A Computed Tomography Radiomics Approach for Assessing the T Cell‐Inflamed Gene Expression Profile in Hepatocellular Carcinoma and Its Association With Immunotherapy Response

ABSTRACT

Background

The T cell‐inflamed gene expression profile (GEP) provides a biomarker for immunotherapy response across various cancers. However, specific biomarkers for immunotherapy response in hepatocellular carcinoma (HCC) are lacking. In this study, we established a computed tomography (CT)‐based radiomics model to predict the T cell‐inflamed GEP and immunotherapy response in HCC.

Methods

This retrospective, multicenter study included 270 patients with HCC, who were divided into training (n = 227) and test sets (n = 43). All included patients had available contrast‐enhanced CT and RNA sequencing data. The T cell‐inflamed GEP consisted of 18 genes derived from RNA sequencing data, and patients were classified into GEP‐high and GEP‐low groups. A support vector machine was used to establish the radiomics model. The immunotherapy dataset comprised patients who underwent contrast‐enhanced CT and received anti‐programmed cell death protein 1/programmed cell death ligand 1 immunotherapy at three hospitals (209 lesions). This immunotherapy dataset was used to evaluate the association between the predicted GEP status and immunotherapy response. Additionally, overall survival was compared between patients with and without predicted GEP‐high lesions in the immunotherapy dataset.

Results

The areas under the receiver operating characteristic curve of the radiomics model for predicting GEP in the training and test sets were 0.824 (95% confidence interval [CI] = 0.757–0.892) and 0.827 (95% CI: 0.674–0.980), respectively. In the immunotherapy dataset, the response rate of liver lesions at 6 months was significantly higher in the GEP‐high group than in the GEP‐low group (48.1% vs. 9.3%, p < 0.001). The areas under the receiver operating characteristic curve of the radiomics model for predicting immunotherapy response at 6 months was 0.750 (95% CI: 0.614–0.886). Median overall survival among patients with GEP‐high lesions was 17.5 months, which notably exceeded the value of 8.3 months for patients without GEP‐high lesions.

Conclusions

This study provides the first CT‐based radiomics model for predicting the T cell‐inflamed GEP in HCC. This model highlights the potential of the T cell‐inflamed GEP as a noninvasive biomarker to help guide immunotherapy selection.

The T cell‐inflamed gene expression profile (GEP) provides a biomarker for immunotherapy response across various cancers. We established a CT‐based radiomics model to predict the T cell‐inflamed GEP and immunotherapy response in hepatocellular carcinoma (HCC). This multicenter study included 270 HCC patients (training set: 227; test set: 43) with contrast‐enhanced CT and RNA sequencing data. A support vector machine was used to establish the radiomics model. The immunotherapy dataset (209 lesions) was used to evaluate the association between predicted GEP status and immunotherapy response. The AUCs for predicting GEP in training and test sets were 0.824 and 0.827, respectively. The response rate at 6 months was significantly higher in the GEP‐high group (48.1% vs. 9.3%, p < 0.001). Median overall survival was 17.5 months for patients with GEP‐high lesions versus 8.3 months without. This study provides the first CT‐based radiomics model for predicting the T cell‐inflamed GEP in HCC, highlighting its potential as a noninvasive biomarker to guide immunotherapy selection.

Abbreviations

AFP

α‐fetoprotein

ALT

alanine transaminase

AST

aspartate transaminase

AUC

area under the receiver operating characteristic curve

BCLC

Barcelona Clinic Liver Cancer

CECT

contrast‐enhanced computed tomography

CI

confidence interval

CR

complete response

CT

computed tomography

DGPH

Dongguan People's Hospital

FAHSYSU

The First Affiliated Hospital, Sun Yat‐sen University

GEP

gene expression profile

GEP^hi

GEP‐high

GEP^lo

GEP‐low

GGT

γ‐glutamyltransferase

HBsAg

hepatitis B surface antigen

HBV

hepatitis B virus

HCC

hepatocellular carcinoma

ICC

intraclass correlation coefficient

ICI

immune checkpoint inhibitor

MRI

magnetic resonance imaging

MSI

microsatellite instability

OR

odds ratio

OS

overall survival

PD

progressive disease

PD‐1

programmed cell death protein 1

PD‐L1

programmed cell death‐ligand 1

PLT

platelet

PR

partial response

Rad score

radiomics score

RNA‐seq

RNA sequencing

ROC

receiver operating characteristic

SD

stable disease

SVM

support vector machine

SYSUCC

Sun Yat‐sen University Cancer Center

TCGA

The Cancer Genome Atlas

TMB

tumor mutational burden

1. Background

Hepatocellular carcinoma (HCC) accounts for approximately 90% of liver cancers [1]. Globally, liver cancers represent the fourth leading cause of cancer‐related death [2]. It is estimated that 50%–60% of patients with HCC will require systemic therapies during their lifespan [1]. Immunotherapy has emerged as a promising therapeutic option for HCC. Treatment with immune checkpoint inhibitors (ICIs) has achieved overall survival (OS) of 16.4 months in the first‐line setting, as evidenced by the CheckMate 040 and KEYNOTE‐224 trials [3, 4, 5]. However, fewer than 30% of patients with HCC respond to ICIs. Therefore, there is an urgent need for effective methods to identify patients most likely to benefit from immunotherapy.

Programmed cell death‐ligand 1 (PD‐L1) expression, CD8⁺ T cell infiltration, microsatellite instability (MSI), and the tumor mutational burden (TMB) are reportedly associated with the response to ICIs [6, 7, 8, 9]. Nevertheless, these biomarkers only provide limited insight into the biology of the tumor–immune interface. For example, PD‐L1 expression might reflect only a component of T cell‐related biology that is correlated with a positive immune microenvironment. The T cell‐inflamed gene expression profile (GEP), which is related to antigen presentation, cytotoxic activity, chemokine expression, and adaptive immune response, can be used to evaluate the complexities of tumor and host immune cell interactions within the tumor–immune microenvironment [10]. It has been widely accepted as an effective biomarker for predicting clinical response to ICIs across various cancers. Higher response rates and longer progression‐free survival have been demonstrated in tumors with a higher T cell‐inflamed GEP [9, 10, 11]. However, the T cell‐inflamed GEP is based on the mRNA expression of 18 immune‐related genes isolated from tumor tissues obtained via surgery or biopsy. Given the invasive nature of these procedures and the infeasibility of multiple biopsies over the course of the disease, it is particularly important to develop a noninvasive method to assess the T cell‐inflamed GEP in patients with HCC. This would facilitate the pretreatment prediction of patients who could benefit from ICI therapy and permit the longitudinal evaluation of changes in the tumor–immune microenvironment associated with immunotherapy response.

Several studies have reported that radiomics features can be used to predict tumor gene expression, pathological classification, molecular subtyping, or treatment response [12, 13, 14, 15]. A radiomics signature for CD8⁺ T cells has emerged as a promising imaging biomarker for predicting the immune phenotype of tumors and inferring clinical outcomes for ICI‐treated patients [12]. However, this study was a pan‐cancer analysis featuring only nine patients with HCC. To our knowledge, no studies have specifically focused on the efficacy of radiomics computed tomography (CT) for assessing the T cell‐inflamed GEP in HCC.

In this study, we developed and validated a radiomics model for predicting the T cell‐inflamed GEP score based on pretreatment CT images, and the model's ability to predict immunotherapy response was evaluated in patients with HCC treated with anti‐programmed cell death protein 1 (PD‐1) immunotherapy.

2. Methods

2.1. Study Design and Patient Cohorts

This study was approved by the Human Research Committee of the First Affiliated Hospital, Sun Yat‐sen University (approval number: [2021]244), Sun Yat‐sen University Cancer Center (approval number: B2025‐782‐01), and Dongguan People's Hospital (approval number: KYKT2022‐069‐C1) and was conducted following the Declaration of Helsinki and its subsequent revisions. The requirement for informed consent was waived because of the retrospective nature of this study.

This retrospective multicenter study consisted of two steps (Figure 1, and Supporting Figure S1).

Figure 1.

Study design and patient cohorts. Left panel: The FAHSYSU cohort formed the training dataset, consisting of 227 patients with both CECT and RNA‐sequencing data, used to develop the radiomics model for predicting the T cell‐inflamed GEP. The TCGA cohort formed the external test dataset, including 43 patients with both CECT and RNA‐sequencing data, used to validate the radiomics model. Right panel: The immunotherapy dataset included 40 patients with 209 evaluable lesions from three hospitals, used to evaluate the association of the radiomics model with immunotherapy response. The radiomics model, developed from the training and test datasets, was applied to each lesion in the immunotherapy cohort to predict the status of the T cell‐inflamed GEP (GEP‐high vs. GEP‐low). Subsequently, the immunotherapy response rate of target lesions was compared between the predicted GEP‐high group and the GEP‐low group. CT images in this figure are from de‐identified patients in the training cohort and the immunotherapy cohort. The organs, human body, and ball illustrations in this figure were created by the authors. CECT, contrast‐enhanced computed tomography; CT, computed tomography; DGPH, Dongguan People's Hospital; FAHSYSU, The First Affiliated Hospital, Sun Yat‐sen University; GEP, gene expression profile; GEP^hi, GEP‐high; GEP^lo, GEP‐low; RNA‐seq, RNA sequencing; SYSUCC, Sun Yat‐sen University Cancer Center; TCGA, The Cancer Genome Atlas.

First, patients with HCC who underwent contrast‐enhanced CT (CECT) before hepatectomy were included in the training dataset. We performed RNA sequencing (RNA‐seq) on tumor tissues to analyze the T cell‐inflamed GEP, which consisted of the following genes: TIGIT, CD27, CD8A, PDCD1LG2(PD‐L2), LAG3, CD274(PD‐L1), CXCL6, CMKLR1, NKG7, CCL5, PSMB10, IDO1, CXCL9, HLA‐DQA1, CD276, STAT1, HLA‐DRB1, and HLA‐E. Based on the GEP score cutoff from previous research [10], patients were categorized into the GEP‐high (GEP^hi) and GEP‐low (GEP^lo) groups. A radiomics model based on CECT was developed to predict the T cell‐inflamed GEP in the training dataset and validated using a dataset from The Cancer Genome Atlas (TCGA). Second, this model was confirmed via application to patients with HCC who received anti‐PD‐1/PD‐L1 immunotherapy.

To develop the radiomics model in the first phase, we included patients with histologically diagnosed HCC who underwent CECT within 1 month prior to hepatectomy, which was conducted at The First Affiliated Hospital, Sun Yat‐sen University (FAHSYSU) between May 2018 and May 2020. The exclusion criteria were as follows (Supporting Figure S2): prior anticancer treatment before hepatectomy; presence of other malignancies; extrahepatic metastasis; multiple tumor lesions; lack of preserved snap‐frozen tumor tissue; and unqualified snap‐frozen specimens for RNA‐seq. To validate the radiomics model, patients with histologically confirmed HCC who underwent CECT between November 1992 and November 2006 were included from the TCGA database via the Genomic Data Commons (https://gdc.cancer.gov/), and corresponding CT images were retrieved from The Cancer Imaging Archive (https://www.cancerimagingarchive.net/; Supporting Figure S2). Patients with multiple tumor lesions and those who received other locoregional therapies were excluded. Finally, the training dataset included 227 patients (227 lesions), and the test dataset included 43 patients (43 lesions), with all patients having available CECT and RNA‐seq data. The immunotherapy dataset included patients with histologically or radiologically diagnosed HCC who received anti‐PD‐1/PD‐L1 and underwent CECT within 1 month prior to immunotherapy at FAHSYSU, Sun Yat‐sen University Cancer Center, and Dongguan People's Hospital between January 2019 and January 2021 (Supporting Figure S2). Patients without measurable liver lesions and those who received other anticancer treatments within 6 months after immunotherapy were excluded. The immunotherapy dataset consisted of 40 patients with 209 evaluable tumor lesions.

2.2. RNA‐seq and Gene Expression Profiling

Details about RNA sequencing and gene expression profiling are provided in the Supporting Materials.

2.3. CECT Protocol

All patients in the training, test, and immunotherapy datasets underwent pretreatment CECT using one of the CT scanners listed in Supporting Table S1.

2.4. Radiomics Preprocessing, Feature Extraction, and Model Building

Whole‐tumor segmentation was manually performed on portal venous phase CECT images by two independent radiologists. In total, 1130 quantitative radiomics features, including shape, first‐order, and texture features, were extracted from each segmented lesion using the pyradiomics package (version 3.0.1). Feature robustness was assessed by calculating the interobserver intraclass correlation coefficient (ICC). For feature selection, we first identified features that were significantly different between the GEP^hi and GEP^lo groups and then applied a decision tree algorithm to select the most informative features. Finally, a support vector machine (SVM) classifier was employed to build the radiomics model for predicting the T cell‐inflamed GEP status. Detailed methodological descriptions are provided in the Supporting Materials.

2.5. Immunotherapy Dataset

The radiomics model for GEP prediction developed from the training dataset and validated in the test datasets was directly applied to pretreatment CECT images in the immunotherapy dataset. Lesions were categorized into the predicted GEP^hi or predicted GEP^lo group based on the radiomics signature. The difference in treatment response rates between these two groups was evaluated. As the previous model was built based on a single lesion and in consideration of the possibility of mixed responses from different lesions to immunotherapy, we compared each lesion's relative change in diameter between baseline and follow‐up images according to the Response Evaluation Criteria in Solid Tumors version 1.1 criteria at 3 and 6 months [16]. An objective response was defined as the achievement of a complete response (CR) or partial response (PR), whereas no response was defined as stable disease or progressive disease. Tumor response assessments were independently performed on target lesions by two experienced radiologists (Shi‐Ting Feng and Zhenpeng Peng, both with 20 years of experience in liver imaging), who were blinded to the patients' clinical information. Any discrepancies in assessment results were resolved by consensus. For patient‐level classification, a patient was assigned to the GEP^hi group if at least one lesion was predicted to be GEP^hi; otherwise, the patient was classified as GEP^lo.

The immunotherapy regimens consisted of either monotherapy with an anti‐PD‐1/PD‐L1 agent or combination therapy with an anti‐PD‐1/PD‐L1 agent plus lenvatinib/bevacizumab. Specifically, 10 patients received ICI monotherapy, whereas the remaining 30 patients received combination therapy. The agents used included anti‐PD‐1 monoclonal antibodies (pembrolizumab, sintilimab, or tislelizumab) and an anti‐PD‐L1 monoclonal antibody (atezolizumab). The regimen (monotherapy vs. combination) and the specific agent were selected by the treating physician based on standard clinical guidelines and the patient's individual condition. Patients were followed up after every two cycles of immunotherapy during the treatment period and then every 3 months after the end of treatment, or earlier if clinically indicated. OS was defined as the interval between the date of initiation of immunotherapy and that of death or the last follow‐up. Follow‐up was censored on February 25, 2024.

2.6. Radiomics‐Based Clinical Model Development

The differences in clinical data between the GEP^hi and GEP^lo groups in the training dataset were compared. Clinical features exhibiting significant differences were integrated with the radiomics score (Rad score) to construct a combined radiomics‐based clinical model. Clinical variables significant at p < 0.1 were included in the multivariable logistic regression model.

2.7. Statistical Analysis

The independent‐samples t‐test or Mann–Whitney U‐test was used to compare clinical characteristics and the radiomics signature between the GEP^hi and GEP^lo groups for continuous variables. The chi‐squared test or Fisher's exact test was used for between‐group comparisons of categorical variables. Variables exhibiting a potential association with the GEP status (p < 0.1 in univariable between‐group comparisons) were selected as candidates. These variables, along with the Rad score, were entered simultaneously into a multivariable logistic regression model using the forced‐entry method to build the combined radiomics‐based clinical model. The consistency of the volume of interest selected by the two radiologists was assessed using the ICC [17], with ICC ≥ 0.8 denoting high consistency, ICC = 0.5–0.79 denoting moderate consistency, and ICC < 0.5 denoting low consistency. The Delong test was used to test whether the predictive performance of these two models was statistically different. OS curves were generated using the Kaplan–Meier method and compared using the log‐rank test. The median follow‐up was calculated using the reverse Kaplan–Meier method. Because the censoring proportion was low, the restricted mean follow‐up duration was also calculated. Statistical analyses were performed using IBM SPSS Statistics version 25 (IBM, Armonk, United States) and GraphPad Prism 8 (GraphPad Software Inc., Boston, United States). A two‐sided significance level of 0.05 was used.

3. Results

3.1. Baseline Patient Characteristics

The baseline characteristics of patients in the training, test, and immunotherapy datasets are summarized in Table 1, Supporting Table S2, and Supporting Table S3, respectively. The training dataset included 227 patients for radiomics model development (Table 1), of which 67 patients (29.5%) were classified into the GEP^hi group. The TCGA dataset included 43 eligible patients for model validation (Supporting Table S2), including nine patients (20.9%) assigned to the GEP^hi group. The immunotherapy dataset consisted of 40 patients with 209 lesions (Supporting Table S3). The lesions were primarily located in the liver (n = 116, 55.5%) and lung (n = 70, 33.5%).

Table 1.

Baseline characteristics of patients in the training dataset.

Variables	Training dataset (n = 227)
Sex
Male	197 (86.8)
Female	30 (13.2)
Age (years)	52.3 ± 11.6
HBsAg
Negative	28 (12.3)
Positive	199 (87.7)
Child–Pugh
A	215 (94.7)
B	12 (5.3)
PLT count (×109/L)
> 100	197 (86.8)
≤ 100	30 (13.2)
ALT (U/L)
> 40	84 (37.0)
≤ 40	143 (63.0)
AST (U/L)
> 40	96 (42.3)
≤ 40	131 (57.7)
GGT (U/L)
> 50	121 (53.5)
≤ 50	105 (46.5)
AFP (ng/mL)
≤ 20	83 (36.7)
> 20	143 (63.3)
Tumor size (cm)
≤ 5	108 (47.6)
> 5	119 (52.4)
BCLC stage
0/A	143 (63.0)
B/C	84 (37.0)
GEP group
GEP‐high	67 (29.5)
GEP‐low	160 (70.5)

Note: Data are presented as frequency (percentage) and mean ± standard deviation. Numbers may not add up to the total because of missing values (GGT, n = 1; AFP, n = 1).

Abbreviations: AFP, α‐fetoprotein; ALT, alanine aminotransaminase; AST, aspartate aminotransaminase; BCLC, Barcelona Clinic Liver Cancer; GEP, gene expression profile; GGT, γ‐glutamyltransferase; HBsAg, hepatitis B surface antigen; PLT, platelet.

3.2. Radiomics Feature Selection

Of the 1130 radiomics features extracted from the training dataset, 576 features that significantly differed between the GEP^hi and GEP^lo groups were selected for further analysis. Subsequently, seven features were chosen using a decision tree with a maximum depth of 3. The Gini index was used to measure the contribution degree of each feature, and features with contribution degree > 0 were selected. All selected features were texture features. Details of these features are listed in Supporting Table S4, and their heatmaps are presented in Supporting Figure S3.

3.3. Interobserver Reproducibility of Radiomics Feature Extraction

The median interobserver ICC for the 1130 radiomics features was 0.998. The interobserver ICC was ≥ 0.8 for 95.04% of the features, between 0.5 and 0.79 for 4.43% of the features, and < 0.5 for 0.53% of the features. The seven selected features also demonstrated robust reproducibility, with interobserver ICCs ranging from 0.849 to 1.000 (all p < 0.001).

3.4. Radiomics Model Development and Validation

The radiomics model for GEP was developed using an SVM classifier on the training dataset and subsequently applied to the test dataset. Each patient's Rad score in the training and test cohorts is presented in Figure 2a,b, respectively. The cutoff determined by the maximum Youden index was 0.255. Receiver operating characteristic (ROC) curves for the training and testing datasets are shown in Figure 2c,d, respectively. In the training cohort, the area under the ROC curve (AUC) was 0.824 (95% confidence interval [CI] = 0.757–0.892), with accuracy of 80.2%, sensitivity of 74.6%, and specificity of 82.5%. In the test cohort, the AUC was 0.827 (95% CI: 0.674–0.980), with accuracy of 69.8%, sensitivity of 88.9%, and specificity of 64.7%. The calibration curves for the model in both the training and test cohorts were aligned with the standard lines (Figure 2e,f).

Figure 2.

Performance of the CT‐based radiomics model in predicting T cell‐inflamed GEP for HCC. (a) Each patient's Rad score in the training dataset and (b) test dataset. (c) The ROC curves for the training and (d) test datasets. (e) The calibration curves for the model in the training dataset and (f) test dataset. (g) The decision curves for the training and (h) test datasets. AUC, area under the receiver operating characteristic curve; CT, computed tomography; GEP, gene expression profile; HCC, hepatocellular carcinoma; Rad score, radiomics score; ROC, receiver operating characteristic curve.

3.5. Radiomics‐Based Clinical Model Development and Comparison with the Radiomics Model

Among all clinical variables and the Rad score, only the Rad score emerged as an independent predictor of the GEP status (p < 0.001; Table 2). The ROC curve for the radiomics‐based clinical model in predicting GEP is presented in Supporting Figure S4a. The AUC for this model was 0.831 (95% CI: 0.766–0.896). There was no significant difference in the AUC between the radiomics‐based clinical model and the radiomics model (p = 0.772). The calibration curve for the radiomics‐based clinical model in the training cohort was aligned with the standard lines (Supporting Figure S4b). Decision curve analysis indicated that the radiomics‐based clinical model did not achieve a higher overall net benefit than the radiomics model, suggesting the high performance of the radiomics model in predicting GEP (Supporting Figure S4c).

Table 2.

Multivariable logistic regression model of the radiomics score and clinical variables.

Variables	OR	95% CI	p
Radiomics score per 1%	1.05	(1.03, 1.06)	< 0.001
Sex		(0.58, 4.17)	0.386
Male	1.00
Female	1.55
PLT count (×109/L)		(0.20, 1.46)	0.224
≤ 100	1.00
> 100	0.54
ALT (U/L)		(0.33, 1.52)	0.373
≤ 40	1.00
> 40	0.71
Tumor size (cm)		(0.49, 2.12)	0.952
≤ 5	1.00
> 5	1.02
HBV DNA (IU/mL)		(0.25, 1.22)	0.144
≤ 100	1.00
> 100	0.55

Abbreviations: ALT, alanine aminotransaminase; CI, confidence interval; HBV, hepatitis B virus; OR, odds ratio; PLT, platelet.

3.6. Application of the Radiomics Model to the Immunotherapy Dataset

Detailed information regarding the scanning equipment used at the three centers is presented in Supporting Table S1. Approximately one‐fifth (n = 37, 18.4%) of the lesions in the immunotherapy dataset exhibited treatment response after 3 months of immunotherapy, whereas 26 lesions (18.3%) displayed a response after 6 months (Table 3). The radiomics model was applied to each lesion from patients undergoing immunotherapy. Across all sites, 39.8% of the lesions (n = 80) were predicted to be GEP^hi by the radiomics model. Within the predicted GEP^hi group, CR or PR was achieved in 21.3% of the lesions, compared with 16.5% in the predicted GEP^lo group, after 3 months of immunotherapy (p = 0.398). After 6 months, 42.3% of the lesions (n = 60) were predicted to be GEP^hi. The response rate in the predicted GEP^hi group was significantly higher than that in the predicted GEP^lo group (30.0% vs. 9.7%, p < 0.05; Figure 3a).

Table 3.

Clinical response of each lesion in the immunotherapy dataset between different predictive T cell‐inflamed GEP groups.

Variables	Lesion in all sites		p a	Lesion in the liver		p a
	GEPhi	GEPlo		GEPhi	GEPlo
Three‐month clinical responseb	80 (100.0)	121 (100.0)	0.398	39 (100.0)	75 (100.0)	0.123
CR	0 (0.0)	0 (0.0)		0 (0.0)	0 (0.0)
PR	17 (21.3)	20 (16.5)		11 (28.2)	12 (16.0)
SD	43 (53.8)	70 (57.9)		11 (28.2)	36 (48.0)
PD	20 (25.0)	31 (25.6)		17 (43.6)	27 (36.0)
Six‐month clinical response	60 (100.0)	82 (100.0)	0.002	27 (100.0)	43 (100.0)	< 0.001
CR	3 (5.0)	1 (1.2)		3 (11.1)	0 (0.0)
PR	15 (25.0)	7 (8.5)		10 (37.0)	4 (9.3)
SD	23 (38.3)	42 (51.2)		3 (11.1)	17 (39.5)
PD	19 (31.7)	32 (39.0)		11 (40.7)	22 (51.2)

Note: Data are presented as frequency (percentage).

Abbreviations: CR, complete response; GEP, gene expression profile; GEP^hi, GEP‐high; GEP^lo, GEP‐low; PD, progressive disease; PR, partial response; SD, stable disease.

^a The difference was compared between CR + PR and SD + PD in the GEP^hi and GEP^lo groups.

^b Eight lesions in two patients were examined only at the 6‐month time point.

Figure 3.

Association of predicted T cell‐inflamed GEP status by radiomics model with immunotherapy response and overall survival of HCC patients. (a, b) The response rate of (a) all the lesions and (b) liver lesions in the predicted GEP^hi group was significantly higher than in the predicted GEP^lo group after 6 months of immunotherapy. (c) Kaplan–Meier curves illustrating the overall survival of patients with or without lesions predicted to be GEP^hi in the immunotherapy dataset. The predictive GEP groups were defined according to the radiomics model based on pre‐treatment CT images for HCC patients treated by anti‐PD‐1 immunotherapy. The organs and human body illustrations in this figure were created by the authors. CR, complete response; CT, computed tomography; GEP, gene expression profile; GEP^hi, GEP‐high; GEP^lo, GEP‐low; HCC, hepatocellular carcinoma; PD‐1, programmed cell death protein 1; PR, partial response.

In the analysis of liver lesions within the immunotherapy dataset, 34.2% of the lesions (n = 39) were classified into the GEP^hi group. After 3 months of immunotherapy, 28.2% of these lesions responded, versus 16.0% of lesions in the predicted GEP^lo group (p = 0.123). After 6 months of immunotherapy, 38.6% (n = 27) of the liver lesions were predicted to be GEP^hi. The response rate in the predicted GEP^hi group was significantly higher than that in the predicted GEP^lo group (48.1% vs. 9.3%, p < 0.001; Figure 3b, and Table 3). Given the significantly different response rates of liver lesions at 6 months after immunotherapy between the predicted GEP^hi and GEP^lo groups, we further evaluated the performance of the predictive GEP^hi group in predicting liver lesion response rates. The AUC for the GEP status predicted by the radiomics model in classifying immunotherapy response was 0.750 (95% CI: 0.614–0.886).

The restricted mean follow‐up duration was 29.95 months (95% CI: 24.31–35.59). Median OS was 11.7 months (95% CI: 5.0–18.5). Patients with predicted GEP^hi lesions had median OS of 17.5 months (95% CI: 9.1–25.9), versus 8.3 months among patients without GEP^hi lesions (95% CI: 6.1–10.4 months, p < 0.05; Figure 3c).

4. Discussion

With the increasingly recognized role of immunotherapy in the treatment of HCC, effective pretreatment biomarkers for identifying potential responders are essential. In our study, we integrated RNA transcriptomic data with CECT images to develop a radiomics model to predict the T cell‐inflamed GEP status in patients with HCC. This model was validated using an external test dataset (TCGA cohort), and it displayed potential in predicting the immunotherapy response among patients treated with anti‐PD‐1 therapy in a clinical cohort.

The use of immunotherapy has led to a paradigm shift in HCC management, but the response rate remains relatively low [18]. Many biomarkers, such as PD‐L1, TMB, and MSI, have not proven predictive of treatment response. The T cell‐inflamed GEP, which includes genes closely related to antigen presentation, cytotoxic activity, chemokine expression, and adaptive immune response, provides a comprehensive evaluation of tumor microenvironment features [10]. It has been utilized as a biomarker for predicting ICI responsiveness in multiple large‐scale clinical trials [9, 10, 11]. To our knowledge, this is the first study to establish a CT radiomics model for predicting T cell‐inflamed GEP in HCC. In our study, the Rad score consisted of several textural features that quantified the complexity of the texture (original_glcm_Imc1), the homogeneity or heterogeneity of the tumor image (log‐sigma‐1‐0‐mm‐3D_glcm_Idn, original_gldm_DependenceEntropy, log‐sigma‐1‐0‐mm‐3D_glszm_GrayLevelNonUniformityNormalized), and the distribution of the different gray levels (wavelet‐LHH_glszm_SmallAreaLowGrayLevelEmphasis, wavelet‐HLH_glrlm_HighGrayLevelRunEmphasis, log‐sigma‐1‐0‐mm‐3D_glrlm_LongRunHighGrayLevelEmphasis), which are increasingly recognized as imaging features of the underlying histological diversity, including variations in cellular density, necrosis, stromal content, and vascularity. Based on these features, we established an effective model for predicting the T cell‐inflamed GEP in HCC. Previous studies explored the relationship between CT texture features and the tumor microenvironment. In Sun's study [12], homogeneous tumors, lower CT density, and peripheral rings reflected a high CD8⁺ T cell score in various advanced solid tumors. In Trebeschi et al.'s study [19], advanced melanoma and non‐small‐cell lung cancer with more nonuniform morphological profiles, heterogeneous density patterns, and compact borders were more likely to respond to immunotherapy. As an SVM classifier was used in the modeling in our study, it was challenging to clarify the monotonicity between features and the T cell‐inflamed GEP, making it difficult to intuitively understand their relationship [20]. However, the AUC for the external test set was similar to that in the training dataset, supporting the performance of the model.

Our study sought to connect CT‐based radiomics features with RNA transcriptomic profiles in patients with HCC and successfully demonstrated that radiomics can effectively predict the response to immunotherapy in routine clinical practice. Although six studies have investigated the use of radiologic or histologic imaging features as biomarkers to predict immune infiltration or immune expression signatures in HCC [14, 21, 22, 23, 24, 25], none evaluated the potential of these imaging biomarkers to directly predict immunotherapy response and patient survival. Zeng et al. developed and validated artificial intelligence‐based pathology models to predict the activation of several immune and inflammatory gene signatures in patients with HCC and achieved high performance (AUC = 0.81–0.92) [22]. This is the only published HCC study to assess the association between imaging data and transcriptomic data, but it focused on histologic data that cannot be acquired before immunotherapy. Moreover, it remains unclear whether their models can be used to identify immunotherapy responders. In our study, the AUC of the radiomics model for predicting the response of liver lesions to immunotherapy was 0.750 in multicenter data, supporting the model's potential utility in predicting tumor microenvironment features and clinical outcomes of immunotherapy for patients with HCC. The 6‐month endpoint was selected to capture the delayed clinical benefit typical of anti‐PD‐1 therapy, which likely accounts for the weaker association observed at 3 months. In Sun et al.'s study [21], the ability of gadoxetic acid‐enhanced magnetic resonance imaging (MRI) features to predict tumor‐infiltrating CD8⁺ T cells and PD‐L1 expression in HCC was examined. Although different combinations of MRI features were associated with the objective response to immunotherapy in 34 patients, only six patients responded, and the median follow‐up was only 8 months. Moreover, these models were based on imaging features assessed by radiologists, which could be influenced by investigator‐related factors to some extent.

Our radiomics model was developed and validated exclusively using intrahepatic HCC lesions from surgical cohorts. This is because patients undergoing resection typically present with intrahepatic lesions, whereas extrahepatic (e.g., pulmonary) lesions were rarely available for initial model development. Patients receiving systemic immunotherapy often have advanced disease with distant metastases. To address a common clinical scenario and explore the model's broader utility, we applied it to lesions in various anatomical sites, including the liver and lungs, within the immunotherapy cohort. This was based on the biological hypothesis that the T cell‐inflamed GEP, as a signature of the tumor–immune microenvironment, might be reflected by inherent tumor radiomics phenotypes that are potentially transferable across different anatomical sites for the same cancer type. The results provide preliminary support for this approach. Lesions predicted as GEP^hi exhibited a higher response rate across all sites, with the strongest predictive performance observed for liver lesions. However, we fully acknowledge the potential limitations of this extrapolation. Differences in organ parenchyma and background tissue density between the liver and other sites could influence radiomics feature interpretation. Furthermore, biological heterogeneity between primary and metastatic lesions might not be fully captured by a model trained solely on intrahepatic tumors. The observed site‐dependent performance underscores the need for caution and suggests that the model's predictive power is likely optimal for intrahepatic HCC. Future studies using larger, multicenter cohorts that include explicitly annotated lesions from various anatomical sites are essential to rigorously validate the model's generalizability.

Our study had several limitations. First, the sample size of the test dataset was relatively small because of the requirements for both CT images and RNA‐seq data for each patient. In the multicenter immunotherapy cohort, although the real‐world setting strengthened clinical relevance, only a limited number of patients achieved an objective response (n = 9). Therefore, although the survival trend was biologically reasonable, it should be considered exploratory given the small number of objective responses and limited statistical power. Second, this was a retrospective study, and the CT images were acquired from different scanners in different centers, which might have influenced the extraction of imaging features to some extent. Nevertheless, the performance of our model, developed from the training cohort, was validated in the external test dataset, implying that it is generalizable to different datasets, scanners, and populations. Notably, the model was developed using Python 3.6 and Scikit‐learn 0.22, which are currently legacy versions. This dependency could potentially limit the direct reproducibility of the model, and future validation is recommended in updated computational environments.

5. Conclusions

Our findings suggest that the CT‐based radiomics model can effectively predict the T cell‐inflamed GEP in HCC. The predicted GEP was correlated with immunotherapy response and patient survival. These results suggest that this CT‐based radiomics model can potentially support treatment decision‐making for patients with HCC.

Author Contributions

Mimi Tang: writing – original draft, conceptualization, formal analysis, data curation, methodology. Kaiyu Sun: conceptualization, funding acquisition. Dasheng Wu: methodology, software, formal analysis, writing – original draft. Yujian Zou: data curation. Jian Zhou: data curation. Bin Li: methodology, formal analysis. Huanjing Hu: methodology, formal analysis. Zhenpeng Peng: writing – review and editing, supervision. Jifei Wang: writing – review and editing, funding acquisition. Bingsheng Huang: conceptualization, investigation, validation, methodology. Shi‐Ting Feng: funding acquisition, conceptualization, supervision, writing – review and editing, project administration, resources. Shuling Chen: conceptualization, funding acquisition, writing – review and editing, resources, project administration.

Ethics Statement

This study was approved by the Human Research Committee of the First Affiliated Hospital, Sun Yat‐sen University (approval number: [2021]244), Sun Yat‐sen University Cancer Center (approval number: B2025‐782‐01), and Dongguan People's Hospital (approval number: KYKT2022‐069‐C1) and was conducted following the Declaration of Helsinki and its subsequent revisions. The macroscopic specimen photograph in Figure 1 and Supporting Figure S1 were obtained from surgical specimens at The First Affiliated Hospital, Sun Yat‐sen University and photographed by the authors. The image has been de‐identified. Use of the photograph was approved by the Institutional Review Board of The First Affiliated Hospital, Sun Yat‐sen University (approval number: [2021]244).

Consent

The requirement for informed consent was waived by the Human Research Committee of the First Affiliated Hospital, Sun Yat‐sen University (No. [2021]244), Sun Yat‐sen University Cancer Center (No. B2025‐782‐01), and Dongguan People's Hospital (No. KYKT2022‐069‐C1) due to the retrospective nature of the study. For the macroscopic specimen photograph presented in Figure 1 and Supporting Figure S1, informed consent for publication was obtained from the patient as part of the signed informed consent for surgical treatment and the research use of specimens.

Conflicts of Interest

Professor Shi‐Ting Feng is the member of the Health Care Science Editorial Board. To minimize bias, he was excluded from all editorial decision‐making related to the acceptance of this article for publication. The remaining authors declare no conflict of interest.

Supporting information

Figure S1: Flow chart of the study. CT images in this figure are from de‐identified patients in the training cohort and the immunotherapy cohort. The ball illustrations in this figure were created by the authors. Abbreviations: CECT, contrast‐enhanced computed tomography; CR, complete response; CT, computed tomography; GEP^hi, GEP‐high; GEP^lo, GEP‐low; PR, partial response; VOI, volume of interest

Figure S2: Flow diagram of the study population. The exclusion criteria of the three patient cohorts (training cohort, test cohort and immunotherapy cohort). Abbreviations: CT, computed tomography; FAHSYSU, The First Affiliated Hospital, Sun Yat‐sen University; HCC, hepatocellular carcinoma; PD‐1, programmed cell death protein 1; TCGA, The Cancer Genome Atlas

Figure S3: Heatmaps of the radiomics features. (a) 1130 radiomic features extracted from the training dataset. (b) 576 features selected by t‐test or Mann‐Whitney U‐test. (c) 7 features selected by the decision tree

Figure S4: Performance of the radiomics‐based clinical model in predicting T cell‐inflamed GEP for HCC. (a) ROC curve of the radiomics‐based clinical model. (b) Calibration curves of the radiomics‐based clinical model in the training cohort were well‐matched with the standard lines. (c) Decision curves of the radiomics‐based clinical model. Abbreviations: AUC, area under the receiver operating characteristic curve; GEP, gene expression profile; HCC, hepatocellular carcinoma; ROC, receiver operating characteristic curve

Supporting File 1

Supporting File 2

Tang M., Sun K., Wu D., et al., “A Computed Tomography Radiomics Approach for Assessing the T Cell‐Inflamed Gene Expression Profile in Hepatocellular Carcinoma and Its Association With Immunotherapy Response,” Health Care Science 0 (2026): 1‐10. 10.1002/hcs2.70106.

Data Availability Statement

The datasets used and analyzed during the current study are available from the corresponding author upon reasonable request.