Figure 6.

Attention heatmap drawn by gradient-weighted class activation mapping for the model interpretation. (A) CT images, tumor segmentations, and corresponding attention heatmaps in a 60-year-old woman with low-malignant GIST (the first column is arterial phase, the second column is venous phase, and the third column is delayed phase). (B) CT images, tumor segmentations, and corresponding attention heatmaps in a 66-year-old woman with intermediate-malignant GIST. (C) CT images, tumor segmentations, and corresponding attention heatmaps in a 43-year-old woman with high-malignant GIST. The red and yellow regions represent areas activated by the DLM and have the greatest predictive significance; the blue backgrounds reflect areas with weaker predictive values. GIST, gastrointestinal stromal tumor; DLM, deep learning model.