| XAI Method | Characteristics | Description of Pneumonia CXR Image | Spatial Resolution |
| LIME | *A method of approximating the prediction result of a black box model with another model. *Approximate classifiers are not always correct. | *Multiple lesions (GGO, consolidation, etc.) in the image can be effectively visualized. *Highlight which superpixel region of the image is most important for classification. | *The spatial resolution is adjusted according to the number of superpixels, allowing for a high degree of flexibility in feature investigation. |
| Occlusion Sensitivity | *A method of making data partially obstructive and measuring the importance of a region using a black box model. *The importance of combination features cannot be expressed. | *Multiple lesions (GGO, consolidation, etc.) in the image can be visualized. *Highlight which region of the image is most important for classification. | *The spatial resolution is adjusted according to the mask size and the size of the stride, allowing for a high degree of flexibility in feature investigation. |
| Grad-CAM | *A method to make the black box model itself have an evidence for judgment. *Describes the regions that affects the final score. | *Focuses on a wide range of a major lesion and therefore cannot identify (visualize) multiple lesions in the image [26] . *Highlight the pixels that contribute to change the final decision. | *The spatial resolution of the feature map in the final convolution layer is low because it is 7 × 7 (in the case of ResNet-50). |