Introduction: Hepatocellular carcinoma (HCC) is a major global health burden, ranking as the third leading cause of cancer-related deaths worldwide. Despite liver resection being a curative option in the Asia-Pacific region, more than half of the patients experience recurrence within five years post-surgery. Current methods for predicting recurrence, including histological microvascular invasion (MVI) and clinical scoring systems, are inadequate for preoperative prognostication. Dr Wai-Kay Seto's team at The University of Hong Kong, China, in collaboration with experts from The Education University of Hong Kong,China, and Queen Elizabeth Hospital Hong Kong, China, performed a large multicenter study to explore the use of deep learning in HCC prognostication. The team trained, validated, and externally tested a novel multimodal deep learning model – Recurr-NET, which integrates preoperative imaging and clinical data to predict HCC recurrence.Study Design
This multicenter study involved 1,231 patients with histology-confirmed HCC who underwent liver resection. The internal cohort, comprising patients from four centers in Chinese Hong Kong, was divided into training (80%) and internal validation (20%) sets, while the external testing cohort was from a center in Chinese Taiwan. Inclusion criteria required preoperative triphasic CT scans within two months of surgery. Patients with incomplete imaging, non-HCC diagnoses, or poor-quality scans were excluded. The study aimed to validate Recurr-NET, a multimodal deep learning model, against existing histological and clinical prediction methods.
Methods
The development of Recurr-NET leveraged triphasic CT imaging and comprehensive clinical data. The model employed a residual network (ResNet) structure and a random survival forest (RSF) to extract granular imaging features, which were combined with clinical parameters. Three variations of the model were created: Recurr-NETCT, using CT images only; Recurr-NETLITE, incorporating basic clinical parameters; and the full Recurr-NET, integrating all available data. Statistical analyses included AUROC evaluations, Kaplan-Meier survival analysis, and subgroup analyses to assess model performance and stratification capabilities.
Results
Recurr-NET demonstrated significantly superior predictive accuracy compared to MVI and established clinical scores. In the internal validation cohort, Recurr-NET achieved AUROC values ranging from 0.770 to 0.857, outperforming MVI (0.518–0.590) and clinical scores like ERASL and Shim (0.523–0.587). Similar results were observed in the external testing cohort, with AUROC values of 0.758–0.798 for Recurr-NET versus 0.557–0.615 for MVI. The model also accurately stratified recurrence risk, predicting 2-year recurrence probabilities of 72.5% (internal) and 65.3% (external), compared to MVI’s 50.0% and 46.6%. At five years, Recurr-NET predicted recurrence rates of 86.4% (internal) and 81.4% (external), compared to 62.5% and 63.8% for MVI.
Discussion
The model’s ability to predict mortality risk further highlights its clinical utility. Recurr-NET identified high-risk patients with significantly greater liver-related mortality at two years (28.3% vs. 11.8% for MVI) and five years (69.1% vs. 29.9%). It consistently outperformed MVI in subgroup analyses across age, viral hepatitis status, cirrhosis, and steatosis categories. Importantly, the model maintained robust performance in external validation, ensuring generalizability across diverse patient populations. By integrating imaging and clinical data, Recurr-NET addresses both early recurrence driven by tumor characteristics and late recurrence influenced by patient factors. Unlike MVI, which relies on postoperative histology, Recurr-NET provides actionable preoperative prognostication, enabling tailored surveillance and therapeutic strategies.
Conclusion
Recurr-NET represents a significant advancement in the preoperative management of HCC. Its preoperative prognostication capabilities offer a new paradigm in managing HCC, particularly in selecting candidates for neoadjuvant therapies or liver transplantation. Its reliance on standard CT imaging and clinical data ensures cost-effective and practical implementation. Future developments, such as an online calculator for wider accessibility, could further enhance its clinical adoption. As the first deep learning model to incorporate survival analysis for HCC recurrence and mortality, Recurr-NET sets a new standard for leveraging AI in cancer management.
