Scholarlink

The issue of pneumonia remains a topic of concern globally, and consequently, proper diagnosis is extremely critical to improving the prognosis of patients suffering from it. This work evaluates the use of deep learning models, ResNet-50, EfficientNet, and MobileNet, for the class of pneumonia from chest X-ray images. We employed a dataset of about 8,000 images and performed experiments to determine the extent to which model performance depended on data augmentation and the chosen network architecture. The experimental results showed that MobileNet achieved the highest accuracy (97.51%) without the use of augmentations, whereas EfficientNet attained peak accuracy (97.12%) with the use of augmentations. Furthermore, we applied Gradient-Weighted Class Activation Maps, which added value to the model’s ability to be understood by the users. The results showed that with the aid of Explainable Artificial Intelligence (XAI) techniques, deep learning automation of pneumonia detection holds great promise. Further work should be conducted on multi-class classification and external validation, and it is generally more useful to employ sophisticated means of explanation to ensure the model is more useful and reliable in a clinical setting.

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

Deep Learning-Based Pneumonia Identification Using 3D Gradient-Weighted Class Activation Mapping Visualization

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Kornprom Pikulkaew,
  • Suphakit Awiphan

摘要

The issue of pneumonia remains a topic of concern globally, and consequently, proper diagnosis is extremely critical to improving the prognosis of patients suffering from it. This work evaluates the use of deep learning models, ResNet-50, EfficientNet, and MobileNet, for the class of pneumonia from chest X-ray images. We employed a dataset of about 8,000 images and performed experiments to determine the extent to which model performance depended on data augmentation and the chosen network architecture. The experimental results showed that MobileNet achieved the highest accuracy (97.51%) without the use of augmentations, whereas EfficientNet attained peak accuracy (97.12%) with the use of augmentations. Furthermore, we applied Gradient-Weighted Class Activation Maps, which added value to the model’s ability to be understood by the users. The results showed that with the aid of Explainable Artificial Intelligence (XAI) techniques, deep learning automation of pneumonia detection holds great promise. Further work should be conducted on multi-class classification and external validation, and it is generally more useful to employ sophisticated means of explanation to ensure the model is more useful and reliable in a clinical setting.