Scholarlink

<p>With the growing use of deep learning models in medical imaging, it is critical to interpret these models to ensure precise diagnoses. A recent study using the SARS-COV2 dataset evaluated the efficiency of several Explainable Artificial Intelligence (XAI) techniques—LIME, GradCAM, SHAP, and GradCAM++—on deep learning models for detecting anomalies in COVID-19 CT scans. The study assessed five models based on accuracy, precision, recall, and F1-score. ResNet50 outperformed other models, achieving the highest accuracy of 95.26%, along with the best precision (96.24%), recall (94.14%), and F1-score (96.17%). The study further compared the effectiveness of different XAI techniques applied to ResNet50. Among the techniques, GradCAM++ provided the best interpretability, achieving a confidence score of 99.41% and the lowest prediction entropy of 5.21, outperforming GradCAM, LIME, and SHAP. GradCAM achieved a confidence score of 98.82% with a prediction entropy of 9.25, while LIME and SHAP showed lower confidence scores of 95.23% and 78.91% respectively, with higher prediction entropy values. The attribution maps revealed that the models focused on different regions of the CT scans. The quantitative analysis highlighted significant differences in model performance and XAI technique efficiency. These findings can guide the selection of suitable XAI techniques and deep learning models for similar applications in the future.</p>

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

Quantitative Assessment of XAI Methods for COVID-19 Detection: A Comparative Approach

  • 期刊论文

  • 发表日期：

    2025年1月30日

查看全文
  • Reenu Rajpoot,
  • Sweta Jain,
  • Vijay Bhaskar Semwal,
  • Deepankar Singh

摘要

With the growing use of deep learning models in medical imaging, it is critical to interpret these models to ensure precise diagnoses. A recent study using the SARS-COV2 dataset evaluated the efficiency of several Explainable Artificial Intelligence (XAI) techniques—LIME, GradCAM, SHAP, and GradCAM++—on deep learning models for detecting anomalies in COVID-19 CT scans. The study assessed five models based on accuracy, precision, recall, and F1-score. ResNet50 outperformed other models, achieving the highest accuracy of 95.26%, along with the best precision (96.24%), recall (94.14%), and F1-score (96.17%). The study further compared the effectiveness of different XAI techniques applied to ResNet50. Among the techniques, GradCAM++ provided the best interpretability, achieving a confidence score of 99.41% and the lowest prediction entropy of 5.21, outperforming GradCAM, LIME, and SHAP. GradCAM achieved a confidence score of 98.82% with a prediction entropy of 9.25, while LIME and SHAP showed lower confidence scores of 95.23% and 78.91% respectively, with higher prediction entropy values. The attribution maps revealed that the models focused on different regions of the CT scans. The quantitative analysis highlighted significant differences in model performance and XAI technique efficiency. These findings can guide the selection of suitable XAI techniques and deep learning models for similar applications in the future.