Prediction of PD-1 Expression and Outcomes of Combined Therapy in Hepatocellular Carcinoma: an MRI-Based Radiomics Approach

Abstract

This study aims to assess the value of clinical and MRI radiomics features in noninvasively predicting programmed cell death protein 1 (PD-1) expression level and the response to anti-PD-1 immunotherapy combined with transcatheter arterial chemoembolization (TACE) in hepatocellular carcinoma (HCC). A total of 107 HCC patients (85 in training set and 22 in validation set) with PD-1 positive (n = 41) and negative (n = 66) were enrolled. Radiomics features were extracted from T2-weighted, fat suppression T2-weighted, contrast-enhanced, and diffusion-weighted images. A clinical model was constructed based on independent clinical risk factors (p < 0.05), while a radiomics model was developed utilizing the optimal radiomics signature. A radiomics nomogram model for predicting PD-1 integrated the significant clinical and radiomics features. The performance of nomogram was evaluated with respect to its discrimination and clinical utility and compared with that of clinical and radiomics prediction models. The radiomics nomogram, combining the clinical factors with the Rad-score, had best prediction performance (area under the curve [AUC]: 0.95 in the training set; AUC: 0.86 in the validation set). Decision curve analysis (DCA) showed that the nomogram exhibited superior accuracy in clinical assessment compared to the other models. The predicted high-risk group of PD-1 had longer overall survival (OS) than the predicted low-risk group after receiving anti-PD-1 therapy combined with TACE. The MRI-based radiomics nomogram performed well for identifying high-risk PD-1 group for combination therapy and may inform personalized treatment decision-making strategies.

Supplementary Information

The online version contains supplementary material available at 10.1007/s10278-024-01381-7.

Introduction

Hepatocellular carcinoma (HCC) often has a poor prognosis because it is frequently diagnosed at an advanced stage and has complex pathogenesis [1]. The available exclusive therapeutic option for advanced HCC patients centers on novel-targeted therapies, particularly tyrosine kinase inhibitors used either independently or in combination with immune checkpoint inhibitors [2]. Recently, the treatment landscape of HCC has been significantly transformed by immunotherapy, exhibiting a response rate of 15–20% in contrast to only a 2% response rate of multi-kinase inhibitor sorafenib [3]. Thus, the emerging role of immunotherapy has substantially changed the therapeutic strategies for advanced HCC.

The efficacy of immunotherapy depends on immunoprofiling and immune status of tumors, with reports indicating only 20–50% response rate among patients with advanced solid tumors [4]. There is therefore a need for the heterogeneous markers to identify patients who are most likely to benefit from immunotherapy. Programmed cell death protein 1 (PD-1) and programmed cell death protein ligand 1 (PD-L1) as immune checkpoint pathway members have attracted increasing attention in immunotherapy [5].

PD-1 has a negative regulatory influence on antigen responses. When binding to PD-L1, it induces the dysfunction or apoptosis of T lymphocytes, thereby suppressing the antitumor immune response and leading to tumor escape [6]. In recent years, drugs blocking the PD-1/PD-L1 pathway have made remarkable achievements in improving patient survival among a variety of malignant tumors [7]. These agents suggest that the evaluation of PD-1 expression can enhance treatment outcomes for HCC patients.

Immunohistochemical staining is a common technique for assessment of molecular distribution and localization in tumors. However, the method to obtain tumor specimens is invasive and limited. Computed tomography (CT) or magnetic resonance imaging (MRI) are noninvasive imaging techniques in diagnosing and staging HCC. Radiomics, a process by which large numbers of digital features are extracted from medical images, is motivated by the concept that biomedical images could indirectly reflect the underlying molecular properties of tumors [8]. The potential utility of multiparametric MRI features for identification of novel predictors and the efficacy of immunotherapy have been confirmed in encouraging studies. Recently, it was shown that qualitative imaging traits from MRI are associated with expression levels of immunohistochemical HCC markers and therapeutic targets [9]. Several reports have explored the correlation between MRI features and PD-1/PD-L1 expression as well as responses to anti-PD-1 or anti-PD-L1 immunotherapy. Nevertheless, a direct relationship between MRI radiomics features and predicting PD-1 expression in anti-PD-1 immunotherapy remains unclear. Developing a precise approach to qualitatively and quantitatively assess the marked heterogeneity using routine MRI could potentially impact the management of HCC patients. Due to the emerging role of immunotherapy, there is a need for advanced HCC characterization to appropriately predict patient prognostication.

Therefore, our study aimed to construct and validate a radiomics prediction model based on preoperative multi-sequence MRI in HCC patients. The stability of this model was validated, and biological interpretation of the heterogeneous radiomics features was performed by immunohistochemical staining in the prospective cohort. In summary, we established a radiomics prediction model integrating clinical prognosis prediction and MRI radiomics features to assess the PD-1 expression and anti-PD-1 immunotherapy outcomes in HCC patients.

Materials and Methods

Patient Enrolment

Our center collected 179 HCC patients between October 2012 and July 2020 at the Affiliated Hospital of Nantong University. This retrospective single-center study received approval from ethics committee and informed consent from all patients. Inclusion criteria encompassed (a) patients who were pathologically diagnosed with HCC, (b) patients with complete clinical data, and (c) patients with available abdominal enhanced MRI within 3 months before surgery (resection or TACE) for HCC. Exclusion criteria were (a) maximum tumor diameter < 5 mm, (b) the absence of diffusion-weighted imaging (DWI), and (c) unsatisfactory image quality due to motion and image artifacts. To minimize the influence of partial volume effects and reflect true tissue-specific image texture, a maximum tumor diameter threshold was set at 5 mm [10]. Ultimately, 107 HCC patients comprised our study population (Fig. 1). The clinical patient characteristics, including sex, age, prothrombin induced by vitamin K absence-II (PIVKA-II), γ-glutamyl transpeptidase isoenzyme II (GGT-II), and alpha-fetoprotein (AFP), were derived from admission records. Details are described in Table 1.

Fig. 1.

Flow chart of HCC patients enrolled in the study

Table 1.

Clinical characteristics of the participants

	[All]	PD-1 (–)	PD-1 ( +)	p-value
	N = 107	N = 66	N = 41
Sex				0.581
Male	90 (84.1%)	54 (81.8%)	36 (87.8%)
Female	17 (15.9%)	12 (18.2%)	5 (12.2%)
Agea (years)	61.0 [57.0; 71.0]	62.5 [58.2; 71.0]	59.0 [56.0; 73.0]	0.440
AFU				0.740
-	84 (78.5%)	53 (80.3%)	31 (75.6%)
+	23 (21.5%)	13 (19.7%)	10 (24.4%)
PIVKA-II				0.002
-	53 (49.5%)	41 (62.1%)	12 (29.3%)
+	54 (50.5%)	25 (37.9%)	29 (70.7%)
GGT-II				0.006
-	59 (55.1%)	44 (66.7%)	15 (36.6%)
+	18 (16.8%)	10 (15.2%)	8 (19.5%)
+  +	30 (28.0%)	12 (18.2%)	18 (43.9%)
HBV-DNA				0.006
-	56 (52.3%)	42 (63.6%)	14 (34.1%)
+	51 (47.7%)	24 (36.4%)	27 (65.9%)
AFP				0.219
-	86 (80.4%)	56 (84.8%)	30 (73.2%)
+	21 (19.6%)	10 (15.2%)	11 (26.8%)
CEA				0.805
-	92 (86.0%)	57 (86.4%)	35 (85.4%)
+	15 (14.0%)	9 (13.6%)	6 (14.6%)
CA199				0.530
-	95 (88.8%)	60 (90.9%)	35 (85.4%)
+	12 (11.2%)	6 (9.09%)	6 (14.6%)
HBsAg				0.784
-	29 (27.1%)	19 (28.8%)	10 (24.4%)
+	78 (72.9%)	47 (71.2%)	31 (75.6%)
HBsAb				0.367
-	82 (76.6%)	53 (80.3%)	29 (70.7%)
+	25 (23.4%)	13 (19.7%)	12 (29.3%)
HBeAg				0.670
-	87 (81.3%)	55 (83.3%)	32 (78.0%)
+	20 (18.7%)	11 (16.7%)	9 (22.0%)
HBeAb				0.003
-	15 (14.0%)	15 (22.7%)	0 (0.00%)
+	92 (86.0%)	51 (77.3%)	41 (100%)
HBcAb				0.200
-	6 (5.61%)	2 (3.03%)	4 (9.76%)
+	101 (94.4%)	64 (97.0%)	37 (90.2%)
Cirrhosis				0.833
-	50 (46.7%)	31 (47.0%)	19 (46.3%)
+	57 (53.3%)	35 (53.0%)	22 (53.7%)
Number:				0.308
Solitary	85 (79.4%)	55 (83.3%)	30 (73.2%)
Multiple	22 (20.6%)	11 (16.7%)	11 (26.8%)
Maxdiameter (cm)				0.508
≤ 3	49 (45.8%)	30 (45.5%)	19 (46.3%)
3–5	26 (24.3%)	14 (21.2%)	12 (29.3%)
> 5	32 (29.9%)	22 (33.3%)	10 (24.4%)
Envelope				0.029
-	17 (15.9%)	15 (22.7%)	2 (4.88%)
+	90 (84.1%)	51 (77.3%)	39 (95.1%)
MVI				0.936
-	93 (86.9%)	58 (87.9%)	35 (85.4%)
+	14 (13.1%)	8 (12.1%)	6 (14.6%)
PVTT				0.001
Absent	85 (79.4%)	60 (90.9%)	25 (61.0%)
Present	22 (20.6%)	6 (9.09%)	16 (39.0%)

AFU, α-L-fucosidase; PIVKA-II, prothrombin induced by vitamin K absence-II; GGT-II, γ-glutamyl transpeptidase isoenzyme II; AFP, α-fetoprotein; Maxdiameter, maximum diameter of tumor; PVTT, portal vein tumor thrombus; MVI, microvascular invasion. ^aMedian [25% interquartile; 75% interquartile]

MRI Image Acquisition

The scan protocols of MRI are provided in the supplementary material methods 1. The MRI exams included multiphasic contrast-enhanced MRI with liver-specific gadoxetic acid injection (Gd-EOB-DTPA), diffusion-weighted images (DWI), and other routine sequences. The developed radiomics analysis pipeline developed in this study is illustrated in Fig. 2.

Fig. 2.

Workflow of radiomics analysis

Qualitative and Quantitative MRI Features Analysis

Two independent abdominal radiologists performed qualitative MRI feature analysis blindly to the clinical and pathological data. The location and size of the HCC lesion that underwent histopathologic sampling were identified on MRI in consensus. Through delineating tumors’ outermost boundaries manually in 3D Slicer software (version 4.13.0; https://www.slicer.org), 3D segmentations isolated the whole tumor region of interest (ROI) while excluding adjacent extra-lesional tissue to avoid partial volume effects. In addition, areas of necrosis are included, if present. ROIs were resampled to 1 × 1 × 1 mm³ voxels to maintain scales and directions in 3D features [11].

Quantitative ROI measurements were conducted using the extension module in 3D Slicer. The features of shape, first-order statistics, gray-level cooccurrence matrix (GLCM), gray-level dependence matrix (GLDM), gray-level run length matrix (GLRLM), gray-level size zone matrix (GLSZM), and wavelet were calculated for each MRI sequence’s ROI. In total, 1130 MR radiomics features were extracted from the ROI from the T2-weighted images (T2WI), fat suppression T2WI, contrast-enhanced images, and DWI. Radiomics data were preprocessed for repeatability using z-score and mean normalization. Twenty MRI images were randomly chosen to measure for the interobserver and intra-observer agreements in order to improve reproducibility and repeatability of radiomics features. Two radiologists independently segmented ROI on multi-sequence MRI during the same period. Besides, one radiologist performed the same produce for delineating ROI again 2 weeks later. Intraclass correlation coefficient (ICC) was taken to evaluate interobserver and intra-observer reproducibility in the extracted radiomics features. An ICC > 0.75 indicated good agreement. Moreover, Spearman correlation analysis was used to detect hidden correlations in texture parameters. These radiomics features which were demonstrated strong agreement (ICC > 0.75, Spearman ρ < 0.9) were integrated into the radiomics model.

Immunohistochemistry

Tissue microarray immunohistochemical was used to investigate the expression of PD-1 in tumor and peritumoral tissues. One-hundred seventy-nine cases of HCC patients’ tissues were prepared into 2 tissue microarrays, and immunohistochemical staining was taken to detect PD-1 expression in immunocompetent T cells. The details of immunohistochemical staining are shown in supplementary material. Then patients were classified into two groups based on PD-1 expression level. The positive group comprised patients with tumors that showed PD-1 protein overexpression, while the negative group showed the absence of protein overexpression (Fig. 3).

Fig. 3.

Examples of tumor and peritumoral tissues stained with PD-1 antibody on one piece of immunohistochemical microarray with HCC patient samples (A). B1, PD-1-positive tumor tissue under a 4 × microscope; B2, at 10 × ; C1, PD-1-positive peritumoral tissue under a 4 × microscope; C2, at 10 × ; D1, PD-1-negative tumor tissue under a 4 × microscope; D2, at 10 × ; E1, PD-1-negative peritumoral tissue under a 4 × microscope; E2, at 10 × . B and C are from a case. D and E are from another case

Development of the Predictive Model

All participants were randomly divided into the training and testing sets using an 8:2 ratio. Dimension reduction and feature selection were conducted to minimize overfitting from radiomics features during modeling. First, features with ICC scores above 0.75 were selected for the model training. Second, we used least absolute shrinkage and selection operator (LASSO) algorithm to reduce dimension and the bias. Then LASSO regression was applied to identify the optimal predictive features for model building. This shrinkage approach is effective in reducing the bias and improving the prediction accuracy [12]. Finally, a few radiomics features with prominent contributions to the predictive model were selected. The regularization parameters of each selected features were weighted by LASSO logistic regression, and a radiomics score (Rad-score) was calculated.

For the clinical model, the individual clinical variables, including sex, age, vitamin K absence-II (PIVKA-II), γ-glutamyl transpeptidase isoenzyme II (GGT-II), and alpha-fetoprotein (AFP), underwent statistical testing (t-test or chi-square tests as appropriate) to identify significant contributors. The significant clinical features (p < 0.05) were integrated into clinical predictive model. Moreover, we constructed a univariable logistic regression model to filtrate relevant clinical features.

To improve the accuracy of prediction, multivariable logistic regression models were created, incorporating multisequence MRI radiomics, clinic, and clinic-radiology. The performance of each model in differentiating PD-1 positive from negative was evaluated by AUC in both the training and validation sets.

Development and Validation of a Radiomics Nomogram Model

Significant clinical variables and radiomics features were integrated into an MRI-based radiomics nomogram model to predict PD-1 expression. The performance characteristics of the nomogram were investigated by a calibration curve. Moreover, the clinical and radiomics usefulness of the nomogram was assessed using a decision curve based on the clinical factors and radiomics signature, and the net benefits were calculated.

To validate the nomogram’s performance, we collected radiomic and clinical data from a cohort of 12 HCC patients who underwent combination therapy and subsequently determined their overall survival (OS). These 12 participants were divided into the high-risk group and the low-risk group based on a scoring system. The OS between predicted positive and negative patients was compared using Kaplan–Meier analysis.

Statistical Analysis

Statistical analysis was conducted with SPSS (version 25), R software (version 4.2.1), and Python (version 3.7.1) with statistical significance set at p < 0.05.

Results

The demographics of the participants are shown in Table 1. The mean patient age was 61 years (SD: 10.2 years; range: 29–91 years). Out of 107 of the HCC patients included, 41 cases were PD-1 + and 66 cases PD-1 − , respectively. A significant difference in clinical factors was observed between the groups regarding PIVKA-II, GGT-II, HBV-DNA, HBeAb, envelope, and PVTT (p < 0.05) in the training set. Based on multiple logistic regression analysis, PIVKA-II (p = 0.0021), GGT-II (p = 0.0007), HBV-DNA (p = 0.0047), and PVTT (p = 0.0024) remained as independent predictors in the clinical factor model. This suggests that elevated levels of PIVKA-II, GGT-II, and HBV-DNA, along with the presence of PVTT, are associated with a positive PD-1 expression profile in HCC tumors. The clinical prediction model exhibited an AUC of 0.84 in the training cohort, while the model yielded an AUC of 0.76 in the validation cohort (Fig. 4).

Fig. 4.

Receiver operating characteristic curves of the clinical model, radiomics signature and radiomics nomogram in the training (A), and test (B) datasets

In total, 1130 radiomics features were extracted from T2WI, contrast-enhanced images, fat-suppressed images, and DWI sequence. Radiomics features with ICCs above 0.75 and Spearman ρ below 0.9 were considered significant. Further details are provided in the supplementary material. Then LASSO logistic regression algorithm was utilized to identify the most valuable features. Finally, six radiomics features were included in the Rad-score calculation to evaluate the performance of radiomics predictive models. The Rad-score was obtained using the following equation:

$$\begin{array}{cc}\hfill Rad-score& =(0.2151)\times waveletHHH\_glszm\_GrayLevelNonUniformityNormalized\hfill \\ \hfill & +(-0.4723)\times wavelet-HHL\_ngtdm\_Contrast\hfill \\ \hfill & +0.3131\times wavelet-LHH\_glszm\_SmallAreaLowGrayLevelEmphasis\hfill \\ \hfill & +(-0.4183)\times wavelet-HHL\_glszm\_SmallAreaEmphasis\hfill \\ \hfill & +0.3197\times wavelet-LHH\_gldm\_DependenceVariance\hfill \\ \hfill & +0.4320\times wavelet-LLL\_firstorder\_90Percentile\hfill \\ \hfill & +(-0.2865)\times log-sigma-6-0-mm-3D\_glrlm\_ShortRunHighGrayLevelEmphasis\hfill \\ \hfill & +(-0.5175)\times wavelet-LHL\_glrlm\_RunEntropy+1.0363\hfill \end{array}$$

The radiomics prediction model demonstrated discriminative ability with AUCs of 0.91 and 0.77 in the training and test cohort, respectively (Fig. 4).

To enhance efficiency and user-friendliness, we integrated significant clinical factors and Rad-score to develop a nomogram model (Fig. 5A). The radiomics nomogram model exhibited compelling diagnostic performances for predicting PD-1 positivity, achieving an AUC of 0.95, an accuracy of 88%, sensitivity and specificity of 88% in the training cohort, and 0.86, 73%, 88%, and 74% in test cohort, outperforming the other traditional models (Table 2). ROC curves illustrating the performance of the three models in the training and validation sets are presented in Fig. 4. The calibration curve demonstrated strong alignment between predicted and observed probabilities in the training and validation sets (Fig. 5B, C). The DCAs based on a nomogram model are presented in Fig. 5D and E. In differentiating PD-1 positive from negative cases, the nomogram model demonstrated superior net benefit across a wide range of reasonable threshold probability ranges.

Fig. 5.

A nomogram (A) incorporating PIVKA-II, GGT-II, HBV-DNA, PVTT, and radiomics signature for preoperative prediction of PD-1 expression. Calibration curves for the radiomics nomogram in the training (B) and validation (C) sets. Calibration curves represent the goodness of fit of the nomogram. Decision curves of the training (D) and test (E) datasets for three models predicting PD-1 expression. The red line, green line, and blue line represent net benefit of the clinical factor model, the radiomics signature, and the radiomics nomogram, respectively. The radiomics nomogram had a higher overall net benefit in differentiating PD-1 positive from negative than the other

Table 2.

Model performances of prediction PD-1 expression in training and test cohorts

Models	Clinical model		Radiomics signature		Radiomics nomogram
Cohort	Training	Test	Training	Test	Training	Test
AUC	0.84	0.76	0.91	0.77	0.95	0.86
95% CI	[0.756–0.923]	[0.554–0.972]	[0.853–0.970]	[0.554–0.981]	[0.902–0.992]	[0.744–0.997]
Accuracy	0.76	0.64	0.81	0.64	0.88	0.73
Sensitivity	0.79	0.88	0.97	0.75	0.88	0.88
Specificity	0.75	0.50	0.71	0.57	0.88	0.74
The comparisons of different models					p-value*
Clinical model	Radiomics signature				0.016
Clinical model	Radiomics nomogram				< 0.001
Radiomics signature	Radiomics nomogram				0.026

AUC area under curve. CI confidence interval. *p-value of Delong’s test between every model in predicting PD-1

Utilizing the clinical scores and Rad-score (cut-off value = 0.40) as indicated by the nomogram, 12 patients were classified into two groups: a 6-patient high-risk group and a 6-patient low-risk group (Fig. 6).

Fig. 6.

Plots depict the radiomics nomogram–predicted probabilities for each individual HCC patient in the training dataset (A: n = 81) and test dataset (B: n = 35). The dashed line represents the cutoff value of 0.40

Kaplan–Meier survival analysis depicted lower mortality rates in the high-risk group compared to the low-risk group (Fig. 7). The predicted high-risk PD-1 HCC patients exhibit a tendency toward longer survival outcomes following combined therapy.

Fig. 7.

Kaplan–Meier curves for OS according to the low- and high-risk PD-1 predictions of the nomogram in 12 advanced HCC patients after combined therapy. Low- and high-risk groups were classified included a six high-risk group and a six low-risk group according to the nomogram. The differences of survival between predicted negative and positive PD-1 groups are significant (log-rank p = 0.02)

Discussion

This study highlights the potential of MRI-based radiomics in noninvasively predicting PD-1-positive characteristics in HCC and its impact on treatment outcome. The increasing interest in HCC’s immunologic hallmarks has spurred novel and targetable immunotherapies [13, 14], notably anti-PD-1 adjuvant therapy combined with TACE, showing promise in tumor control and prolong survival [15]. In our study, several clinical characteristics and MRI radiomic features showed correlation with PD-1 expression in HCC patients. Leveraging these significant features, we developed a multivariate regression model to predict PD-1 expression as immunotherapy targets in HCC. Furthermore, our findings suggest improved efficacy in predictive PD-1-positive HCC patients after TACE combined with PD-1 inhibitor. To our knowledge, there is rare report to combine clinical factors and MRI radiomics for predicting PD-1 risk and TACE outcomes with immunotherapy.

The aberrant activation of the PD-1 signaling pathway has been identified as predictor of poor prognosis and overall survival in HCC [16]. Noninvasive prediction of PD-1 expression through quantitative MRI has obtained increasing concern. In the recent studies, multiple MRI features reflect messages about internal and marginal areas of the tumor, enhancing the accuracy of PD-1 prediction and aiding in the selection of optimal treatment strategy for precision medicine [17]. In the current research, we calculated radiomics features from various MRI sequences, highlighting the potential of multiple sequences to enhance the depiction of tumor heterogeneity [18]. The integration of T2WI and DWI for radiomics feature extraction is based on their distinct advantages in characterizing tumor biology and providing valuable non-contrast information. Studies have shown that radiomics based on T2WI can predict pathologic grade, microvascular invasion, and outcomes of TACE in HCC patients [19, 20]. DWI offers insights into the diffusion of water molecules in tissues, serving as a surrogate marker for cellular density and tissue architecture [21]. Additionally, liver DWI can suppress intravascular signals, resulting in black-blood T2WI with good signal-to-noise ratio and tissue contrast, improving the detectability of lesions, particularly small ones near vessels [22]. Furthermore, recent studies have suggested that contrast-enhanced MRI and DWI may serve as the optimal method to predict molecular expression and measure treatment response in cancer [23, 24]. The integration of T2WI, contrast-enhanced images, and DWI allows the detection of morphological information, tumor microvessels, and water diffusion properties at the same time. In this study, six multiparametric MRI features from T2-weighted, fat-saturated T2-weighted, contrast-enhanced, and DWI demonstrated significant associations with PD-1 expression. Among all six features from multi-sequence MRI, wavelet features were the most heavily weighted. This filter, which is founded on 3D segmentation, can effectively smooth volumetric inputs as well as capture tumor phenotypic characteristics [8]. Moreover, the LoG filter serves as another crucial feature, offering insights into the three-dimensional morphology of the tumor’s internal structure [25].

Clinical biomarkers like GGTII, PIVKA-II, HBVDNA, and PVTT emerge as risk factors in our clinical model. It is reported that GGTII, as a significant biomarker, can increase the sensitivity for diagnosing HCC [26]. It was an interesting finding that elevated GGTII level was linked to PD-1 positive, suggesting its potential as a biomarker for advanced HCC patients undergoing anti-PD-1 immunotherapy combined with TACE. Besides, recent studies have highlighted PIVKA-II as a potential new marker for HCC and a highly specific predictor of the prognosis of HCC patients [27, 28]. Moreover, monitoring the serum levels of PIVKA-II could effectively evaluate and forecast the efficiency of anti-PD-1 immunotherapy [29]. In our research, HBVDNA was proven to be an additional independent predictor by using multiple logistic regression analysis. HBV-infected HCC patients presented highly upregulated PD-1, possibly associated with virus-specific T-cell dysfunction [30]. Although the association between PVTT and PD-1 remains to be elucidated, several studies revealed that TACE with PD-1 inhibitors enhances the therapeutic efficacy for HCC patients with PVTT [31, 32]. Overall, our findings align with previous studies, demonstrating that the measurement of GGTII, PIVKA-II, HBVDNA, and PVTT showed a good performance in accuracy for predicting PD-1 expression.

A comprehensive analysis of clinical and imaging data allows for decoding tumor phenotype. It was reported that clinical and radiomics features showed significant associations with PD-1 expression in HCC [17]. Leveraging these insights, we constructed a comprehensive nomogram integrating both clinical and radiomics models. This nomogram enables direct observation of the probability of PD-1 in each subject, offering a practical tool for clinicians. Incorporating features from the clinical model with those from the radiomics model yielded a better performance than their individual use. Our proposed radiomics nomogram model showed superior performance in differentiating PD-1 positive from negative, with AUCs of 0.95 and 0.86 in the training and validation sets, respectively. Furthermore, to validate its predictive capacity, we conducted Kaplan–Meier survival analyses in 12 advanced HCC patients. Wang et al. [33] have reported that MRI radiomics has the potential to predict preoperatively the survival of HCC patients, which may have some similarity with our findings.

In this study, we used immunohistochemistry of tissue microarrays techniques to detect the expression of PD-1. Tissue microarray images by immunohistochemically stained can mine relationship of biomarker expression with clinical-pathological features, disease stages, and disease prognosis [34]. Invaluable TMAs of 179 HCC patients’ samples enable rapid biomarker validation associated with tumor heterogeneity. TMAs and immunohistochemical techniques not only enable the preservation of these precious tumor materials but also advance the speed, sensitivity, and accuracy of PD-1detection.

Nevertheless, limitations such as the small sample size, the single-center study design, the retrospective analysis inevitably prone to selection bias, and variability in imaging protocols in clinical practice pose challenges. External validation and improvements in reproducibility across multiple institutions are essential for the wider applicability of our findings.

Conclusion

In conclusion, our study demonstrates the efficacy of a radiomics nomogram derived from multi-sequence MRI features. This valuable clinical tool can differentiate PD-1 positive from negative accurately and robustly in a noninvasive way. Patients predicted as positive, who received anti-PD-1 therapy combined with TACE, demonstrated prolonged OS compared with predicted negative patients. Consequently, within a precision medicine context, employing MRI-based radiomics nomogram for noninvasively screening enables the identification of high-risk HCC patients suitable for combination therapy, significantly improving prognosis.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

AUC

Area under curve

CT

Computed tomography

MRI

Magnetic resonance imaging

T2WI

T2-weighted images

DCA

Decision curve analysis

HCC

Hepatocellular carcinoma

ICC

Intraclass correlation coefficients

AFP

α-Fetoprotein

PIVKA-II

Prothrombin induced by vitamin K absence-II

GGT-II

γ-Glutamyl transpeptidase isoenzyme II

PVTT

Portal vein tumor thrombus

PD-1

Programmed cell death protein 1

TACE

Transcatheter arterial chemoembolization

ROI

Regions of interest

LASSO

Least absolute shrinkage and selection operator

ROC

Receiver operating characteristic curve

OS

Overall survival

GLCM

Gray-level cooccurrence matrix

GLDM

Gray-level dependence matrix

GLRLM

Gray-level run length matrix

GLSZM

Gray-level size zone matrix

Data Availability

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

Declarations

Conflict of Interest

The authors declare no competing interests.