Abstract
PURPOSE:
To predict programmed death-ligand 1 (PD-L1) expression of tumor cells in non-small cell lung cancer (NSCLC) patients by using a radiomics study based on CT images and clinicopathologic features.
MATERIALS AND METHODS:
A total of 390 confirmed NSCLC patients who performed chest CT scan and immunohistochemistry (IHC) examination of PD-L1 of lung tumors with clinic data were collected in this retrospective study, which were divided into two cohorts namely, training (n = 260) and validation (n = 130) cohort. Clinicopathologic features were compared between two cohorts. Lung tumors were segmented by using ITK-snap kit on CT images. Total 200 radiomic features in the segmented images were calculated using in-house texture analysis software, then filtered and minimized by least absolute shrinkage and selection operator (LASSO) regression to select optimal radiomic features based on its relevance of PD-L1 expression status in IHC results and develop radiomics signature. Radiomics signature and clinicopathologic risk factors were incorporated to develop prediction model by using multivariable logistic regression analysis. The receiver operating characteristic (ROC) curves were generated and the areas under the curves (AUC) were reckoned to predict PD-L1 expression in both training and validation cohorts.
RESULTS:
In 200 extracted radiomic features, 9 were selected to develop radiomics signature. In univariate analysis, PD-L1 expression of lung tumors was significantly correlated with radiomics signature, histologic type, and histologic grade (p < 0.05, respectively). However, PD-L1 expression was not correlated with gender, age, tumor location, CEA level, TNM stage, and smoking (p > 0.05). For prediction of PD-L1 expression, the prediction model that combines radiomics signature and clinicopathologic features resulted in AUCs of 0.829 and 0.848 in the training and validation cohort, respectively.
CONCLUSION:
The prediction model that incorporates the radiomics signature and clinical risk factors has potential to facilitate the individualized prediction of PD-L1 expression in NSCLC patients and identify patients who can benefit from anti-PD-L1 immunotherapy.
Introduction
As the result of common incidence and poor prognosis, lung cancer is responsible for more than 1.5 million deaths worldwide each year [1]. Although, there is a potential for earlier lung cancer diagnosis through screening with low-dose computed tomography, which has demonstrated a 20% reduction in lung cancer mortality [2]. In China, the first causes of cancer death among both men and women are cancers of the lung and bronchus because of air pollution and smoking [3, 4]. Currently, immunotherapy can become a crucial therapeutic option to improve prognosis for stage IV lung cancer patients; Furthermore, it could be applied to earlier stages lung cancer [5]. First clinical trials with therapies targeting the programmed cell death 1 (PD-1) and programmed death-ligand 1 (PD-L1) have shown promising results in several solid tumors [6]. However, in lung cancer, the diagnostic, prognostic and predictive value of these immunologic factors remains unclear [7]. Although an increased CD4+/CD8+ cell infiltration of the tumor stroma has previously been shown to represent a favorable prognostic factor in NSCLC [8]. The best-characterized immunological checkpoints with a major impact on both cancer growth and cancer therapy are cytotoxic T lymphocyte-associated protein 4 (CTLA-4), PD-1, or PD-L1 [9].
Immunotherapy by immune checkpoint blockade has emerged as a promising therapeutic option for patients with cancers such as melanomas [10], lymphomas [11], and lung cancer [12]. However, the inability to precisely identify the patients who will benefit from these treatments has limited the applicability of immune checkpoint blockade and could put patients at higher risk of immune-related toxicity [13]. Toward identification of predictors of response to anti-PD-L1 monotherapy, tumor expression of PD-L1 has been a primary focus [14]. Expression of PD-L1 is a commonly used biomarker that indicates to physicians whether a patient should or should not receive immune checkpoint blockade therapy [15]. Generally, lung cancer patients with positive PD-L1 expression (PD-L1 express level ≥50%) are sensitive to the immunotherapy. At present, PD-L1 expression of lung cancer is mainly detected by invasive biopsy and using immunohistochemistry (IHC) examination method [16].
Computational medical imaging, known as radiomics, involves the analysis and translation of medical images into quantitative data [17]. High-dimensional imaging data allow an in-depth characterisation of tumour phenotypes, with the underlying hypothesis that imaging reflects not only macroscopic but also the cellular and molecular properties of tissues [18]. The objective of radiomics is to generate image-driven biomarkers that serve as instruments that provide a deeper understanding of cancer biology to better aid clinical decisions [19]. CT-based radiomics features are complementary to biopsies and have the advantage of being non-invasive, which allows prediction of EGFR mutation subtypes and risk of lymph node metastasis in NSCLC [20, 21], and prediction of the malignancy of pulmonary nodules [22], and decoding tumor mutation burden and driver mutations in early stage lung adenocarcinoma [23, 24]. To the best of our knowledge, there is a similar articles about association of tumor PD-1/PD-L1 and textural or radiomics features in 18F-FDG PET in squamous cell carcinoma of the head and neck [25]. The aim of this study was to develop a prediction model that incorporated both the radiomics signature and clinicopathologic risk factors for individual prediction of PD-L1 expression in patients with NSCLC, and it could be helpful to identify the patients who will benefit from the immunotherapy.
Materials and methods
Patients
Patients ethical approval was obtained for this retrospective analysis, and the informed consent requirement was waved. 390 patients with clinicopathologically confirmed NSCLC were collected from January 2018 to August 2019 in our hospital. They were randomly divided into two cohorts, including the training cohort of 260 consecutive patients and the validation cohort of 130 consecutive patients (2 : 1 ratio). The demography data of two cohorts were showed and compared in Table 1. Inclusion criteria: (a) patients who underwent biopsy or surgery of lung tumor; (b) IHC examination of PD-L1 performed; (c) pathologically proven histological type and grade available; and (d) standard chest CT performed before biopsy or surgery. Exclusion criteria: (a) therapy (radiotherapy, chemotherapy or chemoradiotherapy) performed before CT and IHC; (b) histological grade unknown; and (c) CT images with artifacts. Carcinoembryonic antigen (CEA) was derived from medical records. Laboratory analysis of CEA was done via routine blood tests. The threshold value for CEA level was 5 ng/mL according to the normal range used at our institution. The clinicopathologic TNM stage in this study referred to the eighth edition of the TNM classification for lung cancer [26].
Comparisons of clinicopathologic characteristics of 390 NSCLC patients between training cohort and validation cohort
Comparisons of clinicopathologic characteristics of 390 NSCLC patients between training cohort and validation cohort
Note: There were not any significantly statistical differences in clinicopathologic characteristics of patients between training cohort and validation cohort (p > 0.05, respectively). An independent samples t-test was used to assess the difference in numeric variable (age); While Chi-Squared tests were used to compare the differences in categorical variables (gender, tumor location, CEA level, histologic type and grade, smoking history, and TNM stage).
The biopsy or surgery specimen of lung tumor through hematoxylin-eosin (HE) staining were pathologically examined to confirm histologic type and grade under microscope. Furtherly, PD-L1 expression level was measured through IHC test in our study. PD-L1 test kit was 22C3 pharmDx (Dako company) [27]. Pathologists performed back-to-back interpretation of PD-L1 expression, and further reanalysis would be implemented when there was an inconsistency between previous results. In clinical therapy, the inhibitor medicine was found of great value in NSCLC patients with 50% or more of PD-L1 expression in tumor cells [28].
CT image acquisition and reconstruction
All patients underwent routine chest CT scanning using 64-slice spiral CT machine (Siemens company, Germany). The acquisition parameters were as follows: 120 kV of tuber voltage; 200 mAs of tuber current; 5 mm of slice thickness; field of view (FOV), 350×350 mm; matrix, 512×512. Lung-window CT images were reconstructed with reconstruction thickness of 1.25 mm, which were loaded to personal laptop from the picture archiving and communication system (PACS) (Carestream, Canada) for tumor segmentation and radiomics feature extraction because of well differentiation of tumor tissue from adjacent normal lung tissue.
Tumor segmentation and radiomics feature extraction
Tumor segmentation was performed to select primary lesions of NSCLC cases after images acquisition. A region of interest (ROI) was delineated initially around the tumor outline for the largest cross-sectional area on the lung-window CT images by using ITK-snap kit (version 3.6.1) in this procedure (e.g., Fig. 1). Radiomic features extracted from the tumor ROI were calculated automatically using in-house texture analysis software with algorithms implemented in Matlab 2010a (Mathworks, Natick, USA). The ROI was further refined by excluding air or cavity area with a thresholding procedure that removed from any pixels with attenuation values below –50 HU and beyond 300 HU. Finally, 200 features from CT images were acquired for further selection.
CT images and pathological images of a 65-year-old male patient with NSCLC. (A) Tumor segmentation was performed by delineating a ROI (red line) of tumor on non-enhanced lung window CT image with an ITK-SNAP kit. (B) The ROI of tumor was extracted. (C) Contrast enhancement CT image was referred to delineate the tumor precisely. (D) A routine pathological image presented the poorly differentiated adenocarcinoma (HE staining×100). (E) A immunohistochemical pathology image presented the positive expression (expression rate of 90%) of PD-L1, seeing a large number of brown tumor cells. NSCLC: non-small cell lung cancer; ROI: region of interest.
Inter-observer and intra-observer reproducibility of radiomics feature extraction was initially analyzed with 50 randomly chosen images for ROI-based radiomics feature generation in a blind way by two radiologists. Inter- and intra-class correlation coefficients (ICCs) were used to evaluate the intra and inter-observer agreement of features extraction. An ICC greater than 0.75 presents good agreement [29].
Radiomics feature selection and radiomics signature building
The least absolute shrinkage and selection operator (LASSO) method, which is suitable for the regression of high-dimensional data, was used to select the most useful predictive features from the primary data set. A radiomics score (Radscore) was calculated for each patient via a linear combination of selected features that were weighted by their respective coefficients, and then radiomics signature was developed. The potential association of the radiomics signature with PD-L1 expression was first assessed in the training cohort and then validated in the validation cohort.
Development of an individualized prediction model
Univariable logistic regression analysis began with the following clinical candidate predictors: age, gender, tumor site, histology type and grade, TNM stage, CEA level, and smoking status. The radiomics signature and clinicopathologic risk factors with significantly statistical difference were incorporated to develop prediction model by using multivariable logistic regression analysis in a training cohort that consisted of 260 consecutive patients, and the receiver operating characteristic (ROC) curve and the corresponding area under the curve (AUC) was reckoned for the prediction model in the training cohort and validation cohort, respectively.
Statistical analysis
Statistical analysis was conducted with R software (version3.0.1; http://www.Rproject.org). An independent samples t-test was used to assess the differences in numeric variable (age); While Chi-Squared tests were used to compare the difference in categorical variables (gender, tumor site, smoking status, CEA level, TNM stage, and histological type and grade), and the statistical analysis of non-parametric test was used to assess the differences in radiomics score. Lasso binary logistic regression was performed using the “glmnet” package, while the multivariate logistic regression was performed using the “rms” package. Diagnosis efficacy was assessed using the receiver operating characteristic (ROC) curve with area under the curve (AUC). The reported statistical significance levels were set at 0.05.
Result
Clinical characteristics
There were not any significantly statistical differences in the clinical characteristics between the training and validation cohort in 390 NSCLC patients (Table 1). In our univariate analysis, PD-L1 expression was significantly correlated with radiomics signature, histologic type and grade (p < 0.05, respectively). However, it was not correlated with gender, age, tumor location, CEA level, TNM stage, and smoking (p > 0.05, respectively) (Table 2).
Comparisons of clinicopathologic and radiomic characteristics of 390 NSCLC patients between different PD-L1 expression levels
Comparisons of clinicopathologic and radiomic characteristics of 390 NSCLC patients between different PD-L1 expression levels
*Note: There are significantly statistical differences in histologic grade, histologic type and radiomics score between PD-L1 < 50% and PD-L1≥50% cohort (p < 0.05, respectively). An independent samples t-test was used to assess the difference in numeric variable (age); While Chi-Squared tests were used to compare the differences in categorical variables (gender, tumor location, CEA level, histological type and grade, smoking history, TNM stage). The statistical analysis of non-parametric test was used to assess the difference in radiomics score.
Satisfactory interobserver and intra-observer reproducibility of radiomics features extraction was achieved (ICCs > 0.75, respectively). 200 radiomics features were extracted. Finally, 9 radiomics features were selected in this study, which were features with nonzero coefficients in the LASSO logistic regression model (Fig. 2). These features were weighted by their respective coefficients. Radiomics signature was built on the radscore calculation formula, as following:
We used two feature selection methods, maximum relevancy and minimum redundancy (MRMR) and the least absolute shrinkage and selection operator (LASSO) to select the radiomics feature. At first, MRMR was performed to eliminate the redundant and irrelevant features, 30 features were retained. Then LASSO was conducted to choose the optimized subset of features to construct the final model. (A) Tuning parameter Lambda (λ) selection in the LASSO model used 10-fold cross-validation via minimum criteria. The binomial deviance was plotted versus log(λ). Dotted vertical lines were drawn at the optimal value where the binomial deviance was minimum and radiomics features number was 9; (B) LASSO coefficient profiles of the retained 30 features. A coefficient profile plot was produced against the log (λ) sequence. Vertical line was drawn at the value selected using 10-fold cross-validation, where optimal λ resulted in 9 features with nonzero coefficients.
Radscore = – 0.059 + 0.138×mean_100_2.5 – 0.043×SD_25_0 – 0.158 × kurtosis_2.5 + 0.529× correlation_90_0 + 0.303×energy_135_2 – 0.395×correlation_0_1.5 + 0.041×homogenity_0_0 – 0.2×SD_10_2.5 + 0.486×constrast_90_1.5. (Fig. 3).
Nine selected features and the corresponding coefficients. Radiomics signature, namely radiomics score (radscore), was calculated and constructed via summing the selected features weighted by their coefficients.
The area under the curves (AUCs) of ROC for the three models (Radscore, clinicopathologic features, and the combined model) in training set and validation set for predicting PD-L1 express level ≥50% were showed in Fig. 4. The diagnostic efficacy of the combined model was the best, it resulted in the AUC of 0.829, sensitivity of 80.8% and specificity of 77.0% in the training cohort, and it resulted in the AUC of 0.848, sensitivity of 83.3% and specificity of 72.4% in the validation cohort.
Classifiers’ performance on predicting expression status (≥50%) of PD-L1 based on three models (radiomics, clinics, and the combined model). (A) Classifiers’ performance on predicting 50% level of PD-L1 in training set. (B) Classifiers’ performance on predicting 50% level of PD-L1 in validation set. ROC: receiver operating characteristic; AUC: area under the curve.
We developed and validated the individualized pretreatment prediction model of PD-L1 expression status in NSCLC patients by using a radiomics study based on CT images and clinicopathologic features. The optimum model incorporated three items of the radiomics signature, histologic grade, and histologic type. Among them, radiomics signature successfully stratified patients according to PD-L1 expression rate threshold of 50%. In addition, the radiomics signature from multiple individual imaging features demonstrated better discrimination of PD-L1 expression in the validation cohort than in the training cohort. The surprisingly improved discrimination in the validation cohort implied that the radiomics signature was robust. Subsequently, we incorporated the radiomics signature and clinical risk factors into a combined model, which presented a better diagnostic efficacy both in the training cohort and in the validation cohort than the simple radiomics signature or clinical features model (AUC, 0.829 in the training cohort, 0.848 in the validation cohort; respectively).
At present, the radiomics has been widely used in lung cancer patients [20–24]. but they are rarely used to predict PD-L1 expression status of NSCLC based on CT images. In a recent study that assessed PD-L1 expression level by radiomic features from PET/CT images in NSCLC patients, they got the result that radiomic signatures of over 50% PD-L1 expression reached the score of AUC at 0.880 [30], which was consistent with our results (AUC, 0.848). But they did not combine radiomics features with clinical risk factors in the prediction model. In another study that investigated the association between PD-L1 expression and textural features of PET images in 53 oropharyngeal or hypopharyngeal cancer patients, they thought p16 and Ki-67 staining percentages and several PET/CT-derived textural features can provide supplemental information to determine tumor PD-L1 expression in head and neck cancer [25]. Because their sample size was too small compared with our 390 cases, the constructed prediction model of PD-L1 expression by the radiomics couldn’t be robust. Moreover, the density resolution of PET image was not so good as CT image, this could have a great effect on extracting and selecting the meaningful radiomics texture features and lead to draw an unreliable conclusion. An interesting study that explored radiomics features based on CT images for the prediction of epidermal growth factor receptor (EGFR) mutation subtypes in NSCLC patients, demonstrated the potential for radiomics to predict EGFR 19Del and L858R, which may help to guide the selection of personalized targeted therapy programs for patients with NSCLC [20]. The respective areas under the ROC curves of the EGFR 19Del and L858R joint models were 0.792 and 0.775 for the test set. Compared with our results for the combined model both in the training cohort and in the validation cohort, their AUC values were both lower. Furthermore, their research did not seem to have much guiding significance for the emerging immunotherapy of NSCLC patients. Another similar study assessed the predictive power of a radiomic signature based on PET/CT images for EGFR mutational status in NSCLC patients. Their results presented an AUC of 0.805 for a radiomic signature in discriminating between mutant-type of EGFR and wild-type of EGFR cases [31]. Wang et al. established an efficient fusion-positive tumor prediction model that predicted tumor mutation burden (TMB) status and EGFR/TP53 mutations of early stage lung adenocarcinoma. Combining radiomics features with the clinical information yielded the AUC values of 0.671 for TMB, 0.697 and 0.656 for EGFR/TP53 respectively [24]. Gu et al. explored machine learning-based radiomics strategy for prediction of cell proliferation (Ki-67) in NSCLC, random forest-based radiomics classifier achieved the best performance (AUC = 0.776) in predicting Ki-67 expression level [32].
Our study had several limitations. First, the sample size was small, and the results needed to be confirmed with large sample studies and multi-center data. The prediction model built in this study was validated with internal data but not tested with external test data. So, in the future, multi-center and large sample studies are encouraged. Second, in this study, squamous cell carcinoma and adenocarcinoma types of NSCLC were studied, but no rare pathological types were included. Third, we only explored the radiomic features from non-contrast CT images in NCSLC. However, contrast-enhanced CT images may still include more meaningful features in predicting PD-L1 expression level, this need to further investigate and compare among several scanning series of images.
In conclusion, this radiomics study based on CT images and clinicopathologic features showed that the prediction model could predict PD-L1 expression status and provide a novel strategy for clinicians to screen the beneficial patients to anti-PD-L1 immunotherapy in NSCLC.
Footnotes
Acknowledgments
This work was mainly supported by the National Natural Science Foundation of China (grant number 81671743, 81971573), the clinical key diseases diagnosis and therapy special project of Health and Family Planning Commission of Suzhou (LCZX201801), the program for Advanced Talents within Six Industries of Jiangsu province (WSW-057), and the High-level Health Personnel “six-one” Project of Jiangsu province in China (LGY2016035).