Fig. 3

Ultrasound images segmentation and HER2 status-related URFs extraction. a Workflow of DL model. First (step 1), training a deep learning segmentation model. Second (step 2), automatically identify tumor regions and extract radiomic features by the model. b Heatmap shows all radiomic features between positive and negative samples. Darker shades represent higher levels of features. (*stands for significantly differential feature) c Heatmap shows differential radiomic features between positive and negative samples. Violin plots d of significantly differential features in negative and positive samples. e Ultrasound and immunohistochemical images of patients with negative and positive HER2 status. Green and red, HER2-negative and positive images, respectively