Discriminating malignant from benign testicular masses using machine-learning based radiomics signature of appearance diffusion coefficient maps: Comparing with conventional mean and minimum ADC values.

PURPOSE: To develop a machine-learning-based radiomics signature of ADC for discriminating between benign and malignant testicular masses and compare its classification performance with that of minimum and mean ADC. METHODS: A total of ninety-seven patients with 101 histopathologically confirmed testicular masses (70 malignancies, 31 benignities) were evaluated in this retrospective study. Eight hundred fifty-one radiomics features were extracted from the preoperative ADC map of each lesion. The mean and minimum ADC values are part of the radiomics features. Thirty lesions were randomly selected to estimate the reliability of the features. The redundant features were eliminated using univariate analysis (independent t test and Mann-Whitney U test, where appropriate) and Spearman's rank correlation. The least absolute shrinkage and selection operator (LASSO) algorithm was employed for feature selection and radiomics signature generation. The classification performance of the radiomics signature and minimum and mean ADC values were evaluated by receiver operating characteristic (ROC) curve analysis and compared by DeLong's test. RESULTS: The whole lesion-based mean ADC showed no difference between benign and malignant testicular masses (P = 0.070, training cohort; P = 0.418, validation cohort). Compared with the minimum ADC, the ADC-based radiomics signature yielded a higher area under the curve (AUC) in both the training (AUC: 0.904, 95% confidence interval [CI]: 0.832-0.975) and validation cohorts (AUC: 0.868, 95% CI: 0.728-1.00). CONCLUSIONS: Conventional mean ADC values are not always helpful in discriminating between testicular benignities and malignancies. The minimum ADC and radiomics signature might be better alternatives, with the radiomics signature performing better than the minimum ADC.