Scholarlink

The recent outbreak of coronavirus disease 2019 (COVID-19) has proven to be a dangerous virus, causing thousands of deaths and showing catastrophic effects on a large global population. Artificial intelligence (AI)-driven tools, particularly those leveraging chest X-ray (CXR) images, have emerged as effective solutions for early screening and diagnosis. However, upon inspection, many new methods are being developed and focus primarily on accuracy, but most lack explanation of the classification process. Training a model from scratch requires large amounts of annotated data, which is a tedious and time-consuming task, especially in medical imaging. In this chapter, we address a critical question in the accurate diagnosis of COVID-19: Can fine-tuning (FT) pretrained models improve performance and eliminate the need for training models from scratch? To answer this, we shift the focus beyond accuracy and aim to comprehensively evaluate four distinct training strategies using the widely recognized VGG16 and VGG19 architectures for early COVID-19 detection with CXR images. The four training strategies explored are training from scratch, using pretrained models as off-the-shelf feature extractors, model truncation, and block-wise fine-tuning. The dataset used in this work contains a mixture of 4055 chest radiographs with COVID-19, healthy, and pneumonia cases. We also provide a visual interpretation of the different training strategies using Grad-CAM to highlight the areas of interest the model focuses on when making its diagnosis. Extensive experiments demonstrate that (1) our block-wise fine-tuning (FT) strategy significantly enhances classification performance compared to the other three strategies; (2) a moderate level of FT strikes the optimal balance and proves to be more robust than both shallow and deep FT; and (3) the best fine-tuned VGG16 and VGG19 models achieve impressive classification accuracies of 98.73% and 98.44%, respectively, with Grad-CAM visualizations focusing on the infected regions. These findings suggest that block-wise FT yields the highest accuracy and outperforms other existing studies in COVID-19 detection.

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

Optimized Block-Wise Fine-Tuning of VGG Models for Accurate and Explainable Detection of Chest Infectious Diseases Using Chest X-rays

  • 章节

  • 发表日期：

    2026年

查看全文
  • Sohaib Asif,
  • Fuguo Xue,
  • Youqin Chen,
  • Ruxuan Yan,
  • Lingying Zhu,
  • Enyu Wang,
  • Dong Xu

摘要

The recent outbreak of coronavirus disease 2019 (COVID-19) has proven to be a dangerous virus, causing thousands of deaths and showing catastrophic effects on a large global population. Artificial intelligence (AI)-driven tools, particularly those leveraging chest X-ray (CXR) images, have emerged as effective solutions for early screening and diagnosis. However, upon inspection, many new methods are being developed and focus primarily on accuracy, but most lack explanation of the classification process. Training a model from scratch requires large amounts of annotated data, which is a tedious and time-consuming task, especially in medical imaging. In this chapter, we address a critical question in the accurate diagnosis of COVID-19: Can fine-tuning (FT) pretrained models improve performance and eliminate the need for training models from scratch? To answer this, we shift the focus beyond accuracy and aim to comprehensively evaluate four distinct training strategies using the widely recognized VGG16 and VGG19 architectures for early COVID-19 detection with CXR images. The four training strategies explored are training from scratch, using pretrained models as off-the-shelf feature extractors, model truncation, and block-wise fine-tuning. The dataset used in this work contains a mixture of 4055 chest radiographs with COVID-19, healthy, and pneumonia cases. We also provide a visual interpretation of the different training strategies using Grad-CAM to highlight the areas of interest the model focuses on when making its diagnosis. Extensive experiments demonstrate that (1) our block-wise fine-tuning (FT) strategy significantly enhances classification performance compared to the other three strategies; (2) a moderate level of FT strikes the optimal balance and proves to be more robust than both shallow and deep FT; and (3) the best fine-tuned VGG16 and VGG19 models achieve impressive classification accuracies of 98.73% and 98.44%, respectively, with Grad-CAM visualizations focusing on the infected regions. These findings suggest that block-wise FT yields the highest accuracy and outperforms other existing studies in COVID-19 detection.