Fig. 4

Unsupervised and supervised analysis results of the deep radiomic features (DRFs). a Unsupervised hierarchical clustering analysis of the DRFs. Columns are the 110 patients; rows are the 4,096 DRFs. Clinical information is shown in the sidebar. T refers to the tumor size. For breast tumors, bigger than 2 cm are T-positive. N refers to node status, which is positive when the tumor cell spreads into lymph nodes. ER, PR, HER2 refer to estrogen receptor status, progesterone receptor status, and human epidermal growth factor receptor 2 status. Patients seem to be clustered into 2 groups, but these two groups have no obvious clinical difference. b t-SNE visualizes the patient-level DRFs. Each dot is one patient. Different colors are marked in different patient-level clusters manually. c t-SNE visualizes the image-level DRFs. Each dot is one image. We first tracked the dots at image-level t-SNE map to patient-level, and then colored them using the same colors as what we used in coloring the patient-level t-SNE map. d The supervised LASSO model prediction performance of deep radiomic features under different \(\lambda\) s. Different colors represent different clinical characteristics. The x-axis represents the number of deep radiomics features given different \(\lambda\) in the LASSO models. e The supervised LASSO model prediction performance of traditional radiomic features under different \(\lambda\) s. Different colors represent different clinical characteristics. The x-axis represents the number of traditional radiomic features given different \(\lambda\) in the LASSO models. Please note that the feature number is not going up to the total number of features (4,096 or 36) because there were always a lot of features been regularized out under different \(\lambda\) s. The y-axis represents the corresponding area under the curve (AUC) which is a metric used to assess the performance of the prediction. An AUC equals to 1 means a perfect prediction