Objective To investigate the value of multi-phase MRI-based radiomics machine learning models in differentiating small renal cell carcinoma (sRCC) from fat-poor renal angiomyolipoma (fp-AML).

Methods 79 cases of sRCCs and 35 cases of fp-AMLs (diameter ≤ 4 cm) which were confirmed by pathology were retrospectively analyzed. The volume of interest (VOI) of the total tumor was manually delineated on the images of T₂WI (T₂), unenhanced phase (UP), corticomedullary phase (CMP) and nephrographic phase (NP) and then the radiomics of the VOIs were extracted respectively. The training set and the test set were set according to the ratio of 7:3. The t-test, maximal relevance and minimal redundancy (mRMR) and the least absolute shrinkage and selection operator (LASSO) were used to select the radiomics features. The selected features were used to build classification models with logistic regression (LR) and support vector machine (SVM). The receiver operating characteristic (ROC) curve was used to evaluate the classification performances of the models.

Results There were 4, 12, 3, 11 and 15 optimal features obtained from T₂、UP、CMP、NP and the combined four phases, respectively. The radiomics features based on NP or the combined four phases with LR model performed best, AUCs were respectively 0.956, 0.986 in the training set and both were 0.881 in the test set.

Conclusion The multi-phase MRI-based radiomics machine learning model has favorable diagnostic performance in differentiating sRCC from fp-AML.