Scholarlink

Context <p>Skin cancer is a major health burden in the world, and the incidence continues to rise and when diagnosed late, death rates are high. Dermoscopic imaging has emerged as one of the most important non-invasive diagnostic methods for early diagnosis; nevertheless, the interpretation of findings is largely reliant on clinical skills, which is driving the creation of dependable computer-aided diagnostic systems.</p> Existing literature and limitations <p>More recent developments in deep learning, particularly convolutional neural networks (CNNs), have demonstrated promising performance in automated skin lesion classification. However, current methods based on single CNN networks or probability-level ensemble schemes often suffer from severe class imbalance, inter-class visual similarity, and poor generalisation. Such constraints are particularly observable in multi-class systems where minority and malignant lesion categories are often misclassified.</p> Objective of this research <p>The main objective of this research is to address the gaps in current CNN-based and soft-voting ensemble systems by creating a more robust and discriminatory model for classifying multi-class skin lesions.</p> Present research <p>This paper proposes a stacking-based deep ensemble model combining DenseNet121, EfficientNet-B0, and MobileNetV2 as complementary base learners. Each network is trained to produce deep feature embeddings, which are fused at the feature level and fed into an XGBoost meta-classifier to model nonlinear feature interactions and allow the model to learn feature weights.</p> Findings and results <p>The results of extensive experiments on the HAM10000 dataset, comprising 10,015 dermoscopic images of seven lesion types, demonstrate that the proposed model achieves test accuracy of 86.38%, macro F1-score of 0.7485, weighted F1-score of 0.8559, and macro-AUC of 0.9880. The stacking ensemble demonstrated competitive and balanced performance compared with single CNN models and traditional soft-voting ensembles. To validate the robustness and clinical reliability of the proposed framework, additional experiments, including ablation analysis, statistical significance testing, Grad-CAM explainability analysis, and meta-classifier comparison, were conducted.</p> Contributions <p>The paper has three main contributions: (i) it presents a deep feature-level stacking framework that is effective in capturing complementary representations by different CNNs, (ii) it shows the effectiveness of gradient-boosted meta-learning to deal with class imbalance and inter-class ambiguity, and (iii) it presents in-depth experimental evidence of enhanced sensitivity to clinically important minority-classes including AKIEC and melanoma.</p> Conclusion <p>The results demonstrate that deep feature fusion and gradient-boosted stacking can improve classification robustness and balanced performance of multi-class skin lesion classification. The proposed framework presents a clinically meaningful solution for computer-aided dermatology and provides directions for future research, where uncertainty-aware learning, cross-dataset generalisation, and lightweight implementation in the clinical setting should be considered.</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

A stacking-based deep learning ensemble for multi-class skin lesion classification

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月3日

查看全文
  • H Varun Chand,
  • Seema Sabharwal,
  • Weiwei Jiang,
  • Sardar M. N. Islam,
  • Korhan Cengiz

摘要

Context

Skin cancer is a major health burden in the world, and the incidence continues to rise and when diagnosed late, death rates are high. Dermoscopic imaging has emerged as one of the most important non-invasive diagnostic methods for early diagnosis; nevertheless, the interpretation of findings is largely reliant on clinical skills, which is driving the creation of dependable computer-aided diagnostic systems.

Existing literature and limitations

More recent developments in deep learning, particularly convolutional neural networks (CNNs), have demonstrated promising performance in automated skin lesion classification. However, current methods based on single CNN networks or probability-level ensemble schemes often suffer from severe class imbalance, inter-class visual similarity, and poor generalisation. Such constraints are particularly observable in multi-class systems where minority and malignant lesion categories are often misclassified.

Objective of this research

The main objective of this research is to address the gaps in current CNN-based and soft-voting ensemble systems by creating a more robust and discriminatory model for classifying multi-class skin lesions.

Present research

This paper proposes a stacking-based deep ensemble model combining DenseNet121, EfficientNet-B0, and MobileNetV2 as complementary base learners. Each network is trained to produce deep feature embeddings, which are fused at the feature level and fed into an XGBoost meta-classifier to model nonlinear feature interactions and allow the model to learn feature weights.

Findings and results

The results of extensive experiments on the HAM10000 dataset, comprising 10,015 dermoscopic images of seven lesion types, demonstrate that the proposed model achieves test accuracy of 86.38%, macro F1-score of 0.7485, weighted F1-score of 0.8559, and macro-AUC of 0.9880. The stacking ensemble demonstrated competitive and balanced performance compared with single CNN models and traditional soft-voting ensembles. To validate the robustness and clinical reliability of the proposed framework, additional experiments, including ablation analysis, statistical significance testing, Grad-CAM explainability analysis, and meta-classifier comparison, were conducted.

Contributions

The paper has three main contributions: (i) it presents a deep feature-level stacking framework that is effective in capturing complementary representations by different CNNs, (ii) it shows the effectiveness of gradient-boosted meta-learning to deal with class imbalance and inter-class ambiguity, and (iii) it presents in-depth experimental evidence of enhanced sensitivity to clinically important minority-classes including AKIEC and melanoma.

Conclusion

The results demonstrate that deep feature fusion and gradient-boosted stacking can improve classification robustness and balanced performance of multi-class skin lesion classification. The proposed framework presents a clinically meaningful solution for computer-aided dermatology and provides directions for future research, where uncertainty-aware learning, cross-dataset generalisation, and lightweight implementation in the clinical setting should be considered.