Scholarlink

In this paper, we introduce a novel approach to enhance chest X-ray image classification using self-supervised learning (SSL) on the Swin Transformer architecture (CXR-SSL-SwinT). Our method employs contrastive learning to extract valuable features from extensive unlabeled data, thereby improving data utilization and addressing the shortage of annotated X-ray images. We propose improving X-ray image classification efficiency by integrating SSL with classification algorithms. This involves fine-tuning the linear classifier of an unlabeled pretrained model on labeled data, followed by integrating a nonlinear classifier. Experimental results demonstrate the effectiveness of our proposed approach, showing significant improvements in classification performance. Fine-tuning the linear classifier from the ImageNet pretrained model achieved an accuracy of 81.8% (Model-I), while fine-tuning the linear classifier from the pretrained model with unlabeled X-ray images reached an accuracy of 85.2% (Model-SSL). Integrating the nonlinear classifier onto Model-SSL further increased accuracy to 88.1%, representing improvements of +6.3% compared to Model-I and +2.9% for Model-SSL.

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

Combining Self-Supervised Learning with Swin Transformer and Machine Learning Techniques for Chest X-Ray Image Classification

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Tri-Thuc Vo,
  • Thanh-Nghi Do

摘要

In this paper, we introduce a novel approach to enhance chest X-ray image classification using self-supervised learning (SSL) on the Swin Transformer architecture (CXR-SSL-SwinT). Our method employs contrastive learning to extract valuable features from extensive unlabeled data, thereby improving data utilization and addressing the shortage of annotated X-ray images. We propose improving X-ray image classification efficiency by integrating SSL with classification algorithms. This involves fine-tuning the linear classifier of an unlabeled pretrained model on labeled data, followed by integrating a nonlinear classifier. Experimental results demonstrate the effectiveness of our proposed approach, showing significant improvements in classification performance. Fine-tuning the linear classifier from the ImageNet pretrained model achieved an accuracy of 81.8% (Model-I), while fine-tuning the linear classifier from the pretrained model with unlabeled X-ray images reached an accuracy of 85.2% (Model-SSL). Integrating the nonlinear classifier onto Model-SSL further increased accuracy to 88.1%, representing improvements of +6.3% compared to Model-I and +2.9% for Model-SSL.