Scholarlink

In recent years, the utilization of deep learning in medical image diagnosis has advanced significantly, with convolutional neural networks (CNNs) playing a central role in computer-aided diagnosis (CAD). However, the lack of transparency in the decision-making criteria of deep learning models poses a challenge for clinical applications. In this study, three explainability methods—-Grad-CAM, SHAP, and LIME—are applied to analyze the decision basis of CNNs, and the output results of each method are compared. With Grad-CAM, clear differences in predictions have been observed between “No Finding” and “Pneumoconiosis” images. SHAP shows a tendency for pixel concentrations mainly around the contours of the lung fields, while LIME displays superpixels concentrated primarily along the contours of the lung fields. These results suggest that the model may rely on lung shape and size rather than parenchymal fibrosis. Furthermore, a two-sided t-test was conducted to compare the black-pixel (non-lung) ratio between the “No Finding” and “Pneumoconiosis” classes. The p-value is 0.0002, indicating that the black-pixel ratio in the “Pneumoconiosis” class is significantly lower than in the “No Finding” class (i.e., a larger lung area), suggesting that dataset composition may bias the model’s decisions.

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

Comparison of Explainability Algorithms for Pneumoconiosis Classification Deep Learning Models

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Ryo Takeda,
  • Yoshua Kazukuni Nomura,
  • Narufumi Suganuma,
  • Shinichi Yoshida

摘要

In recent years, the utilization of deep learning in medical image diagnosis has advanced significantly, with convolutional neural networks (CNNs) playing a central role in computer-aided diagnosis (CAD). However, the lack of transparency in the decision-making criteria of deep learning models poses a challenge for clinical applications. In this study, three explainability methods—-Grad-CAM, SHAP, and LIME—are applied to analyze the decision basis of CNNs, and the output results of each method are compared. With Grad-CAM, clear differences in predictions have been observed between “No Finding” and “Pneumoconiosis” images. SHAP shows a tendency for pixel concentrations mainly around the contours of the lung fields, while LIME displays superpixels concentrated primarily along the contours of the lung fields. These results suggest that the model may rely on lung shape and size rather than parenchymal fibrosis. Furthermore, a two-sided t-test was conducted to compare the black-pixel (non-lung) ratio between the “No Finding” and “Pneumoconiosis” classes. The p-value is 0.0002, indicating that the black-pixel ratio in the “Pneumoconiosis” class is significantly lower than in the “No Finding” class (i.e., a larger lung area), suggesting that dataset composition may bias the model’s decisions.