Construction and validation of a CT-based radiomics-deep learning signature for non-invasive prediction of PD-L1 expression and immunotherapy outcomes in non-small cell lung cancer

Introduction

Lung cancer is a major global health burden, with approximately 2.48 million new cases annually, accounting for 14.2% of all cancer diagnoses, and is responsible for about 1.82 million deaths, representing 18.7% of cancer-related mortality worldwide (1-3). As such, it remains the leading cause of both cancer incidence and mortality. Non-small cell lung cancer (NSCLC) constitutes the predominant histological subtype, comprising 80–85% of all primary lung cancers (4). The rapid evolution of tumor immunotherapy has ushered lung cancer treatment into an era dominated by immune-based strategies. A growing number of immunotherapeutic regimens are now standard of care, with applications expanding from advanced metastatic disease to the perioperative setting (5,6). Immunotherapy has shown substantial efficacy both as monotherapy and in combination with chemotherapy, radiotherapy, neoadjuvant therapy (NAT), or other targeted agents (7,8). Immune checkpoint inhibitors (ICIs) targeting programmed cell death protein 1 (PD-1) or its ligand (PD-L1) form the cornerstone of NSCLC immunotherapy. However, tumor heterogeneity limits the response rate to anti-PD-1/PD-L1 therapy, which is observed in only 17–21% of patients (9,10). Current patient selection largely relies on the immunohistochemical assessment of PD-L1 expression in biopsy or surgical specimens. This approach, however, has several drawbacks: it is invasive, precluding repeated sampling to monitor dynamic tumor evolution; carries procedural risks that may exacerbate a patient’s condition or be unfeasible; and fails to capture the immune status of the entire tumor due to intratumoral heterogeneity. Thus, there is a pressing need for non-invasive biomarkers to identify patients most likely to benefit from anti-PD-1/PD-L1 immunotherapy.

In response, several non-invasive predictive strategies have been explored. For instance, liquid biopsy techniques attempt to reflect the molecular characteristics and therapeutic response of tumors by analyzing components such as circulating tumor DNA (ctDNA) and exosomes in the blood. Studies have shown that dynamic changes in peripheral blood immune cells (e.g., neutrophil-to-lymphocyte ratio, NLR) can effectively predict the efficacy of immunotherapy. Concurrently, the analysis of exosomal PD-L1 has revealed its critical role in tumor immune evasion and its potential as both a therapeutic target and a predictive biomarker (11,12). On the other hand, molecular imaging-based approaches have also demonstrated promise. For example, deep learning (DL) models based on positron emission tomography/computed tomography (PET/CT) can directly predict PD-L1 status and assess treatment efficacy from images (13). However, these methods are subject to inherent limitations: liquid biopsies are constrained by high detection costs, complex analytical workflows, and issues with result stability, whereas PET/CT suffers from expensive equipment, radiation exposure, and limited accessibility, making it difficult to implement widely in clinical practice. In contrast, routine CT scanning is a standard and ubiquitous tool throughout the entire diagnosis and treatment process of lung cancer. It offers unique advantages such as being non-invasive, radiation-free, low-cost, and highly accessible. Therefore, developing a predictive model that utilizes only existing routine CT images holds the potential to overcome the current bottleneck in non-invasive PD-L1 prediction with minimal additional cost and maximum applicability.

In this context, technological advancements in the field of medical image analysis have made this vision feasible. Amid rapid progress in big data and artificial intelligence (AI), medical diagnosis and treatment are increasingly aligned with the principles of precision medicine. Radiomics and DL enable quantitative extraction of clinically relevant features from medical images, positioning this field at the forefront of medical imaging research (14,15). Previous studies have highlighted the considerable potential of radiomics and DL in predicting molecular subtypes, therapeutic efficacy, and prognosis in NSCLC immunotherapy (16-19). In particular, the application of DL to lung cancer prognosis has evolved rapidly. Early studies predominantly employed standard convolutional neural networks (CNNs), such as ResNet, to automatically learn hierarchical feature representations from CT or PET/CT images for tasks like survival prediction and immunotherapy response assessment. These models demonstrated superior performance over traditional methods by capturing complex, data-driven patterns. More recently, the field has witnessed the emergence of advanced architectures like Vision Transformers, which utilize self-attention mechanisms to model long-range dependencies across entire images, offering new possibilities for comprehensively characterizing tumor heterogeneity and its interactions with the microenvironment. However, the clinical translation of such advanced models faces significant challenges: first, their high computational complexity imposes stringent hardware requirements, and training relies on massive amounts of annotated data—a scarce resource in medical imaging that easily leads to overfitting in small-sample settings; second, and more critically, the issue of interpretability remains prominent (20,21). Although self-attention mechanisms provide different perspectives on feature interaction, the decision-making process of these models is still difficult to translate into logic that clinicians can readily understand.

To address the practical gap between predictive power, model transparency, and clinical applicability, this study innovatively constructs a hybrid radiomics-deep learning signature (RADLsig) model. Compared to approaches that rely solely on either handcrafted radiomic features or end-to-end DL architectures, our proposed solution integrates both technical paradigms. Specifically, RADLsig simultaneously leverages: (I) handcrafted radiomic features based on interpretable imaging phenotypes (e.g., texture, shape, intensity statistics); and (II) deep features automatically extracted by a CNN backbone. This dual-pathway design aims to enhance predictive accuracy through feature complementarity while maintaining model transparency via the interpretable radiomic feature layer, allowing predictions to be partially traced back to quantifiable imaging characteristics.

Furthermore, this study will further evaluate the model’s utility in risk stratification based on patients’ potential response to immunotherapy. Ultimately, by developing and validating RADLsig based on routine pretreatment CT images, this research aims to noninvasively predict PD-L1 expression status in NSCLC patients and assess the model’s clinical utility in predicting immunotherapy response, thereby providing a non-invasive, cost-effective, and reproducible tool for clinical decision-making. We present this article in accordance with the TRIPOD reporting checklist (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-1-1433/rc).

Methods

Study population

This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Shenzhen People’s Hospital (No. LL-KY-2021058) and written informed consent was waived. This single-center retrospective study included two patient cohorts (Cohort 1 and Cohort 2). Cohort 1 consisted of NSCLC patients who underwent PD-L1 immunohistochemistry (IHC) testing and thoracic CT within 3 months before tissue samples were obtained at our institution from January 2021 to May 2025. The inclusion criteria were as follows: (I) pathologically confirmed NSCLC via biopsy or surgery tissue sample; (II) underwent CT scanning within three months prior to surgery; and (III) underwent PD-L1 IHC staining. The exclusion criteria were as follows: (I) received anti-tumor treatment before biopsy or CT examination; (II) the sampled tissue originated from a lymph node or metastatic lesion; and (III) images contained severe artifacts or the region of interest (ROI) could not be delineated. Additionally, to ensure the robust development of the RADLsig, the minimum sample size was calculated on the basis of our previous primary study about CT DL score for PD-L1 using the pmsampsize R package on Riley’s criteria for prediction model (18). The calculation indicated that a minimum of 379 patients is required to precisely estimate model performance. Our final cohort of 531 patients met and exceeded this threshold, providing adequate statistical power for model development and internal validation. Ultimately, cohort 1 included 531 lung cancer patients, who were randomly divided into a training set (n=424) and a validation set (n=107) in an 8:2 ratio. Cohort 2 comprised NSCLC patients who had received ICIs in combination with chemotherapy at our institution from January 2019 to December 2024. The inclusion criteria were as follows: (I) pathologically confirmed NSCLC; (II) underwent a CT scan before treatment; and (III) received at least two cycles of immunotherapy combined with chemotherapy. The exclusion criteria were as follows: (I) did not undergo a CT scan prior to immunotherapy combination therapy; (II) received other treatments prior to immunotherapy combination therapy; and (III) insufficient follow-up or treatment was interrupted due to immune-related adverse events. Ultimately, a total of 145 patients met the inclusion and exclusion criteria. Cohort 3, comprising 128 NSCLC patients from The Cancer Imaging Archive (TCIA) with single-cell RNA sequencing (scRNAseq) analyses, was used to investigate association between the CT-based RADLsig and CD274 count, representing RNA expression of the protein-encoding gene for PD-L1. The inclusion criteria were as follows: (I) pathologically confirmed NSCLC; (II) availability of pretreatment chest CT images; (III) availability of matched single-cell RNA sequencing data containing CD274 gene expression information. Exclusion criteria: CT images with severe artifacts or cases where the volume of interest could not be reliably delineated. The flowchart is shown in Figure 1.

Figure 1 Flowchart of the patient selection process. Cohort 1: consisted of NSCLC patients who underwent PD-L1 IHC testing and thoracic CT within 3 months before tissue samples; Cohort 2: comprised NSCLC patients who had received ICIs in combination with chemotherapy; Cohort 3: comprising 128 NSCLC patients from TCIA with scRNAseq analyses. CT, computed tomography; ICIs, immune checkpoint inhibitors; IHC, immunohistochemistry; NSCLC, non-small cell lung cancer; PD-L1, programmed death-ligand 1; ROI, region of interest; scRNAseq, single-cell RNA sequencing; TCIA, The Cancer Imaging Archive.

Outcomes

Response to immunotherapy was assessed using the immune-related Response Evaluation Criteria in Solid Tumors (iRECIST) (22). Complete response (CR) and partial response (PR) indicate a response to immunotherapy, whereas stable disease (SD) and progressive disease (PD) indicate no response.

An experienced pathologist, who was blinded to imaging findings and clinical information, evaluated histopathological specimens obtained from biopsies or surgery performed prior to primary tumor treatment. PD-L1 expression levels were retrospectively assessed using the VENTANA PD-L1 SP263 IHC assay system (Roche Diagnostics, Basel, Switzerland). The tumor proportion score (TPS) was calculated as follows: TPS = (number of PD-L1-positive tumor cells/total number of tumor cells) ×100%. All counts were performed under 40× magnification. Specimens were categorized into two groups based on TPS: TPS <1% was defined as negative expression and TPS ≥1% was defined as positive. Figure 2 presents a heatmap of PD-L1 expression at different levels. Figure 3 presents a heatmap of different response levels to immunotherapy.

Figure 2 PD-L1 heatmaps. (A) PD-L1 expression positive, TPS ≥1%. (B) PD-L1 expression negative, TPS <1%. The red boxes indicate lesion areas. PD-L1, programmed death ligand-1; TPS, tumor proportion score.

Figure 3 Immunotherapy heatmaps. (A) Response to immunotherapy. (B) Non-response to immunotherapy. The red boxes indicate lesion areas.

CT scans and tumor segmentation

All patients underwent chest CT scanning before treatment. The specific scanning parameters were as follows: collimator 0.625 mm × 128 slices; tube voltage 120 kV; tube current adjusted by automatic exposure control (AEC); reconstructed slice thickness 1.5 mm with an inter-slice spacing of 1 mm; field of view (FOV) 350×350 mm; matrix 512×512. The images were resampled to a voxel size of 1×1×1 mm via linear interpolation. Subsequently, two experienced radiologists manually delineated the tumor regions slice by slice using ITK-SNAP software (Version 3.6.0; https://www.itksnap.org/pmwiki/pmwiki.php) under the lung window setting (window width: 1,500, window level: –500), so as to construct the three-dimensional volume of interest (3D-VOI).

Feature extraction and selection

The radiomics and DL signatures in this study were derived from a standardized preprocessing pipeline. Radiomic features were extracted from the 3D-VOI of interest using the Python-based PyRadiomics platform (https://www.python.org/), and then the features standardized via Z-Score normalization to reduce the impact of feature scale differences. DL features were obtained by extracting the activation maps from the penultimate convolutional layer of a pretrained 3D ResNet-101 model after removing the final fully connected layer. To address the risk of overfitting arising from the resulting high-dimensional feature space relative to the limited training sample size, we implemented a rigorous multi-stage feature dimensionality reduction and selection procedure. All steps of feature preprocessing, selection, and model tuning were performed exclusively within the training cohort to prevent data leakage. The final model and all parameters were then fixed and applied to the held-out validation cohort. First, to assess inter-observer reproducibility, we calculated intraclass correlation coefficients (ICCs) and removed features with poor repeatability (ICCs ≤0.75) and the mean of the remained features was used for further analysis. Second, to eliminate redundant information, Pearson correlation analysis was performed on the remaining features. For any feature pair with a correlation coefficient (r>0.90), the feature exhibiting lower average correlation with all other features was retained. Subsequently, the Mann-Whitney U test was used to select features significantly associated with PD-L1 expression (P<0.05). Finally, the retained features were fed into the least absolute shrinkage and selection operator (LASSO) algorithm with 5-fold cross-validation conducted on the training set and the optimal penalty parameter (λ) selected according to the “one-standard-error” rule, a compact set of key features closely associated with PD-L1 expression in the training set.

Model development and validation

Based on the selected features, a radiomics signature (RAsig) and deep learning signature (DLsig) were constructed through a linear combination of features weighted their coefficients. Then, logistic regression was implemented to construct fusion model for RADLsig. The predictive performance was evaluated in the validation set using area under the receiver operating characteristic (ROC) curve (AUC). The value of the signature for immunotherapy response in clinical situations was assessed in the immunotherapy cohort and gene explanation of the signature underwent the TCIA cohort. Figure 4 illustrates the workflow of this process.

Figure 4 Workflow of this study. AUC, area under the curve; CI, confidence interval; CNN, convolutional neural network; DLsig, deep learning signature; RAsig, radiomics signature; RADLsig, radiomics-deep learning signature.

Statistical analysis

All statistical analyses and graphical outputs were generated using the software SPSS 25.0 (IBM Corp., Armonk, NY, USA), R version 4.1.2 (R Foundation for Statistical Computing, Vienna, Austria), and Python version 3.8.5 (Python Software Foundation, Wilmington, DE, USA). Continuous variables were compared between the training and validation groups using either the Mann-Whitney U test or Student’s t-test. Categorical variables were assessed using either the Chi-squared test or Fisher’s exact test. Model performance was evaluated by calculating the AUC with 95% confidence intervals (CIs) in the validation set. The DeLong test was used to compare AUC values. A two-tailed P value less than 0.05 was considered statistically significant.

Results

Patient characteristics

This study enrolled 804 patients with NSCLC, who were then divided into a training cohort (n=424), validation cohort (n=107), immunotherapy cohort (n=145), and a gene cohort (n=128). Tables 1,2 summarize the baseline clinical characteristics of these cohorts.

Performance analysis of radiomics, DL, and fusion models

A total of 1,834 radiomics features were extracted from the tumor region. The LASSO algorithm revealed that 11 key radiomics features were strong related to PD-L1 expression. Based on these key features, the RAsig obtained AUC of 0.764 (95% CI: 0.672–0.856) to predict PD-L1 expression in the validation set. The ResNet101 extracted 4,096 DL features. After LASSO regression, 6 features were retained to develop the DLsig, which achieved an AUC of 0.858 (95% CI: 0.790–0.927). When the selected radiomics features and DL features were integrated, the overall performance of the fusion model RADLsig achieved an AUC of 0.954 (95% CI: 0.901–0.986). DeLong test showed that the AUC of RADLsig was higher than the AUCs of RAsig and DLsig with statistical significance (P<0.001 and P=0.03, respectively). The ROC curves were shown in Figure 5. In the immunotherapy cohort, among 87 patients predicted by RADLsig to be PD-L1 expression positive, 56 (64.4%) showed response to immunotherapy, whereas only 32.8% (19/58) patients predicted to be PD-L1 expression-negative benefited from immunotherapy. Chi-squared test showed that the PD-L1 expression predicted by the RADLsig was closely associated with immunotherapy response (Figure 6, χ²=12.688 P<0.001). In the TICA cohort, RADLsig was statistically correlated with CD274 count (r=0.337, P=0.02).

Figure 5 ROC curves for the model. AUC area under the curve; CI, confidence interval; DLsig, deep learning signature; RAsig, radiomics signature; RADLsig, radiomics-deep learning signature; ROC, receiver operating characteristic.

Figure 6 Correlation between PD-L1 expression predicted by RADLsig and immunotherapy response. PD-L1, programmed death ligand-1; RADLsig, radiomics-deep learning signature.

Discussion

In this study, we derived a RADLsig from routine pre-treatment CT scans of NSCLC patients to noninvasively predict PD-L1 expression, which achieved high predictive performance with AUC of 0.954. In the clinical setting, the response to immunotherapy was higher in the NSCLC patients who were predicted to be PD-L1 positive by RADLsig than that in the PD-L1 negative patients. The signature also correlated to CD274 count, representing the RNA expression of the protein-encoding gene for PD-L1.

Immunotherapy, especially ICIs, has substantially changed the therapeutic strategies for NSCLC patients. Due to the heterogeneity of NSCLCs, the more precise selection of patients who would be likely to benefit from immunotherapy is necessary. Currently, IHC for PD-L1 expression in NSCLCs is the mainstay biomarker for patient selection. However, several limitations of this tissue sample-dependent procedure hinder its broad application in the clinical setting. Our primary study showed the potential of the CT-based RADLsig to be a noninvasive imaging biomarker to resolve this problem.

Previous studies have explored and validated the accuracy of radiomics models derived from chest CT as imaging biomarkers for quantifying and predicting PD-L1 expression status in NSCLC patients, achieving AUCs of 0.76–0.85 (23), a range also supported by recent work in early-stage NSCLC (24). Radiomics requires domain expertise and human engineering to design specific features. Its performance cannot be improved even if the data size is increased. In contrast, DL can develop the features or representations needed for pattern recognition (25-27), demonstrating its utility in complex predictive tasks such as therapy response assessment (28). Therefore, its performance scales to large datasets and can continue to improve with more data. The present study showed that the predictive performance of the DL model was higher than that of the radiomics model. As the radiomics features are expertly designed according to statistics and texture, they also harbor useful information. Recent studies have revealed that a fusion model integrating radiomics and DL could improve predictive performance dramatically (29-31), an approach gaining traction in oncology imaging, including for metastatic prediction (32). This study also showed that the performance of the fusion model was superior to that of the radiomics model and the DL model. Furthermore, we tested the fusion model for predicting the response of NSCLC patients to immunotherapy in the clinical setting. It was demonstrated that patients tested with the model predicting PD-L1-positive NSCLC exhibited a better response to immunotherapy. The correlation of this signature with CD274 might be a biological explanation. To our knowledge, this is the first study to derive a CT-based RADLsig for PD-L1 expression of NSCLC patients with testing in immunotherapy patients and a gene-based explanation.

The RADLsig achieved a high AUC of 0.954 in our internal validation set. While this result is promising, we acknowledge that such a high value warrants rigorous scrutiny to exclude the possibility of overfitting or data leakage, particularly in the absence of an external test set. We have strictly enforced data segregation throughout our analytical pipeline. The high performance may be attributed to the synergistic effect of combining complementary information from handcrafted radiomic features and DL features. Furthermore, this signature demonstrated strong predictive value for immunotherapy response in an independent clinical cohort and showed a significant correlation with CD274 expression at the transcriptomic level, providing external validation of its biological and clinical relevance. Nevertheless, this study has some limitations. First, the internal validation was conducted using a random split of a single-center retrospective dataset. While this approach assesses model stability within the same data distribution, it does not test the model’s robustness against temporal shifts or evolving clinical practices, which is a more rigorous test of generalizability. Second, the single-center retrospective design itself, coupled with the lack of external validation, which may limit the generalizability of its findings. Future research should adopt a prospective cohort study design, expand sample size through multicenter collaboration, and include more representative subjects to provide a more comprehensive validation of existing radiomics models. Third, radiomics studies based solely on conventional non-contrast CT may struggle to comprehensively assess tumor tissue heterogeneity. Subsequent research should incorporate contrast-enhanced CT imaging data to more accurately evaluate the potential impact of tumor heterogeneity on study outcomes. Finally, due to the relatively short follow-up periods for some cases, the completeness of data for key prognostic indicators such as disease-free survival (DFS) and progression-free survival (PFS) was insufficient, compromising the reliability of survival analysis results. Therefore, extending follow-up periods to obtain more comprehensive clinical outcome data is necessary to validate the prognostic assessment value of predictive models.

Conclusions

In conclusion, this study demonstrated that a CT-based RADLsig could predict PD-L1 expression and response to immunotherapy in NSCLC patients noninvasively, which showed the potential to be an imaging biomarker to aid the selection of patients for immunotherapy and fuel clinical translation of personalized treatment.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-1-1433/rc

Data Sharing Statement: Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-1-1433/dss

Peer Review File: Available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-1-1433/prf

Funding: This study was supported by The National Nature Science Foundation of China (No. 82172026) and Shenzhen Medical Research Fund (No. C2401005).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-2025-1-1433/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by Institutional Review Board of Shenzhen People’s Hospital (approval No. LL-KY-2021058). Individual consent for this analysis was waived due to the retrospective nature.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Bray F, Laversanne M, Sung H, et al. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 2024;74:229-63. [Crossref] [PubMed]
2. Zheng J, Xu S, Wang G, et al. Applications of CT-based radiomics for the prediction of immune checkpoint markers and immunotherapeutic outcomes in non-small cell lung cancer. Front Immunol 2024;15:1434171. [Crossref] [PubMed]
3. Wang F, Chen W, Chen F, et al. Risk stratification and overall survival prediction in extensive stage small cell lung cancer after chemotherapy with immunotherapy based on CT radiomics. Sci Rep 2024;14:22659. [Crossref] [PubMed]
4. Zhang Y, Vaccarella S, Morgan E, et al. Global variations in lung cancer incidence by histological subtype in 2020: a population-based study. Lancet Oncol 2023;24:1206-18. [Crossref] [PubMed]
5. Ettinger DS, Wood DE, Aisner DL, et al. NCCN Guidelines® Insights: Non-Small Cell Lung Cancer, Version 2.2023. J Natl Compr Canc Netw 2023;21:340-50. [Crossref] [PubMed]
6. Reck M, Rodríguez-Abreu D, Robinson AG, et al. Five-year outcomes with pembrolizumab versus chemotherapy for metastatic non-small-cell lung cancer with PD-L1 tumor proportion score ≥ 50. J Clin Oncol 2021;39:2339-49. [Crossref] [PubMed]
7. de Castro G Jr, Kudaba I, Wu YL, et al. Five-year outcomes with pembrolizumab versus chemotherapy as first-line therapy in patients with non-small-cell lung cancer and programmed death ligand-1 tumor proportion score ≥ 1% in the KEYNOTE-042 study. J Clin Oncol 2023;41:1986-91. [Crossref] [PubMed]
8. Herbst RS, Giaccone G, de Marinis F, et al. Atezolizumab for first-line treatment of PD-L1-selected patients with NSCLC. N Engl J Med 2020;383:1328-39. [Crossref] [PubMed]
9. Topalian SL, Hodi FS, Brahmer JR, et al. Safety, activity, and immune correlates of anti-PD-1 antibody in cancer. N Engl J Med 2012;366:2443-54. [Crossref] [PubMed]
10. Garon EB, Ciuleanu TE, Arrieta O, et al. Ramucirumab plus docetaxel versus placebo plus docetaxel for second-line treatment of stage IV non-small-cell lung cancer after disease progression on platinum-based therapy (REVEL): a multicentre, double-blind, randomised phase 3 trial. Lancet 2014;384:665-73. [Crossref] [PubMed]
11. Wu B, Huang X, Shi X, et al. LAMTOR1 decreased exosomal PD-L1 to enhance immunotherapy efficacy in non-small cell lung cancer. Mol Cancer 2024;23:184. [Crossref] [PubMed]
12. Hwang M, Canzoniero JV, Rosner S, et al. Peripheral blood immune cell dynamics reflect antitumor immune responses and predict clinical response to immunotherapy. J Immunother Cancer 2022;10:e004688. [Crossref] [PubMed]
13. Mu W, Jiang L, Shi Y, et al. Non-invasive measurement of PD-L1 status and prediction of immunotherapy response using deep learning of PET/CT images. J Immunother Cancer 2021;9:e002118. [Crossref] [PubMed]
14. Gillies RJ, Kinahan PE, Hricak H. Radiomics: Images Are More than Pictures, They Are Data. Radiology 2016;278:563-77. [Crossref] [PubMed]
15. Lambin P, Leijenaar RTH, Deist TM, et al. Radiomics: the bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol 2017;14:749-62. [Crossref] [PubMed]
16. Liu Z, Yao Y, Zhao M, et al. Radiomics Models Derived From Arterial-Phase-Enhanced CT Reliably Predict Both PD-L1 Expression and Immunotherapy Prognosis in Non-small Cell Lung Cancer: A Retrospective, Multicenter Cohort Study. Acad Radiol 2025;32:493-505. [Crossref] [PubMed]
17. Fan S, Xie J, Zheng S, et al. Non-invasive CT based multiregional radiomics for predicting pathologic complete response to preoperative neoadjuvant chemoimmunotherapy in non-small cell lung cancer. Eur J Radiol 2025;189:112171. [Crossref] [PubMed]
18. Xu T, Liu X, Chen Y, et al. CT-based deep learning radiomics biomarker for programmed cell death ligand 1 expression in non-small cell lung cancer. BMC Med Imaging 2024;24:196. [Crossref] [PubMed]
19. Zhang X, Zhang G, Qiu X, et al. Non-invasive decision support for clinical treatment of non-small cell lung cancer using a multiscale radiomics approach. Radiother Oncol 2024;191:110082. [Crossref] [PubMed]
20. Han K, Wang Y, Chen H, et al. A Survey on Vision Transformer. IEEE Trans Pattern Anal Mach Intell 2023;45:87-110. [Crossref] [PubMed]
21. Dosovitskiy A, Beyer L, Kolesnikov A, et al. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. Proc Int Conf Learn Representations; 2021.
22. Seymour L, Bogaerts J, Perrone A, et al. iRECIST: guidelines for response criteria for use in trials testing immunotherapeutics. Lancet Oncol 2017;18:e143-52. [Crossref] [PubMed]
23. Peng J, Zou D, Zhang X, et al. A novel sub-regional radiomics model to predict immunotherapy response in non-small cell lung carcinoma. J Transl Med 2024;22:87. [Crossref] [PubMed]
24. Zhuo Y, Wang Q, Zhan Y, et al. Clinicopathological-CT model for predicting PD-L1 expression in resectable early-stage non-small cell lung cancer. J Thorac Dis 2025;17:10954-68. [Crossref] [PubMed]
25. Esteva A, Robicquet A, Ramsundar B, et al. A guide to deep learning in healthcare. Nat Med 2019;25:24-9. [Crossref] [PubMed]
26. Patel AN, Srinivasan K. Deep learning paradigms in lung cancer diagnosis: A methodological review, open challenges, and future directions. Phys Med 2025;131:104914. [Crossref] [PubMed]
27. AlOsaimi HM, Alshilash AM, Al-Saif LK, et al. AI models for the identification of prognostic and predictive biomarkers in lung cancer: a systematic review and meta-analysis. Front Oncol 2025;15:1424647. [Crossref] [PubMed]
28. Wang D, Wang M, Song J, et al. Computed tomography-based texture analysis for predicting adjuvant therapy response in postoperative patients with EGFR-mutant non-small cell lung cancer. J Thorac Dis 2025;17:11078-89. [Crossref] [PubMed]
29. Huang Y, Song X, Chen Y, et al. Intratumoral microbiome-related MRI model for predicting breast cancer shrinkage pattern following neoadjuvant therapy. Radiology 2025;316:e243545. [Crossref] [PubMed]
30. Xu S, Ying Y, Hu Q, et al. Fusion model integrating multi-sequence MRI radiomics and habitat imaging for predicting pathological complete response in breast cancer treated with neoadjuvant therapy. Cancer Imaging 2025;25:108. [Crossref] [PubMed]
31. Zhang Y, Cai J, Cui C, et al. Predicting breast cancer response to neoadjuvant therapy by integrating radiomic and deep-learning features from early-and-peak phases of DCE-MRI. BMC Cancer 2025;25:1747. [Crossref] [PubMed]
32. Duan F, Zu H, Zhang R, et al. Deep learning radiomics and (18)F-FDG PET/CT imaging: mediastinal lymph node characteristics as predictors of metastasis in non-small cell lung cancer. Transl Lung Cancer Res 2025;14:4301-14. [Crossref] [PubMed]