Scholarlink

While deep learning has revolutionized medical image analysis, its high computational demands and opaque decision-making have hindered its clinical adoption. In this study, we systematically compare CNN-based (baseline CNN, EfficientNet, CheXNet) and transformer-based (ViT, Swin, DeiT, MedViT) models for pneumonia detection on a chest X-ray dataset. We employ transfer learning to reduce training overhead and evaluate model performance using accuracy, precision, recall, F1 score, and ROC AUC, achieving an F1 score of 95.93% for the best performing model (Swin Transformer). Subsequently, Integrated Gradients provide transparent visual explanations, highlighting pathological regions associated with pneumonia. Our findings show transformer-based models outperform CNN-based models, and when coupled with XAI techniques, they pave the way for robust and trustworthy adoption of AI in healthcare.

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

Decoding Pneumonia: Evaluating Convolutional Neural Networks and Transformer Architectures Using Integrated Gradients

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Abdullah Haneef,
  • Anas Ahmad,
  • Rahul Katarya

摘要

While deep learning has revolutionized medical image analysis, its high computational demands and opaque decision-making have hindered its clinical adoption. In this study, we systematically compare CNN-based (baseline CNN, EfficientNet, CheXNet) and transformer-based (ViT, Swin, DeiT, MedViT) models for pneumonia detection on a chest X-ray dataset. We employ transfer learning to reduce training overhead and evaluate model performance using accuracy, precision, recall, F1 score, and ROC AUC, achieving an F1 score of 95.93% for the best performing model (Swin Transformer). Subsequently, Integrated Gradients provide transparent visual explanations, highlighting pathological regions associated with pneumonia. Our findings show transformer-based models outperform CNN-based models, and when coupled with XAI techniques, they pave the way for robust and trustworthy adoption of AI in healthcare.