Scholarlink

Basal cell carcinoma (BCC) is the most prevalent type of skin cancer, and its early detection is critical for reducing morbidity. This paper presents a novel ensemble deep learning framework that integrates three high-performance convolutional neural networks—EfficientNet-B3, DenseNet-201, and Inception-ResNet-V2—for accurate BCC detection in dermoscopic images. The ensemble utilizes a soft voting strategy to combine the predictive strengths of each model. Evaluated on the HAM10000 dataset, the proposed model achieved an accuracy of 98.4%, a sensitivity of 100%, and an area under the ROC curve (AUC) of 0.9983—outperforming state-of-the-art approaches. Importantly, the framework operates without handcrafted features or metadata, providing a deployable, patient-agnostic solution. Interpretability is ensured through Grad-CAM visualizations, with a thorough comparative evaluation and discussion on ethical implications for clinical integration. These results support a no-miss strategy for BCC detection and establish a foundation for trustworthy AI-assisted dermatological screening.

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 Robust Ensemble Deep Learning Framework for Highly Sensitive and Interpretable Detection of Basal Cell Carcinoma in Dermoscopic Images

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Antriksh Sharma,
  • Sharad Saxena,
  • Harkiran Kaur

摘要

Basal cell carcinoma (BCC) is the most prevalent type of skin cancer, and its early detection is critical for reducing morbidity. This paper presents a novel ensemble deep learning framework that integrates three high-performance convolutional neural networks—EfficientNet-B3, DenseNet-201, and Inception-ResNet-V2—for accurate BCC detection in dermoscopic images. The ensemble utilizes a soft voting strategy to combine the predictive strengths of each model. Evaluated on the HAM10000 dataset, the proposed model achieved an accuracy of 98.4%, a sensitivity of 100%, and an area under the ROC curve (AUC) of 0.9983—outperforming state-of-the-art approaches. Importantly, the framework operates without handcrafted features or metadata, providing a deployable, patient-agnostic solution. Interpretability is ensured through Grad-CAM visualizations, with a thorough comparative evaluation and discussion on ethical implications for clinical integration. These results support a no-miss strategy for BCC detection and establish a foundation for trustworthy AI-assisted dermatological screening.