprehensive, trustworthy computer-aided diagnosis (CAD) solution for clinical dermatology.</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

ExplainableDerm-EfficientNet na: an enhanced deep learning based- framework for dermoscopic image-based skin cancer classification with SMOTE and trimming technique

  • 期刊论文

  • 发表日期：

    2026年3月7日

查看全文
  • Khouloud Elbedoui,
  • Hiba Mzoughi,
  • Mohamed Ben Slima

摘要

Skin cancer remains one of the most life-threatening diseases worldwide, where early and accurate detection is critical for improving patient outcomes. Traditional diagnostic methods heavily rely on clinical expertise and can be subjective, time-consuming, and inaccessible in resource-limited settings. Recent advances in deep learning (DL) have shown remarkable potential in automating skin lesion classification. However, two significant challenges persist: the severe class imbalance in medical datasets and the inherent “black-box” nature of DL models, which hinders clinical trust and adoption. This paper proposes EXplainableDerm-EfficientNet, a novel DL-based framework that integrates advanced data preprocessing with explainable artificial intelligence (XAI) to address these challenges. Our approach introduces a hybrid data balancing strategy combining the Synthetic Minority Over-sampling Technique (SMOTE) with a novel trimming mechanism to simultaneously mitigate class imbalance and label noise in the widely-used HAM10000 dataset. The framework employs a transfer learning-based EfficientNet-B3 architecture as its core classifier, optimized for the seven-class classification of dermoscopic images. To enhance transparency, we integrate two XAI techniques—Local Interpretable Model-agnostic Explanations (LIME) and SHAP and Mutual Information-based Local Explanations (SMILE)—providing interpretable visual insights into the model’s decision-making process. Experimental results demonstrate that EXplainableDerm-EfficientNet outperforms six state-of-the-art DL models, including ResNet-152, VGG-16, and Inception-V3. Our model achieves superior performance metrics, with an accuracy of 87.87%, precision of 89.05%, recall of 84.87%, and an F1-score of 89.05%. The hybrid SMOTE-Trimming preprocessing contributes to a 3.97% accuracy improvement over the baseline. By delivering both high diagnostic accuracy and model interpretability, EXplainableDerm-EfficientNet presents a comprehensive, trustworthy computer-aided diagnosis (CAD) solution for clinical dermatology.