CT-based radiomics signature to predict CD8+ tumor infiltrating lymphocytes in non-small-cell lung cancer

Abstract

Background

An abundance of CD8+ tumor infiltrating lymphocytes (TILs) in the center of solid tumors is a reliable predictive biomarker for patients eligible for immunotherapy.

Purpose

To develop a computed tomography (CT)-based radiomics signature for a preoperative prediction of an abundance of CD8+ TILs in non-small-cell lung cancer (NSCLC).

Material and Methods

In this retrospective study, 117 consecutive patients with pathologically confirmed NSCLC were included and randomly divided into training (n = 77) and test sets (n = 40). A total of 107 radiomics features were extracted from the three-dimensional volumes of interest of each patient. Least absolute shrinkage and selection operator (LASSO) regression was used to select the strongest features for abundance of CD8+ TILs in NSCLC, and the radiomics score was constructed through a linear combination of these selected features. Receiver operating characteristic (ROC) curve analysis was used to evaluate the predictive performance of the radiomics score.

Results

The radiomics score was associated with an abundance of CD8+ TILs in NSCLC, which achieved an area under the curve (AUC) of 0.83 (95% CI=0.73–0.92) and 0.68 (95% CI=0.54–0.87) in the training and test sets, respectively. The difference was not statistically significant (P = 0.20). The tumors with high CD8+ TILs tended to have heterogeneous dependences (high value of Dependence Non-Uniformity Normalized) and complicated texture (high value of Informational Measure of Correlation 1).

Conclusion

CT-based radiomics features have the ability to predict CD8+ TILs expression levels of an abundance of CD8+ TILs in NSCLC, which was shown to be a potential imaging biomarker for stratifying patients who may benefit from immunotherapy.

Keywords

Non-small cell lung cancer radiomics immunotherapy CD8+ tumor infiltrating lymphocyte

Introduction

Lung cancer is the leading cause of cancer-related mortality, accounting for almost one-quarter of all cancer deaths (1,2). Non-small-cell lung cancer (NSCLC) is the most common form of lung cancer, with a prevalence of approximately 85%. Although the widespread availability of computed tomography (CT) has made it possible to detect early lesions, the majority of NSCLCs are diagnosed at an advanced stage, accounting for the poor prognosis. In recent years, immune checkpoint inhibitors (ICIs), which block inhibitory receptors and harness the host's immune system to mount effective antitumor responses, have shown improvements in overall survival relative to standard chemotherapy and have been rapidly incorporated into the standard management of advanced stage NSCLC (3,4,5).

However, only a small portion (20%–30%) of patients showed a response to immunotherapy (6 –9). Meanwhile, ICIs can arouse a spectrum of immune-related adverse events (irAEs) through immune-mediated injuries or toxicities, which affect multiple organs and systems, including the fatal checkpoint inhibitor pneumonitis (CIP) (10 –12). Therefore, it is essential to select patients who would benefit from ICI treatment. Tumor infiltrating lymphocytes (TILs), particularly the CD8+ TILs, and their immunoregulatory cytokines represent adaptive immunity, execute key effector cytotoxic functions in the tumor immune microenvironment (TIME), and mediate responses to ICI treatment (13). Previous studies showed that pre-existing antitumor responses have been assessed by high infiltration of CD8+ TILs and were correlated with clinical responses to ICIs, which could be used as a biomarker to patient selection (14). Assessment of infiltration and abundance of CD8+ TILs mainly depends on proteomic and genomic analyses, which require biopsies or operations to obtain tissue samples. As NSCLC is a high spatial and temporal heterogeneous disease, small portions of tumor tissue cannot reflect the global characteristics of tumors. Furthermore, the invasion process is often impossible at the advanced stage and sometimes worsens patients' prognoses. Radiomics, with its high-throughput extraction of quantitative digital features from medical imaging data, has shown to be a promising approach for predicting responses to immunotherapy of various tumors (15 –17). As a non-invasive modality, radiomics exhibits several advantages over tissue-dependent methods, such as surveying spatial and temporal heterogeneity and monitoring the treatment response. In this study, we extract radiomics features from preoperative CT images of NSCLCs, build a radiomics signature for CD8+ TILs, and intend to establish a machine learning model to predict the abundance of CD8+ TILs in NSCLCs.

Material and Methods

Patients

The institutional review board approved this retrospective study, and the requirement for informed consent was waived due to its retrospective nature. A total of 318 consecutive patients were found in the hospital information system (HIS) with the following inclusion criteria: (i) CT examinations being performed within three months before surgery; (ii) CT images being successfully retrieved from the picture archiving and communication system (PACS); and (iii) a single nodule or mass, within the resected pulmonary tissue, which could be paired with the tissue sample. Exclusion criteria were as follows: (i) history of any anticancer treatment before this time (n = 3); (ii) the tissue samples were unavailable and unsuitable for immunohistochemistry (IHC) (n = 142); (iii) the slice escaped from the glass slide (n = 53); (iv) inadequate quality for image segmentation (n = 2); and (v) failure to extract radiomics features (n = 1). Finally, 117 patients (50 men, 67 women; mean age = 59.01 ± 11.76 years) were included in this study (Fig. 1). In total, 14 clinical and pathological characteristics, including sex, smoking, age, family history, TNM stage, invasion, differentiation, lymph nodes invasion, pleural invasion, venous invasion, nerve invasion, CYFRA21-1, CEA, and CA125, were obtained from medical records in the hospital.

Immunohistochemistry

All of these patients' postoperative specimens were performed with IHC through this procedure: (i) section dewaxing: 5-μm pathological sections were dewaxed with xylene and rehydrated with graded alcohol solution for a few minutes; (ii) antigen retrieval: antigens were recovered by microwaving under medium power for 8 min, turning off the heat and keeping warm for 8 min, and microwaving under low to medium power for 8 min; (iii) hydrogen peroxide incubation: sections were incubated with 3% hydrogen peroxide for 25 min; (iv) section staining: these sections were stained by anti-CD8 (Clone No.1G2B10; Proteintech, PR China) at 4 °C for a night and then incubated with the secondary goat anti-rabbit antibodies (Jackson ImmunoResearch) at room temperature for 50 min. Fresh 3′3′diamino-benzidine (DAB) was used to visualize CD8 staining, and sections were counterstained with hematoxylin.

Evaluation of CD8+TIL expression

Two experienced pathologists, who were blinded to the patient's clinical information, reviewed the IHC slides. The expression of CD8+ TILs was assessed according to the percentage of positive lymphocytes in the stromal compartment in five randomly selected high-power fields. Patients were dichotomized into high or low infiltration of CD8+ TILs by thresholds defined as medium.

CT-based radiomics feature extraction

CT examinations were performed using a 128-detector CT scanner (Brilliance iCT, Philips Medical Systems, Best, the Netherlands). Thin-slice non-contrast-enhanced CT images with collimation of 0.625 × 128, slice thickness of 1.0 mm with a gap of 0.5 mm, and a standard reconstruction algorithm were downloaded from PACS in a Digital Imaging and Communications in Medicine (DICOM) format and transferred to a PC. The tube voltage was 120 KV, and the tube current was automatically adjusted. A radiologist with three years of experience in thoracic radiology segmented the tumors on each slice manually in lung window setting (level = −600 HU; width = 1200 HU) using ITK-SNAP (http://www.itksnap.org) to obtain a three-dimensional volume of interest (VOI) of the whole tumor, which included both solid components and ground-glass opacity of the tumor. Radiomic features were extracted from the VOI by using the open-source software Pyradiomics package. A total of 107 features were extracted, consisting of 18 first-order statistical features, 14 shape features, 24 gray level co-occurrence matrix (GLCM) features, 16 gray level size zone matrix (GLSZM) features, 16 gray level run length matrix (GLRLM) features, five neighboring gray tone difference matrix (NGTDM) features, and 14 gray level dependence matrix (GLDM) features.

Statistical analysis

Statistical analysis was performed using R software version 4.0.0 (http://www.r-project.org) and SPSS version 22.0.0.1 (IBM Corp, Armonk, NY, USA). A P value <0.05 indicated statistical significance. Univariable analysis was used to select those radiomics features and clinical and pathological characteristics that were associated with the abundance of CD8+ TILs with statistical significance. Then the Pearson correlation coefficient was used to exclude those features highly correlated to each other (if two features had a high correlation [r > 0.8], the one with lower area under the receiver operating characteristic [ROC] curve [AUC] was removed). The 117 patients were divided into training and test sets at a ratio of 0.7:0.3 randomly. The selected features were normalized using z-score normalization to reduce the impact of different scales in the training and validation sets separately. Least absolute shrinkage and selection operator (LASSO) regression with fivefold cross-validation was employed to select the strongest features in the training set. Subsequently, the radiomics score for abundance of CD8+ TILs was formulated using a linear combination of these selected features. The predictive performance of radiomics score for abundance of CD8+ TILs was evaluated in both the training and test sets using AUCs of the ROCs. The AUCs were compared with the DeLong test using the “pROC” package in R.

Results

Patient characteristics

Patient characteristics are listed in Table 1. There was no statistically significant difference between clinicopathological characteristics and CD8+ TILs infiltration level at univariable analysis.

Table 1.

Patient characteristics.

Variable	All patients (n = 117)	Training set (n = 77)			Validation set (n = 40)
Variable	All patients (n = 117)	High CD8 TIL infiltration (n = 25)	Low CD8 TIL infiltration (n = 52)	P value	High CD8 TIL infiltration (n = 16)	Low CD8 TIL infiltration (n = 24)	P value
Sex				0.638			0.747
Female	67 (57.3)	13 (52.0)	30 (57.7)		11 (64.7)	13 (56.5)
Male	50 (42.7)	12 (48.0)	22 (42.3)		6 (35.3.0)	10 (43.6)
Smoking
Presence	24 (20.5)	6 (20.8)	10 (19.2)	0.629	2 (11.8)	6 (26.1)	0.472
Absence	93 (79.5)	19 (79.2)	42 (82.7)		15 (88.2)	17 (73.9)
Age (years)	59.09 ± 11.8	59.36 ± 10.57	59.40 ± 12.61	0.988	57.35 ± 11.37	59.78 ± 12.39	0.530
Family history				0.790			0.184
Presence	17 (14.5)	3 (12.0)	9 (17.3)		4 (23.5)	1 (4.3)
Absence	100 (85.5)	22 (88.0)	43 (82.7)		13 (76.5)	22 (95.7)
TNM stage
I	82 (86.3)	17 (89.5)	36 (80.0)	1.000	14 (100)	14 (82.4)	0.488
II	7 (7.4)	1 (5.3)	4 (8.9)		0 (0.0)	2 (11.8)
III	5 (5.3)	1 (5.3)	4 (8.9)		0 (0.0)	0 (0.0)
IV	1 (1.1)	0 (0.0)	1 (1.6)		0 (0.0)	1 (5.9)
Invasion				0.456			0.097
Microinvasive	21 (18.4)	2 (8.3)	9 (18.0)		7 (41.2)	3 (13.0)
Invasive	93 (81.6)	22 (91.7)	41 (82.0)		10 (58.8)	20 (87.0)
Differentiation				0.301			1.000
poor	5 (23.8)	1 (12.5)	4 (30.8)		0 (0.0)	8 (88.9)
moderate	15 (71.4)	6 (75.0)	9 (75.0)		4 (100.0)	1 (11.1)
Well	1 (4.8)	1 (12.5)	0 (0.0)		0 (0.0)	0 (0.0)
Lymph node invasion				0.421			1.000
Presence	8 (7.5)	1 (4.0)	6 (13.0)		0 (0.0)	1 (5.3)
Absence	99 (92.5)	24 (96)	40 (87.0)		16 (100.0)	18 (94.7)
Pleural invasion				0.589			0.347
Presence	16 (13.7)	2 (8.0)	8 (15.4)		1 (5.9)	5 (21.7)
Absence	101 (88.0)	23 (92.0)	44 (84.6)		16 (94.1)	18 (78.3)
Venous invasion				0.749			-
Presence	9 (7.7)	2 (8.0)	7 (13.5)		0 (0.0)	0 (0.0)
Absence	108 (92.3)	23 (92.0)	45 (86.5)		17 (100.0)	22 (95.7)
Nerve invasion				-
Presence	0 (0)	0 (0.0)	0 (0.0)		0 (0.0)	0 (0.0)	-
Absence	117 (100)	25 (100.0)	52 (100.0)		17 (100)	23 (100.0)
CYFRA21-1				0.134			0.689
＞3.3ng/ml	29 (25.7)	4 (16.7)	17 (33.3)		4 (25.0)	3 (14.3)
≤3.3ng/ml	84 (74.3)	20 (83.3)	34 (66.7)		12 (75.0)	18 (85.7)
CEA				0.197			0.347
＞5 μg/L	16 (13.7)	2 (8.0)	12 (23.1)		0 (0.0)	3 (13.0)
≤5 μg/L	101 (86.3)	23 (92.0)	40 (76.9)		17 (100)	20 (87.0)
CA125				0.390			0.795
＞35U/m	6 (5.2)	0 (0.0)	4 (7.8)		1 (5.9)	3 (13.6)
≤35U/ml	116 (94.8)	24 (100.0)	47 (92.2)		16 (94.1)	19 (86.4)
EGFR				0.184			0.378
Mutation	43 (55.1)	7 (38.9)	22 (57.9)		5 (50)	9 (75)
Wildtype	35 (44.9)	11 (61.1)	16 (42.1)		5 (50)	3 (35)

Values are given as n (%) or mean ± SD.

EGFR, estimated glomerular filtration rate; TIL, tumor infiltrating lymphocyte.

Radiomics score construction and predictive performance

At univariable analysis, 47 features were associated with the infiltration level of CD8+ TILs with statistical significance, while 60 features were removed due to no significant difference between high and low CD8+ TIL infiltration groups. In total, 11 features remained after removing the redundant features using Spearman's correlation coefficient at a threshold value of 0.8. LASSO regression selected seven features to construct the radiomics score (Table 2). The feature with the highest coefficient among the seven selected features was Dependence Nonuniformity Normalized (a measurement for homogeneity with a higher value indicating more heterogeneity), followed by Informational Measure of Correlation 1 (a measurement for complexity of texture with a higher value indicating higher complexity). The radiomics score of tumors with high infiltration of CD+ TILs was higher than that of tumors with low infiltration of CD8+ TILs (−0.12 ± 0.95 vs −1.40 ± 0.22; P < 0.001). Representative cases are shown in Figs. 2 and 3. As indicated, tumors with heterogeneous dependence and complex texture tended to have a high radiomics score. The AUCs of the training and test sets were 0.83 (95% confidence interval [CI] = 0.73–0.92) and 0.68 (95% CI = 0.54–0.87), respectively. The difference was not statistically significant (P = 0.20) (Fig. 4).

Table 2.

The seven radiomics features for radiomics signature development and their clinical meaning.

Radiomic features	λ coefficient	Clinical meaning
Dependence Non-Uniformity Normalized	1.68	Measures the similarity of dependence throughout the image, with a lower value indicating more homogeneity among dependencies in the image.
Informational Measure of Correlation 1	1.55	Assesses the complexity of the texture, with 0 indicating the independent distributions and no mutual information, while −1 indicates complete dependence (not necessarily uniform; low complexity).
Uniformity	0.96	This is a measure of the homogeneity of the image array, where a greater uniformity implies a greater homogeneity or a smaller range of discrete intensity values.
Gray Level Non-Uniformity	−1.12	Measures the similarity of gray-level intensity values in the image, where a lower GLN value correlates with a greater similarity in intensity values.
Strength	−0.52	Strength is a measure of the primitives in an image. Its value is high when the primitives are easily defined and visible, so an image with slow change in intensity but more large coarse differences in gray level intensities.
Contrast	0.84	Contrast is a measure of the local intensity variation, favoring values away from the diagonal. A larger value correlates with a greater disparity in intensity values among neighboring voxels.
Busyness	−0.50	A measure of the change from a pixel to its neighbor. A high value for busyness indicates a “busy” image, with rapid changes of intensity between pixels and its neighborhood.

GLN, Gray Level Non-Uniformity.

Discussion

In this primary study, we observed that a radiomics signature derived from preoperative CT could be used to predict the abundance of CD8+ TILs in NSCLC, which showed the probability that the radiomics signature could be used as an imaging biomarker to stratify patients who are more likely to respond to ICI treatment. The radiomics score was constructed from a linear combination of seven radiomics features, which consisted of Dependence Non-Uniformity Normalized, Informational Measure of Correlation 1, Uniformity, Gray Level Non-Uniformity, Strength, Contrast, and Busyness. Radiomics score of tumors with high CD8+ TILs were higher than those of tumors with low CD8+ TILs (−0.12 ± 0.95 vs −1.40 ± 0.22; P < 0.001). The feature with the highest coefficient among the seven selected features was Dependence Nonuniformity Normalized, followed by Informational Measure of Correlation 1. Dependence Nonuniformity Normalized, which belongs to GLDM features, measures the similarity of dependence throughout the image, with a lower value indicating more homogeneity among dependencies in the image. Informational Measure of Correlation 1 is a GLCM feature, which quantifies the complexity of the texture. Therefore, tumors with heterogeneous dependence and complicated texture tended to have a higher radiomics score. The radiomics score calculated from these selected features through a linear combination obtained AUCs for the CD8+ TIL infiltration level of 0.83 (95% CI = 0.73–0.92) and 0.68 (95% CI = 0.54–0.87) in the training and test sets, respectively.

Over the past two decades, important advancement has shifted the treatment of NSCLC from the empirical use of cytotoxic therapy to personalized medicine according to the genetic alterations and the immune microenvironment status of tumors. Stratifying patients who would benefit from ICI treatment is needed. An abundance of CD8+ TILs was found to be a reliable biomarker to predict the response to immunotherapy in various tumors in previous studies (18 –20). Based on rich or poor infiltration of T lymphocytes in the center of tumors, solid tumors can be defined as either a “T cell inflamed” phenotype (“hot tumors”) or “non-inflamed” (“cold tumors”) (21 –23). The “cold” tumors, which lack lymphocytic infiltration, are refractory to immunotherapy. In contrast, ICIs can rescue anti-tumor T cell activity of the “hot” tumors, which are characterized by high infiltration of cytotoxic lymphocytes (CTLs) expressing PD-1 and leukocytes and tumor cells expressing the immune-dampening PD-1 ligand PD-L1 (24,25). However, tissue samples cannot reflect the spatial and temporal heterogeneity of solid tumors. Radiomics, which extracts a large number of digital features from medical images and harnesses the data mining algorithm to explore phenotypes in the medical image driven by underlying biological patterns, has redefined medical imaging as a data source more than a picture (26 –28). Two recent studies showed that the radiomics score derived from CT images of multiple tumor types could be used to predict abundance of CD8+ TILs and was associated with clinical outcomes of patients treated with immunotherapy (15,17). Both studies demonstrated that tumor homogeneity was associated with high abundance of CD8+ TILs. Researchers have demonstrated that the phenotypes and abundances of CD8+ TILs may vary dramatically across cancer types (29 –32). Therefore, it is necessary to develop radiomics models in different cancer types to facilitate personalized immunotherapy and achieve maximal clinical benefit. To the best of our knowledge, this is the first study to develop a CD8+ TILs radiomics score based on CT of NSCLC. Unlike the studies of Sun and Ligero, our study showed that tumors with heterogeneous dependences and complicated textures had high infiltration of CD8+ TILs. This apparent contradiction may come from the different tumor types used to construct the radiomics score (15,17). Because malignancies are high heterogeneity diseases, different tumors have different driven gene mutations, molecular characteristics, and tumor microenvironments, especially different tissue and organ backgrounds, which may influence tumors' imaging phenotypes dramatically. With the exception of air inflated parenchyma, NSCLC can be detected at the early stage as pure ground-glass opacity and then develop to partly solid nodules or solid masses. Therefore, NSCLC may have different characters from other solid tumors during its evolution. As CD8+ TILs are cytotoxic T cells and can produce IFN-γ, thereby resulting in an inflammatory environment, central infiltration of CD8+ TILs may lead to central necrosis and edema. Therefore, non-uniform texture can be induced. Sanli et al. reported that high tumor heterogeneity on FDG PET/CT compared tumor was positively correlated to overall survival in patients with metastatic melanoma receiving immunotherapy with a hazard ratio (HR) of 0.3 (95% CI = 0.1–0.5; P = 0.02) (33). Yu et al. reported that tumors in the high TILs group had a more non-uniform density and a smoother gradient of the tumor pattern than the low TILs group of triple-negative breast cancer in the mammography (34). Our findings were consistent with the results of these two studies in that tumors with higher CD8+ TILs tend to be heterogeneous. In the present study, the preoperative CT-based radiomics score obtained AUCs of 0.83 (95% CI = 0.73–0.92) and 0.68 (95% CI = 0.54–0.87) to stratify high and low infiltration of CD8+ TILs of NSCLC in the training and validation sets, respectively. The predictive performance was similar to the results of several recent studies on miscellaneous tumors (14,35,36). All these studies, including our present study, showed that the radiomics signature derived from solid tumors could obtain moderate predictive performance for abundance of CD8+ TILs.

The present study has some limitations. First, this small group only included those patients who had surgery at a single center. Although we performed feature selection before radiomics score construction using LASSO regression, our data may suffer from the curse of dimensionality. Therefore, further studies on large datasets from multiple institutions are warranted to reduce the risk of overfitting and increase the generalizability of the model. Second, this study did not take treatment response and prognosis into account due to the retrospective nature. Not all the patients received ICI treatment, and strict follow-up protocol was not completed. Third, the expert-designed radiomics features might not cover potential features associated with the abundance of CD8+ TILs. Fourth, in this study, radiomics features were not extracted from contrast-enhanced images because most tumors were resected at an early stage as ground-glass opacity or subsolid nodules.

In conclusion, we developed a preoperative CT-based radiomics signature to predict the abundance of CD8+ TILs of NSCLC, which showed to be a potential imaging biomarker for stratifying patients who may benefit from ICI treatment. Due to this model being constructed on a small patient group from a single center, further studies with large sample sizes from different institutions are warranted to validate the clinical usefulness of the CT-based radiomics signature for the abundance of CD8+ TILs and responses to ICI treatment.

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the National Natural Science Foundation of China (grant no. 82172026).

ORCID iD

Jingshan Gong

References

Siegel

Miller

Fuchs

, et al. Cancer statistics, 2021. CA Cancer J Clin 2021;71:7–33.

Islami

Goding Sauer

Miller

, et al. Proportion and number of cancer cases and deaths attributable to potentially modifiable risk factors in the United States. CA Cancer J Clin 2018;68:31–54.

Callahan

Postow

Wolchok

. Targeting T cell co-receptors for cancer therapy. Immunity 2016;44:1069–1078.

Guibert

Mazieres

. Nivolumab for treating non-small cell lung cancer. Expert Opin Biol Ther 2015;15:1789–1797.

Hirsch

Scagliotti

Mulshine

, et al. Lung cancer: current therapies and new targeted treatments. Lancet 2017;389:299–311.

Antonia

Villegas

Daniel

, et al. Durvalumab after chemoradiotherapy in stage III non-small-cell lung cancer. N Engl J Med 2017;377:1919–1929.

Brahmer

Reckamp

Baas

, et al. Nivolumab versus docetaxel in advanced squamous-cell non-small-cell lung cancer. N Engl J Med 2015;373:123–135.

Herbst

Baas

Kim

, et al. Pembrolizumab versus docetaxel for previously treated, PD-L1–positive, advanced non-small-cell lung cancer (KEYNOTE-010): a randomised controlled trial. Lancet 2016;387:1540–1550.

Reck

Rodríguez-Abreu

Robinson

, et al. Pembrolizumab versus chemotherapy for PD-L1–positive non–small-cell lung cancer. N Engl J Med 2016;375:1823–1833.

10.

Naidoo

Page

, et al. Toxicities of the anti-PD-1 and anti-PD-L1 immune checkpoint antibodies. Ann Oncol 2015;26:2375–2391.

11.

Naidoo

Wang

Woo

, et al. Pneumonitis in patients treated with anti-programmed death-1/programmed death ligand 1 therapy. J Clin Oncol 2017;35:709–717.

12.

Jardim

de Melo Gagliato

Giles

, et al. Analysis of drug development paradigms for immune checkpoint inhibitors. Clin Cancer Res 2018;24:1785–1794.

13.

Altorki

Markowitz

Gao

, et al. The lung microenvironment: an important regulator of tumour growth and metastasis. Nat Rev Cancer 2019;19:9–31.

14.

Tumeh

Harview

Yearley

, et al. PD-1 blockade induces responses by inhibiting adaptive immune resistance. Nature 2014;515:568–571.

15.

Sun

Limkin

Vakalopoulou

, et al. A radiomics approach to assess tumour-infiltrating CD8 cells and response to anti-PD-1 or anti-PD-L1 immunotherapy: an imaging biomarker, retrospective multicohort study. Lancet Oncol 2018;19:1180–1191.

16.

Sun

Sundahl

Hecht

, et al. Radiomics to predict outcomes and abscopal response of patients with cancer treated with immunotherapy combined with radiotherapy using a validated signature of CD8 cells. J Immunother Cancer 2020;8:e001429. doi: 10.1136/jitc-2020-001429.

17.

Ligero

Garcia-Ruiz

Viaplana

, et al. A CT-based radiomics signature is associated with response to immune checkpoint inhibitors in advanced solid tumors. Radiology 2021:299:109–119.

18.

Thompson

Zahurak

Murphy

, et al. Patterns of PD-L1 expression and CD8 T cell infiltration in gastric adenocarcinomas and associated immune stroma. Gut 2017;66:794–801.

19.

Wang

Liu

, et al. The prognostic value of PD-L1 expression for non-small cell lung cancer patients: a meta-analysis. Eur J Surg Oncol 2015;41:450–456.

20.

Huang

Wang

Luo

, et al. Relationship between PD-L1 expression and CD8 + T-cell immune responses in hepatocellular carcinoma. J Immunother 2017;40:323–333.

21.

Sharma

Allison

. The future of immune checkpoint therapy. Science 2015;348:56–61.

22.

Demaria

Coleman

Formenti

. Radiotherapy: changing the game in immunotherapy. Trends Cancer 2016;2:286–294.

23.

Bonaventura

Shekarian

Alcazer

, et al. Cold tumors: a therapeutic challenge for immunotherapy. Front Immunol 2019;10:168.

24.

Binnewies

Roberts

Kersten

, et al. Understanding the tumor immune microenvironment (TIME) for effective therapy. Nat Med 2018;24:541–550.

25.

Chen

, et al. Application of PD-1 blockade in cancer immunotherapy. Comput Struct Biotechnol J 2019;17:661–674.

26.

Mayerhoefer

Materka

Langs

, et al. Introduction to radiomics. J Nucl Med 2020;61:488–495.

27.

Gillies

Kinahan

Hricak

. Radiomics: images are more than pictures, they are data. Radiology 2016;278:563–577.

28.

Aerts

Velazquez

Leijenaar

, et al. Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach. Nat Commun 2014;5:4006.

29.

Wang

Jia

, et al. The prognostic role of circulating CD8(+) T cell proliferation in patients with untreated extensive stage small cell lung cancer. J Transl Med 2019;17:402.

30.

Combes

Samad

Tsui

, et al. Discovering dominant tumor immune archetypes in a pan-cancer census. Cell 2022;185:184–203.e19.

31.

Krekorian

Fruhwirth

Srinivas

, et al. Imaging of T-cells and their responses during anti-cancer immunotherapy. Theranostics 2019;9:7924–7947.

32.

Zheng

Qin

, et al. Pan-cancer single-cell landscape of tumor-infiltrating T cells. Science 2021;374:abe6474.

33.

Sanli

Leake

Odu

, et al. Tumor heterogeneity on FDG PET/CT and immunotherapy: an imaging biomarker for predicting treatment response in patients with metastatic melanoma. AJR Am J Roentgenol 2019. doi: 10.2214/AJR.18.19796.

34.

Meng

Chen

, et al. Correlation between mammographic radiomics features and the level of tumor-infiltrating lymphocytes in patients with triple-negative breast cancer. Front Oncol 2020;10:803–814.

35.

Bian

Liu

, et al. Preoperative radiomics approach to evaluating tumor-infiltrating CD8( + ) T cells in patients with pancreatic ductal adenocarcinoma using noncontrast magnetic resonance imaging. J Magn Reson Imaging 2022;55:803–814.

36.

Wen

Yang

Zhu

, et al. Pretreatment CT-based radiomics signature as a potential imaging biomarker for predicting the expression of PD-L1 and CD8 + TILs in ESCC. Onco Targets Ther 2020;13:12003–12013.