Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

CT radiomics-based long-term survival prediction for locally advanced non-small cell lung cancer patients treated with concurrent chemoradiotherapy using features from tumor and tumor organismal environment

CT radiomics-based long-term survival prediction for locally advanced non-small cell lung cancer... Background Definitive concurrent chemoradiotherapy (CCRT ) is the standard treatment for locally advanced non- small cell lung cancer (LANSCLC) patients, but the treatment response and survival outcomes varied among these patients. We aimed to identify pretreatment computed tomography-based radiomics features extracted from tumor and tumor organismal environment ( TOE) for long-term survival prediction in these patients treated with CCRT. Methods A total of 298 eligible patients were randomly assigned into the training cohort and validation cohort with a ratio 2:1. An integrated feature selection and model training approach using support vector machine combined with genetic algorithm was performed to predict 3-year overall survival (OS). Patients were stratified into the high-risk and low-risk group based on the predicted survival status. Pulmonary function test and blood gas analysis indicators were associated with radiomic features. Dynamic changes of peripheral blood lymphocytes counts before and after CCRT had been documented. Results Nine features including 5 tumor-related features and 4 pulmonary features were selected in the predictive model. The areas under the receiver operating characteristic curve for the training and validation cohort were 0.965 and 0.869, and were reduced by 0.179 and 0.223 when all pulmonary features were excluded. Based on radiomics- Nai-Bin Chen, Mai Xiong and Rui Zhou contributed equally as co- first authors. *Correspondence: Chuan-Miao Xie xiechm@sysucc.org.cn Hui Liu liuhuisysucc@126.com Full list of author information is available at the end of the article © The Author(s) 2022. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data. Chen et al. Radiation Oncology (2022) 17:184 Page 2 of 12 derived stratification, the low-risk group yielded better 3-year OS (68.4% vs. 3.3%, p < 0.001) than the high-risk group. Patients in the low-risk group had better baseline FEV1/FVC% (96.3% vs. 85.9%, p = 0.046), less Grade ≥ 3 lymphopenia during CCRT (63.2% vs. 83.3%, p = 0.031), better recovery of lymphopenia from CCRT (71.4% vs. 27.8%, p < 0.001), lower incidence of Grade ≥ 2 radiation-induced pneumonitis (31.6% vs. 53.3%, p = 0.040), superior tumor remission (84.2% vs. 66.7%, p = 0.003). Conclusion Pretreatment radiomics features from tumor and TOE could boost the long-term survival forecast accuracy in LANSCLC patients, and the predictive results could be utilized as an effective indicator for survival risk stratification. Low-risk patients might benefit more from radical CCRT and further adjuvant immunotherapy. Trial registration: retrospectively registered. Keywords Locally advanced non-small cell lung cancer, Radiomics, Machine learning, Long-term survival prediction, Tumor organismal environment. Introduction investigate the relationship of tumor and TOE, due to its Definitive concurrent chemoradiotherapy (CCRT) is the accessibility. standard treatment for patients with unresectable locally In this study, we utilized computed tomography (CT) advanced non-small cell lung cancer (LANSCLC). In the images before CCRT to develop an image-based machine past two decades, concomitant regimens achieved prom- learning framework to analyze the relationship of pri- ising tumor local control and long-term survival. With mary lung tumor and bilateral lungs for long-term sur- improved outcome, the maintenance of an adequate pul- vival prediction in LANSCLC. To balance the training monary function is essential to ensure acceptable qual- accuracy and predictive capability using relative small ity of life and adjuvant immunotherapy. However, many number of patient samples, an integrated feature selec- patients with LANSCLC are diagnosed with pre-existing tion and model training (IFSMT) approach was devel- lung comorbidities, which significantly increases the risk oped to extract the most critical quantitative radiomic for radiation-induced lung toxicity (RILT) [1, 2]. features from both tumor and lungs. A radiomic-based Most existing RILT prediction models largely focused risk stratification was built to distinguish high-risk and on clinical prognostic factors (CPFs) and dose-volume low-risk patients and provided evidence for clinical deci- histogram parameters [3–5], but remained insufficient. sion making. Recently, machine learning methods have been reported to improve the capacity of the predictive modelling [6–9], Methods compared with logistic regression widely used in normal Study population tissue complication probability model. Consecutive patients irradiated for lung cancer from Moreover, radiomics analysis, attempting to identify September 2011 to April 2019 in our institution were computational biomarkers potentially hidden within retrospectively screened. Inclusion criteria included: (1) high-throughput imaging data [10, 11], has been dem- histologically confirmed NSCLC; (2) unresectable stage onstrated the added predictive value for overall survival III disease (AJCC/UICC 8th staging criteria) proven by (OS) [12–14] or RILT [8, 9]. However, most of them rely chest and upper abdominal CT, brain magnetic reso- on the radiomic information from tumor or its surround- nance imaging (MRI), bone scan and/or positron emis- ing peritumoral region, few studies have been designed sion tomography-computed tomography (PET-CT); (3) based on the radiomics analysis of tumor organismal definitive radiotherapy with concurrent chemotherapy environment (TOE). was administered; (4) stay followed-up no less than 6 Similar to other published reports [15, 16], our previ- months since the start of radiotherapy (unless death ous study [17] indicated that pulmonary function test or disease progression was documented); (5) complete (PFT) was significantly related to patients’ long-term clinical records. Patients that met the inclusion criteria survival. However, it failed to predict progression-free were randomly assigned into the training and validation survival (PFS). Even though patients with worse FEV1/ cohort, with the numbers at a ratio 2:1. FVC% or DLCO% showed a high objective response rate (ORR) to CCRT, their survival outcomes were still poor, Planning CT image acquisition hinting that TOE, the status of lungs in the case of LAN- The four-dimensional (4D) planning CT scan was per - SCLC might play an indispensable role in the prognostic formed 1–2 weeks prior to treatment, using multiple CT prediction after CCRT. As some patients could not tol- simulation positioning machines with varied parameter erate well with PFT, radiomics analysis using machine settings in our institution (detailed in Additional File 1). learning method might be an effective technique to Ten phases of the breathing cycle were reconstructed, Chen et al. Radiation Oncology (2022) 17:184 Page 3 of 12 Fig. 1 Schematic overview of the integrated feature selection and model training (IFSMT ) approach. IFSMT approach consisted of five steps: (1) volumet - ric data pre-processing; (2) delineation; (3) feature extraction; (4) integrated feature selection and model training; (5) model validation using leave-one- out cross-validation (LOOCV ). including: 0%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, medical records, including blood tests, PFT, blood gas and 90%. The segmentation and radiomics were then per - analysis (BGA) and radiologic tests. All included patients formed on the 20% phase (middle exhale phase) with a received regular radiologic follow-up, including chest consistent mediastinum/lung window level setting. and upper abdominal CT and brain MRI performed every 3 ~ 6 months in the first 2 years, and every 6 ~ 12 Radiotherapy and concurrent chemotherapy months thereafter. PET-CT, bone scan, and biopsy were Patients were positioned supine and immobilized in a recommended if clinically required. The responses to vacuum pad. They were scanned from the Atlas to the CCRT were first assessed by an independent radiation second lumbar vertebra level with 0.3-0.5  cm thickness oncologist and confirmed by a senior physician at 4 ~ 6 slices to obtain the stimulation CT images. The respira - weeks post CCRT, based on Response Evaluation Crite- tion motion was recorded by performing 4DCT scan- ria in Solid Tumors 1.1. Another senior radiologist was ning. The maximum intensity projection images were consulted for disagreement. Therapeutic toxicities were reconstructed using the images collected in 10 phases of graded and recorded according to Common Terminology respiratory cycle. Gross tumor volume (GTV) was delin- Criteria for Adverse Events 4.0. eated to cover the tumor and involved regional nodes visible on each phase of the 4DCT. The total volumes of OS modelling procedures GTVs across the 10 respiratory phases CT composed the The whole procedures were illustrated in Fig.  1. For both internal target volume (ITV). Planning target volumes cohorts of patients, the regions of interest (ROIs) corre- (PTVs) were created by expanding GTV and clinical tar- sponding to GTV and lungs were delineated by an auto- get volume with 6  mm. Lungs were delineated accord- contouring software tool CezanneDraw™ v1.0 (Homology ing to the atlases for organs at risk (OARs) in thoracic Medical, Ningbo, China, 2020) using the CT slices and radiation therapy [18], but GTV was excluded from the manually modified by radiation oncologists if necessary. lung delineation. A dose of 60-76  Gy was prescribed to One 3D bounding box was fitted for each ROI. And inside PTV-GTV in 22–33 fractions, with 2-3  Gy per fraction the bounding box, the CT values of the ROI voxels were performed once daily, using intensity modulated radia- retained while the values of other voxels were marked tion therapy technique. The dose constraints for OARs by zero. CT values of voxels in each bounding box were were: V20 < 35% for lungs; mean lung dose < 19 Gy; maxi- then interpolated to a resolution of 1  mm×1  mm×5  mm mun dose (Dmax) of esophagus < 105% prescription dose; and resampled into 400 discrete values (called bins) with Dmax of spinal cord < 46 Gy; V30 < 40% for heart. absolute discretization from − 1000 to 3000 Hounsfield All patients received platinum-based double agents units, leading to a fixed bin size of 10 Hounsfield units. weekly or every three weeks. The regimens included A total of 92 tumor-related and lung-related features docetaxel/paclitaxel/etopside/pemetrexed plus platinum. were then computed for both ROIs and used as the input feature pool for the machine learning framework by the Evaluation and follow-up LIFEx software (version 3.44) [19]. The imaging-based The baseline characteristics of each patient before entry features covered two categories of texture features and were reviewed attentively and extracted from their first order features. The texture features consisted of four Chen et al. Radiation Oncology (2022) 17:184 Page 4 of 12 Fig. 2 Schematic overview of the genetic algorithm (GA) in the integrated feature selection and model training (IFSMT ) approach. A chromosome is scored with LOOCV-SVM. The chromosomes of higher scores may go through mutation and crossover to make new ones to replace those of lower scores. Collect the chromosome of best score from each generation into a group. And the best one in the group is the result of the model. Abbreviation: LOOCV, leave-one-out cross-validation; SVM, support vector machine sub-categories of matrix based texture features. These generation, the chromosomes of higher scores may go matrices included the grey-level co-occurrence matrix through mutation, partially changing feature encoding, (GLCM), neighborhood grey-level different matrix and crossover, partially exchanging feature encoding, to (NGLDM), grey-level run length matrix (GLRLM) and make new ones to replace those of lower scores. Collect grey-level zone length matrix (GLZLM). The first order the chromosome of best score from each generation into features included indices from shape, indices from histo- a group. And the best one in the group is the result of the gram and conventional indices. model. Manual reconfiguration of SVM is not included in The machine learning based classification method used the model. to predict the two-class 3-year survival status for each Once the optimal set of features was determined, the individual patient was support vector machine (SVM) SVM models were trained again on the training cohort. [20]. The SVM mapped the features of training data into In this study, after extensive experimental comparisons, a high-dimensional feature space through a kernel func- the linear kernel was chosen for SVM and optimal hyper tion and utilizes a hyper-plane to optimally separate the parameters of the SVM (C, ε and γ) were determined training data points into two categories. To reduce the through exhaustive search in the parametric space. possibility of overfitting, only a subset of features from Receiver operating characteristics (ROC) curves were the feature pool could be selected for the input of SVM. obtained by varying threshold of the decision variable, In this study, the IFSMT approach was developed to the signed distance to decision hyper-plane. Area under maximize the fitting accuracy and minimize the over - curve (AUC) for each ROC was calculated for training fitting potential. This posteriori approach applied the cohort. The trained models were then used to predict the genetic algorithm (GA) for the feature selection, which survival status for each individual patient in the valida- was illustrated in Fig. 2 and Additional File 2. A chro- tion cohort, and ROCs and their corresponding AUCs mosome represents a feature template working with were also calculated. All the above feature selection and SVM of certain configuration for diagnosing purpose. machine learning approaches were implemented on the The SVM is implemented in leave-one-out cross-valida - cloud-based clinical data service platform iRAAS® v2.0 tion (LOOCV) fashion to score a chromosome. In each (Homology Medical, Ningbo, China, 2020). Chen et al. Radiation Oncology (2022) 17:184 Page 5 of 12 Table 1 OS Training and validation accuracy using all selected To assess the importance of each selected feature to the features or without pulmonary features accurate prediction of the clinical outcome, a one-by-one OS status accuracy Training cohort Validation cohort feature evaluation procedure was designed. This proce - All Without All With- dure tested the importance of each feature by deleting features pulmo- features out pul- each feature from the selected feature set and calculating nary monary the reduction of the AUC for the model trained with the features features original selected features except this specific feature. This AUC 0.965 0.786 0.869 0.646 reduction of model performance was used as the impor- TNR (%) 95.00 74.00 92.86 75.00 tance weight (IW) of this feature. All the selected features TPR (%) 90.00 68.00 82.86 50.00 were then sorted according to their IWs. To further assess F1 score 0.923 0.701 0.892 0.625 the importance organismal features, the AUC for the Average accuracy (%) 92.50 71.00 87.86 62.50 model trained with the original selected features exclud- Overall accuracy (%) 92.50 71.00 85.71 57.14 Abbreviations: OS, overall survival; AUC, area under curve; TNR, true negative ing all the lung-related features were also calculated. rate; TPR, true positive rate Statistical methods OS was defined as the time from radiotherapy start to the in the whole cohort, with the median age of 59 years last follow up, which ended at November 30th, 2021, or (range, 28–81 years). Squamous cell carcinoma was the death. A t-test was used to determine if there was signifi - predominant histologic type both in the training (46.5%) cant difference between the means of continuous vari - and validation (62.2%) cohorts. ables, while Fisher’s exact test was performed to reveal the difference in distribution between two groups of OS modelling categories variables. The association between radiomic With the median follow-up of 27.7 (range, 4.0 ~ 122.7) features and PFT/BGA indicators was examined using months for all and 67.0 months (range, 36.2 ~ 122.7 Pearson’s correlation coefficient. A p-value < 0.05 (two- months) for event-free patients, our cohort demonstrated sided) were considered as statistically significant. Missing the estimated median OS of 27.6 (95% confidence inter - data were excluded from the statistical analysis. Statis- val (CI), 22.3 ~ 33.0) months, and the 3-year OS rate was tics were performed using SPSS 22.0 (IBM, Chicago, IL, 43.0% (95%CI, 37.3%~48.7%). USA). As shown in Table  1, the overall prediction accuracy To report the model fitting accuracy and the predic - for 3-year survival status was 92.50% and 85.71%, and the tion capability, the true positive rate (TPR), true negative AUC of the ROC was 0.965 and 0.869, respectively, in the rate (TNR), F1 score, overall prediction accuracy, average training and validation cohort. prediction accuracy for the training cohort and valida- tion cohort were calculated based on the SVM model. Stratification of patients in the validation cohort with Herein, death is marked as the positive. The overall pre - machine learning model diction accuracy was expressed as the number correctly In the validation cohort, 60 (61.2%) of 98 patients were predicted patients / the number of all patients; and the stratified into the high-risk group and 38 (44.1%) into the average prediction accuracy = (TPR + TNR)/2. low-risk group. CCRT was more successful in patients in To assess the prognostic value of the survival status the low-risk group than those in the high-risk group. The model, the predicted 3-year survival status was adopted ORR was 84.2% (32/38) and 66.7% (40/60) in the low-risk respectively as the clinical risk estimator to stratify the and high-risk group, respectively (p = 0.003) (Additional patients into the high-risk and low-risk groups. Patients File 4). And the low-risk group yielded better 3-year OS with negative predicted survival status were classified (68.4% versus 3.3%, p < 0.001, log-rank) than the high- into the low-risk group and the others with positive pre- risk group (Fig.  3B). What’s more, the rate of Grade ≥ 2 dicted survival status into the high-risk group. Kaplan- pneumonitis was 31.6% (12/38), versus 53.3% (32/60) Meier curves for both groups were displayed to illustrate (p = 0.040) in the low-risk and high-risk group. The typi - its effectiveness and log-rank test was performed. cal presentation of two patients in the low-risk and high- risk group was illustrated in Fig. 4. Results Patient characteristics Correlation of selected radiomic features to the model A total of 298 LANSCLC patients were included for performance analysis, with 200 in the training cohort and 98 in the val- A total of 9 features were selected in the proposed model, idation cohort. The baseline and treatment-related char - including 5 tumor-related features and 4 lung-related acteristics were comparable between these two cohorts features. In Table  2, the IW of each selected feature for (Additional File 3). There were 57 females and 241 males both training and validation cohorts were listed in the Chen et al. Radiation Oncology (2022) 17:184 Page 6 of 12 Fig. 3 Kaplan-Meier curves for the training and validation cohort, with all selected features (a, b), and without pulmonary features (c, d), respectively order from high to low. The imaging features from lungs Patients in the low-risk group had better baseline ranked at 2nd, 4th, 5th and 8th in the all 9 features in the FEV1/FVC% (median, 96.3% vs. 85.9%, p = 0.046) com- training cohort, and 1st, 3rd, 6th, and 8th in the valida- pared with those in the high-risk group (Fig. 6B). Kaplan- tion cohort. When all pulmonary features were excluded Meier analysis indicated that better baseline FEV1/FVC% from the selected feature set, the AUCs for the training (p = 0.006) and SaO2 (p = 0.039) could exhibit superior and validation cohorts were reduced by 0.179 and 0.223, OS, DLCO% (p = 0.063) had a tendency to be associated respectively (Fig.  5). Figure  4 showed two patients in the with OS, however, pO2 (p = 0.110) and AaDO2 (p = 0.299) low-risk and high-risk groups. failed to predict OS (Additional File 5). Correlation of radiomic features to the PFT/BGA indicators Dynamic changes of lymphocyte counts before and after Pearson correlation analysis (Fig.  6A) demonstrated that CCRT FEV1/FVC% had modest correlation with three pulmo- Although there was no significant difference in lym - nary features (SHAPE_Volume_mL, GLRLM_LRE and phocyte counts before CCRT (median, 1650 vs. 1650 GLRLM_RP) (all Pearson correlation >|0.45|), and mild cells/mm , p > 0.99) between the low-risk and high- correlation with CONV_SUVstd of lungs and GLZLM_ risk group (Additional File 6), patients in the low-risk GLNUz of tumor (all Pearson correlation >|0.25|). Other group had less Grade ≥ 3 lymphopenia (63.2% vs. 83.3%, PFT/BGA indicators and radiomic features were not well p = 0.031) during CCRT, and more patients in the low-risk correlated. group could recover to normal level (≥ 1000 cells/mm ) Chen et al. Radiation Oncology (2022) 17:184 Page 7 of 12 Fig. 4 Two patients from the high-risk (A) and low-risk group (B). From the series CT images, there were discernible distinction observed in tumor and pulmonary status between the two cases. High-risk patient had heterogeneous primary lung tumor and chronic obstructive pneumonia (A-1), while low- risk patient had relatively homogeneous primary lung tumor and better pulmonary condition (B-1). Significant tumor remission was achieved in low-risk patient without obvious radiation pneumonitis after CCRT (B-2), while high-risk patient had stable disease and developed Grade 2 radiation pneumonitis in bilateral lungs (A-2). The GLZLM matrices for tumor, GLRLM matrices along + x axis for lungs, and histograms of HU values for tumor (A-3, B-3) were displayed. It was found that the short homogeneous runs and the non-uniformity of the grey-levels (CT value) were emphasized for high-risk patient compared to those of low-risk patient Chen et al. Radiation Oncology (2022) 17:184 Page 8 of 12 Table 2 Selected features and their importance rank for the training and validation cohort Importance rank Training cohort Validation cohort Selected features Category IW Selected features Category IW 1 GLRLM_SRE Tumor 0.0451 SHAPE_Volume_mL Lung 0.1439 2 SHAPE_Volume_mL Lung 0.0426 GLZLM_GLNUz Tumor 0.1122 3 CONV_SUVstd Tumor 0.0177 GLRLM_RP Lung 0.0776 4 CONV_SUVstd Lung 0.0169 GLRLM_SRE Tumor 0.0643 5 GLRLM_RP Lung 0.0143 HISTO_Entropy_log10 Tumor 0.0541 6 GLZLM_GLNUz Tumor 0.0126 CONV_SUVstd Lung 0.0372 7 HISTO_Entropy_log10 Tumor 0.0121 TLG_mL Tumor 0.0224 8 GLRLM_LRE Lung 0.0036 GLRLM_LRE Lung 0.0066 9 TLG_mL Tumor 0.0006 CONV_SUVstd Tumor <0.0001 Abbreviations: IW, importance weight; GLRLM, grey-level run length matrix; SRE, short-run emphasis; RP, run percentage; GLZLM, grey-level zone length matrix; GLNUz, gray-level non-uniformity for zone; LRE, long-run emphasis; TLG, total lesion glycolysis Fig. 5 The receiver operating characteristics (ROC) curves for the training and validation cohort, with all selected features (a, b), and without pulmonary features (c, d), respectively at 4 ~ 6 weeks post CCRT (71.4% vs. 27.8%, p < 0.001). Discussions Kaplan–Meier curves demonstrated that better recov- Application of radiomics to the long-term survival pre- ery to normal level (≥ 1000 cells/mm ) at 4 ~ 6 weeks post diction for LANSCLC after CCRT is a reasonable exten- CCRT (3-year OS rate, 47.5% versus 14.3%, p = 0.001) was sion under the background of the field-wide adoption of a prognostic factor of OS (Additional File 5). machine learning methods. Other than previous works focused on the features from tumor and peritumoral tis- sue, the relationship between tumor and TOE is increas- ingly attached importance. Significant association was Chen et al. Radiation Oncology (2022) 17:184 Page 9 of 12 Fig. 6 Pearson correlation coefficient heatmap for radiomic features and PFT/BGA indicators ( A), and the distribution of PFT/BGA indicators between the low-risk and high-risk group (B-E). A. The FEV1/FVC% had modest correlation with three pulmonary features (SHAPE_Volume_mL, GLRLM_LRE and GLRLM_RP) (all Pearson correlation >|0.45|), and mild correlation with CONV_SUVstd of lungs and GLZLM_GLNUz of tumor (all Pearson correlation >|0.25|). Other PFT/BGA indicators and radiomic features were not well correlated. B-E. Patients in the low-risk group had better baseline FEV1/FVC% (median, 96.3% vs. 85.9%, p = 0.046) compared with those in the high-risk group. No significant difference of DLCO% (median, 84.3% vs. 77.5%, p = 0.136), pO2 (median, 86.0 vs. 86.0 mmHg, p > 0.999), AaDO2 (median, 18.0 vs. 18.0 mmHg, p > 0.999) and SaO2 (median, 96.8% vs. 96.5%, p = 0.634) was found. Abbreviation: PFT, pulmonary function test; BGA, blood gas analysis found between pulmonary function and radiomic fea- support this finding, implying that the TOE, herein the tures extracted from the lungs of CT images [21–23]. In pulmonary environment, might have a significant impact current study, the long-term survival forecast accuracy of in LANSCLC patients with large tumor burden and lim- LANCLC patients after CCRT was demonstrated to be ited pulmonary function. Accordingly, the relatively lon- boosted by integrating primary tumor characteristics and ger OS for patients with healthier pulmonary status could pulmonary features from pretreatment CT images. Based possibly contribute to their more tolerance to radical on the CT-based predictive model, patients could be CCRT and less incidence of severe lung toxicities. precisely stratified into the low-risk and high-risk group PFT have been reported to predict the risk of RILT before treatment, which should be considered in individ- after CCRT [26–29]. Our previous work showed that ualized treatment decision-making process. FEV1/FVC% and DLCO% were prognostic factors for From the importance rank of the selected features, long-term survival but not for PFS [17], implying that it could be confirmed that two features from tumor, long-term survival outcomes might not be achievable GLRLM_SRE and GLZLM_GLNUz which represent the due to detriment of pulmonary function even though inhomogeneity of CT images [19], remained important patients had good early response to CCRT. To further factors determining OS, which were consistent with pub- interpret the underlying role of these selected radiomic lished literatures [24, 25]. Meanwhile, the ranking of pul- lung features, the correlation between radiomic fea- monary features underlined their indispensable role in tures and PFT/BGA indicators were explored in depth the OS forecast. Our results of the significant difference and it was confirmed that FEV1/FVC% was well corre - between fitting and prediction accuracies with and with - lated with radiomic pulmonary features. This correlation out pulmonary features in model performance further between the pulmonary ventilation function and selected Chen et al. Radiation Oncology (2022) 17:184 Page 10 of 12 radiomic pulmonary features for OS prediction reaffirms What’s more, the predictive OS results using imag- the findings in Occhipinti et al.’s study that the changes ing features in our study with machine learning could in lung function, such as bronchial thickening and hon- be utilized as an effective indicator for the survival risk eycombing, can be mechanistically explained based on stratification of these patients, which could potentially morphological CT features [23]. And it might addition- individualize CCRT regimen and adjuvant treatment ally imply that the tumor not only interacts with cells in from the perspective of personalized medicine. For its immediate vicinity, but also communicates with the example, immunotherapy has evolved into a standard entire host organ [30], just as suggested by a prior study adjuvant treatment option for LANSCLC patients treated [31] that the tumor and TOE could possibly interact in a with definitive CCRT. Based on the promising results of bi-directional way. the phase III PACIFIC study [33, 34], adjuvant immu- In the aspect of methodology, the machine learning notherapy resulted in a significant prolonged PFS and framework in this study used SVM combined with the OS for those patients. To be noticed, the most common proposed IFSMT approach to iteratively select features grade 3 or 4 adverse event in the durvalumab arm was using GA and improve the accuracy of the prediction pneumonia (4.4%), followed by pneumonitis or radiation model. Our avoidance of topical deep-learning frame- pneumonitis (3.4%), and Asian patients seemed to have a works, such as deep convolutional neural network, is due higher rate of any grade pneumonitis (73.6%) and severe to the intrinsic weaknesses of overfitting and blackbox pneumonitis (5.6%) [35]. Thus, based on the survival risk for these frameworks. To ease the problem of overfitting, stratification of LANSCLC patients in this study, low-risk the deep-learning frameworks are more suitable for the patients might have several potential advantages for adju- learning tasks armed with big data as learning samples. vant immunotherapy: (1) supporting role of better pul- However, the number of patients in current study for monary function and quality of life; (2) superior tumor model training was relatively small, which intensively remission with less pulmonary toxicities; (3) less severe restricts the application of deep-learning frameworks lymphopenia during CCRT and better recovery of lym- which may have millions of parameters and thousands phopenia from CCRT. However, for high-risk LANSCLC of decision making variables. The SVM is equivalent to patients who had worse baseline FEV1/FVC%, higher rate an optimized three-layer neural network with only one of Grade ≥ 3 lymphopenia during CCRT, worse recov- hidden layer. This simplified neural network architec - ery of lymphopenia from CCRT, and higher incidence of ture substantially reduces the potential of overfitting. radiation-induced pneumonitis, radical CCRT or further Additionally, in contrast with the problem of blackbox adjuvant immunotherapy might not be feasible because for deep learning framework, the features used in mod- of poor organ functions and high probability of severe eling are explicitly created and selected with the IFSMT complications. Therefore, pretreatment radiomics-based approach. Therefore, each feature had an explicit clinical risk stratification of LANSCLC patients using features or physical meaning relevant to image of a specific ROI, from tumor and TOE could provide direct evidences to which made it easy to apprehend the behind-the-scene effectively support the treatment decision making. mechanism of the survival status prediction and directly It should also be noted that there were a few limita- related the comprehensible clinical and image oriented tions in this study. First, the absence of external valida- indices to the clinical outcome. The effectiveness of tion was the major disadvantage. Nevertheless, multiple IFSMT approach had been demonstrated by high AUC CT simulation machines were available in our institution values achieved for the survival status prediction. (Additional File 1). The high AUC values were generated The most recent work on prognostic model for the sur - from these different scanners with varied parameter set - vival outcome for NSCLC patients treated with CCRT tings, demonstrating the great robustness of our model. demonstrated that pretreatment CT texture features Besides, Zhao et al. considered that radiomic features provided prognostic information beyond CPFs [12]. in lung cancer were reproducible over a wide range of However, it didn’t provide the result in terms of AUC imaging settings [36]. Multicenter validations with larger or employ the validation cohort. In another predictive samples are warranted for the ultimate application of this model conducted by Dehing-Oberije C et al. [32], which model clinically. Second, there might be some variabil- used CPF indices only, the AUC was 0.74 for the train- ity in multiple observer delineations in our study. E et al. ing cohort, 0.75 and 0.76 for the two separate valida- reported that although the ROIs delineation tended to be tion cohorts. The improvement of model performance different between individual experts, an overall high AUC by imaging features in current study is discernible with value could still be achieved [37]. Third, we focused only the AUCs of 0.965 and 0.869 for the training and valida- on the radiomic analysis of pretreatment planning CT in tion cohort, which could be attributed to inclusion of the this study, and other imaging modalities, such as PET-CT image-based pulmonary features. [38] and MRI, still need to be investigated as to whether Chen et al. Radiation Oncology (2022) 17:184 Page 11 of 12 Authors’ contributions they could also yield complementary information which HL and CMX contributed to the conception and design of the study. NBC, MX, RZ, BQ, YFL, SZ, CC, QWL, BW, JYG and KQP contributed to acquisition of data. would facilitate more accurate predictive models. NBC, MX, RZ, YZ, BQ, YFL and HHJ analyzed and interpreted data. NBC, MX and RZ drafted the article. HL and CMX revised it critically and made the final Conclusion approval of the version to be submitted. All authors read and approved the final manuscript. Pretreatment CT-based radiomics features from tumor and TOE could improve the long-term survival fore- Funding cast accuracy in LANSCLC patients treated with CCRT This work was supported by National Natural Science Foundation of China [Grant Number 82073328], National Key R&D Program of China [Grant Number using machine learning. The predictive results could be 2018YFC0116800] and Suzhou Municipal Science and Technology Program utilized as an effective indicator for the stratification of [Grant Number SYS2018009]. The funders had no role in study design; in the these patients into the low-risk and high-risk groups. It collection, analysis and interpretation of data; in the writing of the report; and in the decision to submit the article for publication. was further confirmed that patients in the low-risk group had better baseline FEV1/FVC%, less severe lymphope- Data Availability nia during CCRT, better recovery of lymphopenia from The datasets used and analysed during the current study are available from the corresponding author on reasonable request. CCRT, lower incidence of radiation-induced pneumo- nitis, superior tumor remission and long-term survival, Declarations which might suggest more benefit for these patients from radical CCRT or further adjuvant immunotherapy. Ethics approval and consent to participate This study was conducted based on the ethical standards of the Declaration Abbreviations of Helsinki. It was reviewed and approved by the review board of Sun Yat-sen CI confidence interval University Cancer Center. Since it was an anonymous retrospective study, we CCRT concurrent chemoradiotherapy required and were granted a waiver of written informed-consent. LANSCLC locally advanced non-small cell lung cancer RILT radiation-induced lung t oxicity Consent for publication CPF clinical prognostic factor OS overall survival Not applicable. TOE tumor organismal environment PFT pulmonary function test Competing interests PFS progression-free survival The authors declare that they have no competing interests. ORR objective response rate CT computed tomography IFSMT integrated feature selection and Author details model training Department of Radiation Oncology, State Key Laboratory of Oncology MRI magnetic resonance imaging in South China, Collaborative Innovation Center for Cancer Medicine, Sun PET-CT positr on emission tomography-computed tomography Yat-sen University Cancer Center, No.651 Dongfeng Road East, 4D four-dimensional 510060 Guangzhou, China GTV gross tumor volume Department of Imaging Diagnosis and Interventional Center, State ITV internal target volume Key Laboratory of Oncology in South China, Collaborative Innovation PTV planning target volume Center for Cancer Medicine, Sun Yat-sen University Cancer Center, No.651 OAR organ at risk Dongfeng Road East, 510060 Guangzhou, Guangdong, China BGA blood gas analysis Department of Cardiac Surgery, The First Affiliated Hospital of Sun Yat- ROI regions of interest sen University, Guangzhou, Guangdong, China GLCM grey-level co-occurrence matrix Department of Pulmonary and Critical Care Medicine, The First Affiliated NGLDM neighborhood grey-level different matrix Hospital of Sun Yat-sen University, Guangzhou, Guangdong, China GLRLM grey-level run length matrix Homology Medical Technologies Inc., Ningbo, Zhejiang, China GLZLM grey-level zone length matrix Guangzhou Xinhua University, Guangzhou, Guangdong, China SVM support vector machine GA genetic algorithm Received: 7 June 2022 / Accepted: 28 September 2022 LOOCV leave-one-out cross-validation ROC receiver operating characteristics AUC area under curve IW importance weight TPR true positive rate TNR true negative rate References 1. Zhou Z, Song X, Wu A, Liu H, Wu H, Wu Q, et al. Pulmonary emphysema is a Supplementary Information risk factor for radiation pneumonitis in NSCLC patients with squamous cell The online version contains supplementary material available at https://doi. carcinoma after thoracic radiation therapy. Sci Rep. 2017;7(1):2748. org/10.1186/s13014-022-02136-w. 2. Li F, Zhou Z, Wu A, Cai Y, Wu H, Chen M, et al. Preexisting radiological inter- stitial lung abnormalities are a risk factor for severe radiation pneumonitis in patients with small-cell lung cancer after thoracic radiation therapy. Radiat Additional File 1. 4DCT image acquisition Oncol. 2018;13(1):82. Additional File 2. Genetic algorithm 3. Roach M 3rd, Gandara DR, Yuo HS, Swift PS, Kroll S, Shrieve DC, et al. Radiation pneumonitis following combined modality therapy for lung cancer: analysis Additional File 3 Patient characteristics of prognostic factors. J Clin Oncol. 1995;13(10):2606–12. 4. Bradley JD, Hope A, El Naqa I, Apte A, Lindsay PE, Bosch W, et al. A nomogram Additional File 4. Tumor response after CCRT in the low-risk and high-risk to predict radiation pneumonitis, derived from a combined analysis of RTOG group 9311 and institutional data. Int J Radiat Oncol Biol Phys. 2007;69(4):985–92. Additional File 5. Additional File 6. Chen et al. Radiation Oncology (2022) 17:184 Page 12 of 12 5. Hope AJ, Lindsay PE, El Naqa I, Alaly JR, Vicic M, Bradley JD, et al. Modeling measured by quantitative CT and CT-based radiomics in COPD. Respir Res. radiation pneumonitis risk with clinical, dosimetric, and spatial parameters. 2019;20(1):101. Int J Radiat Oncol Biol Phys. 2006;65(1):112–24. 24. Ganeshan B, Panayiotou E, Burnand K, Dizdarevic S, Miles K. Tumour hetero- 6. Valdes G, Solberg TD, Heskel M, Ungar L, Simone CB 2. Using machine learn- geneity in non-small cell lung carcinoma assessed by CT texture analysis: a ing to predict radiation pneumonitis in patients with stage I non-small cell potential marker of survival. Eur Radiol. 2012;22(4):796–802. lung cancer treated with stereotactic body radiation therapy. Phys Med Biol. 25. Bradley JD, Ieumwananonthachai N, Purdy JA, Wasserman TH, Lockett MA, 2016;61(16):6105–20. nd. . Graham MV, et al. Gross tumor volume, critical prognostic factor in patients 7. Yakar M, Etiz D, Metintas M, Ak G, Celik O. Prediction of Radiation Pneumonitis treated with three-dimensional conformal radiation therapy for non-small- With Machine Learning in Stage III Lung Cancer: A Pilot Study. Technol Cancer cell lung carcinoma. Int J Radiat Oncol Biol Phys. 2002;52(1):49–57. Res Treat. 2021;20:15330338211016373. 26. Lopez Guerra JL, Gomez D, Zhuang Y, Levy LB, Eapen G, Liu H, et al. Change 8. Bourbonne V, Da-Ano R, Jaouen V, Lucia F, Dissaux G, Bert J, et al. Radiomics in diffusing capacity after radiation as an objective measure for grading analysis of 3D dose distributions to predict toxicity of radiotherapy for lung radiation pneumonitis in patients treated for non-small-cell lung cancer. Int J cancer. Radiother Oncol. 2021;155:144–50. Radiat Oncol Biol Phys. 2012;83(5):1573–9. 9. Puttanawarut C, Sirirutbunkajorn N, Tawong N, Jiarpinitnun C, Khachonkham 27. Park YH, Kim JS. Predictors of radiation pneumonitis and pulmonary function S, Pattaranutaporn P, et al. Radiomic and Dosiomic Features for the Prediction changes after concurrent chemoradiotherapy of non-small cell lung cancer. of Radiation Pneumonitis Across Esophageal Cancer and Lung Cancer. Front Radiation Oncol J. 2013;31(1):34–40. Oncol. 2022;12:768152. 28. Torre-Bouscoulet L, Muñoz-Montaño WR, Martínez-Briseño D, Lozano-Ruiz FJ, 10. Lee G, Lee HY, Park H, Schiebler ML, van Beek EJR, Ohno Y, et al. Radiomics Fernández-Plata R, Beck-Magaña JA, et al. Abnormal pulmonary function tests and its emerging role in lung cancer research, imaging biomarkers and clini- predict the development of radiation-induced pneumonitis in advanced cal management: State of the art. Eur J Radiol. 2017;86:297–307. non-small cell lung Cancer. Respir Res. 2018;19(1):72. 11. Thawani R, McLane M, Beig N, Ghose S, Prasanna P, Velcheti V, et al. Radiomics 29. Zhou Y, Yan T, Zhou X, Cao P, Luo C, Zhou L, et al. Acute severe radiation and radiogenomics in lung cancer: A review for the clinician. Lung cancer pneumonitis among non-small cell lung cancer (NSCLC) patients with (Amsterdam Netherlands). 2018;115:34–41. moderate pulmonary dysfunction receiving definitive concurrent chemo - 12. Fried DV, Tucker SL, Zhou S, Liao Z, Mawlawi O, Ibbott G, et al. Prognostic radiotherapy: Impact of pre-treatment pulmonary function parameters. value and reproducibility of pretreatment CT texture features in stage III non- Strahlentherapie und Onkologie: Organ der Deutschen Rontgengesellschaft small cell lung cancer. Int J Radiat Oncol Biol Phys. 2014;90(4):834–42. [et al]. 2020;196(6):505–14. 13. Sun W, Jiang M, Dang J, Chang P, Yin FF. Eec ff t of machine learning methods 30. Dieterich LC, Bikfalvi A. The tumor organismal environment: Role in on predicting NSCLC overall survival time based on Radiomics analysis. tumor development and cancer immunotherapy. Sem Cancer Biol. Radiation Oncol (London England). 2018;13(1):197. 2020;65:197–206. 14. Lu MT, Ivanov A, Mayrhofer T, Hosny A, Aerts H, Hoffmann U. Deep Learning 31. Khorrami M, Prasanna P, Gupta A, Patil P, Velu PD, Thawani R, et al. Changes to Assess Long-term Mortality From Chest Radiographs. JAMA Netw open. in CT Radiomic Features Associated with Lymphocyte Distribution Predict 2019;2(7):e197416. Overall Survival and Response to Immunotherapy in Non-Small Cell Lung 15. Duijm M, van der Voort van Zyp NC, Granton PV, van de Vaart P, Mast ME, Cancer. Cancer Immunol Res. 2020;8(1):108–19. Oomen-de Hoop E, et al. Prognostic factors of local control and disease free 32. Dehing-Oberije C, Yu S, De Ruysscher D, Meersschout S, Van Beek K, Lievens survival in centrally located non-small cell lung cancer treated with stereotac- Y, et al. Development and external validation of prognostic model for 2-year tic body radiation therapy. Acta Oncol. 2020;59(7):809–17. survival of non-small-cell lung cancer patients treated with chemoradio- 16. Kang J, Ning MS, Feng H, Li H, Bahig H, Brooks ED, et al. Predicting 5-Year therapy. Int J Radiat Oncol Biol Phys. 2009;74(2):355–62. Progression and Survival Outcomes for Early Stage Non-small Cell Lung 33. Antonia SJ, Villegas A, Daniel D, Vicente D, Murakami S, Hui R, et al. Dur- Cancer Treated with Stereotactic Ablative Radiation Therapy: Development valumab after Chemoradiotherapy in Stage III Non-Small-Cell Lung Cancer. N and Validation of Robust Prognostic Nomograms. Int J Radiat Oncol Biol Phys. Engl J Med. 2017;377(20):1919–29. 2020;106(1):90–9. 34. Antonia SJ, Villegas A, Daniel D, Vicente D, Murakami S, Hui R, et al. Overall 17. Qiu B, Xiong M, Luo Y, Li Q, Chen N, Chen L, et al. Hypofractionated Intensity Survival with Durvalumab after Chemoradiotherapy in Stage III NSCLC. N Modulated Radiation Therapy With Concurrent Chemotherapy in Locally Engl J Med. 2018;379(24):2342–50. Advanced Non-Small Cell Lung Cancer: A Phase II Prospective Clinical Trial 35. Peng L, Wu YL. Immunotherapy in the Asiatic population: any differences (GASTO1011). Pract Radiat Oncol. 2021;11(5):374–83. from Caucasian population? J Thorac disease. 2018;10(Suppl 13):1482-s93. 18. Kong FM, Ritter T, Quint DJ, Senan S, Gaspar LE, Komaki RU, et al. Consider- 36. Zhao B, Tan Y, Tsai WY, Schwartz LH, Lu L. Exploring Variability in CT Char- ation of dose limits for organs at risk of thoracic radiotherapy: atlas for lung, acterization of Tumors: A Preliminary Phantom Study. Translational Oncol. proximal bronchial tree, esophagus, spinal cord, ribs, and brachial plexus. Int J 2014;7(1):88–93. Radiat Oncol Biol Phys. 2011;81(5):1442–57. 37. Lu EL, Li L, Yang L, Schwartz H, Zhao LH. B. Radiomics for Classification of 19. Nioche C, Orlhac F, Boughdad S, Reuzé S, Goya-Outi J, Robert C, et al. LIFEx: Lung Cancer Histological Subtypes Based on Nonenhanced Computed A Freeware for Radiomic Feature Calculation in Multimodality Imaging to Tomography. Acad Radiol. 2019;26(9):1245–52. Accelerate Advances in the Characterization of Tumor Heterogeneity. Cancer 38. Cook GJ, Yip C, Siddique M, Goh V, Chicklore S, Roy A, et al. Are pretreatment Res. 2018;78(16):4786–9. 18F-FDG PET tumor textural features in non-small cell lung cancer associated 20. Smola AJ, Scholkopf B. A tutorial on support vector regression. Stat Comput. with response and survival after chemoradiotherapy? Journal of nuclear 2004;14(3):199–222. medicine: official publication. Soc Nuclear Med. 2013;54(1):19–26. 21. Wu MT, Chang JM, Chiang AA, Lu JY, Hsu HK, Hsu WH, et al. Use of quantita- tive CT to predict postoperative lung function in patients with lung cancer. Publisher’s Note Radiology. 1994;191(1):257–62. Springer Nature remains neutral with regard to jurisdictional claims in 22. Lafata KJ, Zhou Z, Liu JG, Hong J, Kelsey CR, Yin FF. An Exploratory Radiomics published maps and institutional affiliations. Approach to Quantifying Pulmonary Function in CT Images. Sci Rep. 2019;9(1):11509. 23. Occhipinti M, Paoletti M, Bartholmai BJ, Rajagopalan S, Karwoski RA, Nardi C, et al. Spirometric assessment of emphysema presence and severity as http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Radiation Oncology Springer Journals

CT radiomics-based long-term survival prediction for locally advanced non-small cell lung cancer patients treated with concurrent chemoradiotherapy using features from tumor and tumor organismal environment

Loading next page...
 
/lp/springer-journals/ct-radiomics-based-long-term-survival-prediction-for-locally-advanced-WiOQX6VOSr

References (41)

Publisher
Springer Journals
Copyright
Copyright © The Author(s) 2022
eISSN
1748-717X
DOI
10.1186/s13014-022-02136-w
Publisher site
See Article on Publisher Site

Abstract

Background Definitive concurrent chemoradiotherapy (CCRT ) is the standard treatment for locally advanced non- small cell lung cancer (LANSCLC) patients, but the treatment response and survival outcomes varied among these patients. We aimed to identify pretreatment computed tomography-based radiomics features extracted from tumor and tumor organismal environment ( TOE) for long-term survival prediction in these patients treated with CCRT. Methods A total of 298 eligible patients were randomly assigned into the training cohort and validation cohort with a ratio 2:1. An integrated feature selection and model training approach using support vector machine combined with genetic algorithm was performed to predict 3-year overall survival (OS). Patients were stratified into the high-risk and low-risk group based on the predicted survival status. Pulmonary function test and blood gas analysis indicators were associated with radiomic features. Dynamic changes of peripheral blood lymphocytes counts before and after CCRT had been documented. Results Nine features including 5 tumor-related features and 4 pulmonary features were selected in the predictive model. The areas under the receiver operating characteristic curve for the training and validation cohort were 0.965 and 0.869, and were reduced by 0.179 and 0.223 when all pulmonary features were excluded. Based on radiomics- Nai-Bin Chen, Mai Xiong and Rui Zhou contributed equally as co- first authors. *Correspondence: Chuan-Miao Xie xiechm@sysucc.org.cn Hui Liu liuhuisysucc@126.com Full list of author information is available at the end of the article © The Author(s) 2022. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data. Chen et al. Radiation Oncology (2022) 17:184 Page 2 of 12 derived stratification, the low-risk group yielded better 3-year OS (68.4% vs. 3.3%, p < 0.001) than the high-risk group. Patients in the low-risk group had better baseline FEV1/FVC% (96.3% vs. 85.9%, p = 0.046), less Grade ≥ 3 lymphopenia during CCRT (63.2% vs. 83.3%, p = 0.031), better recovery of lymphopenia from CCRT (71.4% vs. 27.8%, p < 0.001), lower incidence of Grade ≥ 2 radiation-induced pneumonitis (31.6% vs. 53.3%, p = 0.040), superior tumor remission (84.2% vs. 66.7%, p = 0.003). Conclusion Pretreatment radiomics features from tumor and TOE could boost the long-term survival forecast accuracy in LANSCLC patients, and the predictive results could be utilized as an effective indicator for survival risk stratification. Low-risk patients might benefit more from radical CCRT and further adjuvant immunotherapy. Trial registration: retrospectively registered. Keywords Locally advanced non-small cell lung cancer, Radiomics, Machine learning, Long-term survival prediction, Tumor organismal environment. Introduction investigate the relationship of tumor and TOE, due to its Definitive concurrent chemoradiotherapy (CCRT) is the accessibility. standard treatment for patients with unresectable locally In this study, we utilized computed tomography (CT) advanced non-small cell lung cancer (LANSCLC). In the images before CCRT to develop an image-based machine past two decades, concomitant regimens achieved prom- learning framework to analyze the relationship of pri- ising tumor local control and long-term survival. With mary lung tumor and bilateral lungs for long-term sur- improved outcome, the maintenance of an adequate pul- vival prediction in LANSCLC. To balance the training monary function is essential to ensure acceptable qual- accuracy and predictive capability using relative small ity of life and adjuvant immunotherapy. However, many number of patient samples, an integrated feature selec- patients with LANSCLC are diagnosed with pre-existing tion and model training (IFSMT) approach was devel- lung comorbidities, which significantly increases the risk oped to extract the most critical quantitative radiomic for radiation-induced lung toxicity (RILT) [1, 2]. features from both tumor and lungs. A radiomic-based Most existing RILT prediction models largely focused risk stratification was built to distinguish high-risk and on clinical prognostic factors (CPFs) and dose-volume low-risk patients and provided evidence for clinical deci- histogram parameters [3–5], but remained insufficient. sion making. Recently, machine learning methods have been reported to improve the capacity of the predictive modelling [6–9], Methods compared with logistic regression widely used in normal Study population tissue complication probability model. Consecutive patients irradiated for lung cancer from Moreover, radiomics analysis, attempting to identify September 2011 to April 2019 in our institution were computational biomarkers potentially hidden within retrospectively screened. Inclusion criteria included: (1) high-throughput imaging data [10, 11], has been dem- histologically confirmed NSCLC; (2) unresectable stage onstrated the added predictive value for overall survival III disease (AJCC/UICC 8th staging criteria) proven by (OS) [12–14] or RILT [8, 9]. However, most of them rely chest and upper abdominal CT, brain magnetic reso- on the radiomic information from tumor or its surround- nance imaging (MRI), bone scan and/or positron emis- ing peritumoral region, few studies have been designed sion tomography-computed tomography (PET-CT); (3) based on the radiomics analysis of tumor organismal definitive radiotherapy with concurrent chemotherapy environment (TOE). was administered; (4) stay followed-up no less than 6 Similar to other published reports [15, 16], our previ- months since the start of radiotherapy (unless death ous study [17] indicated that pulmonary function test or disease progression was documented); (5) complete (PFT) was significantly related to patients’ long-term clinical records. Patients that met the inclusion criteria survival. However, it failed to predict progression-free were randomly assigned into the training and validation survival (PFS). Even though patients with worse FEV1/ cohort, with the numbers at a ratio 2:1. FVC% or DLCO% showed a high objective response rate (ORR) to CCRT, their survival outcomes were still poor, Planning CT image acquisition hinting that TOE, the status of lungs in the case of LAN- The four-dimensional (4D) planning CT scan was per - SCLC might play an indispensable role in the prognostic formed 1–2 weeks prior to treatment, using multiple CT prediction after CCRT. As some patients could not tol- simulation positioning machines with varied parameter erate well with PFT, radiomics analysis using machine settings in our institution (detailed in Additional File 1). learning method might be an effective technique to Ten phases of the breathing cycle were reconstructed, Chen et al. Radiation Oncology (2022) 17:184 Page 3 of 12 Fig. 1 Schematic overview of the integrated feature selection and model training (IFSMT ) approach. IFSMT approach consisted of five steps: (1) volumet - ric data pre-processing; (2) delineation; (3) feature extraction; (4) integrated feature selection and model training; (5) model validation using leave-one- out cross-validation (LOOCV ). including: 0%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, medical records, including blood tests, PFT, blood gas and 90%. The segmentation and radiomics were then per - analysis (BGA) and radiologic tests. All included patients formed on the 20% phase (middle exhale phase) with a received regular radiologic follow-up, including chest consistent mediastinum/lung window level setting. and upper abdominal CT and brain MRI performed every 3 ~ 6 months in the first 2 years, and every 6 ~ 12 Radiotherapy and concurrent chemotherapy months thereafter. PET-CT, bone scan, and biopsy were Patients were positioned supine and immobilized in a recommended if clinically required. The responses to vacuum pad. They were scanned from the Atlas to the CCRT were first assessed by an independent radiation second lumbar vertebra level with 0.3-0.5  cm thickness oncologist and confirmed by a senior physician at 4 ~ 6 slices to obtain the stimulation CT images. The respira - weeks post CCRT, based on Response Evaluation Crite- tion motion was recorded by performing 4DCT scan- ria in Solid Tumors 1.1. Another senior radiologist was ning. The maximum intensity projection images were consulted for disagreement. Therapeutic toxicities were reconstructed using the images collected in 10 phases of graded and recorded according to Common Terminology respiratory cycle. Gross tumor volume (GTV) was delin- Criteria for Adverse Events 4.0. eated to cover the tumor and involved regional nodes visible on each phase of the 4DCT. The total volumes of OS modelling procedures GTVs across the 10 respiratory phases CT composed the The whole procedures were illustrated in Fig.  1. For both internal target volume (ITV). Planning target volumes cohorts of patients, the regions of interest (ROIs) corre- (PTVs) were created by expanding GTV and clinical tar- sponding to GTV and lungs were delineated by an auto- get volume with 6  mm. Lungs were delineated accord- contouring software tool CezanneDraw™ v1.0 (Homology ing to the atlases for organs at risk (OARs) in thoracic Medical, Ningbo, China, 2020) using the CT slices and radiation therapy [18], but GTV was excluded from the manually modified by radiation oncologists if necessary. lung delineation. A dose of 60-76  Gy was prescribed to One 3D bounding box was fitted for each ROI. And inside PTV-GTV in 22–33 fractions, with 2-3  Gy per fraction the bounding box, the CT values of the ROI voxels were performed once daily, using intensity modulated radia- retained while the values of other voxels were marked tion therapy technique. The dose constraints for OARs by zero. CT values of voxels in each bounding box were were: V20 < 35% for lungs; mean lung dose < 19 Gy; maxi- then interpolated to a resolution of 1  mm×1  mm×5  mm mun dose (Dmax) of esophagus < 105% prescription dose; and resampled into 400 discrete values (called bins) with Dmax of spinal cord < 46 Gy; V30 < 40% for heart. absolute discretization from − 1000 to 3000 Hounsfield All patients received platinum-based double agents units, leading to a fixed bin size of 10 Hounsfield units. weekly or every three weeks. The regimens included A total of 92 tumor-related and lung-related features docetaxel/paclitaxel/etopside/pemetrexed plus platinum. were then computed for both ROIs and used as the input feature pool for the machine learning framework by the Evaluation and follow-up LIFEx software (version 3.44) [19]. The imaging-based The baseline characteristics of each patient before entry features covered two categories of texture features and were reviewed attentively and extracted from their first order features. The texture features consisted of four Chen et al. Radiation Oncology (2022) 17:184 Page 4 of 12 Fig. 2 Schematic overview of the genetic algorithm (GA) in the integrated feature selection and model training (IFSMT ) approach. A chromosome is scored with LOOCV-SVM. The chromosomes of higher scores may go through mutation and crossover to make new ones to replace those of lower scores. Collect the chromosome of best score from each generation into a group. And the best one in the group is the result of the model. Abbreviation: LOOCV, leave-one-out cross-validation; SVM, support vector machine sub-categories of matrix based texture features. These generation, the chromosomes of higher scores may go matrices included the grey-level co-occurrence matrix through mutation, partially changing feature encoding, (GLCM), neighborhood grey-level different matrix and crossover, partially exchanging feature encoding, to (NGLDM), grey-level run length matrix (GLRLM) and make new ones to replace those of lower scores. Collect grey-level zone length matrix (GLZLM). The first order the chromosome of best score from each generation into features included indices from shape, indices from histo- a group. And the best one in the group is the result of the gram and conventional indices. model. Manual reconfiguration of SVM is not included in The machine learning based classification method used the model. to predict the two-class 3-year survival status for each Once the optimal set of features was determined, the individual patient was support vector machine (SVM) SVM models were trained again on the training cohort. [20]. The SVM mapped the features of training data into In this study, after extensive experimental comparisons, a high-dimensional feature space through a kernel func- the linear kernel was chosen for SVM and optimal hyper tion and utilizes a hyper-plane to optimally separate the parameters of the SVM (C, ε and γ) were determined training data points into two categories. To reduce the through exhaustive search in the parametric space. possibility of overfitting, only a subset of features from Receiver operating characteristics (ROC) curves were the feature pool could be selected for the input of SVM. obtained by varying threshold of the decision variable, In this study, the IFSMT approach was developed to the signed distance to decision hyper-plane. Area under maximize the fitting accuracy and minimize the over - curve (AUC) for each ROC was calculated for training fitting potential. This posteriori approach applied the cohort. The trained models were then used to predict the genetic algorithm (GA) for the feature selection, which survival status for each individual patient in the valida- was illustrated in Fig. 2 and Additional File 2. A chro- tion cohort, and ROCs and their corresponding AUCs mosome represents a feature template working with were also calculated. All the above feature selection and SVM of certain configuration for diagnosing purpose. machine learning approaches were implemented on the The SVM is implemented in leave-one-out cross-valida - cloud-based clinical data service platform iRAAS® v2.0 tion (LOOCV) fashion to score a chromosome. In each (Homology Medical, Ningbo, China, 2020). Chen et al. Radiation Oncology (2022) 17:184 Page 5 of 12 Table 1 OS Training and validation accuracy using all selected To assess the importance of each selected feature to the features or without pulmonary features accurate prediction of the clinical outcome, a one-by-one OS status accuracy Training cohort Validation cohort feature evaluation procedure was designed. This proce - All Without All With- dure tested the importance of each feature by deleting features pulmo- features out pul- each feature from the selected feature set and calculating nary monary the reduction of the AUC for the model trained with the features features original selected features except this specific feature. This AUC 0.965 0.786 0.869 0.646 reduction of model performance was used as the impor- TNR (%) 95.00 74.00 92.86 75.00 tance weight (IW) of this feature. All the selected features TPR (%) 90.00 68.00 82.86 50.00 were then sorted according to their IWs. To further assess F1 score 0.923 0.701 0.892 0.625 the importance organismal features, the AUC for the Average accuracy (%) 92.50 71.00 87.86 62.50 model trained with the original selected features exclud- Overall accuracy (%) 92.50 71.00 85.71 57.14 Abbreviations: OS, overall survival; AUC, area under curve; TNR, true negative ing all the lung-related features were also calculated. rate; TPR, true positive rate Statistical methods OS was defined as the time from radiotherapy start to the in the whole cohort, with the median age of 59 years last follow up, which ended at November 30th, 2021, or (range, 28–81 years). Squamous cell carcinoma was the death. A t-test was used to determine if there was signifi - predominant histologic type both in the training (46.5%) cant difference between the means of continuous vari - and validation (62.2%) cohorts. ables, while Fisher’s exact test was performed to reveal the difference in distribution between two groups of OS modelling categories variables. The association between radiomic With the median follow-up of 27.7 (range, 4.0 ~ 122.7) features and PFT/BGA indicators was examined using months for all and 67.0 months (range, 36.2 ~ 122.7 Pearson’s correlation coefficient. A p-value < 0.05 (two- months) for event-free patients, our cohort demonstrated sided) were considered as statistically significant. Missing the estimated median OS of 27.6 (95% confidence inter - data were excluded from the statistical analysis. Statis- val (CI), 22.3 ~ 33.0) months, and the 3-year OS rate was tics were performed using SPSS 22.0 (IBM, Chicago, IL, 43.0% (95%CI, 37.3%~48.7%). USA). As shown in Table  1, the overall prediction accuracy To report the model fitting accuracy and the predic - for 3-year survival status was 92.50% and 85.71%, and the tion capability, the true positive rate (TPR), true negative AUC of the ROC was 0.965 and 0.869, respectively, in the rate (TNR), F1 score, overall prediction accuracy, average training and validation cohort. prediction accuracy for the training cohort and valida- tion cohort were calculated based on the SVM model. Stratification of patients in the validation cohort with Herein, death is marked as the positive. The overall pre - machine learning model diction accuracy was expressed as the number correctly In the validation cohort, 60 (61.2%) of 98 patients were predicted patients / the number of all patients; and the stratified into the high-risk group and 38 (44.1%) into the average prediction accuracy = (TPR + TNR)/2. low-risk group. CCRT was more successful in patients in To assess the prognostic value of the survival status the low-risk group than those in the high-risk group. The model, the predicted 3-year survival status was adopted ORR was 84.2% (32/38) and 66.7% (40/60) in the low-risk respectively as the clinical risk estimator to stratify the and high-risk group, respectively (p = 0.003) (Additional patients into the high-risk and low-risk groups. Patients File 4). And the low-risk group yielded better 3-year OS with negative predicted survival status were classified (68.4% versus 3.3%, p < 0.001, log-rank) than the high- into the low-risk group and the others with positive pre- risk group (Fig.  3B). What’s more, the rate of Grade ≥ 2 dicted survival status into the high-risk group. Kaplan- pneumonitis was 31.6% (12/38), versus 53.3% (32/60) Meier curves for both groups were displayed to illustrate (p = 0.040) in the low-risk and high-risk group. The typi - its effectiveness and log-rank test was performed. cal presentation of two patients in the low-risk and high- risk group was illustrated in Fig. 4. Results Patient characteristics Correlation of selected radiomic features to the model A total of 298 LANSCLC patients were included for performance analysis, with 200 in the training cohort and 98 in the val- A total of 9 features were selected in the proposed model, idation cohort. The baseline and treatment-related char - including 5 tumor-related features and 4 lung-related acteristics were comparable between these two cohorts features. In Table  2, the IW of each selected feature for (Additional File 3). There were 57 females and 241 males both training and validation cohorts were listed in the Chen et al. Radiation Oncology (2022) 17:184 Page 6 of 12 Fig. 3 Kaplan-Meier curves for the training and validation cohort, with all selected features (a, b), and without pulmonary features (c, d), respectively order from high to low. The imaging features from lungs Patients in the low-risk group had better baseline ranked at 2nd, 4th, 5th and 8th in the all 9 features in the FEV1/FVC% (median, 96.3% vs. 85.9%, p = 0.046) com- training cohort, and 1st, 3rd, 6th, and 8th in the valida- pared with those in the high-risk group (Fig. 6B). Kaplan- tion cohort. When all pulmonary features were excluded Meier analysis indicated that better baseline FEV1/FVC% from the selected feature set, the AUCs for the training (p = 0.006) and SaO2 (p = 0.039) could exhibit superior and validation cohorts were reduced by 0.179 and 0.223, OS, DLCO% (p = 0.063) had a tendency to be associated respectively (Fig.  5). Figure  4 showed two patients in the with OS, however, pO2 (p = 0.110) and AaDO2 (p = 0.299) low-risk and high-risk groups. failed to predict OS (Additional File 5). Correlation of radiomic features to the PFT/BGA indicators Dynamic changes of lymphocyte counts before and after Pearson correlation analysis (Fig.  6A) demonstrated that CCRT FEV1/FVC% had modest correlation with three pulmo- Although there was no significant difference in lym - nary features (SHAPE_Volume_mL, GLRLM_LRE and phocyte counts before CCRT (median, 1650 vs. 1650 GLRLM_RP) (all Pearson correlation >|0.45|), and mild cells/mm , p > 0.99) between the low-risk and high- correlation with CONV_SUVstd of lungs and GLZLM_ risk group (Additional File 6), patients in the low-risk GLNUz of tumor (all Pearson correlation >|0.25|). Other group had less Grade ≥ 3 lymphopenia (63.2% vs. 83.3%, PFT/BGA indicators and radiomic features were not well p = 0.031) during CCRT, and more patients in the low-risk correlated. group could recover to normal level (≥ 1000 cells/mm ) Chen et al. Radiation Oncology (2022) 17:184 Page 7 of 12 Fig. 4 Two patients from the high-risk (A) and low-risk group (B). From the series CT images, there were discernible distinction observed in tumor and pulmonary status between the two cases. High-risk patient had heterogeneous primary lung tumor and chronic obstructive pneumonia (A-1), while low- risk patient had relatively homogeneous primary lung tumor and better pulmonary condition (B-1). Significant tumor remission was achieved in low-risk patient without obvious radiation pneumonitis after CCRT (B-2), while high-risk patient had stable disease and developed Grade 2 radiation pneumonitis in bilateral lungs (A-2). The GLZLM matrices for tumor, GLRLM matrices along + x axis for lungs, and histograms of HU values for tumor (A-3, B-3) were displayed. It was found that the short homogeneous runs and the non-uniformity of the grey-levels (CT value) were emphasized for high-risk patient compared to those of low-risk patient Chen et al. Radiation Oncology (2022) 17:184 Page 8 of 12 Table 2 Selected features and their importance rank for the training and validation cohort Importance rank Training cohort Validation cohort Selected features Category IW Selected features Category IW 1 GLRLM_SRE Tumor 0.0451 SHAPE_Volume_mL Lung 0.1439 2 SHAPE_Volume_mL Lung 0.0426 GLZLM_GLNUz Tumor 0.1122 3 CONV_SUVstd Tumor 0.0177 GLRLM_RP Lung 0.0776 4 CONV_SUVstd Lung 0.0169 GLRLM_SRE Tumor 0.0643 5 GLRLM_RP Lung 0.0143 HISTO_Entropy_log10 Tumor 0.0541 6 GLZLM_GLNUz Tumor 0.0126 CONV_SUVstd Lung 0.0372 7 HISTO_Entropy_log10 Tumor 0.0121 TLG_mL Tumor 0.0224 8 GLRLM_LRE Lung 0.0036 GLRLM_LRE Lung 0.0066 9 TLG_mL Tumor 0.0006 CONV_SUVstd Tumor <0.0001 Abbreviations: IW, importance weight; GLRLM, grey-level run length matrix; SRE, short-run emphasis; RP, run percentage; GLZLM, grey-level zone length matrix; GLNUz, gray-level non-uniformity for zone; LRE, long-run emphasis; TLG, total lesion glycolysis Fig. 5 The receiver operating characteristics (ROC) curves for the training and validation cohort, with all selected features (a, b), and without pulmonary features (c, d), respectively at 4 ~ 6 weeks post CCRT (71.4% vs. 27.8%, p < 0.001). Discussions Kaplan–Meier curves demonstrated that better recov- Application of radiomics to the long-term survival pre- ery to normal level (≥ 1000 cells/mm ) at 4 ~ 6 weeks post diction for LANSCLC after CCRT is a reasonable exten- CCRT (3-year OS rate, 47.5% versus 14.3%, p = 0.001) was sion under the background of the field-wide adoption of a prognostic factor of OS (Additional File 5). machine learning methods. Other than previous works focused on the features from tumor and peritumoral tis- sue, the relationship between tumor and TOE is increas- ingly attached importance. Significant association was Chen et al. Radiation Oncology (2022) 17:184 Page 9 of 12 Fig. 6 Pearson correlation coefficient heatmap for radiomic features and PFT/BGA indicators ( A), and the distribution of PFT/BGA indicators between the low-risk and high-risk group (B-E). A. The FEV1/FVC% had modest correlation with three pulmonary features (SHAPE_Volume_mL, GLRLM_LRE and GLRLM_RP) (all Pearson correlation >|0.45|), and mild correlation with CONV_SUVstd of lungs and GLZLM_GLNUz of tumor (all Pearson correlation >|0.25|). Other PFT/BGA indicators and radiomic features were not well correlated. B-E. Patients in the low-risk group had better baseline FEV1/FVC% (median, 96.3% vs. 85.9%, p = 0.046) compared with those in the high-risk group. No significant difference of DLCO% (median, 84.3% vs. 77.5%, p = 0.136), pO2 (median, 86.0 vs. 86.0 mmHg, p > 0.999), AaDO2 (median, 18.0 vs. 18.0 mmHg, p > 0.999) and SaO2 (median, 96.8% vs. 96.5%, p = 0.634) was found. Abbreviation: PFT, pulmonary function test; BGA, blood gas analysis found between pulmonary function and radiomic fea- support this finding, implying that the TOE, herein the tures extracted from the lungs of CT images [21–23]. In pulmonary environment, might have a significant impact current study, the long-term survival forecast accuracy of in LANSCLC patients with large tumor burden and lim- LANCLC patients after CCRT was demonstrated to be ited pulmonary function. Accordingly, the relatively lon- boosted by integrating primary tumor characteristics and ger OS for patients with healthier pulmonary status could pulmonary features from pretreatment CT images. Based possibly contribute to their more tolerance to radical on the CT-based predictive model, patients could be CCRT and less incidence of severe lung toxicities. precisely stratified into the low-risk and high-risk group PFT have been reported to predict the risk of RILT before treatment, which should be considered in individ- after CCRT [26–29]. Our previous work showed that ualized treatment decision-making process. FEV1/FVC% and DLCO% were prognostic factors for From the importance rank of the selected features, long-term survival but not for PFS [17], implying that it could be confirmed that two features from tumor, long-term survival outcomes might not be achievable GLRLM_SRE and GLZLM_GLNUz which represent the due to detriment of pulmonary function even though inhomogeneity of CT images [19], remained important patients had good early response to CCRT. To further factors determining OS, which were consistent with pub- interpret the underlying role of these selected radiomic lished literatures [24, 25]. Meanwhile, the ranking of pul- lung features, the correlation between radiomic fea- monary features underlined their indispensable role in tures and PFT/BGA indicators were explored in depth the OS forecast. Our results of the significant difference and it was confirmed that FEV1/FVC% was well corre - between fitting and prediction accuracies with and with - lated with radiomic pulmonary features. This correlation out pulmonary features in model performance further between the pulmonary ventilation function and selected Chen et al. Radiation Oncology (2022) 17:184 Page 10 of 12 radiomic pulmonary features for OS prediction reaffirms What’s more, the predictive OS results using imag- the findings in Occhipinti et al.’s study that the changes ing features in our study with machine learning could in lung function, such as bronchial thickening and hon- be utilized as an effective indicator for the survival risk eycombing, can be mechanistically explained based on stratification of these patients, which could potentially morphological CT features [23]. And it might addition- individualize CCRT regimen and adjuvant treatment ally imply that the tumor not only interacts with cells in from the perspective of personalized medicine. For its immediate vicinity, but also communicates with the example, immunotherapy has evolved into a standard entire host organ [30], just as suggested by a prior study adjuvant treatment option for LANSCLC patients treated [31] that the tumor and TOE could possibly interact in a with definitive CCRT. Based on the promising results of bi-directional way. the phase III PACIFIC study [33, 34], adjuvant immu- In the aspect of methodology, the machine learning notherapy resulted in a significant prolonged PFS and framework in this study used SVM combined with the OS for those patients. To be noticed, the most common proposed IFSMT approach to iteratively select features grade 3 or 4 adverse event in the durvalumab arm was using GA and improve the accuracy of the prediction pneumonia (4.4%), followed by pneumonitis or radiation model. Our avoidance of topical deep-learning frame- pneumonitis (3.4%), and Asian patients seemed to have a works, such as deep convolutional neural network, is due higher rate of any grade pneumonitis (73.6%) and severe to the intrinsic weaknesses of overfitting and blackbox pneumonitis (5.6%) [35]. Thus, based on the survival risk for these frameworks. To ease the problem of overfitting, stratification of LANSCLC patients in this study, low-risk the deep-learning frameworks are more suitable for the patients might have several potential advantages for adju- learning tasks armed with big data as learning samples. vant immunotherapy: (1) supporting role of better pul- However, the number of patients in current study for monary function and quality of life; (2) superior tumor model training was relatively small, which intensively remission with less pulmonary toxicities; (3) less severe restricts the application of deep-learning frameworks lymphopenia during CCRT and better recovery of lym- which may have millions of parameters and thousands phopenia from CCRT. However, for high-risk LANSCLC of decision making variables. The SVM is equivalent to patients who had worse baseline FEV1/FVC%, higher rate an optimized three-layer neural network with only one of Grade ≥ 3 lymphopenia during CCRT, worse recov- hidden layer. This simplified neural network architec - ery of lymphopenia from CCRT, and higher incidence of ture substantially reduces the potential of overfitting. radiation-induced pneumonitis, radical CCRT or further Additionally, in contrast with the problem of blackbox adjuvant immunotherapy might not be feasible because for deep learning framework, the features used in mod- of poor organ functions and high probability of severe eling are explicitly created and selected with the IFSMT complications. Therefore, pretreatment radiomics-based approach. Therefore, each feature had an explicit clinical risk stratification of LANSCLC patients using features or physical meaning relevant to image of a specific ROI, from tumor and TOE could provide direct evidences to which made it easy to apprehend the behind-the-scene effectively support the treatment decision making. mechanism of the survival status prediction and directly It should also be noted that there were a few limita- related the comprehensible clinical and image oriented tions in this study. First, the absence of external valida- indices to the clinical outcome. The effectiveness of tion was the major disadvantage. Nevertheless, multiple IFSMT approach had been demonstrated by high AUC CT simulation machines were available in our institution values achieved for the survival status prediction. (Additional File 1). The high AUC values were generated The most recent work on prognostic model for the sur - from these different scanners with varied parameter set - vival outcome for NSCLC patients treated with CCRT tings, demonstrating the great robustness of our model. demonstrated that pretreatment CT texture features Besides, Zhao et al. considered that radiomic features provided prognostic information beyond CPFs [12]. in lung cancer were reproducible over a wide range of However, it didn’t provide the result in terms of AUC imaging settings [36]. Multicenter validations with larger or employ the validation cohort. In another predictive samples are warranted for the ultimate application of this model conducted by Dehing-Oberije C et al. [32], which model clinically. Second, there might be some variabil- used CPF indices only, the AUC was 0.74 for the train- ity in multiple observer delineations in our study. E et al. ing cohort, 0.75 and 0.76 for the two separate valida- reported that although the ROIs delineation tended to be tion cohorts. The improvement of model performance different between individual experts, an overall high AUC by imaging features in current study is discernible with value could still be achieved [37]. Third, we focused only the AUCs of 0.965 and 0.869 for the training and valida- on the radiomic analysis of pretreatment planning CT in tion cohort, which could be attributed to inclusion of the this study, and other imaging modalities, such as PET-CT image-based pulmonary features. [38] and MRI, still need to be investigated as to whether Chen et al. Radiation Oncology (2022) 17:184 Page 11 of 12 Authors’ contributions they could also yield complementary information which HL and CMX contributed to the conception and design of the study. NBC, MX, RZ, BQ, YFL, SZ, CC, QWL, BW, JYG and KQP contributed to acquisition of data. would facilitate more accurate predictive models. NBC, MX, RZ, YZ, BQ, YFL and HHJ analyzed and interpreted data. NBC, MX and RZ drafted the article. HL and CMX revised it critically and made the final Conclusion approval of the version to be submitted. All authors read and approved the final manuscript. Pretreatment CT-based radiomics features from tumor and TOE could improve the long-term survival fore- Funding cast accuracy in LANSCLC patients treated with CCRT This work was supported by National Natural Science Foundation of China [Grant Number 82073328], National Key R&D Program of China [Grant Number using machine learning. The predictive results could be 2018YFC0116800] and Suzhou Municipal Science and Technology Program utilized as an effective indicator for the stratification of [Grant Number SYS2018009]. The funders had no role in study design; in the these patients into the low-risk and high-risk groups. It collection, analysis and interpretation of data; in the writing of the report; and in the decision to submit the article for publication. was further confirmed that patients in the low-risk group had better baseline FEV1/FVC%, less severe lymphope- Data Availability nia during CCRT, better recovery of lymphopenia from The datasets used and analysed during the current study are available from the corresponding author on reasonable request. CCRT, lower incidence of radiation-induced pneumo- nitis, superior tumor remission and long-term survival, Declarations which might suggest more benefit for these patients from radical CCRT or further adjuvant immunotherapy. Ethics approval and consent to participate This study was conducted based on the ethical standards of the Declaration Abbreviations of Helsinki. It was reviewed and approved by the review board of Sun Yat-sen CI confidence interval University Cancer Center. Since it was an anonymous retrospective study, we CCRT concurrent chemoradiotherapy required and were granted a waiver of written informed-consent. LANSCLC locally advanced non-small cell lung cancer RILT radiation-induced lung t oxicity Consent for publication CPF clinical prognostic factor OS overall survival Not applicable. TOE tumor organismal environment PFT pulmonary function test Competing interests PFS progression-free survival The authors declare that they have no competing interests. ORR objective response rate CT computed tomography IFSMT integrated feature selection and Author details model training Department of Radiation Oncology, State Key Laboratory of Oncology MRI magnetic resonance imaging in South China, Collaborative Innovation Center for Cancer Medicine, Sun PET-CT positr on emission tomography-computed tomography Yat-sen University Cancer Center, No.651 Dongfeng Road East, 4D four-dimensional 510060 Guangzhou, China GTV gross tumor volume Department of Imaging Diagnosis and Interventional Center, State ITV internal target volume Key Laboratory of Oncology in South China, Collaborative Innovation PTV planning target volume Center for Cancer Medicine, Sun Yat-sen University Cancer Center, No.651 OAR organ at risk Dongfeng Road East, 510060 Guangzhou, Guangdong, China BGA blood gas analysis Department of Cardiac Surgery, The First Affiliated Hospital of Sun Yat- ROI regions of interest sen University, Guangzhou, Guangdong, China GLCM grey-level co-occurrence matrix Department of Pulmonary and Critical Care Medicine, The First Affiliated NGLDM neighborhood grey-level different matrix Hospital of Sun Yat-sen University, Guangzhou, Guangdong, China GLRLM grey-level run length matrix Homology Medical Technologies Inc., Ningbo, Zhejiang, China GLZLM grey-level zone length matrix Guangzhou Xinhua University, Guangzhou, Guangdong, China SVM support vector machine GA genetic algorithm Received: 7 June 2022 / Accepted: 28 September 2022 LOOCV leave-one-out cross-validation ROC receiver operating characteristics AUC area under curve IW importance weight TPR true positive rate TNR true negative rate References 1. Zhou Z, Song X, Wu A, Liu H, Wu H, Wu Q, et al. Pulmonary emphysema is a Supplementary Information risk factor for radiation pneumonitis in NSCLC patients with squamous cell The online version contains supplementary material available at https://doi. carcinoma after thoracic radiation therapy. Sci Rep. 2017;7(1):2748. org/10.1186/s13014-022-02136-w. 2. Li F, Zhou Z, Wu A, Cai Y, Wu H, Chen M, et al. Preexisting radiological inter- stitial lung abnormalities are a risk factor for severe radiation pneumonitis in patients with small-cell lung cancer after thoracic radiation therapy. Radiat Additional File 1. 4DCT image acquisition Oncol. 2018;13(1):82. Additional File 2. Genetic algorithm 3. Roach M 3rd, Gandara DR, Yuo HS, Swift PS, Kroll S, Shrieve DC, et al. Radiation pneumonitis following combined modality therapy for lung cancer: analysis Additional File 3 Patient characteristics of prognostic factors. J Clin Oncol. 1995;13(10):2606–12. 4. Bradley JD, Hope A, El Naqa I, Apte A, Lindsay PE, Bosch W, et al. A nomogram Additional File 4. Tumor response after CCRT in the low-risk and high-risk to predict radiation pneumonitis, derived from a combined analysis of RTOG group 9311 and institutional data. Int J Radiat Oncol Biol Phys. 2007;69(4):985–92. Additional File 5. Additional File 6. Chen et al. Radiation Oncology (2022) 17:184 Page 12 of 12 5. Hope AJ, Lindsay PE, El Naqa I, Alaly JR, Vicic M, Bradley JD, et al. Modeling measured by quantitative CT and CT-based radiomics in COPD. Respir Res. radiation pneumonitis risk with clinical, dosimetric, and spatial parameters. 2019;20(1):101. Int J Radiat Oncol Biol Phys. 2006;65(1):112–24. 24. Ganeshan B, Panayiotou E, Burnand K, Dizdarevic S, Miles K. Tumour hetero- 6. Valdes G, Solberg TD, Heskel M, Ungar L, Simone CB 2. Using machine learn- geneity in non-small cell lung carcinoma assessed by CT texture analysis: a ing to predict radiation pneumonitis in patients with stage I non-small cell potential marker of survival. Eur Radiol. 2012;22(4):796–802. lung cancer treated with stereotactic body radiation therapy. Phys Med Biol. 25. Bradley JD, Ieumwananonthachai N, Purdy JA, Wasserman TH, Lockett MA, 2016;61(16):6105–20. nd. . Graham MV, et al. Gross tumor volume, critical prognostic factor in patients 7. Yakar M, Etiz D, Metintas M, Ak G, Celik O. Prediction of Radiation Pneumonitis treated with three-dimensional conformal radiation therapy for non-small- With Machine Learning in Stage III Lung Cancer: A Pilot Study. Technol Cancer cell lung carcinoma. Int J Radiat Oncol Biol Phys. 2002;52(1):49–57. Res Treat. 2021;20:15330338211016373. 26. Lopez Guerra JL, Gomez D, Zhuang Y, Levy LB, Eapen G, Liu H, et al. Change 8. Bourbonne V, Da-Ano R, Jaouen V, Lucia F, Dissaux G, Bert J, et al. Radiomics in diffusing capacity after radiation as an objective measure for grading analysis of 3D dose distributions to predict toxicity of radiotherapy for lung radiation pneumonitis in patients treated for non-small-cell lung cancer. Int J cancer. Radiother Oncol. 2021;155:144–50. Radiat Oncol Biol Phys. 2012;83(5):1573–9. 9. Puttanawarut C, Sirirutbunkajorn N, Tawong N, Jiarpinitnun C, Khachonkham 27. Park YH, Kim JS. Predictors of radiation pneumonitis and pulmonary function S, Pattaranutaporn P, et al. Radiomic and Dosiomic Features for the Prediction changes after concurrent chemoradiotherapy of non-small cell lung cancer. of Radiation Pneumonitis Across Esophageal Cancer and Lung Cancer. Front Radiation Oncol J. 2013;31(1):34–40. Oncol. 2022;12:768152. 28. Torre-Bouscoulet L, Muñoz-Montaño WR, Martínez-Briseño D, Lozano-Ruiz FJ, 10. Lee G, Lee HY, Park H, Schiebler ML, van Beek EJR, Ohno Y, et al. Radiomics Fernández-Plata R, Beck-Magaña JA, et al. Abnormal pulmonary function tests and its emerging role in lung cancer research, imaging biomarkers and clini- predict the development of radiation-induced pneumonitis in advanced cal management: State of the art. Eur J Radiol. 2017;86:297–307. non-small cell lung Cancer. Respir Res. 2018;19(1):72. 11. Thawani R, McLane M, Beig N, Ghose S, Prasanna P, Velcheti V, et al. Radiomics 29. Zhou Y, Yan T, Zhou X, Cao P, Luo C, Zhou L, et al. Acute severe radiation and radiogenomics in lung cancer: A review for the clinician. Lung cancer pneumonitis among non-small cell lung cancer (NSCLC) patients with (Amsterdam Netherlands). 2018;115:34–41. moderate pulmonary dysfunction receiving definitive concurrent chemo - 12. Fried DV, Tucker SL, Zhou S, Liao Z, Mawlawi O, Ibbott G, et al. Prognostic radiotherapy: Impact of pre-treatment pulmonary function parameters. value and reproducibility of pretreatment CT texture features in stage III non- Strahlentherapie und Onkologie: Organ der Deutschen Rontgengesellschaft small cell lung cancer. Int J Radiat Oncol Biol Phys. 2014;90(4):834–42. [et al]. 2020;196(6):505–14. 13. Sun W, Jiang M, Dang J, Chang P, Yin FF. Eec ff t of machine learning methods 30. Dieterich LC, Bikfalvi A. The tumor organismal environment: Role in on predicting NSCLC overall survival time based on Radiomics analysis. tumor development and cancer immunotherapy. Sem Cancer Biol. Radiation Oncol (London England). 2018;13(1):197. 2020;65:197–206. 14. Lu MT, Ivanov A, Mayrhofer T, Hosny A, Aerts H, Hoffmann U. Deep Learning 31. Khorrami M, Prasanna P, Gupta A, Patil P, Velu PD, Thawani R, et al. Changes to Assess Long-term Mortality From Chest Radiographs. JAMA Netw open. in CT Radiomic Features Associated with Lymphocyte Distribution Predict 2019;2(7):e197416. Overall Survival and Response to Immunotherapy in Non-Small Cell Lung 15. Duijm M, van der Voort van Zyp NC, Granton PV, van de Vaart P, Mast ME, Cancer. Cancer Immunol Res. 2020;8(1):108–19. Oomen-de Hoop E, et al. Prognostic factors of local control and disease free 32. Dehing-Oberije C, Yu S, De Ruysscher D, Meersschout S, Van Beek K, Lievens survival in centrally located non-small cell lung cancer treated with stereotac- Y, et al. Development and external validation of prognostic model for 2-year tic body radiation therapy. Acta Oncol. 2020;59(7):809–17. survival of non-small-cell lung cancer patients treated with chemoradio- 16. Kang J, Ning MS, Feng H, Li H, Bahig H, Brooks ED, et al. Predicting 5-Year therapy. Int J Radiat Oncol Biol Phys. 2009;74(2):355–62. Progression and Survival Outcomes for Early Stage Non-small Cell Lung 33. Antonia SJ, Villegas A, Daniel D, Vicente D, Murakami S, Hui R, et al. Dur- Cancer Treated with Stereotactic Ablative Radiation Therapy: Development valumab after Chemoradiotherapy in Stage III Non-Small-Cell Lung Cancer. N and Validation of Robust Prognostic Nomograms. Int J Radiat Oncol Biol Phys. Engl J Med. 2017;377(20):1919–29. 2020;106(1):90–9. 34. Antonia SJ, Villegas A, Daniel D, Vicente D, Murakami S, Hui R, et al. Overall 17. Qiu B, Xiong M, Luo Y, Li Q, Chen N, Chen L, et al. Hypofractionated Intensity Survival with Durvalumab after Chemoradiotherapy in Stage III NSCLC. N Modulated Radiation Therapy With Concurrent Chemotherapy in Locally Engl J Med. 2018;379(24):2342–50. Advanced Non-Small Cell Lung Cancer: A Phase II Prospective Clinical Trial 35. Peng L, Wu YL. Immunotherapy in the Asiatic population: any differences (GASTO1011). Pract Radiat Oncol. 2021;11(5):374–83. from Caucasian population? J Thorac disease. 2018;10(Suppl 13):1482-s93. 18. Kong FM, Ritter T, Quint DJ, Senan S, Gaspar LE, Komaki RU, et al. Consider- 36. Zhao B, Tan Y, Tsai WY, Schwartz LH, Lu L. Exploring Variability in CT Char- ation of dose limits for organs at risk of thoracic radiotherapy: atlas for lung, acterization of Tumors: A Preliminary Phantom Study. Translational Oncol. proximal bronchial tree, esophagus, spinal cord, ribs, and brachial plexus. Int J 2014;7(1):88–93. Radiat Oncol Biol Phys. 2011;81(5):1442–57. 37. Lu EL, Li L, Yang L, Schwartz H, Zhao LH. B. Radiomics for Classification of 19. Nioche C, Orlhac F, Boughdad S, Reuzé S, Goya-Outi J, Robert C, et al. LIFEx: Lung Cancer Histological Subtypes Based on Nonenhanced Computed A Freeware for Radiomic Feature Calculation in Multimodality Imaging to Tomography. Acad Radiol. 2019;26(9):1245–52. Accelerate Advances in the Characterization of Tumor Heterogeneity. Cancer 38. Cook GJ, Yip C, Siddique M, Goh V, Chicklore S, Roy A, et al. Are pretreatment Res. 2018;78(16):4786–9. 18F-FDG PET tumor textural features in non-small cell lung cancer associated 20. Smola AJ, Scholkopf B. A tutorial on support vector regression. Stat Comput. with response and survival after chemoradiotherapy? Journal of nuclear 2004;14(3):199–222. medicine: official publication. Soc Nuclear Med. 2013;54(1):19–26. 21. Wu MT, Chang JM, Chiang AA, Lu JY, Hsu HK, Hsu WH, et al. Use of quantita- tive CT to predict postoperative lung function in patients with lung cancer. Publisher’s Note Radiology. 1994;191(1):257–62. Springer Nature remains neutral with regard to jurisdictional claims in 22. Lafata KJ, Zhou Z, Liu JG, Hong J, Kelsey CR, Yin FF. An Exploratory Radiomics published maps and institutional affiliations. Approach to Quantifying Pulmonary Function in CT Images. Sci Rep. 2019;9(1):11509. 23. Occhipinti M, Paoletti M, Bartholmai BJ, Rajagopalan S, Karwoski RA, Nardi C, et al. Spirometric assessment of emphysema presence and severity as

Journal

Radiation OncologySpringer Journals

Published: Nov 16, 2022

Keywords: Locally advanced non-small cell lung cancer; Radiomics; Machine learning; Long-term survival prediction; Tumor organismal environment.

There are no references for this article.