Predicting the presence of adherent perinephric fat using MRI radiomics combined with machine learning

•Adherent perinephric fat presents a major surgical challenge in renal tumor removal.•Preoperative prediction of this condition has a significant clinical benefit.•The current methods of evaluation are qualitative and subject to readers.•Machine learning trained with MRI data can provide a quantitative method. Adherent perinephric fat (APF) poses significant challenges to surgical procedures. This study aimed to evaluate the usefulness of machine learning algorithms combined with MRI-based radiomics features for predicting the presence of APF. Patients with renal cell carcinoma who underwent surgery between April 2019 and February 2022 at Chonnam National University Hwasun Hospital were retrospectively screened, and 119 patients included. Twenty-one and seventeen patients were set aside for the internal and external test sets, respectively. Pre-operative T1-weighted MRI acquired at 60 s following a contrast injection (T1w-60) were collected. For each T1w-60 data, two regions of interest (ROIs) were manually drawn: the perinephric fat tissue and an aorta segment on the same level as the targeted kidney. Preprocessing steps included resizing voxels, N4 Bias Correction filtering, and aorta-based normalization. For each patient, 851 radiomics features were extracted from the ROI of perinephric fat tissue. Gender and BMI were added as clinical factors. Least Absolute Shrinkage and Selection Operator was adopted for feature selection. We trained and evaluated five models using a 4-fold cross validation. The final model was chosen based on the highest mean AUC across four folds. The performance of the final model was evaluated on the internal and external test sets. A total of 15 features were selected in the final set. The final model achieved the accuracy, sensitivity, specificity, and AUC of 81% (95% confidence interval, 61.9–95.2%), 72.7% (42.9–100%), 90% (66.7–100%), and 0.855 (0.615–1.0), respectively on the internal test set, and 88.2% (70.6–100%), 100% (100–100%), 80% (50%–100%), 0.971 (0.871–1.0), respectively on the external test set. Our study demonstrated the feasibility of machine learning algorithms trained with MRI-based radiomics features for APF prediction. Further studies with a multi-center approach are necessary to validate our findings.