AI-Guided Surveillance and Curative Salvage of Recurrence After Hepatocellular Carcinoma Resection or Ablation

Overview

Hepatocellular carcinoma, the most common type of primary liver cancer, can recur after liver resection or thermal ablation performed with curative intent. Follow-up imaging is commonly scheduled at fixed intervals, although the risk of recurrence differs among patients and may change over time. This study will test whether a locked artificial intelligence system can use routinely collected clinical, laboratory, and imaging information to recommend when the next surveillance scan should occur. Adults with no radiological evidence of viable hepatocellular carcinoma after microscopically margin-negative liver resection or complete radiofrequency or microwave ablation will be randomly assigned in a 1:1 ratio to artificial intelligence-guided risk-adapted surveillance or fixed-interval surveillance. In the artificial intelligence-guided group, participants classified as having high, intermediate, or low current recurrence risk will generally undergo the next protocol-scheduled contrast-enhanced imaging examination after 3, 4, or 6 months, respectively. Participants in the control group will undergo protocol-scheduled imaging every 4 months. The risk thresholds are designed so that the expected total number of protocol-scheduled imaging examinations is approximately comparable between the two groups over 24 months. The artificial intelligence system provides surveillance recommendations only. It does not diagnose recurrence, determine eligibility for liver transplantation, or select anticancer treatment. Clinically indicated examinations may be performed at any time in either group. When recurrence is confirmed, participants in both groups will undergo the same protocol-defined multidisciplinary evaluation, and potentially curative-intent treatment may be considered when clinically appropriate. The primary purpose is to determine whether artificial intelligence-guided surveillance increases the probability of being alive at 24 months without having a recurrence that is no longer amenable to protocol-defined curative-intent treatment. The study evaluates the timing and allocation of surveillance rather than an adjuvant anticancer treatment and is not intended to prevent the biological occurrence of recurrence.

Recurrence is common after curative-intent treatment of hepatocellular carcinoma. Surveillance after liver resection or thermal ablation is therefore an essential component of long-term care. However, conventional surveillance commonly applies similar imaging intervals to patients whose recurrence risks differ substantially and may change over time. A risk-adapted strategy could allocate surveillance examinations to periods when timely detection is most likely to preserve an opportunity for curative-intent management, while avoiding unnecessary examinations during lower-risk periods. This is a pragmatic, multicenter, open-label, parallel-group, patient-level randomized superiority trial with blinded central outcome assessment. Eligible participants will have undergone curative-intent microscopically margin-negative (R0) liver resection or complete radiofrequency or microwave ablation for hepatocellular carcinoma. Before randomization, protocol-specified contrast-enhanced computed tomography or magnetic resonance imaging must demonstrate no viable residual tumor. Complete response after thermal ablation will be established radiologically; absence of viable enhancement is not considered evidence that microscopic disease is absent. Participants will be randomly assigned in a 1:1 ratio to an artificial intelligence-guided risk-adapted surveillance strategy or a fixed-interval surveillance strategy. Randomization will be implemented centrally and stratified by study center and initial curative-intent treatment modality. Participants and treating clinicians cannot be masked to surveillance timing. Independent central reviewers and the committee adjudicating the primary outcome will remain unaware of randomized assignment. In the artificial intelligence-guided group, a locked artificial intelligence system will use protocol-defined baseline and longitudinal clinical, laboratory, and imaging information to estimate the participant's current risk of recurrence. Separate modality-specific components will be used for participants treated by resection and those treated by thermal ablation. Resection-specific pathological information will not be assumed to be available for participants treated by ablation. At each protocol-defined reassessment, participants classified as having high, intermediate, or low current risk will generally have their next protocol-scheduled contrast-enhanced imaging examination arranged after 3, 4, or 6 months, respectively. Prespecified changes in clinical status, tumor-marker trajectories, or other protocol-defined findings may trigger expedited diagnostic evaluation. Symptoms or clinically concerning findings must be evaluated irrespective of the artificial intelligence score or scheduled interval. Artificial intelligence outputs are advisory and may be overridden by the treating clinician for documented clinical reasons. Participants assigned to the control group will undergo contrast-enhanced surveillance imaging every 4 months according to a fixed protocol-defined schedule. The artificial intelligence thresholds and interval-mapping rules will be calibrated before enrollment so that the expected aggregate number of protocol-scheduled cross-sectional imaging examinations through Month 24 is approximately comparable between study groups. This resource-matching objective applies at the study-group level and does not mean that each participant will undergo the same number of examinations. All protocol-scheduled and unscheduled imaging examinations, false-positive investigations, radiation and contrast exposure, and related health-care use will be recorded. Except for the timing of protocol-scheduled surveillance imaging, procedures for evaluating suspected recurrence and access to recurrence management will be harmonized between the two groups. A positive artificial intelligence output or abnormal tumor-marker result will not by itself establish recurrence. Suspected recurrence will require guideline-concordant imaging and, when clinically indicated, pathological confirmation. After recurrence is confirmed, the treating multidisciplinary team will assess management using the same protocol-defined criteria in both groups. Potentially curative-intent options may include repeat liver resection, thermal ablation, or liver transplantation in appropriately selected patients. These treatments are not assigned or guaranteed by the artificial intelligence system or by the trial. Eligibility depends on recurrence extent, liver function, performance status, technical feasibility, patient preference, institutional capability, and applicable organ-allocation policies. Participants for whom curative-intent treatment is not appropriate may receive other guideline-concordant locoregional or systemic treatment. Treatment of underlying liver disease, including antiviral therapy when indicated, will continue according to the same clinical standards in both groups. Before enrollment of the first participant, the artificial intelligence model parameters, input variables, variable definitions, preprocessing procedures, handling of missing inputs, modality-specific components, risk thresholds, surveillance-interval mapping, output format, and software version will be finalized and locked. The deployed version will be documented using version identifiers, timestamps, and a reproducible code or model hash. Trial outcomes will not be used to retrain, recalibrate, or otherwise modify the model during the primary trial period. System inputs, outputs, recommended intervals, clinician overrides, scheduled examinations, and subsequent software changes will be recorded in audit logs. The principal clinical endpoint is curative-strategy failure-free survival through Month 24. A primary endpoint event is the first occurrence of recurrent hepatocellular carcinoma that a blinded independent central multidisciplinary adjudication committee determines is not amenable to protocol-defined curative-intent treatment, or death from any cause, whichever occurs first. A recurrence adjudicated as amenable to a protocol-defined curative-intent strategy is not considered a primary endpoint event at that time, and the participant will remain under follow-up. Secondary assessments will characterize the occurrence and timing of any recurrence, tumor burden and stage at recurrence, eligibility for and receipt of curative-intent salvage treatment, outcomes following salvage treatment, overall survival, conventional recurrence-free survival, surveillance burden, unscheduled diagnostic procedures, false-positive investigations, safety, health-related quality of life, adherence, and health-care resource use. Conventional recurrence-free survival is not the primary endpoint because different surveillance intervals can change the recorded time of recurrence detection without changing the biological time at which recurrence develops. The primary analysis will follow the intention-to-treat principle according to randomized assignment. Prespecified analyses will account for study center, initial treatment modality, missing data, adherence, clinician overrides, and heterogeneity across clinically relevant subgroups. Trial conduct, data quality, participant safety, and the occurrence of non-curatively treatable recurrence will be centrally monitored in accordance with the protocol and statistical analysis plan.