Fig. 1

Design of the study. A: The process of building semantic segmentation model. a: Manual segmentation of tumor and suspicious lymph node regions in original images. b: Training Deeplab v3 and U-Net using manually segmented datasets. c: Automatic segmentation of datasets for classification using the trained semantic segmentation model. d. Extension by 5, 10, 20, 40, and 60 pixels outward from the segmented region to form 5 new datasets. B: The process of building classification model for predicting tumor prognosis