Fig. 5

The Grad-CAM images that were generated based on EfficientNet-B0. A. A patient with nasopharyngeal carcinoma with clinical stage T4N2 was found to have tumor recurrence at 15 months after treatment (ground truth is high risk). A1 represents the original image of the patient. A2 to A8 represent Grad-CAM images generated by EfficientNet-B0 using the original image, expand 60 image, expand 40 image, expand 20 image, expand 10 image, expand 5 image, and Deeplab seg image, respectively. B. A patient with nasopharyngeal carcinoma with clinical stage T2N3 and no tumor recurrence at 43 months of follow-up (ground truth is low risk). B1 represents the original image of the patient. B2 to B8 represent Grad-CAM images generated by EfficientNet-B0 using the original image, expand 60 image, expand 40 image, expand 20 image, expand 10 image, expand 5 image, and Deeplab seg image, respectively. The yellow bright area indicates the region considered by the model to be most relevant to the prognosis of the tumor, followed by the green color. As it can be seen in the A2 and B2 plots that the DL model based on the original image classifies high- and low-risk patients on the basis of features that appear to be unreasonable in the physician’s experience as the bright yellow areas are concentrated around the brainstem. This situation occurs in a very large number of cases