Radiomics-based prediction of response to immune checkpoint inhibitor treatment for solid cancers using computed tomography: a real-world study of two centers
BMC Cancer volume 22, Article number: 1241 (2022)
Immune checkpoint inhibitors (ICIs) represent an approved treatment for various cancers; however, only a small proportion of the population is responsive to such treatment. We aimed to develop and validate a plain CT-based tool for predicting the response to ICI treatment among cancer patients.
Data for patients with solid cancers treated with ICIs at two centers from October 2019 to October 2021 were randomly divided into training and validation sets. Radiomic features were extracted from pretreatment CT images of the tumor of interest. After feature selection, a radiomics signature was constructed based on the least absolute shrinkage and selection operator regression model, and the signature and clinical factors were incorporated into a radiomics nomogram. Model performance was evaluated using the training and validation sets. The Kaplan–Meier method was used to visualize associations with survival.
Data for 122 and 30 patients were included in the training and validation sets, respectively. Both the radiomics signature (radscore) and nomogram exhibited good discrimination of response status, with areas under the curve (AUC) of 0.790 and 0.814 for the training set and 0.831 and 0.847 for the validation set, respectively. The calibration evaluation indicated goodness-of-fit for both models, while the decision curves indicated that clinical application was favorable. Both models were associated with the overall survival of patients in the validation set.
We developed a radiomics model for early prediction of the response to ICI treatment. This model may aid in identifying the patients most likely to benefit from immunotherapy.
Recent research has demonstrated that immunotherapy with immune checkpoint inhibitors (ICIs) is significantly effective in patients with various types of cancer. Several ICIs, including anti-programmed cell death protein–1 (PD1) and anti-programmed cell death protein ligand–1 (PDL1) antibodies, have been approved as treatments for solid cancers, such as lung, liver, gastrointestinal, bladder, and kidney cancers and melanoma [1,2,3]. Although ICIs have shown promise in the context of cancer management, some challenges with optimizing their benefits remain. Notably, the response to ICI treatment is often low in clinical environments, highlighting the need to identify patients most likely to benefit from ICIs, to reduce risks and maximize efficiency.
At present, several research groups have focused on developing effective predictors of ICI response. Some studies have identified clinical characteristics such as body mass index (BMI), sex, and age as promising predictors given their association with the response to immunotherapy [4,5,6]. However, predictors derived from the tumor microenvironment remain the most accepted indicators. In clinical practice, tumor PDL1 expression, microsatellite instability status, and tumor mutational burden are commonly recommended as criteria for guiding the selection of patients for immunotherapy [7,8,9,10]. Despite their important roles in cancer immunotherapy, these criteria have exhibited varying predictive performance across different cancer types and different degrees of invasiveness, limiting their clinical utilization.
Radiomics is an emerging technique that can extract quantitative high-dimensional data from medical images . Studying tumor-derived data generated via radiomics provides crucial information related to tumor biology. With the radiomics data, novel imaging biomarkers could be developed for patient classification and prediction of treatment response, which makes a better implementation of personalized management in cancer . As the response to immunotherapy highly relies on the tumor microenvironment, it will be possible to make an early prediction by using the radiomics features linked to the tumor biology. To examine the feasibility of predicting immunotherapy outcomes using radiomics-based tools, we conducted a real-world study on ICI-treated patients. In this study, we developed and validated a plain CT-based model for predicting the response to ICI treatment in patients with solid cancers and explored the clinical significance of the constructed model, intending to provide support for personalized cancer treatment.
The study design and basic workflow are illustrated in Fig. 1a.
Patients and data collection
This retrospective study was approved by the Institutional Research Ethics Committee of our hospital and the requirement for informed consent was waived from patients. Patients who received ICIs at two medical centers (Institution I and Institution II) from October 2019 to October 2021 were included. Inclusion criteria were as follows: (1) diagnosis of solid cancer and receipt of ICIs as first- or later-line treatment, (2) availability of complete clinical data, (3) availability of CT images obtained within 14 days before ICI treatment, and (4) at least one CT re-evaluation within 6 months after treatment. The exclusion criteria were as follows: (1) receipt of concurrent or subsequent local combination treatments, (2) presence of multiple primary malignancies, (3) absence of measurable lesions based on the Response Evaluation Criteria in Solid Tumors (RECIST) 1.1 , and (4) previous immunotherapy. The data of the enrolled patients were divided into training and validation sets at a 4:1 ratio using stratified random sampling. The following clinical data were obtained from medical records: sex, age, BMI, tumor type, clinical stage, number of metastases (defined as the number of metastatic regions per patient), and treatment information (e.g., ICI type, combination therapy, and line of therapy).
The treatment response defined based on the RECIST was used as the patient label to train the models. A dichotomous response status was used to label patients as responders or non-responders. Each patient was evaluated for a response during the first CT follow-up, which was performed at least 1 month after the initiation of treatment, to ensure sufficient follow-up time and evaluate consistency. Patients who exhibited a complete or partial response according to RECIST 1.1 were considered “responders,” while those who exhibited stable or progressive disease after treatment were considered “non-responders.” Overall survival (OS) was also assessed during follow up. In this study, OS was defined as the time from ICI initiation to death from any cause or the last follow up in surviving patients.
Acquisition of CT images
CT scans obtained within 14 days before treatment were used for analysis. Plain CT scans were acquired using spiral scanners (GE Medical Systems, Philips, and Siemens) and reconstructed in the axial plane using a tube voltage of 100–120 kVp, slice thickness of 1–8 mm, and pixel spacing of 0.62–0.98 mm. Further details concerning the imaging reconstruction parameters are included in Additional file Table S1.
Tumor segmentation and extraction of radiomics features
Tumor segmentation was performed using 3D slicer software (version 4.11.20210226, https://www.slicer.org). Two experienced radiologists segmented the regions of interest by manually delineating the boundaries of the target lesion slice by slice. The most representative lesion was selected as the region of interest for segmentation to ensure a single lesion per patient. CT images were pre-processed before feature extraction. All CT images and the corresponding segmentation data were normalized by resampling to a pixel size of 1 × 1 × 1 mm3 and discretized to a bin width of 25 Hounsfield units (HUs). In addition to the original images, wavelet-transformed images were used for feature extraction. During wavelet transformation, high- and low-pass filters were applied on the x, y, and z axes to produce eight decomposition images, which were labeled as LLL, LLH, LHL, LHH, HLL, HLH, HHL, and HHH. L and H represented low- and high-frequency signals in each direction, respectively. No limitations on feature classes were set for each extraction. Subsequently, 851 features (107 original features and 744 wavelet features) were extracted from the plain CT images using an open-source Python package: Pyradiomics (https://pyradiomics.readthedocs.io/en/latest/index.html, version 3.0.1, Computational Imaging & Bioinformatics Lab, Harvard Medical School). The Pyradiomics adheres to the Imaging Biomarker Standardization Initiative (IBSI) for the most of part. Z-score normalization with the same mean and standard deviation were performed to the extracted features in both the training and validation sets.
Intra-observer and inter-observer reliability of the radiomics features
Thirty randomly selected CT images were segmented by two radiologists during the same period to assess inter-observer reliability. Intra-observer reliability was assessed by analyzing de novo segmentations of 20 randomly selected images performed by one radiologist after a 2-week interval. The same radiologist performed the remaining image segmentations. Intra-class correlation coefficients (ICCs) were calculated for each feature to determine inter- and intra-observer agreement, with values > 0.75 denoting good agreement. Only features with good inter- and intra-observer agreement were retained.
Feature selection and construction of the radiomics signature
The feature selection and model construction were performed by using the treatment response we defined during outcome evaluation. The Student’s t-test was first performed for each extracted feature with ICCs > 0.75 to identify the features that differed significantly between responders and non-responders. Levene’s test was used to assess the equality of variances before performing t-tests. For the features that didn’t meet a normal distribution, Mann–Whitney U test was used for identification of differential features. Significant features (p < 0.05) based on the results of t-test/Mann–Whitney U test were maintained for further selection. In the training set, all significant features were entered into a least absolute shrinkage and selection operator (LASSO) regression algorithm (α = 1) to select the most predictive features with nonzero coefficients. The tuning regularization parameter, λ, was determined using 10-fold cross-validation based on the minimum criteria. As different types of cancer were included in this study, the location of the target lesions was also used as a variable in these analyses to eliminate the influence of organ heterogeneity on textural patterns. The radiomics signature was constructed by linearly combining the selected features with their nonzero coefficients. Subsequently, a radiomics score (radscore) was calculated for each patient based on the radiomics signature.
Development of a radiomics-based nomogram
Univariate logistic regression analysis was performed on the potential predictors of treatment response, including the clinical factors and radscore, for the overall population. Variables exhibiting a significant association (p < 0.05) with treatment response in the univariate analysis were used to develop a multivariate logistic regression model in the training set. A radiomics-based nomogram was constructed using the radscore and clinical factors to visualize the individualized results of this model. A nomogram score (nomoscore) was calculated for each patient based on a linear combination of selected variables and their coefficients obtained from the multivariate regression analysis.
Evaluation of predictive performance and model validation
Hosmer–Lemeshow H test was used to evaluate the calibration of the two models (radscore and nomogram) based on the training and validation data. The area under the curve (AUC) of the receiver operating characteristic (ROC) curve was used to evaluate discriminative ability. Difference between two ROC curves was tested using the DeLong’s method. The sensitivity, specificity, and accuracy of the two models were calculated for both datasets based on the optimal cut-off values determined from the training set by maximizing the Youden index. The clinical utility of the models was evaluated using decision curve analysis, which can quantify the net benefits of models at different threshold probabilities .
Exploration of the associations between established models and patient outcomes
The radscores and nomoscores were compared between responders and non-responders, and the differences between the groups were visualized using a boxplot. Survival analysis grouped by the radscores and nomoscores was performed for patients in the validation set to clarify the association between each model and OS.
All statistical analyses were performed using R (version 4.0.2; https://www.r-project.org) and Python (version 3.6.5; https://www.python.org). Continuous data are presented as the mean ± standard deviation, while categorical data are presented as numbers with percentages. The Student’s t-test was used to analyze normally distributed quantitative data, and Levene’s test was used to assess the equality of variances. The Mann–Whitney U-test was used to analyze the non-normally distributed data. Chi-square tests and Fisher’s exact tests were used to compare categorical data. Survival curves were plotted using the Kaplan–Meier method, and the difference between the two curves was assessed using the log-rank test. A Cox proportional hazards model was used to calculate the hazard ratio for OS. All statistical tests were two-sided, and statistical significance was set at p < 0.05. The following external R packages were used: (1) the “irr” package for intra-class correlation coefficient score calculation; (2) the “glmnet” package for LASSO logistic regression; (3) the “pROC” and “ROCR” packages for ROC analysis and DeLong’s test; (4) the “rms” package for the nomogram; (5) the “ResourceSelection” package for the Hosmer–Lemeshow H test; (6) the “ggDCA” package for decision curve analysis; and (7) the “survival” and “survminer” packages for survival analysis.
A total of 152 patients were included in this study, including 128 from institution I and 24 from institution II. During patient selection, 32 patients were excluded for the following reasons: unavailable pretreatment CT scans, combination with local treatment, multiple primary malignancies, no measurable lesions, and non-evaluation of treatment response due to missing follow-up. Using randomization, the data of 122 patients were allocated to the training set, while those of 30 patients were allocated to the validation set. The details are shown in Fig. 1b.
Among the included patients, 119 were men, while 33 were women. The mean age and BMI (±standard deviation) were 59.62 ± 10.07 years and 22.74 ± 3.52 kg/m2, respectively. Five types of cancer were observed, the most prevalent of which was lung cancer (n = 78), followed by gastric cancer (n = 39), hepatocellular carcinoma (n = 16), esophageal cancer (n = 13), and colorectal cancer (n = 6). Most patients had advanced disease (stage III/IV, n = 143), with the number of metastases ranging from one to five. Six types of target lesions were employed for data analysis, including those of the esophagus, liver, lung, stomach, lymph nodes, and others. Of these patients, 34.9% (n = 53) showed a clinical response to immunotherapy, whereas 65.1% (n = 99) did not. The patient treatment characteristics and detailed information are listed in Table 1.
Feature selection and construction of the radiomics signature
A total of 851 radiomic features were originally extracted from the pretreatment CT scans. During the intra- and inter-observer reliability analyses, 157 features were excluded due to low agreement. Thus, 694 radiomic features remained for selection. After the difference test, 89 features (including target lesion location [liver]) that differed significantly between responders and non-responders in the overall population were retained. Using the training set, the LASSO regression model identified 14 features with nonzero coefficients using an optimal regulation weight (λ) of 0.01352951 based on the minimum criterion (Fig. 2). These 14 features were finally used to construct the radiomics signature, and the radscore was calculated according to the following formula (Additional file Table S2), in which the 14 features were replaced by letters a through n:
radscore = 0.254 × a – 0.027 × b + 0.546 × c + 0.026 × d – 5.469E-11 × e – 0.116 × f – 0.170×g + 0.235 × h + 0.318 × i – 0.286 × j – 0.237 × k – 0.664 × l – 0.444 × m – 0.166 × n − 0.706.
Difference analysis indicated that responders had significantly higher radscores than non-responders based on the training and validation sets (Additional file Fig. S1).
Development of the radiomics nomogram
In the overall population, univariate logistic regression indicated that radscore, stage, number of metastases, tumor type (gastric cancer vs. hepatocellular carcinoma), tumor type (lung cancer vs. hepatocellular carcinoma), and line of therapy (> 3 vs. 1) were significantly associated with treatment response (Additional file Table S3). Tumor type and line of therapy were excluded from the multivariate analysis, as not all comparisons of the groups yielded significant results. Using the training set, a radiomics nomogram incorporating the radscore, stage, and number of metastases was developed using multivariate logistic regression (Fig. 3). The following formula was used to calculate each patient’s nomoscore (Additional file Table S4):
nomoscore = 1.523 × radscore – 0.538 × stage – 0.259 × number of metastases + 2.520.
Difference analysis indicated that responders had significantly higher nomoscores than non-responders based on the training and validation sets (Additional file Fig. S2).
Model performance and validation
Calibration evaluation by the Homer–Lemeshow test indicated that there were no evidence of poor fit for two models and both the radscore and nomogram exhibited good accuracy in predicting the treatment response in both the training (radscore, p = 0.1532; nomogram, p = 0.9576) and validation sets (radscore, p = 0.3063; nomogram, p = 0.5241). Figure 4 shows the ROC curves for both models based on the training and validation sets. The radscore had AUC values of 0.790 (95% CI [0.705, 0.874]) for the training set and 0.831 (95% CI [0.649, 1]) for the validation set. For the radiomics signature, we also assessed the predictive performance for each included feature individually; the AUC value for a single feature was not favorable (Additional file Table S5). The nomogram had AUC values of 0.814 (95% CI [0.734, 0.893]) for the training set and 0.847 (95% CI [0.662, 1]) for the validation set. The nomoscore exhibited slightly better performance in predicting treatment response than the radscore although the differences were not significant (DeLong test, p = 0.1945 in the training set, and p = 0.5002 in the validation set). The sensitivities, specificities, and accuracies of the two models, along with their optimal cut-off values, are summarized in Table 2.
Decision curve analysis demonstrated a good overall net benefit for the radscore and nomogram for distinguishing responders from non-responders over a wide range of threshold probabilities (Fig. 5). Within a range of reasonable risk thresholds for both the training and validation sets, the nomogram may provide a greater net benefit than the radscore when attempting to predict the response to immunotherapy.
Association with survival
Survival analysis demonstrated that both the radscore and nomogram were significantly associated with OS (Fig. 6). In the validation set, patients in the high radscore (> − 0.478) group exhibited improved OS when compared with those in the low radscore (≤ − 0.478) group (hazard ratio: 0.18; 95% CI [0.04, 0.87], p = 0.01786). Similarly, patients in the high nomoscore (> − 0.132) group also exhibited improved OS when compared with those in the low nomoscore (≤ − 0.132) group (hazard ratio: 0.21; 95% CI [0.04, 0.99], p = 0.03212).
As a high proportion of patients exhibit a poor response to immunotherapy, one of the challenges in cancer immunotherapy is identifying the population most likely to benefit from such treatment. Despite substantial effort to develop markers for predicting ICI response, few attempts have been successful. While several studies have reported satisfactory results using clinical factors to predict ICI response [15,16,17,18], the robustness of these predictors has been questioned given the high intra- and inter-tumoral heterogeneity. No clinical predictor can fully represent the characteristics of the entire tumor microenvironment, limiting their use in clinical practice. Genetic and protein-based markers are characterized by spatial limitations related to sample collection and cannot fully represent the tumor microenvironment. In the current study, we extracted the radiomic features of the entire tumor from CT images obtained prior to ICI treatment for various types of solid cancer. Although a single feature is not powerful enough to predict the response to ICIs, we successfully developed a radiomics signature for identification of responder by combining fourteen radiomics features, which had an AUC of 0.790 (95% CI [0.705, 0.874]). In addition, we developed a robust radiomics nomogram by combining the radscore with significant clinical variables. According to our analysis, the radiomics nomogram increased performance in discriminating between responders and non-responders, with an AUC of 0.814 (95% CI [0.734, 0.893]).
Both the radscore and nomogram were validated for their performance using the validation set. The results exhibited robust predictive performance, with high AUCs for both models during validation. For a specific evaluation of discriminative performance, both models exhibited satisfactory sensitivity, specificity, and accuracy in the training and validation sets. However, the performance of the radiomics nomogram was superior to that of the radscore, especially for specificity. High specificity is crucial for selecting potential responders to immunotherapy, as misjudgment of responders will increase the financial burden and risk of adverse events among patients. Our nomogram had high specificities of 0.872 and 0.905 for the training and validation sets, respectively, demonstrating the potential clinical usefulness of our models. In addition, both the radscore and nomogram were confirmed an association with OS in the validation set, which further highlights their clinical significance for personalized immunotherapy. In the nomogram model, the variables of stage and metastases number were incorporated with negative coefficients, which indicated both variables are risk factors of poor response. This result is consistent with our conventional perspectives in clinical practice. The stage and number of metastases can reflect the degree of extension of the tumor. With the increase of stage and number of metastases, tumor shows a more aggressive phenotype with higher tumor burden and wider spreading. Thus, it is not hard to explain the finding that both the stage and number of metastases are negatively associated with the response of immunotherapy in cancer patients.
While the ability of the radiomics method to successfully predict the response to ICI treatment remains difficult to explain from a mechanistic perspective, we believe that its predictive ability involves a correlation between radiomic features and the tumor microenvironment. The higher-dimensional data obtained via radiomics may reflect tumor heterogeneity at the cellular level [19, 20]. The effectiveness of immunotherapy is mainly influenced by the status of tumor–immune interactions, and radiomics may represent this characteristic of the tumor microenvironment. Sun et al. identified an association between a radiomics-based signature and immune infiltration in solid tumors . In addition, previous studies have successfully established predictive models for the analysis of tumor PDL-1 expression in various cancers, including non-small cell lung cancer, esophageal squamous cell carcinoma, and head and neck squamous cell carcinoma [22,23,24]. Both immune infiltration and PDL-1 are key factors involved in cancer immunotherapy; thus, the successful detection of these features on images represents a reasonable means by which radiomics can predict ICI response based on CT images.
Some recent studies have also attempted to predict the response to immunotherapy using radiomics. However, our study has several strengths when compared with these studies. While some radiomics-based models have demonstrated good performance in predicting ICI response, all were developed for lung cancer [25,26,27], making it difficult to determine whether these models are applicable to other types of cancer. ICI immunotherapy is an intervention with a wide range of applications, and responses are not limited to specific cancers. Thus, a pan-cancer predictor with robust performance across different cancers is required. In our study, we included patients with five different cancer types. Importantly, no cancer-specified variables were included in the final formula, indicating that the constructed models can be used for several cancer types even when the specific type of cancer remains unknown. This is especially appropriate for patients with advanced disease involving metastasis of an unconfirmed pathological type. Ligero et al. also developed models using data from patients with different cancer types . Although their models demonstrated acceptable predictive ability for both the training and validation sets, they were developed using clinical trial data, making them difficult to adopt in clinical practice due to the various confounding factors associated with real-world settings. In addition, they defined response to treatment as complete/partial response or stable disease, which may not be common in crucial evaluations. The features included in our radiomics model differed from those used by Ligero et al. , which may be related to differences in the types of images used. In addition to the original images, we also used wavelet-transformed images for feature extraction in our study, as these can provide more information concerning intra-tumor heterogeneity .
Our study also had some limitations. First, the sample was relatively small, especially for the validation set. Second, selection bias was inevitable in this study due to its retrospective nature, which might have influenced the results. Third, no external validation was performed owing to the limited number of patients. In addition, radiomics models have not been validated separately for different cancer types. However, we plan to conduct a prospective study to explore the performance and robustness of radiomics models in real-world practice.
In conclusion, we developed a non-invasive radiomics-based model for early identification of response potential in ICI-treated patients with solid cancers. Radiomics features (with or without clinical factors) were associated with patient response and survival after immunotherapy, thus highlighting the potential of our models to improve clinical decision-making regarding personalized immunotherapy. The results should be adapted into clinical practice with cautions before the further validation from prospective large sample studies.
Availability of data and materials
The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.
Intra-class correlation coefficients
Area under the curve
Body mass index
Immune checkpoint inhibitor
Least absolute shrinkage and selection operator
Programmed cell death protein 1
Programmed cell death protein ligand 1
Response Evaluation Criteria in Solid Tumors
Receiver operating characteristic
Kawazoe A, Shitara K, Boku N, Yoshikawa T, Terashima M. Current status of immunotherapy for advanced gastric cancer. Jpn J Clin Oncol. 2021;51(1):20–7.
Nasser NJ, Gorenberg M, Agbarya A. First line Immunotherapy for Non-Small Cell Lung Cancer. Pharmaceuticals (Basel). 2020;13(11):373.
Vaddepally RK, Kharel P, Pandey R, Garje R, Chandra AB. Review of Indications of FDA-Approved Immune Checkpoint Inhibitors per NCCN Guidelines with the Level of Evidence. Cancers. 2020;12(3):738.
An Y, Wu Z, Wang N, Yang Z, Li Y, Xu B, et al. Association between body mass index and survival outcomes for cancer patients treated with immune checkpoint inhibitors: a systematic review and meta-analysis. J Transl Med. 2020;18(1):235.
Xu Y, Wang Q, Xie J, Chen M, Liu H, Zhan P, et al. The Predictive Value of Clinical and Molecular Characteristics or Immunotherapy in Non-Small Cell Lung Cancer: A Meta-Analysis of Randomized Controlled Trials. Front Oncol. 2021;11:732214.
Conforti F, Pala L, Pagan E, Corti C, Bagnardi V, Queirolo P, et al. Sex-based differences in response to anti-PD-1 or PD-L1 treatment in patients with non-small-cell lung cancer expressing high PD-L1 levels. A systematic review and meta-analysis of randomized clinical trials. ESMO open. 2021;6(5):100251.
Chinese Anti-Cancer Association LCSGoCoO, Chinese Society of Lung Cancer EGoP-LTC. Chinese Expert Consensus on Standards of PD-L1 Immunohistochemistry Testing for Non-small Cell Lung Cancer. Zhongguo Fei Ai Za Zhi. 2020;23(9):733–40.
Chinese Society of Clinical Oncology ECoTV-tT, Chinese Society of Clinical Oncology ECoN-sCLC. Expert Consensus on Tumor Mutational Burden for Immunotherapy in Lung Cancer. Zhongguo Fei Ai Za Zhi. 2021;24(11):743–52.
Wang F, Wang ZX, Chen G, Luo HY, Zhang DS, Qiu MZ, et al. Expert opinions on immunotherapy for patients with colorectal cancer. Cancer Commun (London, England). 2020;40(10):467–72.
Galsky MD, Balar AV, Black PC, Campbell MT, Dykstra GS, Grivas P, et al. Society for Immunotherapy of Cancer (SITC) clinical practice guideline on immunotherapy for the treatment of urothelial cancer. J Immunother Cancer. 2021;9(7):e002552.
Gillies RJ, Kinahan PE, Hricak H. Radiomics: Images Are More than Pictures, They Are Data. Radiology. 2016;278(2):563–77.
Limkin EJ, Sun R, Dercle L, Zacharaki EI, Robert C, Reuzé S, et al. Promises and challenges for the implementation of computational medical imaging (radiomics) in oncology. Ann Oncol. 2017;28(6):1191–206.
Eisenhauer EA, Therasse P, Bogaerts J, Schwartz LH, Sargent D, Ford R, et al. New response evaluation criteria in solid tumours: revised RECIST guideline (version 1.1). Eur J Cancer (Oxford, England : 1990). 2009;45(2):228–47.
Vickers AJ, Elkin EB. Decision curve analysis: a novel method for evaluating prediction models. Med Decis Mak. 2006;26(6):565–74.
Wei XL, Xu JY, Wang DS, Chen DL, Ren C, Li JN, et al. Baseline lesion number as an efficacy predictive and independent prognostic factor and its joint utility with TMB for PD-1 inhibitor treatment in advanced gastric cancer. Ther Adv Med Oncol. 2021;13:1758835921988996.
Wu Y, Wu H, Lin M, Liu T, Li J. Factors associated with immunotherapy respond and survival in advanced non-small cell lung cancer patients. Transl Oncol. 2022;15(1):101268.
Ahn B-C, Pyo KH, Jung D, Xin C-F, Kim CG, Hong MH, et al. A machine learning based prediction model of anti-PD-1 therapy response using noninvasive clinical information and blood markers of lung cancer patients. J Clin Oncol. 2019;37(15_suppl):e14138-e.
Valero C, Lee M, Hoen D, Weiss K, Kelly DW, Adusumilli PS, et al. Pretreatment neutrophil-to-lymphocyte ratio and mutational burden as biomarkers of tumor response to immune checkpoint inhibitors. Nat Commun. 2021;12(1):729.
Moon SH, Kim J, Joung JG, Cha H, Park WY, Ahn JS, et al. Correlations between metabolic texture features, genetic heterogeneity, and mutation burden in patients with lung cancer. Eur J Nucl Med Mol Imaging. 2019;46(2):446–54.
Choi ER, Lee HY, Jeong JY, Choi YL, Kim J, Bae J, et al. Quantitative image variables reflect the intratumoral pathologic heterogeneity of lung adenocarcinoma. Oncotarget. 2016;7(41):67302–13.
Sun R, Limkin EJ, Vakalopoulou M, Dercle L, Champiat S, Han SR, 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(9):1180–91.
Jiang M, Sun D, Guo Y, Guo Y, Xiao J, Wang L, et al. Assessing PD-L1 Expression Level by Radiomic Features From PET/CT in Nonsmall Cell Lung Cancer Patients: An Initial Result. Acad Radiol. 2020;27(2):171–9.
Zheng YM, Zhan JF, Yuan MG, Hou F, Jiang G, Wu ZJ, et al. A CT-based radiomics signature for preoperative discrimination between high and low expression of programmed death ligand 1 in head and neck squamous cell carcinoma. Eur J Radiol. 2022;146:110093.
Wen Q, Yang Z, Zhu J, Qiu Q, Dai H, Feng A, et al. Pretreatment CT-Based Radiomics Signature as a Potential Imaging Biomarker for Predicting the Expression of PD-L1 and CD8+TILs in ESCC. OncoTargets Ther. 2020;13:12003–13.
Yang B, Zhou L, Zhong J, Lv T, Li A, Ma L, et al. Combination of computed tomography imaging-based radiomics and clinicopathological characteristics for predicting the clinical benefits of immune checkpoint inhibitors in lung cancer. Respir Res. 2021;22(1):189.
Wu M, Zhang Y, Zhang J, Zhang Y, Wang Y, Chen F, et al. A Combined-Radiomics Approach of CT Images to Predict Response to Anti-PD-1 Immunotherapy in NSCLC: A Retrospective Multicenter Study. Front Oncol. 2021;11:688679.
Gong J, Bao X, Wang T, Liu J, Peng W, Shi J, et al. A short-term follow-up CT based radiomics approach to predict response to immunotherapy in advanced non-small-cell lung cancer. Oncoimmunology. 2022;11(1):2028962.
Ligero M, Garcia-Ruiz A, Viaplana C, Villacampa G, Raciti MV, Landa J, et al. A CT-based Radiomics Signature Is Associated with Response to Immune Checkpoint Inhibitors in Advanced Solid Tumors. Radiology. 2021;299(1):109–19.
Chen J, Lu S, Mao Y, Tan L, Li G, Gao Y, et al. An MRI-based radiomics-clinical nomogram for the overall survival prediction in patients with hypopharyngeal squamous cell carcinoma: a multi-cohort study. Eur Radiol. 2022;32(3):1548–57.
This study was supported by a grant of the Key Project of Science and Technology in Gansu province (19ZD2WA001), Key Talents Project of Gansu Province (2019RCXM020), Science and Technology Project of Chengguan District of Lanzhou City (2019RCCX0034), Cuiying Scientific and Technological Innovation Program of Lanzhou University Second Hospital (CY2017-ZD01 and CY2020-BJ06), Natural Science Foundation of Gansu Province (21JR1RA129), and Innovation Fund Project of Colleges and Universities in Gansu Province (2021B-026).
Ethics approval and consent to participate
This study was performed in accordance with the Declaration of Helsinki and ethics approval was obtained from the Institutional Research Ethics Committee of Lanzhou University Second Hospital (2019A-323). The requirement for informed consent was waived by our institutional review board (Ethics Committee of Lanzhou University Second Hospital) from all subjects (patients) in this study.
Consent for publication
The authors declare that they have no competing interests.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
image parameters of CT images in each cohort; Table S2. features with their coefficients in the radiomic signature; Table S3. univariate logistic regression analysis for association between related variables and immunotherapy response; Table S4. variables with their coefficients in the radiomics nomogram; Table S5. AUC values for each radiomics feature included in the radiomics signature; Fig. S1. Differences in radiomics scores between responders and non-responders; Fig. S2. Differences in nomogram scores between responders and non-responders.
About this article
Cite this article
Yu, Y., Bai, Y., Zheng, P. et al. Radiomics-based prediction of response to immune checkpoint inhibitor treatment for solid cancers using computed tomography: a real-world study of two centers. BMC Cancer 22, 1241 (2022). https://doi.org/10.1186/s12885-022-10344-6