Scholarlink

In recent years, there has been a rapid increase in interest regarding the widespread application of Convolutional Neural Networks (CNNs) in Computer-Aided Diagnosis (CAD) and image-based diagnosis. However, CNN-based diagnostic approaches still face numerous challenges in terms of interpretability. Previous studies have proposed the use of CycleGAN to analyze the classification processes of CNNs, suggesting its potential to enhance interpretability. CycleGAN is characterized by its ability to transform specific parts of an image without altering the background, allowing it to capture more detailed information such as differences in shapes and patterns within regions, compared to traditional methods like Grad-CAM. This study aims to apply CycleGAN to the disease pneumoconiosis to visualize which parts of the image the classification model focuses on when making its determinations. The results reveal that the brightness across the entire lung field changes before and after transformation, suggesting that the classification model may be focusing on the degree of brightness within the lung region when identifying pneumoconiosis.

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

Explainability of CNN Classification Models Using CycleGAN and Their Application to Medical Imaging

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Taiga Nakajima,
  • Yoshua Kazukuni Nomura,
  • Narufumi Suganuma,
  • Shinichi Yoshida

摘要

In recent years, there has been a rapid increase in interest regarding the widespread application of Convolutional Neural Networks (CNNs) in Computer-Aided Diagnosis (CAD) and image-based diagnosis. However, CNN-based diagnostic approaches still face numerous challenges in terms of interpretability. Previous studies have proposed the use of CycleGAN to analyze the classification processes of CNNs, suggesting its potential to enhance interpretability. CycleGAN is characterized by its ability to transform specific parts of an image without altering the background, allowing it to capture more detailed information such as differences in shapes and patterns within regions, compared to traditional methods like Grad-CAM. This study aims to apply CycleGAN to the disease pneumoconiosis to visualize which parts of the image the classification model focuses on when making its determinations. The results reveal that the brightness across the entire lung field changes before and after transformation, suggesting that the classification model may be focusing on the degree of brightness within the lung region when identifying pneumoconiosis.