Prediction of PD-L1 expression in nonsmall cell lung cancer using artificial intelligence models based on radiomics: a systematic review and meta-analysis

Abstract

Background

Since 1992, the incidence of lung cancer has shown an overall downward trend, with survival rates significantly improving in recent years due to the application of immune checkpoint inhibitors (ICIs). Assessing programmed death ligand 1 (PD-L1) expression status in lung cancer patients is crucial for predicting the efficacy of ICIs and formulating treatment plans. Therefore, there is an urgent need to assess PD-L1 expression status in lung cancer patients, particularly those with nonsmall cell lung cancer (NSCLC). While numerous studies have demonstrated promising results using artificial intelligence (AI) models based on radiomics, current understanding in this area remains incomplete.

Method

We conducted a search of medical databases and screened studies that meet the criteria. We performed feature extraction and data extraction from these studies, analysed them using R software, and generated forest plots. We conducted subgroup analyses based on imaging modality and algorithm model.

Results

A total of 35 studies were included in the analysis. For the Tumor Proportion Score (TPS)1 group, the pooled area under the curve (AUC) for the validation set was 0.782. For the TPS50 group, the pooled AUC for the validation set was 0.798. Subgroup analysis revealed that machine learning performed better when predicting TPS50, while deep learning was superior for predicting TPS1. When predicting TPS, prediction models constructed using 2-fluoro-2-deoxy-d-glucose (¹⁸F) positron emission tomography/computed tomography (CT) and contrast-enhanced CT outperformed those based on noncontrast CT. Conversely, when predicting a TPS, the noncontrast-CT-based prediction model significantly outperformed the other two groups.

Conclusions

Imaging modality and algorithm model are key factors influencing AI prediction models for PD-L1 expression in NSCLC patients.

Shareable abstract

Our findings reveal potential influencing factors in constructing AI models for predicting programmed death-ligand 1 expression https://bit.ly/4aRcfIJ

Introduction

Lung cancer is the second most commonly diagnosed cancer in the United States of America and the leading cause of cancer death [1]. The American Cancer Society estimated that 238 340 new cases of lung cancer would be diagnosed in 2023, and >120 000 people would die from the disease, accounting for ∼20% of all cancer deaths [2]. Since 1992, lung cancer incidence has shown an overall downward trend, with survival rates significantly improving in recent years due to advances in treatment modalities and enhanced early detection. This improvement is particularly notable with the widespread adoption of immunotherapy in lung cancer patients, especially in recent years [3]. Immune checkpoint inhibitors (ICIs), including anti-programmed death-ligand 1 (PD-L1) and anti-cytotoxic T lymphocyte-associated antigen-4 antibodies, are currently the only widely used immunotherapies for lung cancer patients [4, 5]. Whether used alone or in combination with chemotherapy, these agents have substantially improved most clinical efficacy end-points in patients with locally advanced or metastatic nonsmall cell lung cancer (NSCLC) lacking oncogenic driver mutations. They have also demonstrated efficacy in patients with extensive-stage small cell lung cancer [6, 7]. In both treatment-naive and previously treated advanced NSCLC patients, regimens combining two immune checkpoint inhibitors or anti-PD-(L)1 antibodies with chemotherapy significantly improved survival compared to chemotherapy alone, with particularly pronounced effects in PD-L1 high patients (defined as ≥50% PD-L1 expression in tumour cells by immunohistochemistry) [8, 9]. Consequently, evaluating and predicting responses to ICIs has gained increasing attention. However, PD-L1 expression status and regulation in lung cancer patients are critical for tailoring ICIs treatment strategies [10], with PD-L1 expression being recognised as a key factor influencing ICI efficacy [11]. For instance, Du et al. [12] demonstrated that activating PD-L1 transcription and expression enables immune evasion in NSCLC cells. Therefore, there is an urgent need for methods to assess PD-L1 expression status in lung cancer patients, particularly those with NSCLC. Although conventional immunohistochemistry serves as the gold standard for PD-L1 expression assessment, this method still has limitations, such as difficulty in tissue sampling, invasiveness [13] and unavoidable issues with suboptimal biopsy tissue quality [14].

In recent years, with the widespread adoption of computed tomography (CT) in lung cancer screening, an increasing number of predictive models have been developed based on imaging features. This field of research, known as radiomics, primarily involves feature extraction from lesions visible in patient images such as CT and positron emission tomography (PET) scans. Based on these imaging characteristics (including texture features) models are constructed to predict various patient characteristics [15]. The application of radiomics is extensive among NSCLC patients [16, 17], particularly demonstrating promising predictive outcomes for PD-L1 expression in NSCLC [18]. Wang et al. [19] developed a multisource deep-learning model using radiomics to predict PD-L1 expression in NSCLC patients, achieving an area under the receiver operating characteristic curve (AUC) of 0.950 (95% CI 0.906–0.964). Furthermore, Kiss et al. [20] developed a predictive model for PD-L1 expression status in NSCLC patients using PET/CT radiomics features. Their optimal model achieved an AUC of 0.783 (95% CI 0.625–0.942), demonstrating high accuracy (81.25%) and sensitivity (90.00%). Notably, radiomics possesses characteristics such as noninvasiveness and convenience, making it highly significant in predicting PD-L1 expression in NSCLC patients. However, despite its multiple advantages, significant variability exists among results due to heterogeneity between different algorithms and individual patients. For instance, Chen et al. [21] developed a gene-based radiomics prediction model using CT radiomics, achieving an AUC of 0.70 (95% CI 0.57–0.84) for predicting PD-L1 expression in NSCLC patients. Conversely, the CT-based radiomics model developed by Zheng et al. [22] achieved an AUC of 0.892 in NSCLC patients. Furthermore, significant differences were observed between the predictive model constructed using PET/CT radiomics by Jiang et al. [23] and models built solely on CT data. PET/CT-based predictive models demonstrated predictive performance for PD-L1 expression in NSCLC patients with an AUC of 0.97 (95% CI 0.92–1.00). Therefore, we propose that systematic reviews and meta-analyses should further investigate the predictive efficacy of artificial intelligence (AI)-driven radiomics models for PD-L1 expression in NSCLC patients.

This study aims to further investigate the efficacy of AI models based on radiomics in predicting PD-L1 expression in NSCLC patients through systematic review and meta-analysis. First, our research summarises and categorises AI models based on radiomics, further observing the application of different AI models in radiomics, particularly in NSCLC patients. This study primarily analyses different AI algorithm models and imaging modalities to identify the most suitable algorithm models and imaging data for AI development, providing theoretical foundations and evidence-based medical support for future AI model development. Additionally, it evaluates the predictive performance of various AI algorithms for PD-L1 expression status in NSCLC patients across both validation and training datasets. Notably, our study conducts detailed subgroup analyses using data from model validation cohorts classified by Tumor Proportion Score (TPS) as 1% (TPS1) and 50% (TPS50). We aimed to identify the AI algorithm models and imaging data with the best predictive performance through these subgroup analyses. Furthermore, we explored the application prospects of AI algorithm models across different TPS categories.

Method

This study was conducted in accordance with the relevant requirements of the Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines. A protocol was drafted prior to initiating the study and registered on the systematic review registry platform PROSPERO (www.crd.york.ac.uk/PROSPERO/) with identifier number CRD420251154291.

Literature search

Multiple medical literature databases were searched (including Scopus, Web of Science, Embase and PubMed), with the search period set from the inception of each database to 20 August 2025. The search strategy comprised 1) (pulmonary neoplasm) OR (lung neoplasm) OR (lung cancer) OR (cancer of lung) OR (pulmonary cancer); 2) (radiomics) OR (machine learning) OR (artificial intelligence) OR (deep learning) OR (predictive model); 3) (programmed cell death ligand 1) OR (PD-L1) OR (PD1) OR (programmed death 1). The three sets of keywords were combined using AND for parallel retrieval. Two independent authors conducted the searches in English, and their results were subsequently consolidated. No additional search filters, such as document type, were applied.

Study selection

All retrieved literature records were imported into EndNote version 20.0. After utilising the software's built-in duplicate removal feature, duplicate records were manually removed again. Two independent authors screened the literature records by reviewing titles and abstracts. Subsequently, full-text reviews were conducted on the selected records, and those failing to meet requirements (such as incomplete data) were excluded. The exclusion criteria were: 1) reviews, editorials, letters, syntheses, and guidelines; 2) case reports or small case series (studies involving fewer than five patients); 3) articles not reporting on the disease of interest or radiomics (e.g. classifier input features not derived from CT radiomics); and 4) literature with incomplete data. Disagreements during screening were submitted to a third author for resolution. Studies ultimately included had to meet the following criteria: 1) pathologically confirmed NSCLC patients, regardless of specific subtype; 2) PD-L1 expression status assessed exclusively via immunohistochemistry; 3) studies employing AI algorithms to construct predictive models for PD-L1 expression; 4) studies constructing radiomics features from imaging data (including CT, PET or PET/CT) as inputs for AI algorithm models.

Data extraction

Two independent authors extracted data mentioned in the included literature through full-text reading. The extracted data were primarily categorised into study characteristics, technical features and meta-analysis data. Study characteristics included first author and publication year, first author's country, experimental design, imaging modality, number of cases, TPS cut-off, tumour stage, immunohistochemistry assay and pathology type. Additionally, technical feature data encompassed segmentation method, segmentation software, feature extraction software, selected radiomics features, feature selection method, validation and algorithm type. For meta-analysis data, both training and validation set data were extracted. However, to ensure the validity of the assessment and the generalisability of the results, separate meta-analyses were conducted for the training and validation sets. If a study included both internal and external validation cohorts, all validation data were extracted, coded and then meta-analysed. When multiple internal or external validation cohorts existed, all validation data were still extracted for meta-analysis. This approach aimed to further increase the total data volume, making the pooled analysis results more scientifically sound and minimising data bias caused by study heterogeneity.

In this study, the TPS1 model refers to models that classify TPS using a 1% threshold or PD-L1 positive/negative expression; the TPS50 model refers to models that classify using a 50% threshold or PD-L1 high/low expression. Since TPS1 and TPS50 represent two distinct outcomes, we extracted performance metrics from the included studies and conducted separate meta-analyses for each. When studies included models constructed using multiple features, only results from models employing CT-derived artificially constructed radiomics features were extracted. When studies developed multiple models using different feature selection methods or modelling algorithms, all results were included and numbered. Models were compared and selected based on AUC values in validation cohorts. Data extraction was performed independently by two authors, who cross-checked each other's work for accuracy. Disagreements were resolved through discussion with a third author.

Quality assessment

The latest QUADAS-2 scale [24] was used to assess risk of bias. To enhance the reliability of bias risk assessment, two authors independently completed the Quality Assessment of Diagnostic Accuracy Studies (QUADAS)-2 questionnaire, with the final question version determined through discussion. Additionally, the Methodological Radiomics Scoring (METRICS) tool was employed to evaluate the methodological aspects of the studies [25]. The METRICS assessment is based on a series of weighted “yes/no” questions, culminating in a percentage score. This score is used to categorise studies into quality grades ranging from very low to excellent. The entire quality assessment process was performed by two independent authors, with a third author overseeing the process and mediating in cases of disagreement or judgmental differences.

Statistical analysis

To obtain the AUC and its corresponding 95% confidence intervals, we included only studies reporting radiomics data in a 2×2 table format alongside PD-L1 expression status in NSCLC, thereby avoiding violations of statistical assumptions when combining point estimates from different models. Forest plots were generated based on meta-analysis results, and subgroup analyses were conducted according to different AI algorithm types and imaging modalities. Analysis employed a two-variable random-effects model, with interstudy heterogeneity assessed using the Cochran Q test and I² statistic. All data analysis and graphical representations were performed using R software version 4.3.1. Publication bias was evaluated using a Deeks funnel plot.

Results

Systematic search and study selection

A comprehensive database search yielded a total of 966 records. Among these, 157 records were automatically deduplicated by EndNote and manually deduplicated by two independent reviewers. Nine records were excluded due to inaccessibility. Subsequently, the remaining 800 records underwent title and abstract screening, resulting in 37 records being retained. After full-text review, two studies were excluded due to incomplete data. Ultimately, 35 studies were included in the systematic review and meta-analysis (figure 1).

FIGURE 1.

Preferred Reporting Items for Systematic Reviews and Meta-Analyses flowchart.

Overview of the characteristics of studies

We conducted a systematic review of the characteristics of these 35 studies (table 1), most of which originated from China (23 studies), followed by Italy, Germany and Japan, each with two published studies. South Korea, North Korea, Canada, the United States of America, the United Kingdom and France each contributed only one published study. All included studies were retrospective, with six being multicentre and the remaining 29 single-centre. Regarding imaging modalities, 14 studies employed 2-fluoro-2-deoxy-d-glucose (¹⁸F) (FDG) positron emission tomography (PET)/computed tomography (CT) radiomics features for model construction, while 11 utilised contrast-enhanced CT (CECT) radiomics features, including one study using arterial-phase-enhanced CT. The remaining 10 studies utilised noncontrast CT (NCCT). 17 studies reported only the TPS for PD-L1 expression ≥1%, while three studies reported only the TPS for PD-L1 expression ≥50%. Regarding tumour staging of enrolled patients, 21 studies included NSCLC patients at stages I–IV, and two studies did not report patient tumour staging. Regarding PD-L1 immunohistochemical assessment methods, 11 studies used 22C3, six studies used SP263, and eight studies did not report the assessment method used. 19 studies included pathological types encompassing lung adenocarcinoma, lung squamous cell carcinoma and other various NSCLCs. 11 studies explicitly specified that the included pathological types of NSCLC were lung adenocarcinoma and lung squamous cell carcinoma, and five studies exclusively enrolled lung adenocarcinoma patients.

TABLE 1.

Characteristics of studies included in the literature

First author, year [ref.]	Country	Study design	Image type	Cases n	TPS cut-off	Tumour stage	Immunohistochemistry assay	Pathology type#
Monaco, 2022 [18]	Italy	R/SC	FDG PET/CT	86	1%, 50%	III or IV	22C3	LAC, LSC, others
Jiang, 2021 [26]	China	R/MC	NCCT	125	1%	Tis–III	NR	LAC, LSC
Bracci, 2021 [27]	Italy	R/SC	CECT	72	1%, 50%	III–IV	SP263	LAC, LSC
Liu, 2025 [28]	China	R/MC	AECT	204	1%, 50%	III–IV	22C3	LAC, others
Hashimoto, 2024 [29]	Japan	R/SC	CECT	117	1%	I–II	22C3	LAC, LSC, others
Liu, 2023 [30]	China	R/SC	CECT	125	1%	NR	22C3	LAC, LSC, others
Wang, 2023 [31]	China	R/SC	FDG PET/CT	394	1%	I–IV	PharmDX	LAC, LSC, others
Zheng, 2024 [22]	China	R/SC	CECT	220	1%, 50%	I–IV	SP263	LAC, LSC, others
Tian, 2024 [32]	China	R/MC	CECT	200	1%	I–IV	NR	LAC, LSC, others
Peng, 2025 [33]	China	R/SC	FDG PET/CT	143	1%	I–IV	NR	LAC, LSC
Stüber, 2025 [34]	Germany	R/SC	FDG PET/CT	254	1%	I–IV	NR	LAC, LSC, others
Li, 2024 [35]	China	R/SC	FDG PET/CT	136	1%	I–IV	22C3	LAC, LSC
Chen, 2023 [21]	UK	R/MC	CECT	109	1%, 50%	I–IV	NR	LAC, LSC, others
Zhang, 2024 [36]	China	R/SC	CECT	370	1%	I–IV	NR	LAC, LSC, others
Yolchuyeva, 2023 [37]	Canada	R/MC	NCCT	326	1%, 50%	I–IV	22C3	LAC, LSC, others
Zhang, 2022 [38]	Germany	R/SC	FDG PET/CT	58	1%	NR	NR	LAC, LSC, others
Shiinoki, 2022 [39]	Japan	R/SC	NCCT	202	1%, 50%	I–IV	22C3	LAC, LSC, others
Yoon, 2020 [40]	South Korea	R/SC	CECT	153	50%	III–IV	22C3	LAC
Mu, 2021 [41]	USA	R/MC	FDG PET/CT	485	1%	I–IV	22C3	LAC, LSC
Wang, 2021 [42]	China	R/SC	NCCT	1262	1%	I–IV	SP142	LAC, LSC, others
Xu, 2024 [43]	China	R/SC	NCCT	259	1%	I–IV	NR	LAC, LSC, others
Zhao, 2023 [44]	China	R/SC	FDG PET/CT	334	1%	I–IV	22C3	LAC, LSC, others
Wang, 2022 [19]	China	R/SC	NCCT	1135	1%, 50%	I–IV	SP142	LAC, LSC, others
Wen, 2021 [45]	China	R/SC	NCCT	120	50%	III–IV	SP263	LAC
Li, 2021 [46]	China	R/SC	FDG PET/CT	255	1%, 50%	I–IV	22C3	LAC, LSC
Shi, 2022 [47]	China	R/SC	NCCT	839	1%	I	22C3	LAC
Liang, 2024 [48]	China	R/SC	FDG PET/CT	101	1%	NR	SP263	LAC, LSC, others
Zhang, 2025 [49]	China	R/SC	FDG PET/CT	101	1%	I–IV	22C3	LAC
Tian, 2021 [50]	China	R/SC	NCCT	94	50%	IV	SP142	LAC, LSC, others
Zhu, 2020 [51]	China	R/SC	CECT	127	1%, 50%	III–IV	SP263	LAC
Hannequin, 2022 [52]	France	R/SC	FDG PET/CT	212	1%, 50%	I–IV	22C3	LAC, LSC
Lim, 2022 [53]	Republic of Korea	R/SC	FDG PET/CT	312	1%, 50%	NR	SP263	LAC, LSC
Sun, 2020 [54]	China	R/SC	NCCT	390	50%	I–IV	22C3	LAC, LSC
Wang, 2022 [55]	China	R/SC	NCCT	290	1%, 50%	I–IV	SP142	LAC, LSC, others
Jiang, 2020 [23]	China	R/SC	FDG PET/CT	399	1%, 50%	I–IV	SP142	LAC, LSC

TPS: Tumor Proportion Score; R: retrospective; SC: single-centre; FDG: 2-fluoro-2-deoxy-D-glucose (¹⁸F); PET: positron emission tomography; CT: computed tomography; LAC: lung adenocarcinoma; LSC: lung squamous carcinoma; MC: multicentre; NCCT: noncontrast-enhanced CT; Tis: carcinoma in situ; NR: not reported; CECT: contrast-enhanced CT; AECT: arterial-phase-enhanced CT; UK: United Kingdom; USA: United States of America. ^#: others refers to other nonsmall cell lung cancers besides LAC and LSC.

Overview of the technological characteristics

For the technical characteristics of the included studies, most employed manual delineation and segmentation of regions of interest, with only five studies utilising semi-automated segmentation. Additionally, six studies employed automated segmentation, while six studies did not specify the segmentation technique used. The most frequently used segmentation software was ITK-SNAP and 3Dslicer (seven studies), followed by LIFEx (six studies). Pyradiomics was the most frequently used feature extraction software (15 studies), although 13 studies did not report their feature extraction software. Regarding the number of selected radiomics features, most studies did not report this metric. Among those that did report, the selected features varied significantly, with no uniform requirements. Feature selection methods primarily relied on the Mann–Whitney U-test and least absolute shrinkage and selection operator (LASSO) regression, with the Gini index used in one additional study. Additionally, six studies employed external validation sets, while the remaining studies utilised internal validation sets. The algorithmic models primarily fell into two categories: machine learning and deep learning models (table 2).

TABLE 2.

The radiomics analysis methods adopted in the included studies

First author, year [ref.]	Segmentation method	Segmentation software	Feature extraction software	Radiomics features selected n	Feature selection method	Validation	Algorithm type
Monaco, 2022 [18]	Semi-automatic	PET-VCAR	NR (IBSI-compliant algorithms)	NR	Mann–Whitney U-test	Sample split	SVM, LR, classification tree, and KNN
Jiang, 2021 [26]	Manual	ITK-SNAP	Pyradiomics	9	Mann–Whitney U-test	External	RF, DT, LR, AdaBoost, Gaussian process, and SVM
Bracci, 2021 [27]	Manual	LIFEx	LIFEx	6	LASSO	Sample split	Radiomic score
Liu, 2025 [28]	Semi-automatic	3DSlicer	Pyradiomics	19	LASSO	External	Naive Bayes, DT, XGBoost, and RF
Hashimoto, 2024 [29]	Manual	Eclipse	Pyradiomics	448	Mann–Whitney U-test	Sample split	NR
Liu, 2023 [30]	NR	NR	Pyradiomics	150	Mann–Whitney U-test	Sample split	Deep-learning model
Wang, 2023 [31]	Manual	ITK-SNAP	Pyradiomics	26	Mann–Whitney U-test	Sample split	SVM, naive Bayesian, RF, and XGBoost
Zheng, 2024 [22]	Manual	NR	NR	14	LASSO	Sample split	SVM, LDA, and LR
Tian, 2024 [32]	Manual	NR	R software	30	LASSO	External	LASSO regression and LR
Peng, 2025 [33]	Manual	LIFEx	Pyradiomics	93	LASSO	Sample split	Bayes, LR, RF, and SVM
Stüber, 2025 [34]	Manual	NR	Pyradiomics	NR	NR	Sample split	Boosting and RF
Li, 2024 [35]	Manual	3DSlicer	Pyradiomics	693	Mann–Whitney U-test	Sample split	Deep-learning model
Chen, 2023 [21]	NR	3DSlicer	NR	NR	NR	External	NR
Zhang, 2024 [36]	Manual	3DSlicer	Pyradiomics	NR	Mann–Whitney U-test	Sample split	NR
Yolchuyeva, 2023 [37]	Manual	NR	Pyradiomics	NR	LASSO	External	AdaBoost, DT, RF, LR, SVM, KNN, XGBoost
Zhang, 2022 [38]	Manual	LIFEx	LIFEx	NR	NR	Sample split	NR
Shiinoki, 2022 [39]	Manual	3DSlicer	Pyradiomics	NR	LASSO	Sample split	LightGBM
Yoon, 2020 [40]	Semi-automatic	AVIEW Research	NR	NR	LASSO	Sample split	LASSO regression
Mu, 2021 [41]	NR	NR	NR	NR	NR	External	NR
Wang, 2021 [42]	Manual	NR	NR	NR	LASSO	Sample split	Deep-learning model-CNN
Xu, 2024 [43]	Manual	ITK-SNAP	Pyradiomics	11	LASSO	Sample split	Deep-learning model-CNN
Zhao, 2023 [44]	Automatic	LIFEx	LIFEx	124	LASSO	Sample split	LR
Wang, 2022 [19]	NR	NR	Pyradiomics	NR	LASSO	Sample split	Deep-learning model
Wen, 2021 [45]	NR	3DSlicer	NR	462	LASSO	Sample split	NR
Li, 2021 [46]	Automatic	LIFEx	LIFEx	18	LASSO	Sample split	LR
Shi, 2022 [47]	Automatic	NR	NR	9	LASSO	Sample split	NR
Liang, 2024 [48]	Semi-automatic	3DSlicer	3DSlicer	NR	NR	Sample split	Deep-learning model-DenseNet121
Zhang, 2025 [49]	Manual	ITK-SNAP	Pyradiomics	NR	LASSO	Sample split	DT, LDA, LR and SVM
Tian, 2021 [50]	NR	NR	NR	NR	Mann–Whitney U-test	Sample split	Deep-learning model
Zhu, 2020 [51]	Manual	ITK-SNAP	NR	NR	Mann–Whitney U-test	Sample split	Deep-learning model-DenseNet
Hannequin, 2022 [52]	Semi-automatic	LIFEx	NR	NR	NR	Sample split	NR
Lim, 2022 [53]	Manual	LIFEx	LIFEx	5	Gini index	Sample split	RF, NN, Naïve Bayes, LR, AdaBoost, SGD, and SVM
Sun, 2020 [54]	Automatic	ITK-SNAP	NR	9	LASSO	Sample split	NR
Wang, 2022 [55]	Automatic	NR	NR	12	NR	Sample split	NR
Jiang, 2020 [23]	Automatic	ITK-SNAP	Pyradiomics	24	LASSO	Sample split	LR and RF

PET: positron emission tomography; VCAR: volume computer-assisted reading; NR: not reported; IBSI: Image Biomarker Standardisation Initiative; SVM: support vector machine; LR: logistic regression; KNN: K-nearest neighbourhood; RF: random forest; DT: decision tree; AdaBoost: adaptive boosting; LASSO: least absolute shrinkage and selection operator; XGBoost: extreme gradient boosting; LDA: linear discriminant analysis; CNN: convolutional neural network; NN: neural network; SGD: stochastic gradient descent.

Meta-analysis

A meta-analysis was conducted on the 35 included studies. For the TPS1 group, the pooled AUC for the validation set was 0.782 (95% CI 0.747–0.817) (figure 2a), and the overall AUC for the training set was 0.857 (95% CI 0.810–0.904) (figure 2c). For the TPS50 group, the pooled AUC for the validation set was 0.798 (95% CI 0.755–0.842) (figure 2b), and the overall AUC effect for the training set was 0.856 (95% CI 0.776–0.937) (figure 2d). To more intuitively assess the impact of TPS1 and TPS50 on the performance of radiomics-based AI models for predicting PD-L1 expression, we conducted a subgroup analysis of 11 studies that reported both TPS1 and TPS50. The pooled results showed that in the TPS1 subgroup, the predictive performance of the AI models was 0.799 (95% CI 0.713–0.864) (supplementary figure S1A). In the TPS50 subgroup, the predictive performance was 0.811 (95% CI 0.751–0.860) (supplementary figure S1B). Comparison of the two subgroups indicated that although radiomics-based AI models achieved slightly better performance in the TPS50 subgroup, the difference between the two was not substantial. We therefore consider that, when using radiomics-based AI models to predict PD-L1 expression, using TPS50 as the target threshold may be preferable.

FIGURE 2.

Overall effect of the training and validation sets included in the study. a) Overall effect on the validation set for Tumor Proportion Score (TPS)1; b) overall effect on the validation set for TPS50; c) overall effect on the training set for TPS1; d) overall effect on the training set for TPS50. AUC: area under the curve.

Subgroup analysis

To further examine the impact of imaging modalities and algorithmic models on PD-L1 expression prediction, we conducted detailed subgroup analyses. Data were primarily categorised into machine-learning algorithms and deep-learning algorithms based on different algorithmic classifications. For machine-learning algorithms, the AUC summary effect in the TPS1 validation cohort was 0.769 (95% CI 0.727–0.811) (figure 3a). For the TPS50 validation set, the AUC summary result was 0.805 (95% CI 0.750–0.859) (figure 3b). In the deep-learning validation set, the AUC summary effect was significantly better in the TPS1 group than in the TPS50 group (TPS1: 0.821 versus TPS50: 0.779) (figure 3c, d). Additionally, we performed model-based algorithm analysis on the training set data. We found that deep-learning algorithms outperformed machine learning algorithms in the training set (TPS1: 0.892 versus 0.843; TPS50: 0.874 versus 0.846) (supplementary figure S2).

FIGURE 3.

Subgroup analysis of different programmed death-ligand 1 expression levels based on different algorithmic models. a) Total effect of the machine-learning algorithm validation set in the Tumor Proportion Score (TPS)1 group; b) total effect of the machine-learning algorithm validation set in the TPS50 group; c) total effect of the deep-learning algorithm validation set in the TPS1 group; d) total effect of the deep-learning algorithm validation set in the TPS50 group. AUC: area under the curve.

Based on imaging modality, we categorised the data into three groups: the CECT group, the FDG PET/CT group and the NCCT group. When predicting TPS1, the AUC summary results for the CECT group validation set were 0.792 (95% CI 0.770–0.815) (figure 4a), while the FDG PET/CT group yielded an AUC summary of 0.790 (95% CI 0.736–0.844) (figure 4b). The NCCT group demonstrated poorer performance than the previous two groups, with an AUC summary of 0.757 (95% CI 0.663–0.852) (figure 4c). When predicting TPS50, the NCCT group demonstrated the best validation set performance with a pooled AUC of 0.837 (95% CI 0.761–0.912) (figure 4d). The CECT group followed with an AUC summary effect of 0.765 (95% CI 0.755–0.775) (figure 4e). The FDG PET/CT group demonstrated the poorest performance, yielding an AUC summary result of 0.740 (95% CI 0.693–0.788) (figure 4f).

FIGURE 4.

Subgroup analysis of programmed death-ligand 1 expression based on different imaging modalities. a) Overall effect of the Tumor Proportion Score (TPS)1 group on the contrast-enhanced computed tomography (CECT) validation set; b) overall effect of the TPS1 group on the 2-fluoro-2-deoxy-D-glucose (¹⁸F) (FDG) positron emission tomography (PET)/CT validation set; c) overall effect of the TPS1 group on the noncontrast CT (NCCT) validation set; d) overall effect of the TPS50 group on the NCCT validation set; e) overall effect of the TPS50 group on the CECT validation set; f) overall effect of the TPS50 group on the FDG PET/CT validation set. AUC: area under the curve.

To further examine the influence of specific NSCLC histological subtypes on the predictive performance of radiomics-based AI models, we performed a subgroup analysis restricted to studies that enrolled only lung adenocarcinoma. The pooled estimate was 0.734 (95% CI 0.608–0.830) (figure 5a). Compared with the overall predictive performance in the entire NSCLC population, the performance of radiomics-based AI models for predicting PD-L1 expression was lower in lung adenocarcinoma. Therefore, histological subtype may also be a factor affecting the predictive performance of radiomics-based AI models. In addition, to further control for the influence of TPS cut-offs, we conducted another subgroup analysis within the TPS1 group including only studies restricted to lung adenocarcinoma. The pooled estimate was 0.741 (95% CI 0.564–0.864) (figure 5b). These findings suggest that, when predicting PD-L1 expression in lung adenocarcinoma, differences in TPS cut-offs may not be a major determinant of the predictive performance of radiomics-based AI models. Meanwhile, to evaluate whether potential differences in the reference standard arising from different pathological specimen sources might affect model performance, we further extracted and analysed data according to specimen type. In the TPS1 group, five studies used biopsy specimens only and four studies used surgical resection specimens only. The pooled estimate was 0.825 (95% CI 0.711–0.901) for the biopsy-only subgroup (figure 5c) and 0.807 (95% CI 0.542–0.936) for the resection-only subgroup (figure 5d). In the TPS50 group, five studies used biopsy specimens only and two studies used resection specimens only. Although the number of included studies in the resection subgroup was limited, we performed a subgroup analysis to more directly compare the impact of specimen source. The pooled estimate was 0.818 (95% CI 0.702–0.895) for the biopsy-only subgroup (figure 5e) and 0.719 (95% CI 0.670–0.764) for the resection-only subgroup. Overall, these comparisons indicate that differences in pathological specimen source may also contribute to variability in the predictive performance of radiomics-based AI models.

FIGURE 5.

Subgroup analysis results by histological subtype and pathological specimen source. a) Subgroup analysis including only studies that enrolled lung adenocarcinoma; b) subgroup analysis within the Tumor Proportion Score (TPS)1 group including only studies that enrolled lung adenocarcinoma; c) subgroup analysis of studies using biopsy specimens as the reference standard within the TPS1 group; d) subgroup analysis of studies using surgical resection specimens as the reference standard within the TPS1 group; e) subgroup analysis of studies using biopsy specimens as the reference standard within the TPS50 group; f) subgroup analysis of studies using surgical resection specimens as the reference standard within the TPS50 group. AUC: area under the curve.

Discussion

Imaging is one of the fundamental technologies in the medical field, primarily used in clinical practice to assist with screening, diagnosis [56], treatment [57] and follow-up [15]. With the continuous advancement of imaging technology, clinicians have increasingly higher demands for imaging capabilities, particularly regarding functionality and accuracy. Consequently, radiologists have embarked on innovative explorations in radiology, leading to the introduction of radiomics. First developed and applied in 2012, radiomics utilises automated high-throughput techniques to extract a vast array of imaging features from standard diagnostic medical images. Its core hypothesis posits that automated or semi-automated software-assisted quantitative analysis of medical imaging data can rapidly and reproducibly provide supplementary information for decision support. Reliable radiomics methods must achieve identification of diverse quantitative features from medical images, storage of such data across multiple independent databases to form a unified entity (consolidated database) [58], and subsequent data mining to derive clinically relevant outcomes [59]. By analysing medical images such as CT, magnetic resonance and/or PET scans, radiomics features can be extracted for screening, diagnosis [60], follow-up, prognostic assessment [61] and pharmacokinetic/pharmacodynamic studies. In recent years, with the continuous advancement of radiomics, an increasing number of studies have applied radiomics to treatment response prediction [62] and biological expression analysis [63] in NSCLC patients. However, these studies lack standardised and unified frameworks, coupled with insufficient data sharing, making it difficult to compare results and validate radiomics models, ultimately hindering their clinical application. This review aims to elucidate key factors contributing to variability in radiomics research outcomes and provide comprehensive data support and guidance for constructing clinical radiomics models.

This study focuses on predicting PD-L1 expression in NSCLC patients using image-based AI, incorporating a substantial number of studies. Notably, our findings indicate that AI models constructed based on radiomics predict PD-L1 expression in NSCLC patients with an AUC of ∼0.8. The pooled AUC was significantly lower than the AUC values >0.9 reported in some studies [23, 28, 48]. We attribute this discrepancy to patient heterogeneity and technical/operational variations in PD-L1 expression detection via immunohistochemistry [14, 64]. Zheng et al. [65] reported a sensitivity of 0.75 (95% CI 0.70–0.79) and specificity of 0.78 (95% CI 0.73–0.83) for machine-learning prediction of PD-L1 expression in NSCLC patients, which is consistent with our findings. However, our study focused on AUC, thus the two studies complement each other. Furthermore, our research delved deeper into the predictive performance of different imaging modalities and AI algorithms for PD-L1 expression status. It is worth noting that the meta-analysis by Salimi et al. [66] reported pooled AUC values of 0.85 (95% CI 0.81–0.88) and 0.88 (95% CI 0.85–0.91) for TPS1 and TPS50 in CT radiomics, respectively. This discrepancy with our findings is probably attributable to our inclusion of a larger number of studies and a more comprehensive range of imaging modalities.

Furthermore, through further analysis of the algorithms, we found that the algorithm itself may be one of the factors influencing prediction performance. Additionally, it can be observed that machine-learning algorithms perform better in the TPS50 group, while deep-learning algorithms are more suitable for the TPS1 group. This partially aligns with the findings of Ligero et al. [63]. Naturally, our study delves deeper into the predictive performance of algorithms for different PD-L1 expression states in NSCLC patients by classifying PD-L1 expression concentrations. From another perspective, the differing outcomes across algorithms for various PD-L1 expression states further underscores that PD-L1 expression concentration is a significant factor influencing the AUC of radiomics AI prediction models. Compared to the meta-analysis by Salimi et al. [66], our study meticulously subdivided CT scans into CECT and NCCT categories, which may represent a primary reason for the observed differences between our findings. Notably, our study indicates that both FDG PET/CT and CECT-based prediction models outperform NCCT in the TPS1 group. However, the NCCT-based prediction model significantly outperforms the other two groups in the TPS50 group. Therefore, we conclude that imaging modality is also a critical factor influencing AI model prediction performance.

In addition, we conducted subgroup analyses of several factors that may affect the predictive performance of radiomics-based AI models, including the distinction between TPS1 and TPS50, as well as differences in histological subtype and pathological specimen source. Our results showed that the performance difference between the TPS1 and TPS50 subgroups was not substantial. As for histological subtype, because comparisons with other subtypes were lacking, we could not clearly determine whether meaningful differences exist across subtypes. In the future, with a sufficient number of studies, further subgroup analyses by different histological subtypes could be performed. Regarding specimen source, although in the TPS50 subgroup we observed that differences in the source of pathological specimens might influence the final predictive performance, this effect was not evident in the TPS1 subgroup and may also be affected by multiple factors, such as imaging modality, algorithm/model type, and the number of included studies. Therefore, only when more studies become available (and after controlling for algorithm/model type and imaging modality) can the impact of specimen source on radiomics-based AI models be more definitively evaluated. Consequently, when predicting PD-L1 expression, we cannot yet conclude whether pathological specimens from different sources have a significant effect on the performance of radiomics-based AI models. At present, we can only observe a trend suggesting that specimen source may influence model performance. We look forward to more in-depth investigations of these potential influencing factors in the future.

This study aims to provide clinicians with reasonable and accurate guidance for constructing AI models by conducting a pooled analysis of AUC values for AI algorithms predicting PD-L1 expression in NSCLC patients based on radiomics. Despite limitations such as insufficiently systematic subgroup analyses, our research holds significant value. Our findings reveal potential influencing factors in constructing AI models for predicting PD-L1 expression. Furthermore, this study systematically compared the pooled AUC across different imaging modalities and AI algorithm models for predicting PD-L1 expression in NSCLC patients. This represents the first systematic review and meta-analysis to conduct a detailed analysis by imaging type and algorithm model, and currently stands as the most comprehensive systematic review and meta-analysis in this field.

Conclusion

When constructing AI predictive models to forecast PD-L1 expression in NSCLC patients, imaging modality and algorithmic model may be key factors influencing predictive performance.

Points for clinical practice and questions for future research

• Our findings reveal potential influencing factors in constructing AI models for predicting PD-L1 expression.

• This study systematically compared the pooled AUC across different imaging modalities and AI algorithm models for predicting PD-L1 expression in NSCLC patients.

• This represents the first systematic review and meta-analysis to conduct a detailed analysis by imaging type and algorithm model, and currently stands as the most comprehensive systematic review and meta-analysis in this field.

Supplementary material

Please note: supplementary material is not edited by the Editorial Office, and is uploaded as it has been supplied by the author.

Supplementary figure S1

ERR-0299-2025.SUPPLEMENT

Supplementary figure S2

ERR-0299-2025.SUPPLEMENT2