Integrated ultrasound radiomics and clinical data to predict PD-1 blockade efficacy in unresectable hepatocellular carcinoma.

Introduction
Hepatocellular carcinoma (HCC) represents the fourth most prevalent malignancy globally, with China bearing a disproportionately high disease burden. HCC is associated with an unfavorable prognosis, characterized by a notably low 5-year survival rate that typically remains below 15% [1]. A contributing factor is that the majority of HCC patients present with unresectable disease at initial diagnosis, substantially limiting therapeutic options and compromising treatment outcomes [2, 3]. Currently, extensive clinical trials have been conducted to evaluate the efficacy of PD-1 blockades, both as monotherapy and in combination regimens, for unresectable HCC, revealing substantial interpatient variability in treatment response while demonstrating the potential clinical value of this therapeutic strategy [4]. PD-1 serves as a crucial therapeutic target in cancer immunotherapy through its specific interaction with the PD-L1 protein. This molecular interaction plays a pivotal role in modulating immune responses by enabling the immune system to effectively recognize and eliminate malignant cells, thereby potentiating the host’s intrinsic anti-tumor defense mechanisms [5]. However, there is currently a lack of reliable predictive approaches or established biomarkers to accurately forecast the treatment response in patients with unresectable HCC undergoing PD-1 inhibitors therapy [6]. The development of clinically applicable, minimally invasive, and cost-effective predictive tools for real-time assessment of therapeutic efficacy in unresectable HCC patients following immunotherapy represents an urgent unmet medical needed.
Computed tomography (CT) and magnetic resonance imaging (MRI) are well-established tools for evaluating therapeutic response and monitoring tumor progression. However, both modalities have inherent limitations: CT exposes patients to ionizing radiation and provides suboptimal soft-tissue contrast, whereas MRI, despite superior tissue characterization, is limited by long acquisition times, motion and metal artifacts, inter-scanner variability, and difficulty in distinguishing viable tumor from post-treatment fibrosis or necrosis [7–9]. In contrast, ultrasonography has emerged as a widely adopted alternative for both tumor detection and treatment response assessment in clinical practice. For instance, Sun et al. [10] and Liang et al. [11] developed comprehensive nomograms by integrating quantitative parameters derived from pretreatment contrast-enhanced ultrasound (CEUS) with baseline clinicopathological characteristics, enabling accurate prediction of therapeutic outcomes in advanced HCC patients undergoing combined anti-PD-1 and anti-VEGF treatment. Ultrasound elastography, particularly shear wave elastography (SWE), enables quantitatively assessment of tissue stiffness, with evidence showing that elevated tumor stiffness values measured by SWE correlate with reduced disease-free survival rates in patients diagnosed with early-stage invasive breast cancer [12]. Tumor stiffness has also emerged as a useful biomarker for chemotherapy response assessment. Accumulating evidence indicates that elastography-based stiffness measurements serve as a robust predictive tool for early chemotherapy responses evaluation [13]. Notably, our team’s preliminary investigations have demonstrated that both intratumoral and peritumoral stiffness parameters, as quantified by ultrasound elastography, exhibit significant predictive value for antiviral therapeutic outcomes in HCC [14].
The advent of artificial intelligence technology has revolutionized predictive modeling in HCC management, with a growing body of research leveraging imaging genomics and machine learning algorithms to forecast comprehensive treatment outcomes [15, 16]. Furthermore, innovative approaches integrating neutrophil-to-lymphocyte ratio analysis with radiomic features and machine learning techniques have demonstrated promising potential in predicting post-therapeutic survival outcomes in HCC patients [17, 18]. Nevertheless, these predictive models are constrained by several limitations, including suboptimal accuracy, limited model generalizability, and predominant reliance on single-center datasets without robust external validation. In recent years, convolutional neural networks (CNNs) have emerged as a powerful deep learning tool for accurate classification of various cancer subtypes. Notably, Hong et al. successfully implemented CNN-based algorithms to differentiate histological subtypes of uterine endometrial cancer through automated analysis of whole-slide histopathological images. Similarly, Zhou et al. demonstrated the efficacy of CNN architectures in distinguishing triple-negative breast cancer from non-triple-negative breast cancer subtypes with high diagnostic accuracy [19, 20]. However, there is currently no study regarding the application of CNNs for predicting PD-1 inhibitor treatment response in unresectable HCC using ultrasound imaging data.
To establish the clinical utility of ultrasound imaging in predicting anti-PD-1 treatment response for unresectable HCC, we designed a two-phase longitudinal study leveraging advanced artificial intelligence technologies. The retrospective phase incorporated patients with unresectable HCC receiving PD-1 inhibitor-based therapy, enabling the development of a novel multimodal predictive framework integrating ultrasound radiomics with clinical indicators. Subsequently, we conducted a prospective study involving patients undergoing liver resection following PD-1 inhibitor treatment, to validate the clinical applicability and biological relevance of this model in assessing therapeutic response.

Method

Population
From May 2020 to May 2024, a cohort of 1400 patients with histopathologically confirmed unresectable HCC was enrolled at the Eastern Hepatobiliary Surgery Hospital. Unresectability was defined according to previously established criteria [21]. All patients received PD-1 inhibitor-based combination therapy as their primary treatment regimen. The inclusion criteria comprised: (1) administration of PD-1 inhibitors either as monotherapy or in combination with other therapeutic modalities; (2) completion of at least 3 treatment cycles (21-day per cycle) of PD-1 inhibitor therapy; (3) age range between 18 and 75 years; (4) Child-Pugh classification A or B liver function; (5) Eastern Cooperative Oncology Group Performance Status (ECOG PS) score of 0–1; (6) presence of at least one measurable lesion according to RECIST v1.1 criteria; (7) Barcelona Clinic Liver Cancer (BCLC) stage B or C; and (8) adequate hematologic, hepatic, and renal function. The exclusion criteria encompassed: (1) presence of severe comorbidities or life-threatening systemic diseases; (2) Child-Pugh class C liver dysfunction; (3) inability to adhere to the study protocol requirements; and (4) voluntary withdrawal of informed consent during the study period.
For model development, eligible patients were randomly split into a training set and an internal validation set at a predefined ratio. External validation was independently performed at two tertiary medical centers: Jiaxing First People’s Hospital and Yueyang Hospital of Integrated Traditional Chinese and Western Medicine. Furthermore, the predictive model was applied to a distinct subgroup of patients undergoing combined PD-1 inhibitor and transarterial chemoembolization (TACE) therapy at Jiaxing First People’s Hospital, Yueyang Hospital of Integrated Traditional Chinese and Western Medicine and Taiyuan People’s Hospital. Although the model was originally developed in a cohort receiving PD-1 inhibitor-based regimens without TACE, we applied it to this subgroup to assess the generalizability and robustness of the predictive framework. We hypothesized that key imaging and clinical features predictive of PD-1 blockade efficacy may remain relevant in patients receiving combination locoregional therapy. The comprehensive patient screening and enrollment workflow is illustrated in Fig. 1. (Ethical Approval Number: EHBHKY2024-K020-P001).

PD-1 inhibitor-dominant treatment
During the treatment period, enrolled patients received PD-1 inhibitor-based regimens, including camrelizumab (200 mg), sintilimab (200 mg), tislelizumab (200 mg), toripalimab (240 mg), penpulimab (200 mg), or pembrolizumab (200 mg) [22]. Among these, some patients were administered combination therapy with PD-1 inhibitors plus lenvatinib (8 mg for body weight < 60 kg or 12 mg for body weight ≥ 60 kg, orally once daily), while others patients received PD-1 inhibitors in conjunction with TACE performed under local anesthesia via the right femoral artery approach. To verify the model’s ability to capture PD-1-specific efficacy signals and avoid potential confounding from combination partners, subgroup validation was conducted specifically in these two combination therapy cohorts. Treatment response was systematically evaluated at 3-month intervals through comprehensive assessments including ultrasound imaging, laboratory analyses, and contrast-enhanced CT performed when clinically indicated, including suspected tumor progression, evaluation for potential downstaging to resectable status, or assessment of atypical imaging findings or complications. For patients achieving successful tumor downstaging to resectable status, surgical intervention followed by postoperative consolidation therapy was implemented. In cases demonstrating disease progression or loss of clinical benefit as determined by multidisciplinary team, second-line therapeutic regimens were recommended according to established clinical guidelines [23]. In clinical practice, treatment strategies for unresectable HCC are diverse, and PD-1 inhibitors often require combination with other therapies to optimize efficacy. Local therapies such as TACE can complement the systemic immune effects of PD-1 inhibitors, while combination regimens of targeted agents like lenvatinib with PD-1 inhibitors have become standard treatments for advanced unresectable HCC. The inclusion of the combination therapy subgroup focuses on evaluating the model’s performance within mainstream combination regimens, validating its predictive robustness to ensure applicability across a broad patient population.

Follow-up and clinical observation indicators
All patients were systematically followed up through standardized outpatient protocols. The surveillance regimen included comprehensive physical examinations and serial laboratory investigations, with particular emphasis on AFP level monitoring and ultrasonographic assessments conducted at 3–6 month intervals. Advanced imaging modalities, including contrast-enhanced CT and MRI, were routinely performed on a biannual basis. In cases demonstrating radiological evidence of tumor recurrence, therapeutic interventions such as TACE or ablation therapy were considered, with treatment selection guided by tumor-specific characteristics and multidisciplinary team evaluation. All statistical analyses were based on prospectively collected data up to December 1, 2023.
Imaging changes involve alterations in the size of intrahepatic and extrahepatic tumors as well as tumor and vascular thrombus necrosis, along with fluctuations in quantitative AFP levels. Other key indicators include objective response rate (ORR), disease control rate (DCR), conversion to surgical resection rate, and pathological complete response (pCR) rate. Adverse reactions to be monitored encompass hypertension, hand-foot syndrome, diarrhea, rash, cutaneous capillary proliferation, exfoliative dermatitis, electrolyte imbalance and significant organ toxicities.

Ultrasound examination
Conventional ultrasound diagnostics were performed using the Acuson Sequoia diagnostic ultrasound system (Siemens Healthineers, Mountain View, California, USA) equipped with a 5C1 convex abdominal transducer. Patients were required to fast for at least 8 h before the procedure. Liver ultrasonography focusing on solid liver anomalies was performed by an experienced sonographer with over 15 years of expertise. To ensure consistency and quality control, a second specialist proficient in ultrasound evaluations using the same equipment and transducer independently reviewed all images and performed additional assessments when necessary. All baseline and follow-up examinations were thus either conducted or confirmed by this team to maintain uniformity in imaging acquisition and interpretation. Patients are required to undergo an ultrasound examination before treatment, and follow-up outpatient or inpatient ultrasound examinations should be performed after 3–4 treatment cycles.

Ultrasonic image data processing
We propose a multimodal drug efficacy prediction model that integrates imaging omics and clinical indicators (Fig. 2). For the automated processing of HCC ultrasound images, we employed techniques such as image cleaning and data augmentation to improve image quality and ensure dataset balance. Specifically, denoising and filtering methods, including median filtering and Gaussian filtering, were applied to remove noise and artifacts, thereby enhancing image clarity and interpretability. These techniques effectively reduce interference while preserving the structural details of the target region, ultimately improving the overall quality of HCC ultrasound images. Prior to deep learning training, ultrasound imaging data underwent preprocessing, which involved standardizing the image size to 256 × 256 pixels and cropping the central 224 × 224 pixel region. Experienced radiologists delineated the region of interest (ROI) on the original image data using 3D Slicer, followed by a thorough review of their annotations. The ROIs were then superimposed onto the original images to emphasize critical features, enabling more efficient and accurate modeling of the ultrasound imaging data.

Response assessment and ultrasound-based radiomics model construction
The treatment response to PD-1 inhibitors was evaluated through independent assessment by two senior radiologists according to RECIST v.1.1 criteria, with tumor response defined as either complete response or partial response.
For ultrasound image analysis in HCC modeling, we implemented three established deep learning architectures: MobileNet, DenseNet, and Inception networks. MobileNet is a lightweight convolutional neural network specifically optimized for mobile and embedded vision applications, with advantages of low computational cost and efficient feature extraction for medical images [24]. DenseNet is characterized by a dense connectivity pattern that facilitates feature reuse throughout the network, enabling effective capture of fine-grained details in ultrasound images of HCC [25]. The Inception network architecture adopts parallel multi-scale convolutional operations, which can realize comprehensive feature extraction across different levels and scales, adapting to the heterogeneous imaging features of HCC lesions [26].
Training neural networks for image analysis requires large datasets, a challenge in medical imaging due to limited sample sizes. Our results showed that even efficient architectures like MobileNet performed suboptimally under such constraints. To address this, we designed a lightweight convolutional neural network with inception modules to enhance multi-scale feature extraction. Inspired by VGG, our model incorporates convolutional, pooling, activation, and fully connected layers. A key innovation is a parallel convolutional structure within convolutional layers. The architecture consists of three convolutional modules, an inception layer, and a fully connected layer with softmax activation, achieving state-of-the-art performance in our application. We developed a predictive model for PD-1 blockade efficacy by evaluating multiple machine learning algorithms, including Random Forest, K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Logistic Regression, and Decision Tree. To integrate heterogeneous radiomic and clinical data, we applied a multimodal approach with a soft voting ensemble strategy. Each model generated class probabilities, which were aggregated via weighted averaging, with the final prediction based on the highest averaged probability score. (Training neural networks for image analysis typically demands extensive datasets, a significant hurdle in medical imaging due to the limited sample sizes. Our findings revealed that even highly efficient architectures such as MobileNet underperform under these constraints. To overcome this challenge, we engineered a lightweight convolutional neural network enhanced with inception modules to improve multi-scale feature extraction. Drawing inspiration from the VGG architecture, our model integrates a series of convolutional, pooling, activation, and fully connected layers. A distinctive feature of our design is the incorporation of a parallel convolutional structure within the convolutional layers. The architecture is composed of three convolutional modules, an inception layer, and a fully connected layer utilizing softmax activation, which collectively achieve state-of-the-art performance in our specific application. Furthermore, we developed a predictive model to assess the efficacy of PD-1 blockade by evaluating a range of machine learning algorithms, including Random Forest, K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Logistic Regression, and Decision Tree. To effectively combine heterogeneous radiomic and clinical data, we employed a multimodal approach coupled with a soft voting ensemble strategy. This method involves each model generating class probabilities, which are then aggregated through weighted averaging. The final prediction is determined by the highest averaged probability score, ensuring a robust and comprehensive analysis.)

Assessing the efficacy of the model
Evaluation metrics play a crucial role in assessing the performance of machine learning models, providing objective measures to evaluate and refine report generation models. These metrics are indispensable for objectively gauging model performance and guiding their development and optimization. In this study, the following metrics were employed to determine the relative performance of each model. These include calculating relationships between true positives (TP), false negatives (FN), true negatives (TN), and false positives (FP) to derive key metrics such as accuracy (ACC), precision, recall, F1 score, and area under the curve (AUC) (Supplement Table 3).

Statistical methods
For data analysis, IBM SPSS Statistics 28 and the R software (version 4.0.3, R Foundation for Statistical Computing, Vienna, Austria) were employed. Continuous variables were expressed as mean ± standard deviation and compared using the t-test or the Mann-Whitney U test, as appropriate. Categorical variables were categorized based on clinical outcomes and summarized as counts and percentages, with comparisons performed using Fisher’s exact test or the chi-square test. A SVM approach was applied to evaluate all potential influencing factors. The multivariable SVM model incorporated only factors that demonstrated statistical significance (P < 0.05), selected through a stepwise forward selection process. Odds ratios (ORs) were calculated for these variables, and only those with an AUC greater than 0.7 were considered significant. The performance of the constructed nomogram was assessed using the concordance index (C-index), and the model was validated using an independent validation subset.

Differential expression analysis
The “Limma” package in R software was used to perform differential gene expression analysis on the dataset derived from liver tissue samples of patients with response and non-response to PD-1 blockade therapy. Genes with a fold change (FC) > 2 were defined as significantly upregulated differentially expressed genes (DEGs), while genes with FC < 0.5 were classified as significantly downregulated DEGs. A statistical significance threshold of P < 0.05 was applied to screen for DEGs with biological significance.

Topological and functional enrichment analysis
The Bioinformatics and Evolutionary Genomics Web Tools (https://bioinformatics.psb.ugent.be/webtools/Venn/) were used to conduct topological analysis on the upregulated and downregulated DEGs from the dataset of liver tissue samples of patients with response and non-response to PD-1 blockade therapy. The DAVID tool (https://david.ncifcrf.gov/tools.jsp) was employed for functional enrichment analysis of the screened DEGs. This comprehensive analysis included Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis and Gene Ontology (GO) term enrichment analysis.

Results

Demographics of patients
Between May 2020 and May 2024, among 1400 initially screened patients with unresectable HCC, a total of 793 patients were enrolled in the retrospective multicenter cohort for model development and validation after applying the inclusion and exclusion criteria. This cohort was further divided into a training set (n = 494) and a testing set (n = 219). Additionally, 60 prospectively enrolled patients formed the external validation set, and 80 patients receiving PD-1 inhibitor combined with TACE therapy constituted the TACE combination subgroup.
We conducted a comprehensive demographic analysis of patients in both the training and testing cohorts, as detailed in Table 1. Supplementary Table 1 presents the baseline characteristics of patients in the two subgroup analyses within the training set. Comparative analysis of multiple demographic variables, including age, height, and weight, revealed no statistically significant differences between the training and testing groups (P > 0.05). In our subsequent analytical approach, we focused on incorporating imaging indicators that demonstrated significant intergroup variations. Furthermore, Supplementary Table 2 provides detailed baseline information for patients who underwent liver resection following PD-1 inhibitor therapy.

Internal validation of the efficacy prediction model
Our methodology comprised three sequential modeling approaches for predicting PD-1 inhibitor efficacy. Initially, we developed an ultrasound imaging-based predictive model, employing both Stochastic Gradient Descent (SGD) and Adaptive Moment Estimation (Adam) for parameter optimization. The resulting state-of-the-art (Sota) model demonstrated exceptional predictive performance, achieving an AUC of 0.714, F1 score of 0.889, accuracy of 0.812, precision of 0.812, and recall of 0.982 (Table 2), significantly outperforming comparative models. Subsequently, we constructed a clinical indicator-based prediction framework, implementing four distinct machine learning algorithms: KNN, Logistic Regression, SVM, and Decision Tree models. Comparative analysis revealed that Logistic Regression yielded the optimal AUC (0.632), while SVM attained the highest accuracy (0.769) despite its relatively lower AUC (0.518). Notably, Logistic Regression consistently outperformed other models across multiple metrics, including F1 score, recall, and precision. Finally, we implemented an integrated multimodal prediction framework through ensemble learning with soft voting, combining the predictive capabilities of both imaging and clinical models. This integrated approach achieved an AUC of 0.743 during internal validation, demonstrating superior performance compared to both the imaging-based Sota model (AUC: 0.714) and the Logistic Regression model (AUC: 0.632). The enhanced predictive capability of our integrated model can be attributed to the synergistic combination of diverse model architectures, effectively leveraging their complementary strengths to improve both accuracy and robustness (Table 2; Fig. 3).

Furthermore, our study cohort comprised two distinct treatment subgroups: 130 patients receiving PD-1 inhibitor therapy in combination with molecular targeted agents, and 70 patients undergoing PD-1 inhibitor treatment alongside locoregional therapy (Supplementary Table 1). Application of our predictive models to these subgroups revealed that the ensemble learning approach maintained consistent and robust predictive performance across both treatment modalities, achieving comparable AUC values of 0.677 and 0.683 for the molecular targeted therapy and locoregional therapy subgroups, confirming that combination therapy does not obscure PD-1’s specific contribution, this is further supported by the model’s ability to outperform single-modality models, which rely on non-PD-1-specific signals, respectively (Supplement Fig. 1).

External validation of the multimodal predictive model
During external validation, the logistic regression model demonstrated optimal performance with clinical data, achieving an AUC of 0.655 and an accuracy of 0.700. In the imaging omics domain, the SOTA model maintained its superior predictive capability, attaining an AUC of 0.654 and accuracy of 0.725, thereby confirming its robustness against distributional shifts across different datasets. Comparative analysis revealed that imaging omics consistently outperformed clinical data, a phenomenon potentially attributable to its more comprehensive characterization of PD-1 blockade efficacy and the enhanced capacity of deep learning architectures to capture complex, nonlinear feature representations. The implementation of ensemble learning strategies yielded additional performance enhancements, achieving an AUC of 0.692. This integrated approach not only surpassed the predictive capabilities of individual modalities but also substantiated its effectiveness as a robust framework for predicting PD-1 blockade therapeutic outcomes (Table 3; Fig. 4).

Subgroup verification results
For external validation, we applied our predictive model to an independent cohort comprising 60 patients receiving combined PD-1 blockade and TACE therapy (Supplementary Table 2). This analysis was performed to evaluate whether the predictive model, developed in a PD-1 inhibitor monotherapy cohort, retained its predictive value in a combination therapy context. The validation outcomes, including comprehensive performance metrics and the ROC curve, are detailed in Table 4; Fig. 5, respectively. Notably, the ensemble learning approach demonstrated enhanced predictive capability, achieving an AUC of 0.641. This integrated methodology not only surpassed the performance of individual predictive modalities but also established its clinical utility in evaluating therapeutic responses to the combined PD-1 blockade and TACE treatment regimen.

Preliminary exploration of the molecular mechanism underlying the effectiveness of PD-1 blockade
To elucidate the mechanistic basis of PD-1 blockade efficacy as predicted by our model, we conducted a comprehensive molecular analysis using liver tissue samples obtained from 20 patients (10 predicted responders and 10 predicted non-responders) prior to treatment initiation. To biologically interpret the imaging features predictive of PD-1 blockade efficacy, we performed RNA sequencing on these pretreatment liver biopsies to explore potential molecular pathways associated with the radiomics-based prediction. RNA sequencing analysis revealed distinct gene expression patterns, identifying 158 significantly upregulated genes (represented by orange scatter points in Fig. 6) and 183 downregulated genes (depicted as blue scatter points), alongside 16,539 genes showing no significant differential expression (gray scatter points). We observed that genes involved in necroptosis pathways (CYLD, RIPK1, TRADD, BIRC3, etc.) showed higher expression in predicted responders, suggesting that radiomics signatures reflecting tissue heterogeneity and vascularity may correspond to necroptotic activity within the tumor microenvironment. Subsequent pathway enrichment analysis using both KEGG and GO databases demonstrated significant involvement of key genes (CYLD, TRADD, H2AC19, CAPN2, TNFRSF10B, RIPK1, H2AC18, BIRC3) in necroptosis-related pathways. KEGG analysis specifically identified these genes as being significantly enriched in the “Necroptosis” pathway. GO enrichment analysis further revealed their participation in multiple programmed cell death processes. These molecular insights suggest that the identified genes play crucial regulatory roles in programmed necrosis and its associated signaling cascades, potentially serving as key mediators of therapeutic response to PD-1 blockade.

Discussion
This retrospective study analyzed 713 patients with unresectable HCC who underwent PD-1 inhibitor-based combination therapy, establishing a novel ultrasound radiomics model for predicting tumor response to PD-1 blockade therapy. The ensemble learning model demonstrated superior predictive performance in internal validation, achieving an AUC of 0.743, which significantly outperformed single-modality models, indicating robust generalization capability and cross-dataset reliability. External validation further confirmed the model’s clinical applicability, with an AUC of 0.692, maintaining consistent predictive accuracy across diverse patient populations and real-world settings. Our study introduces an innovative efficacy prediction framework that integrates imaging omics with clinical indicators through ensemble learning. For ultrasound image processing, we implemented advanced computer vision networks and developed an optimized Sota model based on the Inception architecture, specifically tailored for HCC ultrasound analysis [27]. Clinical feature selection was performed using Lasso regression, followed by comprehensive evaluation through multiple machine learning algorithms. The final integrated model synergistically combined both imaging and clinical modalities through ensemble learning techniques [28]. Notably, this study is the first to apply ultrasound imaging for evaluating the therapeutic efficacy of HCC patients receiving PD-1 inhibitor-based therapy, and this method demonstrates significant advantages over previous studies using other modalities. Compared with traditional HCC imaging methods such as CT and MRI, ultrasound avoids ionizing radiation, offers higher cost-effectiveness, enables real-time bedside assessment, and is more patient-friendly for advanced HCC patients with limited mobility. In contrast to invasive biomarker detection, ultrasound, as a non-invasive modality, eliminates the risk of complications associated with sample collection and allows for dynamic observation of tumor changes induced by PD-1 inhibitors. Furthermore, previous radiomics studies based on CT/MRI often lack generalizability due to differences in scanning protocols across centers. However, the ultrasound-based model in this study, combined with standardized preprocessing, exhibits stronger cross-center applicability. This innovative approach provides valuable insights for guiding PD-1 inhibitor therapy in HCC, offering a non-invasive and potentially widely applicable predictive tool for clinical decision-making.
From an architectural perspective, our model leverages convolutional neural networks to effectively extract complex, high-dimensional imaging features, ensuring comprehensive and precise feature representation [29]. The incorporation of diverse machine learning algorithms for clinical indicator analysis further enhances model robustness through increased algorithmic diversity. The implementation of soft voting ensemble learning enables synergistic integration of multimodal predictions, effectively mitigating individual modality limitations and substantially improving both predictive accuracy and model stability. Regarding predictive capabilities, our framework optimally utilizes the complementary strengths of imaging and clinical data modalities [30, 31]. A significant innovation of this study is the implementation of Grad-CAM for model interpretability. This visualization technique, employed for the first time in PD-1 inhibitor efficacy assessment (Supplementary Fig. 2), provides critical insights into the pathological regions driving model predictions, thereby enhancing both clinical interpretability and decision-making confidence [32].
Previous methodologies for assessing PD-1 inhibitor efficacy in HCC treatment have primarily focused on four distinct approaches: (1) biomarker profiling, (2) imaging-based assessment, (3) clinical parameter analysis, and (4) multimodal combination strategies. Among these, the quantitative evaluation of PD-1 and PD-L1 protein expression levels in tumor tissue has emerged as a particularly promising predictive biomarker, offering valuable insights into the likelihood of therapeutic response to PD-1 blockade therapy [33]. Analysis of additional predictive biomarkers, including tumor mutational burden (TMB) and tumor-infiltrating immune cell profiles, represents a valuable complementary strategy for predicting treatment response [34]. Advanced imaging modalities, including contrast-enhanced CT, multiparametric MRI, and PET-CT, have been employed to quantitatively assess treatment response through detailed evaluation of tumor morphology, density, and volumetric changes [35]. Recent radiomics approaches have enabled the identification of specific imaging biomarkers that demonstrate significant correlation with therapeutic outcomes [36, 37]. Furthermore, comprehensive clinical assessment incorporating disease staging, liver functional reserve (Child-Pugh classification), and prior treatment history has proven valuable in predicting PD-1 inhibitor responsiveness. The integration of patient-specific factors, including performance status and comorbidity profiles, further enhances treatment outcome prediction [38]. Despite these advancements, previous predictive models have been constrained by methodological limitations, particularly small cohort sizes and insufficient external validation. Our study addresses these limitations through a large, multicenter dataset and rigorous external validation, demonstrating consistent model performance across diverse patient populations and confirming the robustness of our predictive model.
Ultrasound imaging offers distinct advantages over CT and MRI in tumor response assessment, particularly in terms of real-time evaluation, cost-effectiveness, and absence of ionizing radiation [39, 40]. The tumor microenvironment (TME), comprising complex biochemical and biophysical elements surrounding malignant cells, plays a pivotal role in modulating tumor behavior and therapeutic resistance [41, 42]. A critical component of TME remodeling involves the collagen-rich extracellular matrix (ECM), whose structural alterations significantly influence tumor initiation, progression, and treatment resistance mechanisms [43, 44]. Cellular morphology serves as a fundamental indicator of cellular adaptation to microenvironmental stimuli [45], with morphological variations profoundly affecting cellular functions and intercellular dynamics within the TME. These include ECM tropism, adhesion properties, and mechanotransduction signaling [46]. Ultrasound radiomics has emerged as a powerful tool for non-invasive characterization of both tumor microenvironmental features and cellular morphological patterns, providing unique insights into tumor biology and treatment response. Advanced ultrasound modalities, such as CEUS and elastography, offer additional functional and biomechanical information, including tumor perfusion and tissue stiffness. However, these techniques are technically demanding and highly operator-dependent, requiring experienced sonographers and standardized protocols to ensure reliable and reproducible measurements. By combining conventional ultrasound with these advanced techniques under controlled conditions, it is possible to capture subtle tumor characteristics that may reflect underlying molecular processes and improve the predictive performance of radiomics-based models.
This study demonstrates the predictive value of pre-treatment ultrasound imaging for PD-1 inhibitor efficacy, potentially attributable to its ability to capture subtle tumor characteristics associated with treatment response. In multicenter external validation encompassing data from three independent institutions, our model achieved an AUC of 0.692, significantly outperforming single-modality predictive approaches. These findings not only provide robust scientific evidence for understanding PD-1 inhibitor-induced apoptosis mechanisms in HCC but also validate the clinical utility of our predictive framework. Furthermore, we prospectively applied this predictive model to evaluate treatment outcomes in patients receiving combined PD-1 inhibitor and TACE therapy. The model demonstrated substantial clinical relevance, achieving 64.1% accuracy in identifying responders to this combination treatment regimen, thereby supporting its potential for guiding therapeutic decision-making in clinical practice. The identification of necroptosis-related pathways as a potential mechanistic basis for PD-1 blockade efficacy provides novel insights into the molecular determinants of immunotherapy response. The significant enrichment of key genes (CYLD, TRADD, RIPK1, BIRC3) in both the KEGG necroptosis pathway and multiple GO terms related to programmed cell death suggests that necroptosis may serve as a critical regulator of anti-tumor immunity following PD-1 blockade. Our findings suggest that the ultrasound radiomics features capturing intra-tumoral heterogeneity might indirectly reflect necroptosis-related molecular processes, which are known to modulate immune infiltration and sensitivity to PD-1 blockade. This finding aligns with emerging evidence highlighting the immunogenic nature of necroptosis and its potential to enhance anti-tumor immune responses through the release of damage-associated molecular patterns (DAMPs) and subsequent activation of dendritic cells [47–49].
Several limitations of this study should be acknowledged. First, although this study focuses on evaluating the efficacy of PD-1 inhibitor combination therapy, there may be potential selection bias between the retrospective training cohort and the prospective validation cohort. Despite the two cohorts being sourced from different institutions, their baseline characteristics are highly similar. This may be attributed to the uniform inclusion criteria adopted by all participating centers for unresectable HCC patients receiving PD-1 blockade therapy, which inadvertently reduces sample heterogeneity. Such similarity may limit the generalizability and reproducibility of the prospective study results, as the model may be more applicable to patients with specific characteristics rather than a broader clinical population. Second, this study has not yet analyzed the ultrasound features of PD-1 responders. In subsequent studies, we will supplement PD-1 expression data based on the existing predictive model and further explore the relationship between imaging features and PD-1 expression, thereby enhancing the interpretability and predictive value of the model. Third, we also recognize that there is room for improvement in the current diagnostic performance of the model, as represented by the AUC values. In the future, we will further improve the robustness and predictive accuracy of the model by expanding the sample size, optimizing the radiomic feature extraction algorithm, and refining the selection of clinical variables. Finally, the study participants were exclusively derived from the Chinese population, which may limit the extrapolation of the study results to other ethnic groups. Additionally, the molecular mechanisms underlying the association between ultrasound imaging features and tumor sensitivity to PD-1 inhibitors have not been fully elucidated.

Conclusion
This study represents a significant advancement in precision oncology for unresectable HCC through the development and validation of an ultrasound-based predictive model for PD-1 inhibitor therapy response. Our comprehensive validation strategy, encompassing both internal and external cohorts, demonstrated the model’s robust predictive performance. Notably, the model’s capability to accurately predict pathological necrosis areas in treatment responders provides compelling evidence of its biological relevance and mechanistic alignment with PD-1 inhibitor-induced tumor cell death pathways. The integration of advanced deep learning architectures and AI methodologies has yielded a predictive tool with substantial clinical utility. The model’s demonstrated accuracy in predicting both therapeutic response and pathological outcomes underscores its foundation in the fundamental mechanisms of PD-1-mediated tumor cell killing. These findings not only validate the model’s scientific basis but also highlight its potential as a valuable clinical decision-support tool.

Supplementary Information