Multiparametric MRI-Based Machine Learning Radiomics Prognostic Models for Multifocal Hepatocellular Carcinoma Beyond Milan Criteria: A Retrospective Study.

Abstract

Purpose: To develop machine learning radiomics models for preoperative risk stratification of multifocal hepatocellular carcinoma (MHCC) beyond Milan criteria.

Methods: Patients with pathologically proven MHCC beyond Milan criteria between January 2015 and January 2019 were retrospectively included. Radiomic features were extracted from tumor, peritumor, and tumor-peritumor regions using multiparametric MRI (mpMRI). An unsupervised spectral clustering algorithm was used to identify radiomics-based patient subtypes. Radiomics risk scores (RRS) for overall survival (OS) and recurrence-free survival (RFS) were generated using supervised extreme gradient boosting (XGBoost)-LASSO Cox proportional hazard regression analysis. The Concordance index (C-Index) was used to evaluate the model performance in the training and validation sets.

Results: A total of 156 patients were divided into training (n = 78) and validation (n = 78) groups. Two distinct unsupervised subtypes were identified using spectral clustering, and subtype B was associated with worse OS and poor RFS. Incorporating radiomics predictors into the conventional preoperative clinical-radiological features improved the OS prediction performance (training set: from 0.616 to 0.712; validation set: from 0.522 to 0.710), and RFS prediction (training set: from 0.653 to 0.735; validation set: from 0.574 to 0.698). The combined models showed good predictive performance for 5-year OS (AUC, 0.77) and RFS (AUC, 0.81) in the training set and for 5-year OS (AUC, 0.75) and RFS (AUC, 0.76) in the validation set.

Conclusion: Two preoperative models combining mpMRI-based clinico-radiological and radiomics predictors effectively predicted outcomes for patients with MHCC beyond the Milan criteria.