Precision Recurrence Risk Assessment in Early-stage Hepatocellular Carcinoma

Multimodal Deep Learning Models for Predicting Recurrence Pattern in Hepatocellular Carcinoma: A Multicenter Retrospective Development and Validation Study

This retrospective observational study aims to evaluate whether artificial intelligence (AI) models can predict aggressive recurrence in patients who underwent liver resection for early-stage hepatocellular carcinoma (HCC). The main question it seeks to answer is:

Can deep learning models combining preoperative MRI, postoperative pathology slides, and clinical data accurately identify HCC patients at high risk of aggressive recurrence after surgery?

To answer this, the investigators will analyze existing medical data (preoperative MRIs, postoperative whole-slide images, and clinical records) from 579 patients across two medical centers. All data will be anonymized before analysis, and no additional interventions are required from participants.

This study may help clinicians stratify high-risk patients who could benefit from closer surveillance or adjuvant therapies

Study Overview

• Status: Completed
• Conditions: Hepatocellular Carcinoma (HCC)
• Intervention / Treatment: Procedure: liver resection

• Study Type: Observational
• Enrollment (Actual): 579

Contacts and Locations

Study Locations

• China
  • Hubei
    • Wuhan, Hubei, China, 430030
      • Tongji Hospital

Participation Criteria

Eligibility Criteria

• Ages Eligible for Study: Child; Adult; Older Adult
• Accepts Healthy Volunteers: No

• Sampling Method: Non-Probability Sample

Study Population

This retrospective multicenter study analyzed 579 patients with early-stage hepatocellular carcinoma (HCC) who underwent curative liver resection at two tertiary academic medical centers in China. The study population consisted of:

Primary Cohort (Training/Validation):

462 patients from Tongji Hospital (2018-2021)

External Test Cohort:

117 patients from Sun Yat-sen Memorial Hospital (2021-2022) All patients met strict inclusion criteria: curative (R0) resection, preoperative MRI within 1 month before surgery, available postoperative pathology slides, and complete follow-up. The population represents typical early-stage HCC patients eligible for surgical resection in endemic areas.

Description

Inclusion Criteria:

• Patients who underwent curative liver resection (R0) for pathologically confirmed primary HCC
• BCLC stage 0-A at diagnosis
• Availability of preoperative contrast-enhanced MRI performed within 1 month before surgery
• Availability of postoperative H&E-stained whole slide images (WSIs) with adequate tumor representation
• Complete clinical follow-up data (minimum 2 years if no recurrence)

Exclusion Criteria:

• R1/R2 resection (micro/macroscopically positive margins)
• Missing or poor-quality preoperative MRI (motion artifacts/insufficient contrast enhancement)
• Received neoadjuvant or adjuvant therapy (to avoid treatment confounding)
• Incomplete follow-up (loss to follow-up or missing recurrence status)
• Non-curative procedures (e.g., palliative resection)

Study Plan

How is the study designed?

Number of groups / cohorts

2

Cohorts and Interventions

Group / Cohort	Intervention / Treatment
TJ Cohort (Training/Validation); Internal cohort from Tongji Hospital (2018-2021) used for model training and validation. Includes 462 patients with early-stage HCC who underwent curative resection. Data: preoperative MRI, clinical variables, and postoperative pathology slides. No interventions beyond standard care.	Procedure: liver resection; This is a retrospective observational study analyzing existing clinical data; no experimental interventions were administered. The study evaluates the predictive performance of two deep learning models (preoperative and postoperative) using standard-of-care medical data collected during routine clinical practice, including: Preoperative contrast-enhanced MRI scans Postoperative hematoxylin and eosin (H&E)-stained whole slide images Clinical variables (laboratory results, pathology reports, and demographic data) All data were collected as part of standard diagnostic and treatment protocols for hepatocellular carcinoma (HCC) patients undergoing liver resection. No additional interventions or modifications to clinical care were implemented for study purposes. The artificial intelligence models were applied to previously acquired, de-identified data to predict aggressive recurrence patterns
SYSMH Cohort (External Test); Independent external test cohort from Sun Yat-sen Memorial Hospital (2021-2022). Includes 117 patients with early-stage HCC meeting identical inclusion criteria. Used to validate generalizability of multimodal DL models. Data anonymized; no additional interventions.	Procedure: liver resection; This is a retrospective observational study analyzing existing clinical data; no experimental interventions were administered. The study evaluates the predictive performance of two deep learning models (preoperative and postoperative) using standard-of-care medical data collected during routine clinical practice, including: Preoperative contrast-enhanced MRI scans Postoperative hematoxylin and eosin (H&E)-stained whole slide images Clinical variables (laboratory results, pathology reports, and demographic data) All data were collected as part of standard diagnostic and treatment protocols for hepatocellular carcinoma (HCC) patients undergoing liver resection. No additional interventions or modifications to clinical care were implemented for study purposes. The artificial intelligence models were applied to previously acquired, de-identified data to predict aggressive recurrence patterns

What is the study measuring?

Primary Outcome Measures

Outcome Measure	Measure Description	Time Frame
Aggressive Recurrence Pattern	Defined as first recurrence exceeding Milan criteria within 2 years after liver resection.	2 years after surgery

Secondary Outcome Measures

Outcome Measure	Measure Description	Time Frame
Recurrence-Free Survival (RFS)	Time from surgery date to radiologically confirmed recurrence or last follow-up (until July 30, 2024).	From surgery until first recurrence or July 30, 2024
Overall Survival (OS)	Time from surgery date to death from any cause or last follow-up.	From surgery until death or July 30, 2024

Collaborators and Investigators

• Sponsor: Tongji Hospital
• Collaborators: Sun Yat-Sen Memorial Hospital of Sun Yat-Sen University

Study record dates

Study Major Dates

• Study Start (Actual): August 25, 2023
• Primary Completion (Actual): July 30, 2024
• Study Completion (Actual): July 30, 2024

Study Registration Dates

• First Submitted: June 12, 2025
• First Submitted That Met QC Criteria: June 12, 2025
• First Posted (Actual): June 22, 2025

Study Record Updates

• Last Update Posted (Actual): June 29, 2025
• Last Update Submitted That Met QC Criteria: June 27, 2025
• Last Verified: June 1, 2025

More Information

Terms related to this study

• Keywords: artificial intelligence; deep learning; multimodal; recurrence pattern
• Additional Relevant MeSH Terms: Pathologic Processes; Neoplasms by Site; Neoplasms; Disease Attributes; Neoplasms by Histologic Type; Digestive System Neoplasms; Digestive System Diseases; Liver Diseases; Neoplasms, Glandular and Epithelial; Adenocarcinoma; Liver Neoplasms; Recurrence; Carcinoma; Carcinoma, Hepatocellular
• Other Study ID Numbers: TJ-IRB20230863

Drug and device information, study documents

• Studies a U.S. FDA-regulated drug product: No
• Studies a U.S. FDA-regulated device product: No