Proof-of-concept use of machine learning to predict tumor recurrence of early-stage hepatocellular carcinoma before therapy using baseline magnetic resonance imaging

Background and Aims: Patients with early-stage hepatocellular carcinoma (HCC) are likely to experience tumor recurrence after initial therapy using thermal ablation, surgical resection or orthotopic liver transplantation (OLT), with some reports demonstrating recurrence rates up to 80% five years post-treatment depending on therapy option. Successful identification of patients who will likely recur could help optimize follow-up imaging, influence choice of adjuvant therapy, and could even improve transplant allocation. Currently, nobiomarkers exist that can reliably predict recurrence. The combination of computational power and advanced machine learning-based algorithms allow in-depth exploration of imaging data to discover features predictive of recurrence.This study analyzed baseline magnetic resonance imaging (MRI) using state-of-the-art machine learning algorithms to predict HCC recurrence after first-line therapy with thermal ablation, surgical resection or OLT. Method: This was a HIPAA-compliant, IRB-approved retrospective study with 120 patients who underwent either thermal ablation, surgical resection or OLT as first-line,stand-alone treatment for HCC between 2005and 2018. Multiparametric contrast-enhanced MRI was analyzed by our machine learning model, which combined two readily available algorithms: VGG16 and XGBoost. The imaging dataset was labelled in six categories, with time-to-recurrence cutoffs at 1, 2, 3, 4, 5, or6 years. Areaunderthereceiveroperatingcharacteristiccurves (AUC-ROC) was used to evaluate algorithm performance.Recurrence-free survival (RFS)was evaluated by Kaplan-Meier analysis, and survival curves were compared using the log-rank test. Results: Of the 120 patients, 44 (36.7%) recurred at the time of analysis. Surgical resection(n = 32, 26.7%),thermal ablation (n = 29,24.2%), and OLT(n = 59, 49.1%) were the treatment modalities. With1, 2, 3, 4, 5 and 6 years as the time-to-recurrence cutoff, AUC-ROCvalues were, 0.74, 0.73, 0.71, 0.82, 0.71 and 0.79, respectively. Our algorithm was able to predict RFS with statistical significance at the 2, 4, 5, and 6-year time-to-recurrence cutoffs based on Kaplan-Meier analysis (log-rank p < 0.001). Conclusion: Our study shows that machine learning-based algorithms can predict tumor recurrence of early-stage HCC prior to therapy using baseline MRI. Future developments could improve model sensitivity using larger (and external) patient cohorts and including clinical variables. Figure: Kaplan-Meier analysis of Recurrence-free-survival (RFS) at each time-to-recurrence cutoff timepoints according to algorithm predictions. Each event on the green curve represents prediction inaccuracy (false-negative) and each event on the red curve represents an accurate prediction of recurrence (positive predictive value).