NCT07417800

Construction and Clinical Validation of a Predictive Model for Postoperative Adjuvant Therapy in Hepatocellular Carcinoma Based on Whole-Slide Digital Pathological Images and Deep Learning

• Status: Not Yet Recruiting
• Phase: N/A
• Study type: Interventional
• Enrollment: 1,000 (estimated)
• Sponsor: Second Affiliated Hospital, School of Medicine, Zhejiang University · Academic / Other
• Sex: All
• Age: 18 Years
• Healthy volunteers: Not accepted

Summary

Hepatocellular Carcinoma (HCC) is a common global malignancy, ranking 6th in incidence and 3rd in mortality, causing \~480,000 annual deaths. China accounts for over 45% of global cases, bearing a heavy disease burden. Radical resection is key for long-term survival in early-stage patients, but the 5-year postoperative recurrence rate reaches 50%-70%, limiting prognosis . Postoperative adjuvant therapies like Transarterial Chemoembolization (TACE) and Tyrosine Kinase Inhibitors (TKIs, e.g., sorafenib, lenvatinib) are widely used for high-risk recurrence patients TACE is suitable for intermediate-stage HCC by embolizing tumor vessels and perfusing chemo drugs ; multitarget TKIs inhibit pathways like VEGFR/PDGFR for anti-angiogenesis and anti-proliferation, serving as standard advanced HCC treatment . However, TACE has only 50%-60% objective response rate, with some patients suffering liver damage ; TKIs extend Recurrence-Free Survival (RFS) by 3-5 months in high-risk patients but have \<20% response rate in unselected populations, and \>50% incidence of grade 3-4 adverse events (hypertension, hand-foot skin reaction, proteinuria), leading to 20% treatment discontinuation. Currently, no efficient biomarkers exist for identifying beneficiaries, so treatment decisions rely on clinical experience (tumor size, vascular invasion), resulting in poor individualization, medical resource waste, and extra patient burden. Recent studies show the Tumor Immune Microenvironment (TIME) affects TACE/TKI sensitivity . TIME features (immune cell infiltration like CD8⁺ T cells, PD-L1 expression, spatial structure) correlate with treatment response. For example, immune-inflammatory TIME (high CD8⁺ T cell density) may improve response, while immune-exempt/desert phenotypes indicate resistance . However, TIME assessment relies on high-cost, complex technologies (mIHC, spatial transcriptomics) with poor standardization, limiting clinical use. AI (especially deep learning) enables mining deep pathological info from routine HE-stained Whole Slide Imaging (WSI, generated postoperatively for all HCC patients without extra sampling). WSI's cellular/tissue details map TIME features-models like CNN/ViT can predict "HE morphology → immune status" . HE-WSI deep learning models have high accuracy in predicting MSI (AUC 0.88) in colorectal cancer 18, PD-L1 (AUC 0.80) and TMB (AUC 0.91) in non-small cell lung cancer , and HCC recurrence risk (AUC 0.82)/immune infiltration (AUC 0.78) . Yet no studies focus on "postoperative adjuvant therapy efficacy prediction" with multicenter validation. Thus, building an HCC postoperative adjuvant therapy prediction model via HE-WSI and deep learning can clarify TIME's role and overcome tech limitations. This project integrates multicenter clinicopathological data and AI to establish/validate TACE/TKI efficacy prediction models, providing a reliable tool for HCC postoperative treatment decisions.

Detailed description

\# I. Study Background and Objectives Hepatocellular Carcinoma (HCC) is a common malignant tumor worldwide, ranking 6th in terms of incidence and 3rd in mortality globally. It causes approximately 480,000 deaths each year, with China accounting for over 45% of global cases, representing an extremely heavy disease burden 1. Radical surgical resection is the primary means for patients with early-stage liver cancer to achieve long-term survival. However, the 5-year postoperative recurrence rate is as high as 50%-70%, which severely limits patient prognosis 2. Postoperative adjuvant therapy has become a key strategy to delay recurrence and improve survival. Among such therapies, Transarterial Chemoembolization (TACE) and Tyrosine Kinase Inhibitors (TKIs, such as sorafenib and lenvatinib) have been widely used in the treatment of patients at high risk of recurrence 1,3. TACE induces ischemic necrosis by locally embolizing tumor-feeding blood vessels combined with chemotherapy drug perfusion, and is suitable for intermediate-stage liver cancer 4,5. Multitarget TKI drugs, on the other hand, can systematically inhibit signaling pathways such as Vascular Endothelial Growth Factor Receptor (VEGFR) and Platelet-derived Growth Factor Receptor (PDGFR), exerting anti-angiogenic and anti-tumor proliferation effects, and have become the standard treatment for advanced liver cancer 3. Nevertheless, both treatment modalities have significant limitations: the objective response rate of TACE is usually only 50%-60%, and some patients cannot benefit from it and may even experience liver function damage 6. Although TKIs can extend the Recurrence-Free Survival (RFS) of high-risk postoperative patients by 3-5 months, the treatment response rate in unselected populations is less than 20%, and the incidence of grade 3-4 adverse reactions (such as hypertension, hand-foot skin reaction, and proteinuria) is over 50%, leading to treatment discontinuation in 20% of patients due to intolerance to toxic side effects 3,7,8. Currently, the clinic lacks an efficient and reliable biomarker system for identifying potential benefit populations. As a result, treatment decisions still rely on clinical experience (e.g., based on traditional pathological features such as tumor size and vascular invasion), resulting in limited individualization, waste of medical resources, and additional treatment burden on patients. Recent studies have shown that the Tumor Immune Microenvironment (TIME) is a key biological basis affecting the therapeutic sensitivity of TACE and TKIs 3. The compositional characteristics of TIME, including immune cell infiltration (such as CD8⁺ T cells, Tregs, CTLs, and M2-type tumor-associated macrophages), the expression of immune checkpoint molecules (such as Programmed Cell Death Ligand 1, PD-L1), and spatial structure, are closely related to treatment response 9-12. For example, an immune-inflammatory TIME (high CD8⁺ T cell density and formation of tertiary lymphoid structures) may be positively correlated with tumor necrosis after TACE and TKI treatment response; in contrast, an immune-exempt or desert phenotype often indicates treatment resistance 13-15. However, current TIME assessment mostly relies on high-resolution technologies such as Multiplex Immunohistochemistry (mIHC) and spatial transcriptomics. Although these methods can depict the microenvironment in detail, they have application bottlenecks such as high cost, complex operation, high requirements for sample quality, and low standardization, which limit their clinical promotion. Breakthroughs in artificial intelligence, especially deep learning technology, have provided a new approach for mining in-depth pathological information from conventional Hematoxylin and Eosin (HE)-stained Whole Slide Imaging (WSI). As a routine data for postoperative pathological diagnosis (each HCC patient will have HE-stained WSI generated after surgery without additional sampling or testing), WSI contains cellular morphology and tissue structure details that have been proven to map the core characteristics of TIME. For instance, through architectures such as Convolutional Neural Network (CNN) and Vision Transformer (ViT), morphological patterns related to CD8⁺ T cell infiltration density and PD-L1 expression can be automatically identified, enabling cross-modal prediction of "HE morphology → immune status" 16,17. The latest studies have confirmed that deep learning models based on HE-WSI can predict Microsatellite Instability (MSI) in colorectal cancer with high accuracy (AUC 0.88) 18, PD-L1 expression (AUC 0.80) and Tumor Mutation Burden (TMB, AUC 0.91) in non-small cell lung cancer 19,20. In the field of liver cancer, studies have also used WSI deep learning to predict postoperative recurrence risk (AUC 0.82) and immune cell infiltration in liver cancer tissues (AUC 0.78) 21,22. However, no studies have focused on the clinical pain point of "predicting the efficacy of postoperative adjuvant therapy (TACE/TKI)", and there is a lack of multicenter, large-sample clinical validation. Therefore, the construction of a predictive model for postoperative adjuvant therapy of HCC based on HE-stained WSI and deep learning algorithms not only helps to analyze the mechanistic role of TIME in treatment response, but also breaks through the limitations of existing detection technologies such as high cost, long cycle, and reliance on special platforms, promoting the development of liver cancer adjuvant therapy decisions towards precision and accessibility. This project intends to integrate multicenter clinicopathological data with artificial intelligence algorithms to establish a digital pathological model suitable for predicting the efficacy of TACE and TKI, and conduct rigorous clinical validation, aiming to provide a scientific and reliable decision-making tool for postoperative treatment selection of HCC. \# II. Specific Procedures and Workflow This study is a comprehensive research combining diagnostic trials and interventional studies, divided into two major parts and multiple phases. Its purpose is to develop and validate an artificial intelligence-based predictive system for postoperative adjuvant therapy of Hepatocellular Carcinoma (HCC). The overall study follows the standardized process for the development of Artificial Intelligence (AI)-based Software as a Medical Device (SaMD), including retrospective data collection, model development and training, retrospective validation, prospective observational validation, and final prospective interventional study. * Part 1: Construction and Validation of Tumor Microenvironment Recognition Model * Development and Training Phase Public datasets (e.g., The Cancer Genome Atlas, TCGA) and a retrospective dataset from the research center (n=1500, with 400 cases already collected) will be used as the training set. The core task of this phase is to develop a deep learning model (e.g., an architecture based on Vision Transformer or ResNet) for automatically identifying, segmenting, and quantifying key tumor microenvironment features from Hematoxylin and Eosin (HE)-stained Whole Slide Images (WSIs). These features include but are not limited to the degree of immune cell infiltration, stromal proportion, vascular invasion, and necrotic areas. All images will be independently annotated in a blinded manner by at least two senior pathologists. Disagreements will be resolved through consensus or arbitration by a third senior pathologist to form the gold standard. * External Validation Phase A retrospective WSI dataset from 8-10 domestic cooperative medical centers will be used to conduct preliminary validation of the model's generalization ability. This phase aims to evaluate the robustness of the model under different scanning equipment, different slide preparation and staining processes, and different interpretation habits of pathologists. * Part 2: Construction of Treatment Response Prediction Model and Multilevel Clinical Validation This part is the core of the study. It aims to construct a classification model that can predict patients' response to different postoperative adjuvant treatment regimens (surgery alone, surgery + TACE, surgery + TACE + TKI) based on the output of the model in Part 1 (i.e., the quantified tumor microenvironment features) combined with necessary clinical variables (such as TNM stage, liver function Child-Pugh classification, and AFP level). This part adopts a strict hierarchical validation strategy: * Model Training Phase Retrospective data from the research center (n=1500) will be used. Patients included must meet the following criteria: ① pathologically confirmed HCC; ② received radical resection; ③ have complete postoperative adjuvant therapy and follow-up data. Definition of treatment response: the primary endpoint is Recurrence-Free Survival (RFS), and the secondary endpoints include Overall Survival (OS) and Objective Response Rate (ORR). Training will be conducted using multi-class machine learning algorithms (such as gradient boosting trees and deep neural networks) to output the treatment regimen category from which the patient is most likely to benefit. * Internal Validation Phase An independent validation set (n≈500, non-overlapping with the training set) will be reserved from the center's data through random division. It will be used to initially evaluate the discriminative performance of the model and perform hyperparameter tuning to prevent overfitting. * External Validation Phase 1 (Retrospective Validation and Model Calibration) Retrospective data from multiple centers across the country (8-10 centers, n=3000) will be collected. The main purposes of this phase are: ① to validate the model performance on a larger-scale and more heterogeneous external dataset; ② to calibrate or conduct limited iterative optimization of the model based on the validation results to form the final finalized predictive model. External Validation Phase 2 (Large-Scale Retrospective Validation) The finalized model will be used for validation on a new, ultra-large-scale (n=10,000) retrospective dataset (from 3-5 centers). This will further consolidate the level of evidence and provide a solid basis for prospective studies. External Validation Phase 3 (Prospective Observational Study) A total of 1000 eligible postoperative HCC patients will be prospectively and consecutively enrolled in 10-15 centers. Their WSI images will be collected and input into the model to obtain the predicted treatment regimen. However, this study phase will not interfere with actual clinical decisions. Researchers will record the model's prediction results and actual clinical decisions, and observe the patients' actual prognosis through follow-up to evaluate the consistency between the model's predictions and actual outcomes, as well as the performance attenuation of the model in real-world settings. \### External Validation Phase 4 (Prospective Interventional Study) This is the final validation step of the study. A prospective, randomized controlled or pragmatic clinical trial will be conducted in collaboration with 3-5 centers, with approximately 600 patients enrolled. Patients will be randomly divided into two groups: ① Model-Assisted Decision-Making Group: clinicians will develop treatment plans for patients based on the model's predicted optimal treatment regimen; ② Standard Treatment Group: treatment plans will be developed entirely by clinicians based on existing guidelines and clinical experience. The primary comparison will be the RFS between the two groups, so as to provide the highest level of evidence-based medicine to demonstrate the clinical utility of model-assisted decision-making and its value in improving prognosis.

Conditions

• Hepatocellular Carcinoma (HCC)
• Artificial Intelligent
• Adjuvant Chemoradiotherapy
• TACE
• Lenvatinib
• Liver Surgery

Interventions

Type	Name	Description
PROCEDURE	TACE	Transcatheter arterial chemoembolization (TACE) serves as a promising preventive intervention for hepatocellular carcinoma (HCC), especially in high-risk populations like those with cirrhosis or recurrent small lesions. By delivering chemotherapeutic agents and embolic materials via catheters to target vessels, it inhibits tumor angiogenesis and growth. This minimally invasive approach helps reduce HCC occurrence and improve long-term prognosis, with well-managed safety profiles in clinical practice.
PROCEDURE	TACE combined with Lenvatinib	The combination of transcatheter arterial chemoembolization (TACE) and lenvatinib is an emerging preventive strategy for hepatocellular carcinoma (HCC) in high-risk groups. TACE blocks tumor blood supply locally, while lenvatinib inhibits angiogenesis systemically. Their synergistic effect effectively suppresses potential malignant lesions, reduces recurrence risk, and enhances preventive efficacy. This minimally invasive combined regimen, with manageable safety, has shown promising prospects in improving the long-term outcomes of high-risk populations.

Timeline

• Start date: 2026-03-01
• Primary completion: 2029-01-01
• Completion: 2029-12-01
• First posted: 2026-02-18
• Last updated: 2026-02-18