Can magnetic resonance imaging radiomics of the pancreas predict postoperative pancreatic fistula?

Objectives: To evaluate whether a magnetic resonance imaging (MRI) radiomics-based machine learning classifier can predict postoperative pancreatic fistula (POPF) after pancreaticoduodenectomy (PD) and to compare its performance to T1 signal intensity ratio (T1 SIratio). Methods: Sixty-two patients who underwent 3 T MRI before PD between 2008 and 2018 were retrospectively analyzed. POPF was graded and split into clinically relevant POPF (CR-POPF) vs. biochemical leak or no POPF. On T1- and T2-weighted images, 2 regions of interest were placed in the pancreatic corpus and cauda. 173 radiomics features were extracted using pyRadiomics. Additionally, the pancreas-to-muscle T1 SIratio was measured. The dataset was augmented and split into training (70 %) and test sets (30 %). A Boruta algorithm was used for feature reduction. For prediction of CR-POPF models were built using a gradient-boosted tree (GBT) and logistic regression from the radiomics features, T1 SIratio and a combination of the two. Diagnostic accuracy of the models was compared using areas under the receiver operating characteristics curve (AUCs). Results: Five most important radiomics features were identified for prediction of CR-POPF. A GBT using these features achieved an AUC of 0.82 (95 % Confidence Interval [CI]: 0.74 - 0.89) when applied on the original (nonaugmented) dataset. Using T1 SIratio, a GBT model resulted in an AUC of 0.75 (CI: 0.63 - 0.84) and a logistic regression model delivered an AUC of 0.75 (CI: 0.63 - 0.84). A GBT model combining radiomics features and T1 SIratio resulted in an AUC of 0.90 (CI 0.84 - 0.95). Conclusion: MRI-radiomics with routine sequences provides promising prediction of CR-POPF.