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

Study Background Hepatocellular carcinoma (HCC) is one of the most prevalent malignant tumors worldwide, ranking sixth in global incidence and third in mortality, responsible for approximately 480,000 deaths annually. China accounts for more than 45% of global HCC cases, imposing an extremely heavy disease burden. Curative surgical resection is the primary curative approach for achieving long-term survival in patients with early-stage HCC. However, the postoperative recurrence rate reaches 50%-70% within five years after surgery, severely undermining patient prognosis. Postoperative adjuvant therapy has become a core strategy to delay tumor recurrence and improve survival outcomes. Transarterial chemoembolization (TACE) and tyrosine kinase inhibitors (TKIs), including sorafenib and lenvatinib, are widely administered for high-risk postoperative HCC patients.

Nevertheless, both therapeutic modalities have notable clinical limitations. The objective response rate of TACE is only 50%-60%, with a substantial proportion of patients failing to derive clinical benefits and suffering from treatment-related liver function injury. Although TKIs can prolong recurrence-free survival (RFS) by 3 to 5 months in high-risk postoperative HCC populations, the treatment response rate in unselected patients is less than 20%. Additionally, the incidence of grade 3-4 adverse events, such as hypertension, hand-foot skin reaction, and proteinuria, exceeds 50%, leading to treatment discontinuation in approximately 20% of patients due to intolerable toxicities. Currently, there is a lack of efficient and reliable biomarker systems to screen potential treatment-responsive populations. Clinical treatment decisions still rely on empirical indicators, including tumor size and vascular invasion, resulting in limited individualization, wasted medical resources, and unnecessary treatment burdens for patients.

Emerging studies have demonstrated that the tumor immune microenvironment (TIME) serves as a critical biological determinant of therapeutic sensitivity to TACE and TKI in HCC. Preoperative contrast-enhanced CT and MRI can intuitively reflect tumor vascularity, boundary integrity, peritumoral infiltration, and satellite lesion status. These imaging characteristics are closely correlated with postoperative pathological features such as microvascular invasion (MVI) and tumor differentiation, which further determine adjuvant treatment responses. For instance, preoperative MRI radiomic features, including textual heterogeneity and irregular tumor margins, have been proven to be significantly associated with postoperative HCC recurrence risk. Moreover, CT perfusion parameters can accurately predict the degree of tumor necrosis after TACE treatment. Furthermore, the integration of preoperative imaging and postoperative pathological images enables the construction of tumor spatiotemporal evolution models, revealing dynamic biological alterations before and after therapeutic intervention. Therefore, multimodal data integrating preoperative imaging, postoperative pathological slides, textual pathological reports, and clinical indicators can overcome the limitations of single-source data and substantially improve the accuracy of adjuvant therapy response prediction.

Advancements in artificial intelligence, particularly deep learning techniques, have provided a novel approach to excavate in-depth biological information from routine hematoxylin and eosin (H&E) stained whole-slide imaging (WSI). As standard postoperative pathological materials, H&E WSIs are routinely generated for all HCC patients without additional sampling or testing. The cellular and structural morphological details embedded in WSIs can effectively reflect core TIME characteristics. Specifically, convolutional neural networks (CNN) and Vision Transformer (ViT) architectures can automatically identify morphological patterns associated with CD8⁺ T cell infiltration density and PD-L1 expression, realizing cross-modal prediction of immune status based on routine H&E morphology. Recent studies have validated the high diagnostic efficacy of WSI-based deep learning models in multiple malignancies, including the prediction of microsatellite instability (MSI) in colorectal cancer (AUC = 0.88), PD-L1 expression in non-small cell lung cancer (AUC = 0.80), and tumor mutation burden (TMB) (AUC = 0.91). In HCC, WSI-driven deep learning models have achieved favorable performance in predicting postoperative recurrence risk (AUC = 0.82) and immune cell infiltration (AUC = 0.78). However, no studies have focused on the critical clinical challenge of multimodal prediction of TACE and TKI adjuvant therapy responses by integrating preoperative imaging and postoperative pathological data, and large-scale multicenter prospective clinical validation remains absent.

Accordingly, this study intends to integrate multicenter clinicopathological data with artificial intelligence algorithms to construct a novel multimodal predictive model based on preoperative imaging, postoperative H&E-stained WSI, textual pathological reports, and clinical indicators. This model aims to clarify the correlation between preoperative tumor characteristics and postoperative adjuvant treatment responses and develop a clinically applicable digital decision-making tool, promoting multi-dimensional and full-cycle precise management for HCC.

Study Objectives

This study aims to construct and validate a multimodal data platform integrating preoperative contrast-enhanced CT/MRI imaging, postoperative H&E-stained whole-slide imaging (WSI), textual pathological reports, and clinical indicators from multicenter HCC patients. Based on radiomic feature extraction and deep learning algorithms, this study seeks to develop a joint predictive model to preoperatively identify TACE-sensitive and TKI-sensitive HCC subtypes and quantitatively evaluate treatment efficacy. Meanwhile, the study intends to elucidate the spatiotemporal biological evolution patterns of tumors during postoperative adjuvant therapy, including vascular remodeling and dynamic changes in the tumor immune microenvironment. The predictive performance of the model will be validated in prospective clinical cohorts. Furthermore, a visual clinical decision support system will be developed to optimize individualized postoperative adjuvant therapy strategies, reduce the rate of ineffective treatment, and facilitate the advancement of multi-dimensional, full-cycle precise treatment and management for hepatocellular carcinoma.

Study Design This study adopts a hybrid retrospective construction and prospective observational validation design without any clinical intervention. In the retrospective stage, a total of 10,000 patients who underwent curative surgical resection for hepatocellular carcinoma will be enrolled retrospectively. Comprehensive data including postoperative pathological whole-slide images, preoperative and postoperative radiographic images, and essential clinical variables will be systematically collected. These multicenter real-world data will be used to develop a multimodal classification deep learning model capable of predicting patient therapeutic responses to four mainstream postoperative adjuvant treatment strategies, namely simple surgical resection alone, surgery combined with adjuvant TACE, surgery combined with TACE plus systemic therapy, and surgery combined with exclusive systemic therapy.

In the prospective observational validation stage, a total of 1,000 eligible postoperative HCC patients will be consecutively enrolled from 10 to 15 clinical centers. Standardized preoperative contrast-enhanced imaging data, postoperative H&E-stained WSI, and complete clinical data will be collected and input into the established artificial intelligence model to generate individualized postoperative adjuvant therapy prediction schemes. Notably, no trial-related intervention or mandatory guidance on clinical decision-making will be implemented throughout the study, and all actual treatment strategies are independently determined by attending physicians in accordance with clinical guidelines and individual patient conditions. Researchers will strictly record all model-predicted treatment regimens and corresponding actual clinical treatment decisions.

Prospective enrolled patients will be divided into two observational cohorts according to the consistency between model-predicted regimens and real clinical regimens: the prediction-consistent cohort and the prediction-inconsistent cohort. All subjects will receive standardized long-term postoperative follow-up to collect real prognostic outcomes including tumor recurrence, disease-free survival, and overall survival. Through comparative analysis of prognostic differences between the two groups, this study aims to evaluate the consistency between model predictive results and actual clinical outcomes, as well as the generalization performance and performance degradation degree of the multimodal AI model under real-world multicenter clinical settings, so as to comprehensively verify the clinical applicability and stability of the predictive model.

Eligibility Criteria Inclusion Criteria

Patients must meet all of the following criteria for study enrollment: (1) Histopathologically confirmed diagnosis of hepatocellular carcinoma; (2) Age ranging from 18 to 75 years old; (3) Underwent curative R0 surgical resection for primary hepatocellular carcinoma; (4) Possesses qualified postoperative formalin-fixed paraffin-embedded (FFPE) tissue slides with routine H&E staining, which are complete and eligible for digital whole-slide scanning; (5) Has complete and accessible clinicopathological and follow-up data, including but not limited to age, gender, HBsAg status, liver cirrhosis background, preoperative AFP and PIVKA-II levels, tumor size and tumor number, CNLC stage, AJCC 8th TNM stage, BCLC stage, vascular invasion status, tumor differentiation grade, detailed postoperative adjuvant treatment regimens and medication records, as well as postoperative tumor recurrence and survival outcomes; (6) Has preoperative contrast-enhanced CT or MRI images with standardized scanning parameters and no severe artifacts, meeting the quality requirements for radiomic and artificial intelligence analysis.

Exclusion Criteria

Patients satisfying any of the following conditions will be excluded from the study: (1) Received preoperative anti-tumor treatments at non-collaborative medical centers, including but not limited to TACE, targeted therapy, immunotherapy, and radiotherapy, with unavailable original imaging and clinical baseline data; (2) Complicated with other primary malignant tumors of different organs; (3) Underwent non-R0 resection with positive surgical margins (R1 or R2 resection); (4) Poor quality of pathological tissue slides, including severe fading, folding, breakage, or insufficient tissue volume, which fails digital scanning and valid image analysis; (5) Severe missing of key clinical baseline information or postoperative follow-up data; (6) Absent preoperative contrast-enhanced imaging data or unqualified imaging quality with severe artifacts or incomplete scanning sequences, which cannot be used for subsequent image analysis and model extraction.