Intelligent Radiomics For Predicting The Efficacy Of Immunotherapy And Radiation Pneumonitis In Non-small Cell Lung Cancer

Lung cancer is the leading cause of cancer death worldwide,and approximately 85%of lung cancer patients are non-small cell lung cancer(NSCLC).Most NSCLC are locally advanced or even metastatic disease at diagnosis.The prognosis of these patients remains unsatisfactory,with a 5-year overall survival(OS)rate ranging between 15%and 25%for locally advanced non-small cell lung cancer(LA-NSCLC)and only 5%for advanced NSCLC.Therefore,it is crucial to develop new method which can effectively guide the individualized treatment and improve the prognosis of NSCLC patients in the context of precision medicine.Tumor related imaging information,clinical information and genetic information forming a data-centric information science with the advent of the era of big data,which triggering a revolution in medical thinking and methods.The most common medical data is image data.Imaging plays an important role in early diagnosis,efficacy monitoring and prognosis assessment in NSCLC patients.The traditional evaluation of imaging mainly relies on qualitative characteristics,such as tumor density,enhancement pattern,tumor composition,the regularity of tumor resection margin,and the anatomical relationship with surrounding tissues.To a certain extent,these qualitative characteristics cannot meet the requirements of individualized precision.In contrast,radiomics as a rapidly developing field can automatically extract image features and transform image data into a high-dimensional feature space which can be explored.However,the final outcome of radiomics is to generate a series of high-dimensional image features.How to select effective features,eliminate redundant features and use the selected features to realize the classification or prediction task is a technical challenge.Traditional statistical methods are not able to process the high dimensional radiomics features as the number of features may exceed the sample size.But artificial intelligence(AI)algorithm can well solve this problem.AI becomes more and more important in medical field(including medical imaging field).Intelligent radiomics refers to the combination of radiomics and artificial intelligence,which using Al-based machine learning algorithms to analyze image features and construct radiomics biomarkers to solve different medical problems.Besides,intelligent radiomics integrates multi-dimensional factors such as medical imaging,gene,pathology and clinical data in a high-throughput manner.Further explore the changes of microscopic genes or protein patterns through macroscopic images to provide decision support for clinical practice.Artificial intelligence algorithms can be divided into two categories according to the ending labels,namely algorithms for continuous variables and algorithms for categorical variables,corresponding to prognosis problems and diagnosis problems in the medical field.In this study,we used many artificial intelligence methods such as random forest algorithm,support vector machine(SVM)algorithm,consensus clustering algorithm,regression algorithm,gene set enrichment analysis(GSEA)and least absolute shrinkage and selection operator(LASSO)combined with the radiomics feature to solve the two main challenges in NSCLC:immunotherapy(outcome is a continuous variable)and radiation pneumonitis(outcome is a categorical variable).The aim of this study was to explore the application value of intelligent radiomics in NSCLC for survival prognosis and classification diagnosis.In addition,we further studied the imaging mechanism of radiation pneumonitis through animal experiments,and evaluated the application value of intelligent radiomics in animal images.Part I:Intelligent radiomics for predicting the efficacy of immunotherapy in non-small cell lung cancerSection I:Intelligent radiomics for individualized evaluation of immunotherapy in patients with advanced non-small cell lung cancerObjective:To develop an intelligent radiomics signature based on computed tomography(CT)and construct a visual nomogram to predict the efficacy of immunotherapy in advanced non-small cell lung cancer patients.Besides,the molecular biological mechanism of intelligent radiomics signature was further studied by artificial intelligence algorithm.Methods:This study included 4 independent research cohorts.First,we extracted quantitative and reliable image features from the baseline CT images of 79 advanced NSCLC patients treated with immunotherapy in cohort 1.Artificial intelligence methods such as nested 10-fold cross-validation,Cox proportional hazards model and LASSO algorithm were used to construct intelligent radiomics signature for predicting the time to treatment failure(TTF)and OS of advanced NSCLC patients treated with immunotherapy.We then validated the predictive performance of intelligent radiomics signature in cohort 2,and used gene set enrichment analysis and CIBERSORT algorithm to study the molecular biological mechanism related to intelligent radiomics signature.Finally,we explored the value of selected molecular biological mechanisms in the efficacy of immunotherapy among two independent cohorts of immunotherapy.Results:In the holdout test sets,the intelligent radiomics signature had an average concordance index(C-index)of 0.670(P<0.001)for 100 repeats of 10-fold cross validation.The final signature consisting of 6 imaging features was significantly associated with TTF and OS(P<0.001 and P=0.002,respectively),and stratified patients into low vs.high-risk groups with 1-year TTF rates of 55.5%and 5.2%,and 2-year OS rates of 67.5%and 16.3%,respectively.The integrated nomograms combined the intelligent radiomics signature and clinicopathological factors further improved prediction accuracy for survival outcomes.The intelligent radiomics signature was correlated with genes in the epithelial-mesenchymal transition(EMT)pathway and angiogenesis pathway,which was associated with worse prognosis in two independent immunotherapy cohort.Conclusion:1.Intelligent radiomics signature based on CT images can effectively predict the TTF and OS of advanced NSCLC patients treated with immunotherapy.2.Combined intelligent radiomics signature and independent clinicopathological factors into an integrated nomogram improved the clinical utility of intelligent radiomics signature and can be used to guide individualized immunotherapy for patients with advanced NSCLC.3.Intelligent radiomics signature are significantly correlated with molecular biological mechanisms such as EMT and angiogenesis,while EMT and angiogenesis pathways can affect the efficacy of immunotherapy.Section II:Application of intelligent radiomics in patients with locally advanced non-small cell lung cancer and its predictive effect on immunotherapyObjective:To explore the predictive value of intelligent radiomics in inoperable LA-NSCLC patients.Besides,the correlation between intelligent radiomics signature and immune system was studied and validated in LA-NSCLC patients treated with immunotherapy.Methods:Two independent research cohorts were included in this study.First,we conducted two rounds of interactive segmentation and delineation of CT images in 118 LA-NSCLC patients in cohort 1 and extracted high-dimensional radiomics features.Then the consensus clustering algorithm combined with Cox proportional hazards model(CPH)or random survival forest(RSF)were used to select the stable and effective radiomics features.Subsequently,the prediction ability of CPH model and RSF model was evaluated and compared by bootstrap cross validation,and the model with better C-index was selected to construct the intelligent radiomics signature.Immunohistochemical staining was used to evaluate the correlation between intelligent radiomics signature and immune system.Finally,the predictive value of intelligent radiomics signature was verified in 21 LA-NSCLC patients treated with immunotherapy.Results:All imaging features can be divided into four subtypes through consensus clustering algorithm,and one image feature with the strongest prognostic performance was retained in each subtype.The C-index of CPH model based on four independent image features was 0.792,which can still reach 0.743 after cross-validation.As CPH model was more stable than RSF model,it was used to construct intelligent radiomics signature.Kaplan-Meier survival analysis showed that the signature had good predictive ability(P<0.0001)and could successfully divide LA-NSCLC patients into low-risk group and high-risk group,with 1-year OS rate of 77.5%and 46.5%,and 2-year OS rate of 50.9%and 17.7%,respectively.Immunohistochemical staining showed that intelligent radiomics signature were significantly correlated with CD8,CD3 and CD4(P<0.001,P=0.008 and P=0.013).In 21 LA-NSCLC patients treated with immunotherapy,the area under curve(AUC)of intelligent radiomics signature was 0.818[95%confidence interval(CI):0.622-1.000],and Kaplan-Meier survival analysis showed log-rank P=0.0094.Conclusion:1.This study successfully constructed an intelligent radiomics signature which can effectively predict the prognosis of patients with inoperable LA-NSCLC.2.Intelligent radiomics signature was closely related to the immune status.3.In LA-NSCLC patients treated with immunotherapy,the intelligent radiomics signature can predict the efficacy of immunotherapy.Part II Intelligent radiomics for predicting radiation pneumonitis in non-small cell lung cancerObjective:In the first part of this study,we confirmed that intelligent radiomics could effectively solve the survival problem in NSCLC.In this part,we tried to develop an intelligent radiomics signature based on CT images before chemoradiotherapy for predicting radiation pneumonitis(outcome is a categorical variable)in inoperable LA-NSCLC patients to further evaluate the application of intelligent radiomics in categorical problem.Method:This study included 2 independent research cohorts.First,we used radiomics analysis to extract high-dimensional quantitative image features from the CT images of 74 LA-NSCLC patients who underwent concurrent chemoradiotherapy in the training set.Then,the optimal radiomics subset which can predict radiation pneumonitis was screened through support vector machine algorithm and the corresponding SVM classifier was constructed.Subsequently,SVM classifier was tested in external validation set.Besides,we developed an integrated nomogram combining the intelligent radiomics-based SVM classifier and independent clinicopathological factors and evaluated its predictive performance in terms of calibration,discrimination,and clinical utility.Results:We developed an SVM classifier contained 9 image features through radiomics analysis combined with support vector machine algorithm.Receiver operating characteristic(ROC)analysis showed that the SVM classifier could effectively distinguish radiation pneumonitis and had the best classification performance in both training set(AUC:0.738)and external validation set(AUC:0.676).Multivariate logistic analysis showed that the SVM classifier was an independent predictive factor of radiation pneumonitis in inoperable LA-NSCLC patients[hazard ratio(HR):3.33,95%CI:1.67-6.63,P=0.001].The integrated nomogram combined the SVM classifier and clinicopathological factors further improves the prediction performance.The C-index reached 0.862 and 0.731 in the training set and external validation set,respectively.Besides,the integrated nomogram also showed good discrimination and calibration performance.Conclusion:1.The SVM classifier based on CT images can effectively predict radiation pneumonitis in inoperable LA-NSCLC patients,which stand for a high-risk population.2.The integrated nomogram combined intelligent radiomics-based SVM classifier and independent clinicopathological factors showed good classification performance in both training set and external validation set.It can be used as a potential tool to guide individualized treatment in clinical practice.Part III Imaging mechanism of radiation pneumonitis and the application of intelligent radiomics in animal imagesObjective:The application of intelligent radiomics for classification problems was explored in the second part of this study.It is confirmed that CT-based intelligent radiomics signature can effectively predict radiation pneumonitis.However,many studies suggested that 18F-fluorodeoxyglucose positron e1ission tomography(FDG-PET)also can be used to evaluate and predict radiation pneumonitis.But our previous study found that image features of FDG-PET could not be synthesized as an intelligent radiomics signature for predicting radiation pneumonitis in LA-NSCLC patients.In order to clarify the value of FDG-PET in radiation pneumonitis,we studied the molecular mechanism of FDG-PET in the imaging of radiation pneumonitis through animal experiments and tried to explore the application of intelligent radiomics in animal CT images.Methods:In the preliminary experiment,we constructed a rat model of radiation pneumonitis based on precise radiotherapy and a rat model of bacterial infection based on lipopolysaccharide(LPS).In the formal experiment,40 male Wistar rats were randomly divided into 4 groups(10/group):control group,radiotherapy(RT)group,LPS group,and RT+LPS group.All rats underwent micro-PET scans 7 weeks after radiotherapy(or sham radiotherapy).After the PET scans,the lung tissue,blood and bronchoalveolar lavage fluid(BALF)of all rats were collected.Two important proteins in the aerobic glycolysis pathway,pyruvate kinase M2(PKM2)and glucose transporter 1(GLUT1)were evaluated by immunohistochemical staining.Enzyme-linked immunosorbent assay(ELISA)was used to detect interleukin-1(IL-1),IL-6 and transforming growth factor-?(TGF-?)in rat blood and BALF.The concentration of lactate and pyruvate were detected by biochemical experiments.Finally,the CT images of rats with radiation pneumonitis or without radiation pneumonitis were analyzed using intelligent radiomics method.Results:Irradiated rats developed radiation pneumonitis at 7 weeks after RT based on pathology and CT scans.Maximum and mean standardized uptake values(SUVmax and SUVmean)at that time were significantly increased in the LPS group(P<0.001 for both)and the RT+LPS group(P<0.001 for both)relative to control,but were no different in the RT-only group(SUVmax:P=0.156,SUVmean:P=0.304).The combination of RT and LPS increased the expression of the aerobic glycolysis enzyme PKM2(P<0.001)and GLUT1(P=0.004)in lung tissues.LPS alone increased the expression of PKM2(P=0.018),but RT alone did not affect PKM2(P=0.270)or GLUT1(P=0.989).The concentration of lactate was higher in the RT+LPS group than in the control(P=0.002)or RT-only(P=0.021)groups but were no different from those in the LPS-only group(P=0.084).And pyruvate only increased in the RT+LPS group(P=0.024).The unsupervised clustering method based on intelligent radiomics showed that there was a significant difference in rats with or without radiation pneumonitis.T test found that 49 image features showed differences between the two groups(P<0.05)?among which eight image features had P<0.001.Conclusion:1.Aseptic radiation pneumonitis could not be accurately assessed by 18F-FDG-PET.2.The model of radiation pneumonitis accompanied with bacterial infection simulated by LPS could be visualized in FDG-PET.And the mechanism underlying these results may be the Warburg effect.RT alone did not activate the Warburg effect but could exacerbate the pulmonary effects of bacterial infection.3.Intelligent radiomics also showed good application value in animal CT images.