Scholarlink

<p>Lung cancer remains one of the leading causes of cancer related mortality worldwide, emphasising the urgent need for accurate and early diagnostic tools. This study proposes a robust composite deep learning framework for classifying lung cancer subtypes using histopathological images. The architecture integrates DenseNet and EfficientNetB3 for hierarchical feature extraction, followed by XGBoost as the final classification layer to enhance prediction accuracy. Additionally, Class-Selective Image Preprocessing (CSIP) is employed to emphasise relevant tumour regions, thereby enhancing model interpretability and focus. The proposed ensemble model is trained and validated on the publicly available LC25000 dataset, achieving superior performance with an overall accuracy of 99.87%, precision of 99.75%, F1-score of 98.68%, and AUC of 0.996. A comparative analysis with baseline architectures, including ResNet, VGG, and Vision Transformers, demonstrates significant performance gains. Interpretability is further enhanced by using Grad-CAM visualisations to localise key discriminative features across lung cancer classes. To enable real-time clinical applicability, the model is deployed via a HIPAA and GDPR-compliant cloud-based pipeline using Amazon Web Services (AWS). The system supports fast and scalable inference, making it suitable for telemedicine and remote diagnostics. The pipeline also incorporates provisions for federated learning to ensure data privacy in multi-institutional settings. This work offers a comprehensive and deployable solution for computer-aided lung cancer diagnosis, combining deep learning, interpretability, and real-world infrastructure. Future work will include external validation using cross-institutional and multi-center clinical datasets.</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

DEXNet: An Ensemble Model Integrating DenseNet, EfficientNetB3, and XGBoost for Histopathological Lung and Colon Cancer Classification

  • 期刊论文

  • 发表日期：

    2026年4月13日

查看全文
  • Aaryan Gupta,
  • Varun Tiwari,
  • Divyansh Pandey,
  • Jaydeep Kishore

摘要

Lung cancer remains one of the leading causes of cancer related mortality worldwide, emphasising the urgent need for accurate and early diagnostic tools. This study proposes a robust composite deep learning framework for classifying lung cancer subtypes using histopathological images. The architecture integrates DenseNet and EfficientNetB3 for hierarchical feature extraction, followed by XGBoost as the final classification layer to enhance prediction accuracy. Additionally, Class-Selective Image Preprocessing (CSIP) is employed to emphasise relevant tumour regions, thereby enhancing model interpretability and focus. The proposed ensemble model is trained and validated on the publicly available LC25000 dataset, achieving superior performance with an overall accuracy of 99.87%, precision of 99.75%, F1-score of 98.68%, and AUC of 0.996. A comparative analysis with baseline architectures, including ResNet, VGG, and Vision Transformers, demonstrates significant performance gains. Interpretability is further enhanced by using Grad-CAM visualisations to localise key discriminative features across lung cancer classes. To enable real-time clinical applicability, the model is deployed via a HIPAA and GDPR-compliant cloud-based pipeline using Amazon Web Services (AWS). The system supports fast and scalable inference, making it suitable for telemedicine and remote diagnostics. The pipeline also incorporates provisions for federated learning to ensure data privacy in multi-institutional settings. This work offers a comprehensive and deployable solution for computer-aided lung cancer diagnosis, combining deep learning, interpretability, and real-world infrastructure. Future work will include external validation using cross-institutional and multi-center clinical datasets.