Comparison of different machine learning methods in the prediction of early recurrence in HCC patients with Gd-EOB-DTPA-MRI.

Purpose: To develop machine learning models that are driven by Gd-EOB-DTPA-MRI features for the preoperative prediction of early recurrence in HCC and compare them to the previously proposed ERASL-pre method.

Methods: This retrospective study consisted of 311 consecutive patients who underwent curative hepatic resection between January 2013 and July 2021. Among them, 131 patients with early recurrence of HCC and 180 patients without early recurrence of HCC. The MR images were independently reviewed by two radiologists. Logistic regression, classification tree, random forest, extreme gradient boosting (XGBoost), support vector machines (SVM), and neural network (Nnet) were the six machine learning algorithms used. The baseline model was ERASL-pre. Different models' discrimination, calibration, and overall performance were evaluated and compared.

Results: The baseline ERASL-pre obtained AUCs of 0.703 and 0.716, respectively. In comparison to ERASL-pre, the AUCs for logistic regression, random forest, and SVM were higher but not substantially different. XGBoost produced AUCs for Readers 1 and 2 of 0.720 and 0.685, respectively. Nnet achieved marginally lower but not statistically different AUCs in comparison to ERASL-pre, whereas the classification tree achieved the lowest AUCs. The logistic regression model had the optimal overall net benefit across the majority of the range of reasonable threshold probabilities. Good agreement was observed between prediction and observation in the ERASL-pre and the logistic regression model for both readers.

Conclusion: The logistic regression model performed better in predicting an early recurrence of HCC with Gd-EOB-DTPA MRI. In addition, the model is more sensitive than the baseline ERASL-pre model.