Fig. 2: Diagnostic accuracy of the deep learning model to differentiate malignant from benign renal masses.

A Representative CT images of a benign renal mass. B Automatic segmentation of the renal mass and kidney. C Class activation maps (CAMs) of the benign and malignant prediction for the benign renal mass, where red color represents a region more significant to the designated classification. D Representative CT images of a malignant renal mass. E Automatic segmentation of the renal mass and kidney. F CAMs of the benign and malignant prediction for the malignant renal mass. G The receiver operating characteristic (ROC) curves of the deep learning-based model, the radiomics model, and the nephrometry score nomogram on the internal test set. H The ROC curves of the deep learning-based model, the radiomics model, and the nephrometry score nomogram on the external test set. I The ROC curves of the deep learning-based model, the radiomics model, and the nephrometry score nomogram on prospective test set. J The ROC curves of the deep learning-based model and nephrometry score nomogram on TCIA test set. K The ROC curves of the deep learning-based model and performance of seven radiologists on prospective validation set. L The diagnostic accuracy of seven radiologists with (n = 7) or without (n = 7) the assistant of the deep learning-based model. M Clustered observer similarity matrix of the deep learning-based model and seven radiologists. N The diagnostic efficiency of seven radiologists with (n = 7) or without (n = 7) the assistant of the deep learning-based model. In Fig. 2B, E yellow areas represent renal lesions identified by the nnU-Net model, while blue areas represent the kidney tissue also identified by the nnU-Net model. In Fig. 2C, F, the color intensity represents the level of importance for the deep learning model in making its decision. Red (designated as point 1) indicates the most important regions, green (designated as point 0.5) signifies regions of lesser importance, and blue (designated as point 0) denotes regions that are not important for the model’s decision. Source data for ROC curves are provided as a Source Data file.