Scholarlink

<p>Skin cancer diagnosis using dermoscopic images is challenging due to inter-class similarities, intra-class variations, and severe class imbalance. Despite the strong performance of Convolutional Neural Networks (CNNs), their generalization across diverse lesion types depends on the choice of architecture and training strategies. The proposed study introduces a multi-level ensemble framework which integrates the use of bagging, feature-level boosting, and stacking on deep CNN-based feature extractors. The ResNet50, DenseNet121 and EfficientNetB0 backbone architectures are trained with a standard consistent pipeline on the HAM10000 dataset (70/15/15 train/validation/test split). Five independently trained instances per architecture (15 models in total) are used to implement bagging to increase representational diversity. XGBoost is used to perform feature-level boosting on concatenated CNN embeddings, and then stacked with additional meta-learners, such as Support Vector Machine (SVM), XGBoost, and CatBoost, for decision-level fusion. Experimental results show that the proposed framework outperforms single CNNs and single-level ensemble methods in terms of classification accuracy and robustness. Per-class AUC scores have a value of over 0.98 on most lesion categories. Grad-CAM visualizations, ROC curves, and confusion matrices are used to show the interpretability and diagnostic reliability of the system. The top-performing Stack-SVM model has the highest accuracy of 89.75 with per-class AUC of above 0.98 in most lesion types. This work provides a systematic evaluation of multi-level ensemble strategies and their effectiveness in improving performance and resilience, which makes them potentially applicable to clinical decision support systems.</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

Multi-Level Ensemble Learning Using CNN Feature Aggregation for Dermatological Image Classification

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月21日

查看全文
  • H. Varun Chand,
  • Edeh Michael Onyema,
  • Chin-Shiuh Shieh,
  • Weiwei Jiang,
  • S Arun Kumar,
  • Shashi Kant Gupta

摘要

Skin cancer diagnosis using dermoscopic images is challenging due to inter-class similarities, intra-class variations, and severe class imbalance. Despite the strong performance of Convolutional Neural Networks (CNNs), their generalization across diverse lesion types depends on the choice of architecture and training strategies. The proposed study introduces a multi-level ensemble framework which integrates the use of bagging, feature-level boosting, and stacking on deep CNN-based feature extractors. The ResNet50, DenseNet121 and EfficientNetB0 backbone architectures are trained with a standard consistent pipeline on the HAM10000 dataset (70/15/15 train/validation/test split). Five independently trained instances per architecture (15 models in total) are used to implement bagging to increase representational diversity. XGBoost is used to perform feature-level boosting on concatenated CNN embeddings, and then stacked with additional meta-learners, such as Support Vector Machine (SVM), XGBoost, and CatBoost, for decision-level fusion. Experimental results show that the proposed framework outperforms single CNNs and single-level ensemble methods in terms of classification accuracy and robustness. Per-class AUC scores have a value of over 0.98 on most lesion categories. Grad-CAM visualizations, ROC curves, and confusion matrices are used to show the interpretability and diagnostic reliability of the system. The top-performing Stack-SVM model has the highest accuracy of 89.75 with per-class AUC of above 0.98 in most lesion types. This work provides a systematic evaluation of multi-level ensemble strategies and their effectiveness in improving performance and resilience, which makes them potentially applicable to clinical decision support systems.