Abstract
RECURR-NET, A MULTIPHASIC DEEP LEARNING MODEL, IS SUPERIOR TO MICROVASCULAR INVASION IN PREDICTING HEPATOCELLULAR CARCINOMA RECURRENCE AFTER CURATIVE SURGERY: RESULTS FROM INTERNAL VALIDATION AND EXTERNAL TESTING
Background:
Recurrence can occur in over 70% of hepatocellular carcinoma (HCC) patients within 5 years after curative resection. While histological microvascular invasion (MVI) predicts recurrence, it is ascertained from resected specimens and cannot provide pre-operative prognostication. We developed an artificial intelligence deep learning-based model using pre-operative computed tomography (CT) for predicting HCC recurrence.
Methods:
Chinese patients with resected histology-confirmed HCC were recruited from four centers in Hong Kong, and were randomly divided in an 8:2 ratio into training and internal validation groups. We developed Recurr-Net, a multi-phasic residual-network random survival forest deep learning model, incorporating pre-operative triphasic contrast CT images and clinical data (sex, age, comorbidities, blood tests) to predict HCC recurrence. The model was externally tested in an independent cohort from Taiwan. The area-under-curve (AUC), positive- and negative-predictive values (PPV/ NPV) of the model was compared against MVI. Survival analysis was also performed.
Results:
This analysis included 1,254 patients (82.9% male, age at CT 62.2 +/- 10.8 years, median follow-up 7.8 [5.8-10.0] years). 551 (43.9%), 140 (11.2%) and 563 (44.9%) patients were in the training, internal validation, and external testing cohorts respectively. The cumulative HCC recurrence rate at years 2 and 5 were 42.1% and 56.6% respectively. The model was trained for 42 epochs. In the internal validation cohort, Recurr-NET achieved an AUC of 0.823 (95%CI 0.684-0.842, PPV 0.778, NPV 0.753) and 0.801 (95%CI 0.632-0.815, PPV 0.857, NPV 0.547) for predicting recurrence at years 2 and 5, significantly outperforming the predictive value of MVI (year 2 AUC 0.564 [95%CI 0.484-0.653], PPV 0.591, NPV 0.554; year 5 AUC 0.527 [95%CI 0.437-0.624], PPV 0.690, NPV 0.365) (Both p<0.01). In the external testing cohort, Recurr-NET achieved an AUC of 0.787 (95%CI 0.647-0.801, PPV 0.743, NPV 0.755) and 0.753 (95%CI 0.607-0.765, PPV 0.847, NPV 0.507) for predicting recurrence at years 2 and 5, significantly outperforming MVI (year 2 AUC 0.609, 95%CI 0.535-0.682, PPV 0.478, NPV 0.774; year 5 AUC 0.560, 95%CI 0.483-0.642, PPV 0.679, NPV 0.464) (Both p<0.05). In both internal validation and external testing cohorts, Recurr-NET had significantly better discriminative ability than MVI for 2-year (Figures 1A & 1C; Validation: 77.8% vs 59.1%; External testing: 74.3% vs 47.8%) and 5-year recurrence risks (Figures 1B & 1D; Validation 85.7% vs 69.0%; External testing: 84.7% vs 67.9%) (Both p<0.001).
Conclusion:
Recurr-NET was superior to MVI in predicting early and late HCC recurrence, and has potential to emerge as a novel tool for pre-operative prognostication of HCC outcomes.
Funding:
Health and Medical Research Fund, Hong Kong (Ref no: 07182346) and the General Research Fund, Research Grant Council (Ref no: 17100522).
Related Speaker and Session
Rex Wan-Hin Hui, Department of Medicine, School of Clinical Medicine, the University of Hong KongDate: Sunday, November 12th
Time: 4:30 - 6:00 PM EST