Scholarlink

This study proposes a multimodal deep learning framework to enhance early lung cancer diagnosis by integrating tabular clinical data (309 patient records) and histopathological images (15,000 samples). The tabular dataset includes age-stratified groups (youth: 21–39, adults: 40–60, elderly: 61–87), smoking history, alcohol use, and symptoms (shortness of breath, chest pain). A novel mini ConvNeXt architecture, optimized for medical imaging, was trained using advanced data augmentation (rotations, flips, brightness adjustments) to improve robustness. The model achieved 98.7% accuracy and a near-perfect AUC of 0.9999, outperforming traditional CNNs like ResNet-50 (91–93% accuracy) and VGG-16 (88–90%) with 88% fewer parameters. Logistic regression identified smoking history (odds ratio: 3.1) and elderly age (odds ratio: 3.8) as the strongest predictors. Fusion of image and tabular data improved malignancy detection by 12% compared to image-only models. Clinical validation confirmed zero false negatives, critical for early-stage intervention. The framework’s lightweight design and high accuracy make it suitable for integration into electronic health records (EHRs), with recommendations for federated learning to address dataset diversity limitations.

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

Enhancing Lung Cancer Diagnosis Through CNN-Based Analysis of Clinical and Histopathological Data

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Sudesh Pahal,
  • Ayush Prasad,
  • Neha Khokhar,
  • Nidhi Gupta

摘要

This study proposes a multimodal deep learning framework to enhance early lung cancer diagnosis by integrating tabular clinical data (309 patient records) and histopathological images (15,000 samples). The tabular dataset includes age-stratified groups (youth: 21–39, adults: 40–60, elderly: 61–87), smoking history, alcohol use, and symptoms (shortness of breath, chest pain). A novel mini ConvNeXt architecture, optimized for medical imaging, was trained using advanced data augmentation (rotations, flips, brightness adjustments) to improve robustness. The model achieved 98.7% accuracy and a near-perfect AUC of 0.9999, outperforming traditional CNNs like ResNet-50 (91–93% accuracy) and VGG-16 (88–90%) with 88% fewer parameters. Logistic regression identified smoking history (odds ratio: 3.1) and elderly age (odds ratio: 3.8) as the strongest predictors. Fusion of image and tabular data improved malignancy detection by 12% compared to image-only models. Clinical validation confirmed zero false negatives, critical for early-stage intervention. The framework’s lightweight design and high accuracy make it suitable for integration into electronic health records (EHRs), with recommendations for federated learning to address dataset diversity limitations.