Scholarlink

<p>Lung cancer has the highest fatality rate among all cancer types, causing over 2.1&#xa0;million deaths annually. Early and accurate classification of lung cancer subtypes is essential to improve prognosis and select the most effective treatment. However, traditional diagnostic methods based on manual histopathological analysis are time-consuming and prone to human error. Recent advances in deep learning offer powerful feature extraction capabilities for automated lung cancer classification using histopathological images. This study utilizes the LC25000 (lung and colon) and LungHist700 datasets and presents a novel Dual-Backbone Feature Fusion approach that combines a next-generation convolutional neural network (ConvNeXt) for hierarchical spatial feature extraction and Data-efficient Image Transformer (DeiT) for attention-based global context modeling. The framework also incorporates transfer learning and transformer-based attention mechanisms. Classifiers were used to categorize non-small cell lung cancer (NSCLC) into three classes named as adenocarcinoma, squamous cell carcinoma, and benign lung tissue. Using k-fold cross-validation, the proposed method achieved test accuracies of 99.71% on LC25000 lung images, 99.95% on LC25000 colon images, and 89.45% on LungHist700, outperforming existing state-of-the-art models. These results demonstrate the effectiveness of deep learning in lung cancer classification. Future research will focus on expanding datasets, integrating explainable artificial intelligence (XAI), and applying the model in real-time clinical settings.</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

Dual-backbone feature extraction framework for lung cancer classification in histopathology images

  • 期刊论文

  • 发表日期：

    2025年12月29日

查看全文
  • Sarita,
  • Praveen Kumar Shukla,
  • Vijaypal Singh Dhaka,
  • Nayani Jindal

摘要

Lung cancer has the highest fatality rate among all cancer types, causing over 2.1 million deaths annually. Early and accurate classification of lung cancer subtypes is essential to improve prognosis and select the most effective treatment. However, traditional diagnostic methods based on manual histopathological analysis are time-consuming and prone to human error. Recent advances in deep learning offer powerful feature extraction capabilities for automated lung cancer classification using histopathological images. This study utilizes the LC25000 (lung and colon) and LungHist700 datasets and presents a novel Dual-Backbone Feature Fusion approach that combines a next-generation convolutional neural network (ConvNeXt) for hierarchical spatial feature extraction and Data-efficient Image Transformer (DeiT) for attention-based global context modeling. The framework also incorporates transfer learning and transformer-based attention mechanisms. Classifiers were used to categorize non-small cell lung cancer (NSCLC) into three classes named as adenocarcinoma, squamous cell carcinoma, and benign lung tissue. Using k-fold cross-validation, the proposed method achieved test accuracies of 99.71% on LC25000 lung images, 99.95% on LC25000 colon images, and 89.45% on LungHist700, outperforming existing state-of-the-art models. These results demonstrate the effectiveness of deep learning in lung cancer classification. Future research will focus on expanding datasets, integrating explainable artificial intelligence (XAI), and applying the model in real-time clinical settings.