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

Learn More →

A Novel Nomogram Model Based on Cone-Beam CT Radiomics Analysis Technology for Predicting Radiation Pneumonitis in Esophageal Cancer Patients Undergoing Radiotherapy

A Novel Nomogram Model Based on Cone-Beam CT Radiomics Analysis Technology for Predicting... ORIGINAL RESEARCH published: 17 December 2020 doi: 10.3389/fonc.2020.596013 A Novel Nomogram Model Based on Cone-Beam CT Radiomics Analysis Technology for Predicting Radiation Pneumonitis in Esophageal Cancer Patients Undergoing Radiotherapy 1,2 1 3 3 3 3 Feng Du , Ning Tang , Yuzhong Cui , Wei Wang , Yingjie Zhang , Zhenxiang Li and Jianbin Li Department of Radiation Oncology, School of Clinical Medicine, Cheeloo College of Medicine, Shandong University, Jinan, 2 3 China, Department of Radiation Oncology, Zibo Municipal Hospital, Zibo, China, Department of Radiation Oncology, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Jinan, China Purpose: We quantitatively analyzed the characteristics of cone-beam computed Edited by: Tiziana Rancati, tomography (CBCT) radiomics in different periods during radiotherapy (RT) and then Istituto Nazionale dei Tumori (IRCCS), built a novel nomogram model integrating clinical features and dosimetric parameters for Italy predicting radiation pneumonitis (RP) in patients with esophageal squamous cell Reviewed by: Gilles Defraene, carcinoma (ESCC). KU Leuven, Belgium Methods: At our institute, a retrospective study was conducted on 96 ESCC patients for Irina Vergalasova, Rutgers Cancer Institute of New whom we had complete clinical feature and dosimetric parameter data. CBCT images of Jersey, United States each patient in three different periods of RT were obtained, the images were segmented *Correspondence: using both lungs as the region of interest (ROI), and 851 image features were extracted. Jianbin Li lijianbin@msn.com The least absolute shrinkage selection operator (LASSO) was applied to identify candidate radiomics features, and logistic regression analyses were applied to construct the rad- Specialty section: score. The optimal period for the rad-score, clinical features, and dosimetric parameters This article was submitted to Radiation Oncology, were selected to construct the nomogram model and then the receiver operating a section of the journal characteristic (ROC) curve was used to evaluate the prediction capacity of the model. Frontiers in Oncology Calibration curves and decision curves were used to demonstrate the discriminatory and Received: 18 August 2020 clinical benefit ratios, respectively. Accepted: 04 November 2020 Published: 17 December 2020 Results: The relative volume of total lung treated with ≥5 Gy (V5), mean lung dose (MLD), Citation: and tumor stage were independent predictors of RP and were finally incorporated into the Du F, Tang N, Cui Y, Wang W, nomogram. When the three time periods were modeled, the first period was better than Zhang Y, Li Z and Li J (2020) A Novel Nomogram Model Based on Cone- the others. In the primary cohort, the area under the ROC curve (AUC) was 0.700 (95% Beam CT Radiomics Analysis confidence interval (CI) 0.568–0.832), and in the independent validation cohort, the AUC Technology for Predicting Radiation Pneumonitis in Esophageal Cancer was 0.765 (95% CI 0.588–0.941). In the nomogram model that integrates clinical features Patients Undergoing Radiotherapy. and dosimetric parameters, the AUC in the primary cohort was 0.836 (95% CI 0.700– Front. Oncol. 10:596013. 0.918), and the AUC in the validation cohort was 0.905 (95% CI 0.799–1.000). The doi: 10.3389/fonc.2020.596013 Frontiers in Oncology | www.frontiersin.org 1 December 2020 | Volume 10 | Article 596013 Du et al. A Novel Nomogram Predicting RP for ESCC nomogram model exhibits excellent performance. Calibration curves indicate a favorable consistency between the nomogram prediction and the actual outcomes. The decision curve exhibits satisfactory clinical utility. Conclusion: The radiomics model based on early lung CBCT is a potentially valuable tool for predicting RP. V5, MLD, and tumor stage have certain predictive effects for RP. The developed nomogram model has a better prediction ability than any of the other predictors and can be used as a quantitative model to predict RP. Keywords: esophageal cancer, cone beam computed tomography, radiation pneumonitis, prediction model, radiomics capability of lung texture features, which help describe the INTRODUCTION potential RP risk (14, 15). Among malignant tumors, the incidence rate of esophageal At present, cone-beam computed tomography (CBCT) has cancer (EC) is the seventh highest, and the mortality rate is become a routine online method of image-guided radiotherapy sixth worldwide (1). Radiotherapy (RT) is still one of the main (IGRT) for EC. If we can perform quantitative analysis on CBCT treatments for locally advanced EC (2, 3). However, radiation radiomics features in a certain period of RT and then combine pneumonitis (RP) is one of the major toxicities of thoracic these radiomics features with clinical features and dosimetric radiation therapy. If RP occurs, it seriously affects the patient’s parameters to predict RP in EC, it will help guide clinical quality of life and survival prognosis (4). Therefore, it is treatment strategies in a timely manner. imperative for EC patients undergoing RT to identify this Therefore, the initial aim of this study was to investigate toxicity at the earliest possible time. More importantly, the whether the early changes in CBCT radiomics features could be accurate prediction of RP is essential to facilitate individualized used as potential markers for predicting RP. In the present study, radiation dosing that leads to maximized therapeutic gain. At a comprehensive nomogram, which is a conveniently applicable present, the risk assessment of RP is mainly predicted by using predictive model integrating CBCT radiomics features, clinical lung dosimetric parameters (5, 6), such as the relative volume of features, and dosimetric parameters, was built for the total lung irradiated above a specified threshold dose (V )or individualized risk assessment and precise prediction of RP. mean lung dose (MLD): Although several metrics have appeared promising, the results vary across institutions, so these metrics do not seem to be perfect at predicting RP (7, 8). In addition to MATERIALS AND METHODS dosimetric parameters, some clinical features (tumor stage, smoking history, preexisting lung diseases, concurrent Patients chemotherapy, and radiation dose) are also considered to be The entire cohort of this retrospective study was obtained from the related to RP occurrence. However, the consensus on the records of our institutional picture archiving and communication comparative importance of these related predictors remains system (PACS) from January 2017 to June 2019, which was used unavailable at present. Consequently, in order to individually to identify esophageal squamous cell carcinoma (ESCC) patients and precisely discern RP, an accurate predictive model receiving RT treatment. The inclusion criteria were as follows: (1) incorporating multiple types of factors with superior clinical Karnofsky performance score (KPS) ≥70, (2) no previous history utility is urgently needed. of thoracic RT, (3) intensity-modulated radiotherapy (IMRT) and Computed tomography (CT) images play an essential role in received ≥50 Gy RT, and (4) CBCT scan performed at least once a the diagnosis and treatment of RP. As early as the end of the 20th week during RT with the scanning range of the CBCT imaging century, RP could be identified by CT. However, RP cannot be including at least two thirds of the lungs. The exclusion criteria predicted by superficial CT manifestations. Therefore, the focus were as follows: (1) low image quality, (2) general pulmonary of later research is on the accurate prediction of RP (9). In recent infection unrelated to RT, and (3) treatment break of more than 7 years, with the rapid development of radiomics analysis days during RT. A total of 96 consecutive patients with thoracic technology, increasing attention has been paid to the research middle segment ESCC were identified and divided into two of RT effect and side effect predictions based on radiomics cohorts at a 7:3 proportion using computer-generated random features (10–13). Among them, one study found that there is a numbers. Sixty-seven patients were allocated to the primary dose-dependent relationship between the changes in some cohort, and 29 patients were allocated to the verification cohort. radiomics features and RP ≥2 grade determined by extracting Our institutional research ethics board approved this retrospective local lung CT images after RT (12). Another study successfully study (SDTHEC201703014). It waived the need to obtain established a differential model of high- and low-risk RP by informed consent from the patients due to the retrospective analyzing the region of interest (ROI) of the whole lung tissue nature of the investigation (retrospective single-institution before RT (13). In short, radiomics features can capture the cohort study). Frontiers in Oncology | www.frontiersin.org 2 December 2020 | Volume 10 | Article 596013 Du et al. A Novel Nomogram Predicting RP for ESCC rotating the frame at an angle. This is a slow CBCT acquisition Clinical Data and RT Parameters setting. The acquisition time is 67 s, and the patient keeps breathing The clinical data were all acquired from the institute’s medical evenly during this process. Standard body scan conditions were records. Specifically, clinicopathological parameters included voltage (125 kVp), current (80 mA), exposure time (13 ms), age, sex, KPS, smoking status, diabetes history, chronic exposure (680 mAs), rotation angle (178°–182°), pixel matrix size obstructive lung disease (COPD), pathological diagnosis, tumor (384×384), field of view (FOV, 45×18 cm), slice thickness (2.5 mm), location, TNM stage, radiation dose, and concurrent and fan-beam type (half-fan). Among fan-beam types, the half-fan chemoradiotherapy lack thereof. In addition, the lung mode was used for the image acquisition of lung tissue structures dosimetric parameters involved in this study included V5–V40 larger than 24 cm. In this study, lung CBCT image acquisition was (relative volume of total lung treated with ≥5–40 Gy) and MLD. carried out in three different periods, and then the images were In short, the parameters mentioned above were used to establish imported into 3D Slicer (version 4.10.2; http://www.slicer.org) in a a comprehensive nomogram after univariate analysis or least DICOM format to extract and analyze the radiomics features. It absolute shrinkage selection operator (LASSO) feature selection. should be noted that these three different periods were artificially The Eclipse Treatment Planning System (Varian Medical divided according to the experimental design and corresponded to Systems, Palo Alto, CA, Version 13.5.35) was adopted for RT the early stages: the third, fourth, and fifth weeks of RT (PTV planning design. IMRT adopts a fixed-field, static intensity prescription dose range of EC: 18–22 Gy, 27–32 Gy, and 36–44 Gy). modulation technique, and 5–7 fields of coplanar irradiation are uniformly divided according to the specific situation in each Image Segmentation and Feature case. The required target parameters are then set, and the dose Extraction distribution is obtained by inverse calculation of the treatment planning system. The dose distribution is then graded Images from both lungs were segmented by a semiautomatic segmentation method (16, 17) based on a threshold-based (stratified), and each field is decomposed into a series of subfields. IMRT does not include sIMRT or volumetric algorithm. The specific steps are as follows: First, the background was removed to obtain the internal region of the chest. Second, the intensity-modulated arc therapy (VMAT). The target area includes tumor volume (GTV), including CT imaging of visible appropriate threshold was found to segment the lung and the tissues outside the lung contour to the greatest extent. Finally, esophageal tumors and positive lymph nodes. The clinical target volume (CTV) refers to the upper and lower expansion of the the manual segmentation method (18) was used to erase the extra parts outside the large trachea and lung parenchyma to obtain both esophageal tumor by 3 cm and 6 mm around the tumor and related lymphatic drainage area. The planned target volume lungs as the ROI. Image segmentation was performed by an experienced radiologist and then verified by a senior radiologist. (PTV) is formed by CTV extending 8 mm outward. IMRT was administered by a Varian Linac Accelerator with a 6-MVX ray All features were extracted by using the radiomics plug-in in 3D Slicer. A total of 851 radiomics features were extracted, and 95% PTV, and radiation doses of 50–66 Gy (median dose of 60 Gy) and 1.8–2.0 Gy/fraction 5 times/week were prescribed. including 13 morphological features, 18 histogram features, 74 original texture features, and 746 high-order features (wavelet Normal tissue constraints were prioritized in the following order for treatment planning purposes: maximum spinal cord transform features). dose of 45 Gy, relative volume of total lung treated with ≥5Gy Radiomics Feature Selection and (V5) ≤60%, relative volume of total lung treated with ≥20 Gy Radiomics Signature Construction (V20) ≤28%, MLD ≤20 Gy, relative volume of the heart treated First, the extracted radiomics features were preprocessed. Based on with ≥30 Gy (V30) ≤40%, and relative volume of the heart the Spearman rank correlation test, the features with correlations treated with ≥40 Gy (V40) ≤30%. greater than 0.9 and multicollinearity were deleted, and independent features were preliminarily screened. Meanwhile, Follow-up and Evaluation of RP based on the Mann–Whitney U test, the characteristics with Follow-up items included chest CT, physical examination, and significant differences between the RP (≥2 grade) and non-RP clinical symptoms. Patients were evaluated weekly during RT, (<2 grade) groups were screened out. Finally, the LASSO method followed up at 1 month after completion of the initial treatment, (19) was used to select the final features, and the RP prediction and then followed up every 2–3 months until at least 6 months model of rad-score was constructed based on logistic regression after the end of RT. The grading of RP was confirmed by two analysis. The LASSO method minimizes the sum of squared senior oncologists and one radiologist. The National Cancer residuals by using the case in which the sum of the absolute Institute Common Terminology Criteria for Adverse Events 4.03 values of the coefficients is less than the tuning parameter (l). (CTCAE 4.03) was used to evaluate the degree of RP. In the To prevent overfitting of the model, (20–22)duringmodel present study, grade ≥2 was used as the cutoff for diagnosing RP. building, features are selected by constantly adjusting l. With CBCT Scanning Method and Image the increasing penalty, more regression coefficients are reduced Acquisition to zero, (23, 24) and then the remaining nonzero coefficient Using the on-board imager (OBI) system mounted on the Varian is selected. The nonzero coefficient of the selected features is Trilogy medical linear accelerator, the hardware portion included a the rad-score. Each patient’s rad-score is calculated as a linear diagnostic (kV) level X-ray source (KVS) and an amorphous silicon combination of selected features that have their own flat-panel kV detector (KVD). The CBCT image was obtained by coefficient weighting. Frontiers in Oncology | www.frontiersin.org 3 December 2020 | Volume 10 | Article 596013 Du et al. A Novel Nomogram Predicting RP for ESCC In this study, 50 iterations of 10-fold nested cross-validation Construction and Validation of the were utilized, similarly to Xu et al. (25). Random sampling Nomogram was conducted in an attempt to balance the class distributions First, the prediction efficiency of the three different periods was within the cross-validation partitions. The cross-validation compared, and then the best period was selected. Second, 96 loop provides a profile of model performance. It serves to patients were divided into the RP (39 cases) and non-RP (57 estimate how well the LASSO applied to a given set of candidate cases) groups, and 16 clinical features and dosimetric predictors may generalize to other data sets. Model performance parameters were collected. The best clinical features and was assessed by computing the area under the curve (AUC) for dosimetric parameters were determined by LASSO feature each constructed model on a test partition. The inner cross- selection. Finally, a comprehensive nomogram was validation loop was applied to determine the optimal value for l established. The receiver operating characteristic (ROC) curve such that the resulting model was guarded against overfitting. The was used to evaluate the prediction capacity of the model. The value of l for each cross-validation partition was selected by calibration curve was used to determine whether the predicted determining the value that produced the most regularized model and observed probabilities for RP were in concordance. The such that the AUC was within one standard error of the maximum decision curve was performed to evaluate the clinical benefit (26). The use of 50 resampled iterations with 10-fold nested cross- ratio of the nomogram. validation constructs 500 models used to generate a distribution of This research process can be divided into four parts: image AUC values to estimate how well model construction with LASSO acquisition, ROI segmentation, feature extraction, and radiomics generalizes to other data sets. model construction as shown in Figure 1. FIGURE 1 | Flow chart of radiomics. Frontiers in Oncology | www.frontiersin.org 4 December 2020 | Volume 10 | Article 596013 Du et al. A Novel Nomogram Predicting RP for ESCC TABLE 1 | Univariate analysis of baseline clinical features of patients and RP. Statistical Analysis All statistical analyses were based on SPSS 20.0 (IBM, Armonk, NY, Factor N RP c value P value USA) or R software (R Foundation for Statistical Computing, 2 <2 grade ≥2 grade Vienna, Austria, https://www.R-project.org/). The c test or Fisher exact probability test was used to classify data between the two Sex 96 57 39 2.767 0.096 Male 81 51 30 groups. Two independent-sample t tests were used for counting Female 15 6 9 data (continuous data). The Mann–Whitney U test was used to Age (years) 96 1.619 0.203 compare the differences in clinical features between the primary <60 21 15 6 and validation cohorts. The model was evaluated with respect to ≥60 75 42 33 Stage 2.650 0.008 sensitivity, specificity, ROC curve, and 95% confidence interval II 19 15 4 (CI). P values ≤ 0.05 were considered statistically significant. III 48 30 18 IV 29 12 17 Smoking history 96 0.198 0.656 No 54 31 23 RESULTS Yes 42 26 16 COPD 96 1.436 0.231 Analysis of Clinical Features and No 81 46 35 Yes 15 11 4 Dosimetric Parameters Associated With Diabetes 96 0.318 0.573 RP No 88 53 35 Yes 8 4 4 The 96 patients were divided into RP (39 cases) and non-RP (57 Hypertension 96 0.606 0.436 cases) groups, and 9 clinical features and 7 dosimetric parameters No 83 48 35 that might be related to the occurrence of RP were included. Yes 13 9 4 Univariate analysis showed that tumor stage was correlated with Concurrent 96 Chemotherapy ≥2 grade RP (c = 2.650, P = 0.008), and other factors, including No 71 41 30 0.300 0.584 age, sex, concurrent chemoradiotherapy or lack thereof, COPD Yes 25 16 9 status, smoking status, and RT dose, showed no significant Delivered 96 1.867 0.172 differences between the two groups (all Ps > 0.05). V5, V10, Dose (Gy) <60 45 30 15 V15, V20, V30, and MLD of both lungs were associated with the ≥60 51 27 24 occurrence of grade ≥2 RP (all Ps < 0.05). The characteristics of COPD, chronic obstructive lung disease. the enrolled population are listed in Tables 1 and 2. There were no significant differences in age, sex, tumor stage, V5, and MLD between the primary group and the validation TABLE 2 | Single factor analysis of DVH and RP. group, which indicates that the groupings were reasonable (all Ps Lung DVH RP P value c value > 0.05) as shown in Table 3. Seven factors (tumor stage, V5, V10, V15, V20, V30, and MLD) remained after univariate analysis. <2 grade ≥2 grade The LASSO feature selection method was used to screen these V5 48.95 ± 10.56 59.39 ± 10.00 0.00 -4.91 seven factors, and three potential factors (V5, MLD, and tumor V10 33.64 ± 7.70 40.92 ± 7.95 0.00 -4.46 stage) remained as shown in Figures 2A, B. The AUC values of V15 25.34 ± 6.52 30.77 ± 6.96 0.00 -3.85 prediction efficiency for V5, MLD, and tumor stage were 0.698, V20 18.81 ± 5.47 22.47 ± 4.82 0.00 -3.47 0.685, and 0.662, respectively. To observe the overall predictive V30 9.61 ± 4.40 12.16 ± 5.00 0.01 -2.58 performance of V5, MLD, and tumor stage, we established a full V40 3.80 ± 2.49 4.58 ± 3.24 0.21 -1.25 MLD (cGy) 1016.47 ± 218.82 1260.87 ± 267.38 0.00 -4.72 clinical–dosimetric feature combined model. The AUC value of the combined model was 0.764 as shown in Figure 2C. MLD, mean lung dose; V5, V10, V15, V20, V30, V40 = relative volume of total lung treated with ≥5, 10, 15, 20, 30, and 40 Gy. Radiomics Feature Extraction/Selection at TABLE 3 | Comparison of sex, age, tumor stage, V5, and MLD between the Different Periods and Radiomics Signature primary and the verification cohort. Building In the first period (PTV dose: 18–22 Gy), a total of 851 radiomics Factor Primary cohort Verification cohort c value P value features were extracted from the patients. First, correlations greater Age (years) 65.33 ± 9.37 68.62 ± 8.89 -1.64 0.11 than 0.9 features were deleted, resulting in a total of 220 features Sex (N) 67 29 0.11 0.75 remaining. Second, linear features were removed, and 96 features Male 56 25 remained. Then, 21 features remained after using the rank-sum Female 11 4 Stage 3.54 0.17 test. Finally, the remaining two features after LASSO selection were II 13 6 used to build the radiomics model as shown in Figures 3A, B.The III 30 18 two features are original first-order skewness and original GLSZM- IV 24 5 small area emphasis. The model was built as follows: Rad-score = V5 52.35 ± 11.27 55.14 ± 12.01 -1.07 0.29 -0.924 e+00×Skewness - 7.047 e+00×Small Area Emphasis + MLD (Gy) 11.06 ± 2.61 11.38 ± 2.85 -0.52 0.61 Frontiers in Oncology | www.frontiersin.org 5 December 2020 | Volume 10 | Article 596013 Du et al. A Novel Nomogram Predicting RP for ESCC AB C FIGURE 2 | LASSO characteristic selection of clinical features and dosimetric parameters (A, B). ROC curve of V5, MLD, tumor stage, and combined model (C). A B 21 19 9 2 -8 -6 -4 -2 Log Lambda FIGURE 3 | Feature screening of radiomics in the first period. By adjusting the different penalty parameter (l) to obtain a high-performance model, the radiomics characteristics with the highest predictive performance were obtained. Radiomics feature convergence diagram (A). Each curve represents the trajectory of the coefficient of each independent variable (B). 4.5329. Rad-scores for each patient in the primary cohort and after using the rank-sum test. Finally, the remaining six features validation cohort are shown in Figures 4A, B. (gray-level nonuniformity, small dependence low gray-level In the second period (PTV dose: 27–32 Gy), a total of 851 emphasis, cluster shape, uniformity, entropy, and size zone radiomics features were extracted from the patients. First, nonuniformity) after LASSO selection were used to build the correlations greater than 0.9 features were deleted, resulting radiomics model. The model was built as follows: Rad-score = in a total of 222 features remaining. Second, linear features were +4.680 e-07×gray-level nonuniformity + 1.087 e+01×small removed, and 96 features remained. Then, 10 features remained dependence low gray-level emphasis - 7.913 e-04×cluster shape after using the rank-sum test. Finally, the remaining five + 1.401 e+00×uniformity + 1.406 e+00×entropy - 2.207 e-05×size features (voxel volume, smallest axis length, small zone nonuniformity - 4.776 e+00. dependence low gray-level emphasis, large area low gray-level Validation of Radiomics Signature at emphasis, and busyness) after LASSO selection were used to build the radiomics model. The model was built as follows: Different Periods Rad-score = -1.996 e-07×voxel volume - 4.036 e-03×smallest In the first period, the predictive efficacy of the model was as axis length + 5.376 e+01×small dependence low gray-level follows: In the primary cohort, the AUC was 0.700 (95% CI emphasis + 1.718 e-07×large area low gray-level emphasis - 0.568–0.832), the sensitivity was 61.5%, and the specificity was 2.473 e-04×busyness + 1.041 e+00. 75.0%. In the validation cohort, the AUC was 0.765 (95% CI In the third period (PTV dose: 36–44 Gy), a total of 851 0.588–0.941), the sensitivity was 84.6%, and the specificity was radiomics features were extracted from the patients. First, 64.7% as shown in Table 4 and Figures 5A, B. correlations greater than 0.9 features were deleted, resulting in In the second period, the predictive efficacy of the model was a total of 220 features remaining. Second, linear features were as follows: In the primary cohort, the AUC was 0.663 (95% CI removed, and 96 features remained. Then, 43 features remained 0.530–0.797), the sensitivity was 90.6%, and the specificity was Frontiers in Oncology | www.frontiersin.org 6 December 2020 | Volume 10 | Article 596013 Coefficients 0 1000 2500 Du et al. A Novel Nomogram Predicting RP for ESCC FIGURE 4 | Rad-score for each patient in the primary and validation cohorts. Green bars show scores for patients without RP, and orange bars show scores for those with RP (A, B). TABLE 4 | ROC curve parameters of the radiomics model and nomogram. Classification Primary cohort Validation cohort AUC 95% CI Sensitivity Specificity AUC 95% CI Sensitivity Specificity First period 0.700 0.568-0.832 61.5% 75.0% 0.765 0.588-0.941 84.6% 64.7% Second period 0.663 0.530-0.797 90.6% 42.9% 0.604 0.356-0.851 85.7% 50.0% Third period 0.699 0.573-0.826 66.7% 70.3% 0.756 0.561-0.950 66.7% 80.0% Nomogram 0.836 0.700-0.918 96.0% 54.8% 0.905 0.799-1.000 92.9% 73.3% A B FIGURE 5 | ROC curve of radiomics features in the first period of RT (A, B). Frontiers in Oncology | www.frontiersin.org 7 December 2020 | Volume 10 | Article 596013 Du et al. A Novel Nomogram Predicting RP for ESCC 42.9%. In the validation cohort, the AUC was 0.604 (95% CI DISCUSSION 0.356 –0.851), the sensitivity was 85.7%, and the specificity 50.0%. A single index based on lung dosimetric parameters is not the “gold standard” to judge the occurrence of PR risk; however, In the third period, the predictive efficacy of the model was as follows: In the primary cohort, the AUC was 0.699 (95% radiomics can extract image data to characterize the standard tissue structure, including typical lung structures. They may CI 0.573–0.826), the sensitivity was 66.7%, and the specificity was 70.3%. In the validation cohort, the AUC was 0.756 (95% CI produce clinically relevant improvements in predicting treatment-related toxicities (13). This makes up for the 0.561–0.950), the sensitivity was 66.7%, and the specificity was 80.0% as shown in Table 4. deficiency of dose-volume parameter prediction to a great extent. Some previous studies, (12, 13) respectively, report the By comparing the prediction efficiency of the AUC in three periods, it is obvious that the prediction efficiency in the first relationship between the changes in some second- or higher- order eigenvalues of lung cancer after and before RT and the period is better than those in the second and third periods in both the primary and validation cohorts. To reflect the occurrence of RP. Unfortunately, due to the limitations of detection techniques or other factors, it is not possible to importance of the early prediction of RP in clinical practice, the first-period rad-score and three essential features (V5, MLD, establish predictive models for clinical practice. In this study, we used an automated computer extraction algorithm and digital and tumor stage) were used to establish a comprehensive nomogram model. quantitative analysis technology to obtain high-quality information to comprehensively evaluate various characteristics The Incremental Value of the Radiomics of tumor and normal tissue responses (14, 27). More Signature When Added to the importantly, we constructed a comprehensive nomogram Comprehensive Nomogram model based on CBCT radiomics features in combination with The AUC values of dosimetric parameters (V5, MLD) and clinical features and dosimetric parameters to accurately predict clinical features (tumor stage) were 0.698, 0.685, and 0.662, RP in EC patients treated with RT. To the best of our knowledge, respectively. The AUC values of the full clinical–dosimetric this is the first study of the early prediction of RP by using IGRT feature combined model was 0.764. In addition, the AUC to obtain CBCT imaging information in different periods of RT. values of the radiomics signature at three different periods Importantly, this comprehensive nomogram model is superior to were 0.700, 0.663, and 0.699, respectively (primary cohort). It single clinical features and lung dosimetric parameters in can be seen that the single clinical features, dosimetric RP prediction. We selected CBCT images from three different periods and parameters, or full clinical–dosimetric combined model are not ideal in predicting the risk of RP. To this end, we created a extracted the radiomics features. The primary purpose was to find the first radiomics features that can independently predict comprehensive nomogram that integrates dosimetric parameters and clinical features with the radiomics signature from the first RP; however, after selecting the radiomics features in different periods, it is found that each period has its own independent set period. The results show that, in the primary cohort, the AUC of our nomogram was 0.836 (95% CI: 0.700–0.918), and in the of feature parameters related to RP. We believe that, in addition to the influence of radiation dose factors, whether these validation cohort, the AUC was 0.905 (95% CI: 0.799–1.000) as shown in Table 4 and Figures 6B, C. There is no doubt that the characteristics vary with changes in the RT process is still uncertain. It is gratifying that we found the best prediction of comprehensive nomogram, incorporating radiomics features, significantly improves the ability of conventional dosimetric RP to be in the first period of radiomics characteristics. Two important features can be found in the early stage of low-dose RT parameters and clinical features to predict the risk of RP. The graphical form of the nomogram is shown in Figure 6A. More of lung tissue: Although this may differ from our initial expectation of the experimental results, the results are importantly, the calibration curve is produced as shown in Figure 6D. The diagonal dotted line represents an ideal fascinating. This result is similar to the findings of Cunliffe et al. (12) and Jenkins et al (28). They found that AUC values evaluation, and the other two lines next to it represent the performance of the nomogram. A closer fit to the diagonal in low- and medium-dose areas of the lung were different between RP and non-RP patients even though these AUC dotted line indicates a better evaluation. In summary, this calibration curve shows favorable consistency between the values appeared in areas with lower visible changes. These first radiomics features may be able to be used to explain or screen out predicted RP and the actual observation. those susceptible to RP due to intrinsic genetic mutations. How to Make Clinical Decisions In regard to the susceptible population of RP, we must devote The clinical decision curve analysis of the nomogram is shown in attention to the sensitivity of lung tissue to RT. At present, the Figure 6E, which shows the patient’s benefits when the physician radiosensitivity of lung tissue has been reported (29, 30), and it is makes the judgment. It shows that, if the probability of the considered to be a potential influencing factor for RP occurrence. domain value is 10%, the benefit of using the nomogram to This difference in the sensitivity of lung tissue to radiation predict the efficacy of RP is higher than that of radiomics features constitutes our different understanding of the probability of or other parameters alone. In short, this decision curve exhibits RP. In two groups of patients with different radiosensitivity of satisfactory positive net benefits of the nomogram at the lung tissue, we cannot judge the probability of RP by standard threshold probabilities. clinical features and lung dosimetric parameters. However, Frontiers in Oncology | www.frontiersin.org 8 December 2020 | Volume 10 | Article 596013 Du et al. A Novel Nomogram Predicting RP for ESCC BC DE Decision curve FIGURE 6 | (A) The comprehensive nomogram incorporates V5, MLD, tumor stage, and rad-score (the first period) to predict the risk of RP in EC patients. V5: relative volume of total lung treated with ≥5 Gy; MLD: mean lung dose. (B, C) ROC curves of the comprehensive nomogram in the primary and validation cohorts. (D) Calibration curves of the comprehensive nomogram with the addition of V5, MLD, tumor stage, and radiomics features. The diagonal dotted line represents an ideal evaluation, and the other two lines next to it represent the performance of the nomogram. A closer fit to the diagonal dotted line indicates a better evaluation. (E) Decision curves of the radiomics features model and the combination model (comprehensive nomogram) predicting RP. The y-axis represents the net benefit. The red curve represents the comprehensive nomogram, and the green line represents the radiomics features model. The horizontal black line indicates that the assumption is valid. The oblique gray line indicates that the assumption is invalid. Frontiers in Oncology | www.frontiersin.org 9 December 2020 | Volume 10 | Article 596013 Du et al. A Novel Nomogram Predicting RP for ESCC radiomics can analyze the data by extracting features from CT reports not being sufficient to provide specific and safe standard images of the lung, thus providing a powerful method for the doses (34). Chargari et al. (35) find that V5 is a risk factor for noninvasive description of lung tissue radiosensitivity. This may acute or chronic lung toxicity. Cho et al. (6) find that MLD is the be why the radiomics features are superior to the clinical features most related factor that predicts RP rather than V5, V10, or V20. and dosimetric parameters in current studies. In this study, this Some clinical features have emerged as important risk factors advantage in AUC value, sensitivity, and specificity performance contributing to RP progression. Some studies show that is not particularly good, but through our research analysis, smoking is related to the severity of RP (36, 37). Takeda et al. radiomics features of RP risk prediction cannot be ignored. (38) and Kimura et al. (39) report that COPD is a significant risk The successful establishment of the prediction model is based factor for RP in patients with EC after RT. In this study, we find on the standardization of data collection and the rationalization that smoking status, COPD, and concurrent chemoradiotherapy of data processing. First, we should consider that the feature are not correlated with the incidence of RP, and so these factors extraction data are affected by CT parameters (31) because the are not included in our combined model, but this does not mean CT features may be different under different image-acquisition that they are not important. After LASSO logistic regression conditions. In this study, based on the CBCT of the Varian analysis, several significant variables, including V5, MLD, and accelerators in our center, these devices have the same tube tumor stage, were integrated into the nomogram to predict PR. voltage, tube current, exposure time, exposure amount, and pixel The results were as follows: clinical-dose characteristic model matrix size, which can help control for the differences between (AUC values: V5 = 0.698, MLD = 0.685, tumor stage = 0.662), the scanners and acquisition parameters. Second, to develop the radiomics model (primary cohort AUC 0.700, validation cohort radiomics signature, a total of 851 candidate features were AUC 0.765), and nomogram (primary cohort AUC 0.836, reduced to a set of only a few potential descriptors by using validation cohort AUC 0.905). The nomogram demonstrates a the LASSO logistic regression model to realize feature selection better ability to predict RP than the other models. by constantly adjusting the regularization parameter l to make How to use this information in the treatment plan or the weight coefficient of the feature approach zero. The LASSO alternative program to help clinicians is our greatest concern. (20) logistic regression model is suitable for analyzing large sets Fortunately, the goal of radiomics is to develop a decision- of radiomics features with a relatively small sample size, and it is making tool that meets the needs of clinicians. This is because such a tool could combine radiomics features with other patient designed to avoid overfitting high-dimensional data (21, 32). At the same time, the LASSO logistic regression model allows the characteristics to improve the capability of the decision support radiomics signature to be constructed by combining the selected model (15, 40). We show that radiomics features complement features, so it allows the model to more easily identify the most clinical features and lung dosimetric parameters, helping to closely related features in patients with RP. Finally, the nested provide better predictive ability for RP. The clinical decision cross-validation method (25) was used for internal validation to curve of this nomogram shows that the effectiveness of the improve the accuracy of the model. nomogram in predicting RP is higher than that of using It should be noted that the difference in the irradiation mode radiomic characteristics or other parameters alone. In short, (3-D conformal radiation therapy and IMRT) affects the under the threshold probability, the decision curve exhibits a potential dose distribution of the lung, which may affect the satisfactory positive net benefit of the nomogram. selection of clinical features and dosimetric parameters as risk Our results demonstrate the potential value of radiomics characteristics of RP. This can be quickly confirmed by techniques in the risk prediction of RP patients. If more comparing Tucker et al. (33) and Shane et al. (13) where, in clinical variables are included in the nomogram, there will be the former, 75% received 3-D conformal radiation therapy, and more room for future development of this model, and the the latter 83% received IMRT. Therefore, it seems complicated to resulting prediction effect will be better. A recent study (41)by establish a general model with good discriminant performance another of our teams found that subjective global assessment under different technical conditions. score (SGA), pulmonary fibrosis score (PFS), planning target The clinical factors (age, tumor stage, KPS score, chronic volume/total lung volume (PTV/LV), MLD, and ratio of change lung disease, diabetes, chemotherapy lack thereof) and lung regarding systemic immune inflammation index at 4 weeks (4w dosimetric parameters (V5, V10, V20, MLD) related to RP are SII) were potential valuable markers in predicting severe acute reported in previous studies. To provide better help for the radiation pneumonitis (SARP). Subsequently, the team oncologist, we designed a clinical nomogram to combine the developed a nomogram and corresponding risk classification above available RP risk factors with radiomics features. system with superior prediction ability for SARP. In the next Therefore, we aim to establish a combined model, maximizing step, we will consider combining the research results of this team clinical utility and accuracy of prediction ability, and so the with radiomics to establish a new RP prediction model for better initial experimental design was not expected to rely solely on the clinical application. radiomics model as the final prediction model. Of course, Although our study has many strengths, several limitations judging from the AUC value, sensitivity, and specificity of the should be addressed here. First, the sample size is small, which radiomics model in each period of RT, these characteristics can lead to the inability to apply nonlinear classifiers, such as alone are not perfect in predicting RP. Dose-volume histogram neural networks (42, 43). Because a nonlinear classifier uses a (DVH) metrics have been extensively observed and reported to more extensive data set, it is beneficial to improve the accuracy of be correlated with RP despite the current data and research the RP model. Second, our analysis does not account for two-way Frontiers in Oncology | www.frontiersin.org 10 December 2020 | Volume 10 | Article 596013 Du et al. A Novel Nomogram Predicting RP for ESCC or higher-order interactions between features. If interactions diagnosis and treatment of diseases and the prediction between features had been identified, the interaction terms that of complications. were most strongly associated with the outcome interactions would have been selected when we constructed the radiomics signature, and this could have improved performance. However, CONCLUSIONS uncovering the interactions of multiple attributes is a challenging problem. Third, we used a validation cohort that was drawn from CT radiomics has powerful data-processing and analysis abilities. the same institution as the primary cohort, which prevented us In this context, we explored a method to predict RP based on the from investigating the generalizability of the results to other lung CBCT radiomics features for EC patients. More institutions and settings. In addition, there is a lack of sufficient importantly, we used this method to successfully build and external data validation. Fourth, selection bias occurred when validate a novel nomogram with good predictive value, which strict criteria were used, and this may affect the model training. can help clinicians identify high-risk RP patients early and guide For instance, all patients are middle thoracic EC patients, which personalized treatment and clinical decisions. limits the application of this method to patients with cervical, upper, and lower thoracic segment EC radiotherapy. Also, all patients experienced uniform CBCT imaging scanners and DATA AVAILABILITY STATEMENT parameters, which does not guarantee the reproducibility and stability of radiomics features under other conditions. In the The raw data supporting the conclusions of this article will be future, we should conduct a prospective, multicenter, large- made available by the authors, without undue reservation. cohort study to further develop and validate nomograms in terms of predicting RP. As a future study, we will add different types of patients, including those with different EC locations (cervical, upper AUTHOR CONTRIBUTIONS thoracic, lower thoracic segments) and different RT techniques FD and NT are responsible for analyzing data and writing (3DCRT, TOMO, VMAT). We will also include more laboratory papers. JL designed experiments to guide the writing and indicators that may reflect RP, such as inflammatory indexes and revision of papers. WW directed the writing and revision of immune inflammatory indexes. In terms of basic research, we papers. YC, YZ, and ZL were responsible for radiomics diagnosis should also improve the model of radiomics, especially the and radiomics data processing of patients. All authors combination of radiomics and genomics. The former focuses contributed to the article and approved the submitted version. on medical imaging of the normal tissues or tumors and performs diagnosis and prognosis based on quantitative imaging features, and the latter discovers and notes the gene sequences to study the function and structure of genomes of the FUNDING diseases. Besides this, if we can combine available radiation metabolomics (44) with functional CT (45, 46)radiomics Funding was obtained from the National Key Research Program features, it may help us understand the differences in radiation of China (No. 2016YFC0904700), the National Natural sensitivity and tissue cell metabolism in order to establish a more Science Foundation of China (No. 81773287), and The Key robust prediction model. Therefore, it can be predicted that the Research Development Program of Shandong Province combination of multiple omics will be the best plan for the future (No. 2016GSF201093). 6. Cho WK, Oh D, Kim HK, Ahn YC, Noh JM, Shim YM, et al. Dosimetric REFERENCES predictors for postoperative pulmonary complications in esophageal cancer 1. Bray F, Ferlay J, Soerjomataram I, Siegel RL, Torre LA, Jemal A. Global cancer following neoadjuvant chemoradiotherapy and surgery. Radiother Oncol statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide (2019) 133:87–92. doi: 10.1016/j.radonc.2019.01.005 for 36 cancers in 185 countries. CA Cancer J Clin (2018) 68(6):394–424. 7. Rodrigues G, Lock M, D’Souza D, Yu E, Dyk JV. Prediction of radiation doi: 10.3322/caac.21492 pneumonitis by dose-volume histogram parameters in lung cancer-A systematic 2. Ajani JA, D'Amico TA, Bentrem DJ, Chao J, Corvera C, Das P, et al. review. Radiother Oncol (2004) 71(2):127–38. doi: 10.1016/j.radonc.2004.02.015 Esophageal and Esophagogastric Junction Cancers, Version 2.2019,NCCN 8. SenanS,CardenalF,VansteenkisteJ,Stigt J, AkyolF, Neve WD. A Clinical Practice Guidelines in Oncology. J Natl Compr Canc Netw (2019) 17 randomized phase II study comparing induction or consolidation (7)855–83. doi: 10.6004/jnccn.2019.0033 chemotherapy with cisplatindocetaxel,plus radical concurrent 3. Deng W, Lin SH. Advances in radiotherapy for esophageal cancer. Ann Transl chemoradiotherapy with cisplatin-docetaxel,in patients with unresectable Med (2018) 6(4):79. doi: 10.21037/atm.2017.11.28 locally advanced non-small-cell lung cancer. Ann Oncol (2011) 22:553–8. 4. Tonison JJ, Fischer SG, Viehrig M, Welz S, Boeke S, Zwirner K, et al. doi: 10.1093/annonc/mdq388 Radiation pneumonitis after intensity-modulated radiotherapy for 9. Graham M V, Purdy JA, Emami B, Harms W, Bosch W, Lockett MA. Clinical esophageal cancer: institutional data and a systematic review. Sci Rep dose-volume histogram analysis for pneumonitis after 3D treament for non- (2019) 9(1):2255. doi: 10.1038/s41598-018-38414-5 small cell lung cancer(NSCLC). Int J Radiat Oncol Biol Phys (1999) 45(2):323– 5. Zhang Y, Wang H, Huang X, Zhang Q, Ren R, Sun R, et al. Dosimetric 9. doi: 10.1016/s0360-3016(99)00183-2 comparison of TomoDirect, helical tomotherapy, VMAT,and ff-IMRT for 10. Bashir U, Siddique MM, Mclean E, Goh V, Cook GJ. Imaging heterogeneity in upper thoracic esophageal carcinoma. Med Dosim (2019) 44(2):167–72. lung cancer:techniques,applications,and challenges. Am J Roentgenol (2016) doi: 10.1016/j.meddos.2018.05.001 207(3):534–43. doi: 10.2214/AJR.15.15864 Frontiers in Oncology | www.frontiersin.org 11 December 2020 | Volume 10 | Article 596013 Du et al. A Novel Nomogram Predicting RP for ESCC ́ ́ 11. Yip SS, Aerts HJ. Applications and limitations of radiomics. Phys Med Biol 32. Daghir-Wojtkowiak E, Wiczling P, Bocian S, Kubik Ł, Koslinski P, Bocian S, (2016) 61(13):R150–66. doi: 10.1088/0031-9155/61/13/R150 et al. Least absolute shrinkage and selection operator and dimensionality 12. Cunliffe A, Armato SG, Castillo R, Pham N, Guerrero T, Al-Hallaq HA. Lung reduction techniques in quantitative structure retention relationship modeling texture in serial thoracic computed tomography scans:correlation of of retention in hydrophilic interaction liquid chromatography. J Chromatogr radiomics-based features with radiation therapy dose and radiation A (2015) 1403:54–62. doi: 10.1016/j.chroma.2015.05.025 pneumonitis development. Int J Radiat Oncol Biol Phys (2015) 91(5):1048– 33. Tucker S L, Liao Z, Dinh J, Bian SX, Mohan R, Martel MK, et al. Is there an 56. doi: 10.1016/j.ijrobp.2014.11.030 impact of heart exposure on the incidence of radiation pneumonitis? Analysis 13. Krafft SP, Rao A, Stingo F, Briere TM, Court LE, Liao Z, et al. The utility of of data from a large clinical cohort. Acta Oncol (2014) 53(5):590–6. quantitative CT radiomics features for improved prediction of radiation doi: 10.3109/0284186X.2013.831185 pneumonitis. Med Phys (2018) 45(11):5317–24. doi: 10.1002/mp.13150 34. Zhao J, Yorke ED, Li L, Kavanagh BD, Li XA, Das S, et al. Simpe factors 14. Lambin P, Rios-Velazquez E, Leijenaar R, Carvalho S, van Stiphout RG, associated with radiatio-reduced lung toxicity after stereotactic body radiation Granton P, et al. Radiomics: Extracting more information from medical therapy of the thorax:a pooled analysis of 88 studies. Int J Radiat Oncol Biol images using advanced feature analysis. Eur J Cancer (2012) 48(4):441–6. Phys (2016) 95(5):1357–66. doi: 10.1016/j.ijrobp.2016.03.024 doi: 10.1016/j.ejca.2011.11.036 35. Chargari C, Riet F, Mazevet M, Morel E, Lepechoux C, Deutsch E. 15. Gillies RJ, Kinahan PE, Hricak H. Radiomics:images are more than pictures,they Complications of thoracic radiotherapy. Presse Med (2013) 42(9 pt 2):e342– are data. Radiology (2016) 278(2):563–77. doi: 10.1148/radiol.2015151169 51. doi: 10.1016/j.lpm.2013.06.012 16. Parmar C, Rios Velazquez E, Leijenaar R, Jermoumi M, Carvalho S, Mak RH, 36. Tucker S L, Liu HH, Liao Z, Wei X, Wang S, Jin H, et al. Analysis of radiation et al. Robust radiomics feature quantification using semiautomatic volumetric pneumonitis risk using a generalized lyman model. Int J Radiat Oncol Biol segmentation. PloS One (2014) 9(7):e102107. doi: 10.1371/journal.pone.0102107 Phys (2008) 72(2):568–74. doi: 10.1016/j.ijrobp.2008.04.053 17. Doumou G, Siddiques M, Tsoumpas C, Goh V, Cook GJ. The precision of 37. Jin H, Tucker SL, Liu HH, Wei X, Yom SS, Wang S, et al. Dose-volume textural analysis in (18)F-FDG-PET scans of oesophageal cancer. Eur Radiol thresholds and smoking status for the risk of treatment-related pneumonitis (2015) 25(9):2805–12. doi: 10.1007/s00330-015-3681-8 in inoperable non-small cell lung cancer treated with definitive radiotherapy. 18. Coroller TP, GrossmannP,Hou Y, Velazquez ER, Leijenaar RTH, HermannG, et al. Radiother Oncol (2009) 919(3):427–32. doi: 10.1016/j.radonc.2008.09.009 CT-based radiomic signature predicts distant metastasis in lung adenocarcinoma. 38. Takeda A, Kunieda E, Ohashi T, Aoki Y, Oku Y, Enomoto T, et al. Severe Radiother Oncol (2015) 114(3):345–50. doi: 10.1016/j.radonc.2015.02.015 COPD is correlated with mild radiation pneumonitis following stereotactic 19. Tibshirani R. Regression Shrinkage and Selection via the Lasso:a retrospective. body radiotherapy. Chest (2012) 141(4):858–66. doi: 10.1378/chest.11-1193 J Roy Statist Soc Ser B (1996) 58(1):267–88. doi: 10.1111/j.1467-9868. 39. Kimura T, Togami T, Takashima H, Nishiyama Y, Ohkawa M, Nagata Y. 2011.00771.x Radiation pneumonitis in patients with lung and mediastinal tumours:a 20. Gui J, Li H. Penalized Cox regression analysis in the high-dimensional and retrospective study of risk factors focused on pulmonary emphysema. Br J low-sample size settings, with applications to microarray gene expression Radiol (2012) 85(1010):135–41. doi: 10.1259/bjr/32629867 data. Bioinformatics (2005) 21(13):3001–8. doi: 10.1093/bioinformatics/bti422 40. Permuth JB, Choi J, Balarunathan Y, Kim J, Chen DT, Chen L, et al. 21. Hepp T, Schmid M, Gefeller O, Waldmann E, Mayr A. Approaches to Combining radiomic features with a miRNA classififier may improve regularized regression–a comparison between gradient boosting and the prediction of malignant pathology for pancreatic intraductal papillary Lasso. Methods Inf Med (2016) 55(5):422–30. doi: 10.3414/ME16-01-0033 mucinous neoplasms. Oncotarget (2016) 7:85785–97. doi: 10.18632/ 22. Ndhlovu ZM, Chibnik LB, Proudfoot J, Vine S, McMullen A, Cesa K, et al. oncotarget.11768 High-dimensionai immunomonitonitoring models of HIV-1-specific CD8 T- 41. Wang L, Liang S, Li C, Sun X, Pang L, Meng X, et al. A novel nomogram and cell responses accurately identify subjects achieving spontaneous viral control. risk classification system predicting radiation pneumonitis in patients with Blood (2013) 121(5):801–11. doi: 10.1182/blood-2012-06-436295 esophageal cancer receiving radiotherapy. Int J Radiat Oncol Biol Phys (2019) 23. Kumamaru KK, Saboo SS, Aghayev A, Cai P, Quesada CG, George E, et al. CT 105(5):1074–85. doi: 10.1016/j.ijrobp.2019.08.024 pulmonary angiography-based scoring system to predict the prognosis of 42. Dreiseitl S, Ohno-Machado L. Logistic regression and artificial neural network acute pulmonary embolism. J Cardiovasc Comput Tomogr (2016) 10(6):473–9. classifications models: A methodology review. J BioMed Inform (2002) 35(5- doi: 10.1016/j.jcct.2016.08.007 6):352–9. doi: 10.1016/s1532-0464(03)00034-0 24. Vasquez MM, Hu C, Roe DJ, Chen Z, Halonen M, Guerra S. Least absolute 43. Ayer T, Chhatwal J, Alagoz O, Jr CEK, Woods RW, Burnside ES. Informatics shrinkage and selection operator type methods for the identification of serum in radiology:Comparison of logistic regression and artificial neural network biomarkers of overweight and obesity: simulation and application. BMC Med models in breast cancer risk estimation. Radiographics (2010) 30(1):13–22. Res Methodol (2016) 16(1):154. doi: 10.1186/s12874-016-0254-8 doi: 10.1148/rg.301095057 25. Xu C-J, van der Schaaf A, van’t Veld AA, Langendijk JA, Schilstra C. Statistical 44. Coy S L, Cheema AK, Tyburski JB, Laiakis EC, Collins SP, Fornace A Jr. validation of normal tissue complication probability models. Int J Radiat Radiation metabolomics and its potential in biodosimetry. Int J Radiat Biol Oncol Biol Phys (2012) 84(1):e123–9. doi: 10.1016/j.ijrobp.2012.02.022 (2011) 87(8):802–23. doi: 10.3109/09553002.2011.556177 26. Friedman J, Hastie T, Tibshirani R. Regularization Paths for Generalized 45. Hoover DA, Reid RH, Wong E, Stitt L, Sabondjian E, Rodrigues GB, et al. Linear Models via Coordinate Descent. J Stat software (2010) 33(1):1–22. SPECT-based functional lung imaging for the prediction of radiation doi: 10.1016/j.jspi.2009.07.020 pneumonitis: A clinical and dosimetric correlation. J Med Imaging Radiat 27. Aerts HJ, Velazquez ER, Leijenaar RT, Parmar C, Grossmann P, Carvalho S, Oncol (2014) 58(2):214–22. doi: 10.1111/1754-9485.12145 et al. Decoding tumour phenotype by noninvasive imaging using aquantitative 46. Castillo R, Pham N, Ansari S, Meshkov D, Castillo S, Li M, et al. Pre- radiomics approach. Nat Commun (2014) 5:4006. doi: 10.1038/ncomms5006 radiotherapy FDG PET predicts radiation pneumonitis in lung cancer. Radiat 28. Jenkins P, Welsch A. Computed tomography appearance of early radiation Oncol (2014) 9:74. doi: 10.1186/1748-717X-9-74 injury to the lung: Correlation with clinical and dosimetric factors. Int J Radiat Oncol Biol Phys (2011) 81(1):97–103. doi: 10.1016/j.ijrobp.2010.05.017 29. Xu L, Jiang J, Li Y, Zhang L, Li Z, Xian J, et al. Genetic variants of SP-D confer Conflict of Interest: The authors declare that the research was conducted in the susceptibility to radiation pneumonitis in lung cancer patients undergoing thoracic absence of any commercial or financial relationships that could be construed as a radiation therapy. Cancer Med (2019) 8(5):2599–611. doi: 10.1002/cam4.2088 potential conflict of interest. 30. Yin M, Liao Z, Liu Z, Wang L, O'Reilly M, Gomez D, et al. Genetic variants of the nonhomologous end joining gene LIG4 and severe radiation pneumonitis Copyright © 2020 Du, Tang, Cui, Wang, Zhang, Li and Li. This is an open-access in nonsmall cell lung cancer patients treated with definitive radiotherapy. article distributed under the terms of the Creative Commons Attribution License Cancer (2012) 118:528–35. doi: 10.1002/cncr.26214 (CC BY). The use, distribution or reproduction in other forums is permitted, provided 31. Mackin D, Fave XB, Zhang L, Fried D, Yang J, Taylor B, et al. Measuring the original author(s) and the copyright owner(s) are credited and that the original computed tomography scanner variability of radiomics features. Invest Radiol publication in this journal is cited, in accordance with accepted academic practice. No (2015) 50:(11):757–65. doi: 10.1097/RLI.0000000000000180 use, distribution or reproduction is permitted which does not comply with these terms. Frontiers in Oncology | www.frontiersin.org 12 December 2020 | Volume 10 | Article 596013 http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Frontiers in Oncology Pubmed Central

A Novel Nomogram Model Based on Cone-Beam CT Radiomics Analysis Technology for Predicting Radiation Pneumonitis in Esophageal Cancer Patients Undergoing Radiotherapy

Frontiers in Oncology , Volume 10 – Dec 17, 2020

Loading next page...
 
/lp/pubmed-central/a-novel-nomogram-model-based-on-cone-beam-ct-radiomics-analysis-BQBMzJPgA9

References (96)

Publisher
Pubmed Central
Copyright
Copyright © 2020 Du, Tang, Cui, Wang, Zhang, Li and Li
ISSN
2234-943X
eISSN
2234-943X
DOI
10.3389/fonc.2020.596013
Publisher site
See Article on Publisher Site

Abstract

ORIGINAL RESEARCH published: 17 December 2020 doi: 10.3389/fonc.2020.596013 A Novel Nomogram Model Based on Cone-Beam CT Radiomics Analysis Technology for Predicting Radiation Pneumonitis in Esophageal Cancer Patients Undergoing Radiotherapy 1,2 1 3 3 3 3 Feng Du , Ning Tang , Yuzhong Cui , Wei Wang , Yingjie Zhang , Zhenxiang Li and Jianbin Li Department of Radiation Oncology, School of Clinical Medicine, Cheeloo College of Medicine, Shandong University, Jinan, 2 3 China, Department of Radiation Oncology, Zibo Municipal Hospital, Zibo, China, Department of Radiation Oncology, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Jinan, China Purpose: We quantitatively analyzed the characteristics of cone-beam computed Edited by: Tiziana Rancati, tomography (CBCT) radiomics in different periods during radiotherapy (RT) and then Istituto Nazionale dei Tumori (IRCCS), built a novel nomogram model integrating clinical features and dosimetric parameters for Italy predicting radiation pneumonitis (RP) in patients with esophageal squamous cell Reviewed by: Gilles Defraene, carcinoma (ESCC). KU Leuven, Belgium Methods: At our institute, a retrospective study was conducted on 96 ESCC patients for Irina Vergalasova, Rutgers Cancer Institute of New whom we had complete clinical feature and dosimetric parameter data. CBCT images of Jersey, United States each patient in three different periods of RT were obtained, the images were segmented *Correspondence: using both lungs as the region of interest (ROI), and 851 image features were extracted. Jianbin Li lijianbin@msn.com The least absolute shrinkage selection operator (LASSO) was applied to identify candidate radiomics features, and logistic regression analyses were applied to construct the rad- Specialty section: score. The optimal period for the rad-score, clinical features, and dosimetric parameters This article was submitted to Radiation Oncology, were selected to construct the nomogram model and then the receiver operating a section of the journal characteristic (ROC) curve was used to evaluate the prediction capacity of the model. Frontiers in Oncology Calibration curves and decision curves were used to demonstrate the discriminatory and Received: 18 August 2020 clinical benefit ratios, respectively. Accepted: 04 November 2020 Published: 17 December 2020 Results: The relative volume of total lung treated with ≥5 Gy (V5), mean lung dose (MLD), Citation: and tumor stage were independent predictors of RP and were finally incorporated into the Du F, Tang N, Cui Y, Wang W, nomogram. When the three time periods were modeled, the first period was better than Zhang Y, Li Z and Li J (2020) A Novel Nomogram Model Based on Cone- the others. In the primary cohort, the area under the ROC curve (AUC) was 0.700 (95% Beam CT Radiomics Analysis confidence interval (CI) 0.568–0.832), and in the independent validation cohort, the AUC Technology for Predicting Radiation Pneumonitis in Esophageal Cancer was 0.765 (95% CI 0.588–0.941). In the nomogram model that integrates clinical features Patients Undergoing Radiotherapy. and dosimetric parameters, the AUC in the primary cohort was 0.836 (95% CI 0.700– Front. Oncol. 10:596013. 0.918), and the AUC in the validation cohort was 0.905 (95% CI 0.799–1.000). The doi: 10.3389/fonc.2020.596013 Frontiers in Oncology | www.frontiersin.org 1 December 2020 | Volume 10 | Article 596013 Du et al. A Novel Nomogram Predicting RP for ESCC nomogram model exhibits excellent performance. Calibration curves indicate a favorable consistency between the nomogram prediction and the actual outcomes. The decision curve exhibits satisfactory clinical utility. Conclusion: The radiomics model based on early lung CBCT is a potentially valuable tool for predicting RP. V5, MLD, and tumor stage have certain predictive effects for RP. The developed nomogram model has a better prediction ability than any of the other predictors and can be used as a quantitative model to predict RP. Keywords: esophageal cancer, cone beam computed tomography, radiation pneumonitis, prediction model, radiomics capability of lung texture features, which help describe the INTRODUCTION potential RP risk (14, 15). Among malignant tumors, the incidence rate of esophageal At present, cone-beam computed tomography (CBCT) has cancer (EC) is the seventh highest, and the mortality rate is become a routine online method of image-guided radiotherapy sixth worldwide (1). Radiotherapy (RT) is still one of the main (IGRT) for EC. If we can perform quantitative analysis on CBCT treatments for locally advanced EC (2, 3). However, radiation radiomics features in a certain period of RT and then combine pneumonitis (RP) is one of the major toxicities of thoracic these radiomics features with clinical features and dosimetric radiation therapy. If RP occurs, it seriously affects the patient’s parameters to predict RP in EC, it will help guide clinical quality of life and survival prognosis (4). Therefore, it is treatment strategies in a timely manner. imperative for EC patients undergoing RT to identify this Therefore, the initial aim of this study was to investigate toxicity at the earliest possible time. More importantly, the whether the early changes in CBCT radiomics features could be accurate prediction of RP is essential to facilitate individualized used as potential markers for predicting RP. In the present study, radiation dosing that leads to maximized therapeutic gain. At a comprehensive nomogram, which is a conveniently applicable present, the risk assessment of RP is mainly predicted by using predictive model integrating CBCT radiomics features, clinical lung dosimetric parameters (5, 6), such as the relative volume of features, and dosimetric parameters, was built for the total lung irradiated above a specified threshold dose (V )or individualized risk assessment and precise prediction of RP. mean lung dose (MLD): Although several metrics have appeared promising, the results vary across institutions, so these metrics do not seem to be perfect at predicting RP (7, 8). In addition to MATERIALS AND METHODS dosimetric parameters, some clinical features (tumor stage, smoking history, preexisting lung diseases, concurrent Patients chemotherapy, and radiation dose) are also considered to be The entire cohort of this retrospective study was obtained from the related to RP occurrence. However, the consensus on the records of our institutional picture archiving and communication comparative importance of these related predictors remains system (PACS) from January 2017 to June 2019, which was used unavailable at present. Consequently, in order to individually to identify esophageal squamous cell carcinoma (ESCC) patients and precisely discern RP, an accurate predictive model receiving RT treatment. The inclusion criteria were as follows: (1) incorporating multiple types of factors with superior clinical Karnofsky performance score (KPS) ≥70, (2) no previous history utility is urgently needed. of thoracic RT, (3) intensity-modulated radiotherapy (IMRT) and Computed tomography (CT) images play an essential role in received ≥50 Gy RT, and (4) CBCT scan performed at least once a the diagnosis and treatment of RP. As early as the end of the 20th week during RT with the scanning range of the CBCT imaging century, RP could be identified by CT. However, RP cannot be including at least two thirds of the lungs. The exclusion criteria predicted by superficial CT manifestations. Therefore, the focus were as follows: (1) low image quality, (2) general pulmonary of later research is on the accurate prediction of RP (9). In recent infection unrelated to RT, and (3) treatment break of more than 7 years, with the rapid development of radiomics analysis days during RT. A total of 96 consecutive patients with thoracic technology, increasing attention has been paid to the research middle segment ESCC were identified and divided into two of RT effect and side effect predictions based on radiomics cohorts at a 7:3 proportion using computer-generated random features (10–13). Among them, one study found that there is a numbers. Sixty-seven patients were allocated to the primary dose-dependent relationship between the changes in some cohort, and 29 patients were allocated to the verification cohort. radiomics features and RP ≥2 grade determined by extracting Our institutional research ethics board approved this retrospective local lung CT images after RT (12). Another study successfully study (SDTHEC201703014). It waived the need to obtain established a differential model of high- and low-risk RP by informed consent from the patients due to the retrospective analyzing the region of interest (ROI) of the whole lung tissue nature of the investigation (retrospective single-institution before RT (13). In short, radiomics features can capture the cohort study). Frontiers in Oncology | www.frontiersin.org 2 December 2020 | Volume 10 | Article 596013 Du et al. A Novel Nomogram Predicting RP for ESCC rotating the frame at an angle. This is a slow CBCT acquisition Clinical Data and RT Parameters setting. The acquisition time is 67 s, and the patient keeps breathing The clinical data were all acquired from the institute’s medical evenly during this process. Standard body scan conditions were records. Specifically, clinicopathological parameters included voltage (125 kVp), current (80 mA), exposure time (13 ms), age, sex, KPS, smoking status, diabetes history, chronic exposure (680 mAs), rotation angle (178°–182°), pixel matrix size obstructive lung disease (COPD), pathological diagnosis, tumor (384×384), field of view (FOV, 45×18 cm), slice thickness (2.5 mm), location, TNM stage, radiation dose, and concurrent and fan-beam type (half-fan). Among fan-beam types, the half-fan chemoradiotherapy lack thereof. In addition, the lung mode was used for the image acquisition of lung tissue structures dosimetric parameters involved in this study included V5–V40 larger than 24 cm. In this study, lung CBCT image acquisition was (relative volume of total lung treated with ≥5–40 Gy) and MLD. carried out in three different periods, and then the images were In short, the parameters mentioned above were used to establish imported into 3D Slicer (version 4.10.2; http://www.slicer.org) in a a comprehensive nomogram after univariate analysis or least DICOM format to extract and analyze the radiomics features. It absolute shrinkage selection operator (LASSO) feature selection. should be noted that these three different periods were artificially The Eclipse Treatment Planning System (Varian Medical divided according to the experimental design and corresponded to Systems, Palo Alto, CA, Version 13.5.35) was adopted for RT the early stages: the third, fourth, and fifth weeks of RT (PTV planning design. IMRT adopts a fixed-field, static intensity prescription dose range of EC: 18–22 Gy, 27–32 Gy, and 36–44 Gy). modulation technique, and 5–7 fields of coplanar irradiation are uniformly divided according to the specific situation in each Image Segmentation and Feature case. The required target parameters are then set, and the dose Extraction distribution is obtained by inverse calculation of the treatment planning system. The dose distribution is then graded Images from both lungs were segmented by a semiautomatic segmentation method (16, 17) based on a threshold-based (stratified), and each field is decomposed into a series of subfields. IMRT does not include sIMRT or volumetric algorithm. The specific steps are as follows: First, the background was removed to obtain the internal region of the chest. Second, the intensity-modulated arc therapy (VMAT). The target area includes tumor volume (GTV), including CT imaging of visible appropriate threshold was found to segment the lung and the tissues outside the lung contour to the greatest extent. Finally, esophageal tumors and positive lymph nodes. The clinical target volume (CTV) refers to the upper and lower expansion of the the manual segmentation method (18) was used to erase the extra parts outside the large trachea and lung parenchyma to obtain both esophageal tumor by 3 cm and 6 mm around the tumor and related lymphatic drainage area. The planned target volume lungs as the ROI. Image segmentation was performed by an experienced radiologist and then verified by a senior radiologist. (PTV) is formed by CTV extending 8 mm outward. IMRT was administered by a Varian Linac Accelerator with a 6-MVX ray All features were extracted by using the radiomics plug-in in 3D Slicer. A total of 851 radiomics features were extracted, and 95% PTV, and radiation doses of 50–66 Gy (median dose of 60 Gy) and 1.8–2.0 Gy/fraction 5 times/week were prescribed. including 13 morphological features, 18 histogram features, 74 original texture features, and 746 high-order features (wavelet Normal tissue constraints were prioritized in the following order for treatment planning purposes: maximum spinal cord transform features). dose of 45 Gy, relative volume of total lung treated with ≥5Gy Radiomics Feature Selection and (V5) ≤60%, relative volume of total lung treated with ≥20 Gy Radiomics Signature Construction (V20) ≤28%, MLD ≤20 Gy, relative volume of the heart treated First, the extracted radiomics features were preprocessed. Based on with ≥30 Gy (V30) ≤40%, and relative volume of the heart the Spearman rank correlation test, the features with correlations treated with ≥40 Gy (V40) ≤30%. greater than 0.9 and multicollinearity were deleted, and independent features were preliminarily screened. Meanwhile, Follow-up and Evaluation of RP based on the Mann–Whitney U test, the characteristics with Follow-up items included chest CT, physical examination, and significant differences between the RP (≥2 grade) and non-RP clinical symptoms. Patients were evaluated weekly during RT, (<2 grade) groups were screened out. Finally, the LASSO method followed up at 1 month after completion of the initial treatment, (19) was used to select the final features, and the RP prediction and then followed up every 2–3 months until at least 6 months model of rad-score was constructed based on logistic regression after the end of RT. The grading of RP was confirmed by two analysis. The LASSO method minimizes the sum of squared senior oncologists and one radiologist. The National Cancer residuals by using the case in which the sum of the absolute Institute Common Terminology Criteria for Adverse Events 4.03 values of the coefficients is less than the tuning parameter (l). (CTCAE 4.03) was used to evaluate the degree of RP. In the To prevent overfitting of the model, (20–22)duringmodel present study, grade ≥2 was used as the cutoff for diagnosing RP. building, features are selected by constantly adjusting l. With CBCT Scanning Method and Image the increasing penalty, more regression coefficients are reduced Acquisition to zero, (23, 24) and then the remaining nonzero coefficient Using the on-board imager (OBI) system mounted on the Varian is selected. The nonzero coefficient of the selected features is Trilogy medical linear accelerator, the hardware portion included a the rad-score. Each patient’s rad-score is calculated as a linear diagnostic (kV) level X-ray source (KVS) and an amorphous silicon combination of selected features that have their own flat-panel kV detector (KVD). The CBCT image was obtained by coefficient weighting. Frontiers in Oncology | www.frontiersin.org 3 December 2020 | Volume 10 | Article 596013 Du et al. A Novel Nomogram Predicting RP for ESCC In this study, 50 iterations of 10-fold nested cross-validation Construction and Validation of the were utilized, similarly to Xu et al. (25). Random sampling Nomogram was conducted in an attempt to balance the class distributions First, the prediction efficiency of the three different periods was within the cross-validation partitions. The cross-validation compared, and then the best period was selected. Second, 96 loop provides a profile of model performance. It serves to patients were divided into the RP (39 cases) and non-RP (57 estimate how well the LASSO applied to a given set of candidate cases) groups, and 16 clinical features and dosimetric predictors may generalize to other data sets. Model performance parameters were collected. The best clinical features and was assessed by computing the area under the curve (AUC) for dosimetric parameters were determined by LASSO feature each constructed model on a test partition. The inner cross- selection. Finally, a comprehensive nomogram was validation loop was applied to determine the optimal value for l established. The receiver operating characteristic (ROC) curve such that the resulting model was guarded against overfitting. The was used to evaluate the prediction capacity of the model. The value of l for each cross-validation partition was selected by calibration curve was used to determine whether the predicted determining the value that produced the most regularized model and observed probabilities for RP were in concordance. The such that the AUC was within one standard error of the maximum decision curve was performed to evaluate the clinical benefit (26). The use of 50 resampled iterations with 10-fold nested cross- ratio of the nomogram. validation constructs 500 models used to generate a distribution of This research process can be divided into four parts: image AUC values to estimate how well model construction with LASSO acquisition, ROI segmentation, feature extraction, and radiomics generalizes to other data sets. model construction as shown in Figure 1. FIGURE 1 | Flow chart of radiomics. Frontiers in Oncology | www.frontiersin.org 4 December 2020 | Volume 10 | Article 596013 Du et al. A Novel Nomogram Predicting RP for ESCC TABLE 1 | Univariate analysis of baseline clinical features of patients and RP. Statistical Analysis All statistical analyses were based on SPSS 20.0 (IBM, Armonk, NY, Factor N RP c value P value USA) or R software (R Foundation for Statistical Computing, 2 <2 grade ≥2 grade Vienna, Austria, https://www.R-project.org/). The c test or Fisher exact probability test was used to classify data between the two Sex 96 57 39 2.767 0.096 Male 81 51 30 groups. Two independent-sample t tests were used for counting Female 15 6 9 data (continuous data). The Mann–Whitney U test was used to Age (years) 96 1.619 0.203 compare the differences in clinical features between the primary <60 21 15 6 and validation cohorts. The model was evaluated with respect to ≥60 75 42 33 Stage 2.650 0.008 sensitivity, specificity, ROC curve, and 95% confidence interval II 19 15 4 (CI). P values ≤ 0.05 were considered statistically significant. III 48 30 18 IV 29 12 17 Smoking history 96 0.198 0.656 No 54 31 23 RESULTS Yes 42 26 16 COPD 96 1.436 0.231 Analysis of Clinical Features and No 81 46 35 Yes 15 11 4 Dosimetric Parameters Associated With Diabetes 96 0.318 0.573 RP No 88 53 35 Yes 8 4 4 The 96 patients were divided into RP (39 cases) and non-RP (57 Hypertension 96 0.606 0.436 cases) groups, and 9 clinical features and 7 dosimetric parameters No 83 48 35 that might be related to the occurrence of RP were included. Yes 13 9 4 Univariate analysis showed that tumor stage was correlated with Concurrent 96 Chemotherapy ≥2 grade RP (c = 2.650, P = 0.008), and other factors, including No 71 41 30 0.300 0.584 age, sex, concurrent chemoradiotherapy or lack thereof, COPD Yes 25 16 9 status, smoking status, and RT dose, showed no significant Delivered 96 1.867 0.172 differences between the two groups (all Ps > 0.05). V5, V10, Dose (Gy) <60 45 30 15 V15, V20, V30, and MLD of both lungs were associated with the ≥60 51 27 24 occurrence of grade ≥2 RP (all Ps < 0.05). The characteristics of COPD, chronic obstructive lung disease. the enrolled population are listed in Tables 1 and 2. There were no significant differences in age, sex, tumor stage, V5, and MLD between the primary group and the validation TABLE 2 | Single factor analysis of DVH and RP. group, which indicates that the groupings were reasonable (all Ps Lung DVH RP P value c value > 0.05) as shown in Table 3. Seven factors (tumor stage, V5, V10, V15, V20, V30, and MLD) remained after univariate analysis. <2 grade ≥2 grade The LASSO feature selection method was used to screen these V5 48.95 ± 10.56 59.39 ± 10.00 0.00 -4.91 seven factors, and three potential factors (V5, MLD, and tumor V10 33.64 ± 7.70 40.92 ± 7.95 0.00 -4.46 stage) remained as shown in Figures 2A, B. The AUC values of V15 25.34 ± 6.52 30.77 ± 6.96 0.00 -3.85 prediction efficiency for V5, MLD, and tumor stage were 0.698, V20 18.81 ± 5.47 22.47 ± 4.82 0.00 -3.47 0.685, and 0.662, respectively. To observe the overall predictive V30 9.61 ± 4.40 12.16 ± 5.00 0.01 -2.58 performance of V5, MLD, and tumor stage, we established a full V40 3.80 ± 2.49 4.58 ± 3.24 0.21 -1.25 MLD (cGy) 1016.47 ± 218.82 1260.87 ± 267.38 0.00 -4.72 clinical–dosimetric feature combined model. The AUC value of the combined model was 0.764 as shown in Figure 2C. MLD, mean lung dose; V5, V10, V15, V20, V30, V40 = relative volume of total lung treated with ≥5, 10, 15, 20, 30, and 40 Gy. Radiomics Feature Extraction/Selection at TABLE 3 | Comparison of sex, age, tumor stage, V5, and MLD between the Different Periods and Radiomics Signature primary and the verification cohort. Building In the first period (PTV dose: 18–22 Gy), a total of 851 radiomics Factor Primary cohort Verification cohort c value P value features were extracted from the patients. First, correlations greater Age (years) 65.33 ± 9.37 68.62 ± 8.89 -1.64 0.11 than 0.9 features were deleted, resulting in a total of 220 features Sex (N) 67 29 0.11 0.75 remaining. Second, linear features were removed, and 96 features Male 56 25 remained. Then, 21 features remained after using the rank-sum Female 11 4 Stage 3.54 0.17 test. Finally, the remaining two features after LASSO selection were II 13 6 used to build the radiomics model as shown in Figures 3A, B.The III 30 18 two features are original first-order skewness and original GLSZM- IV 24 5 small area emphasis. The model was built as follows: Rad-score = V5 52.35 ± 11.27 55.14 ± 12.01 -1.07 0.29 -0.924 e+00×Skewness - 7.047 e+00×Small Area Emphasis + MLD (Gy) 11.06 ± 2.61 11.38 ± 2.85 -0.52 0.61 Frontiers in Oncology | www.frontiersin.org 5 December 2020 | Volume 10 | Article 596013 Du et al. A Novel Nomogram Predicting RP for ESCC AB C FIGURE 2 | LASSO characteristic selection of clinical features and dosimetric parameters (A, B). ROC curve of V5, MLD, tumor stage, and combined model (C). A B 21 19 9 2 -8 -6 -4 -2 Log Lambda FIGURE 3 | Feature screening of radiomics in the first period. By adjusting the different penalty parameter (l) to obtain a high-performance model, the radiomics characteristics with the highest predictive performance were obtained. Radiomics feature convergence diagram (A). Each curve represents the trajectory of the coefficient of each independent variable (B). 4.5329. Rad-scores for each patient in the primary cohort and after using the rank-sum test. Finally, the remaining six features validation cohort are shown in Figures 4A, B. (gray-level nonuniformity, small dependence low gray-level In the second period (PTV dose: 27–32 Gy), a total of 851 emphasis, cluster shape, uniformity, entropy, and size zone radiomics features were extracted from the patients. First, nonuniformity) after LASSO selection were used to build the correlations greater than 0.9 features were deleted, resulting radiomics model. The model was built as follows: Rad-score = in a total of 222 features remaining. Second, linear features were +4.680 e-07×gray-level nonuniformity + 1.087 e+01×small removed, and 96 features remained. Then, 10 features remained dependence low gray-level emphasis - 7.913 e-04×cluster shape after using the rank-sum test. Finally, the remaining five + 1.401 e+00×uniformity + 1.406 e+00×entropy - 2.207 e-05×size features (voxel volume, smallest axis length, small zone nonuniformity - 4.776 e+00. dependence low gray-level emphasis, large area low gray-level Validation of Radiomics Signature at emphasis, and busyness) after LASSO selection were used to build the radiomics model. The model was built as follows: Different Periods Rad-score = -1.996 e-07×voxel volume - 4.036 e-03×smallest In the first period, the predictive efficacy of the model was as axis length + 5.376 e+01×small dependence low gray-level follows: In the primary cohort, the AUC was 0.700 (95% CI emphasis + 1.718 e-07×large area low gray-level emphasis - 0.568–0.832), the sensitivity was 61.5%, and the specificity was 2.473 e-04×busyness + 1.041 e+00. 75.0%. In the validation cohort, the AUC was 0.765 (95% CI In the third period (PTV dose: 36–44 Gy), a total of 851 0.588–0.941), the sensitivity was 84.6%, and the specificity was radiomics features were extracted from the patients. First, 64.7% as shown in Table 4 and Figures 5A, B. correlations greater than 0.9 features were deleted, resulting in In the second period, the predictive efficacy of the model was a total of 220 features remaining. Second, linear features were as follows: In the primary cohort, the AUC was 0.663 (95% CI removed, and 96 features remained. Then, 43 features remained 0.530–0.797), the sensitivity was 90.6%, and the specificity was Frontiers in Oncology | www.frontiersin.org 6 December 2020 | Volume 10 | Article 596013 Coefficients 0 1000 2500 Du et al. A Novel Nomogram Predicting RP for ESCC FIGURE 4 | Rad-score for each patient in the primary and validation cohorts. Green bars show scores for patients without RP, and orange bars show scores for those with RP (A, B). TABLE 4 | ROC curve parameters of the radiomics model and nomogram. Classification Primary cohort Validation cohort AUC 95% CI Sensitivity Specificity AUC 95% CI Sensitivity Specificity First period 0.700 0.568-0.832 61.5% 75.0% 0.765 0.588-0.941 84.6% 64.7% Second period 0.663 0.530-0.797 90.6% 42.9% 0.604 0.356-0.851 85.7% 50.0% Third period 0.699 0.573-0.826 66.7% 70.3% 0.756 0.561-0.950 66.7% 80.0% Nomogram 0.836 0.700-0.918 96.0% 54.8% 0.905 0.799-1.000 92.9% 73.3% A B FIGURE 5 | ROC curve of radiomics features in the first period of RT (A, B). Frontiers in Oncology | www.frontiersin.org 7 December 2020 | Volume 10 | Article 596013 Du et al. A Novel Nomogram Predicting RP for ESCC 42.9%. In the validation cohort, the AUC was 0.604 (95% CI DISCUSSION 0.356 –0.851), the sensitivity was 85.7%, and the specificity 50.0%. A single index based on lung dosimetric parameters is not the “gold standard” to judge the occurrence of PR risk; however, In the third period, the predictive efficacy of the model was as follows: In the primary cohort, the AUC was 0.699 (95% radiomics can extract image data to characterize the standard tissue structure, including typical lung structures. They may CI 0.573–0.826), the sensitivity was 66.7%, and the specificity was 70.3%. In the validation cohort, the AUC was 0.756 (95% CI produce clinically relevant improvements in predicting treatment-related toxicities (13). This makes up for the 0.561–0.950), the sensitivity was 66.7%, and the specificity was 80.0% as shown in Table 4. deficiency of dose-volume parameter prediction to a great extent. Some previous studies, (12, 13) respectively, report the By comparing the prediction efficiency of the AUC in three periods, it is obvious that the prediction efficiency in the first relationship between the changes in some second- or higher- order eigenvalues of lung cancer after and before RT and the period is better than those in the second and third periods in both the primary and validation cohorts. To reflect the occurrence of RP. Unfortunately, due to the limitations of detection techniques or other factors, it is not possible to importance of the early prediction of RP in clinical practice, the first-period rad-score and three essential features (V5, MLD, establish predictive models for clinical practice. In this study, we used an automated computer extraction algorithm and digital and tumor stage) were used to establish a comprehensive nomogram model. quantitative analysis technology to obtain high-quality information to comprehensively evaluate various characteristics The Incremental Value of the Radiomics of tumor and normal tissue responses (14, 27). More Signature When Added to the importantly, we constructed a comprehensive nomogram Comprehensive Nomogram model based on CBCT radiomics features in combination with The AUC values of dosimetric parameters (V5, MLD) and clinical features and dosimetric parameters to accurately predict clinical features (tumor stage) were 0.698, 0.685, and 0.662, RP in EC patients treated with RT. To the best of our knowledge, respectively. The AUC values of the full clinical–dosimetric this is the first study of the early prediction of RP by using IGRT feature combined model was 0.764. In addition, the AUC to obtain CBCT imaging information in different periods of RT. values of the radiomics signature at three different periods Importantly, this comprehensive nomogram model is superior to were 0.700, 0.663, and 0.699, respectively (primary cohort). It single clinical features and lung dosimetric parameters in can be seen that the single clinical features, dosimetric RP prediction. We selected CBCT images from three different periods and parameters, or full clinical–dosimetric combined model are not ideal in predicting the risk of RP. To this end, we created a extracted the radiomics features. The primary purpose was to find the first radiomics features that can independently predict comprehensive nomogram that integrates dosimetric parameters and clinical features with the radiomics signature from the first RP; however, after selecting the radiomics features in different periods, it is found that each period has its own independent set period. The results show that, in the primary cohort, the AUC of our nomogram was 0.836 (95% CI: 0.700–0.918), and in the of feature parameters related to RP. We believe that, in addition to the influence of radiation dose factors, whether these validation cohort, the AUC was 0.905 (95% CI: 0.799–1.000) as shown in Table 4 and Figures 6B, C. There is no doubt that the characteristics vary with changes in the RT process is still uncertain. It is gratifying that we found the best prediction of comprehensive nomogram, incorporating radiomics features, significantly improves the ability of conventional dosimetric RP to be in the first period of radiomics characteristics. Two important features can be found in the early stage of low-dose RT parameters and clinical features to predict the risk of RP. The graphical form of the nomogram is shown in Figure 6A. More of lung tissue: Although this may differ from our initial expectation of the experimental results, the results are importantly, the calibration curve is produced as shown in Figure 6D. The diagonal dotted line represents an ideal fascinating. This result is similar to the findings of Cunliffe et al. (12) and Jenkins et al (28). They found that AUC values evaluation, and the other two lines next to it represent the performance of the nomogram. A closer fit to the diagonal in low- and medium-dose areas of the lung were different between RP and non-RP patients even though these AUC dotted line indicates a better evaluation. In summary, this calibration curve shows favorable consistency between the values appeared in areas with lower visible changes. These first radiomics features may be able to be used to explain or screen out predicted RP and the actual observation. those susceptible to RP due to intrinsic genetic mutations. How to Make Clinical Decisions In regard to the susceptible population of RP, we must devote The clinical decision curve analysis of the nomogram is shown in attention to the sensitivity of lung tissue to RT. At present, the Figure 6E, which shows the patient’s benefits when the physician radiosensitivity of lung tissue has been reported (29, 30), and it is makes the judgment. It shows that, if the probability of the considered to be a potential influencing factor for RP occurrence. domain value is 10%, the benefit of using the nomogram to This difference in the sensitivity of lung tissue to radiation predict the efficacy of RP is higher than that of radiomics features constitutes our different understanding of the probability of or other parameters alone. In short, this decision curve exhibits RP. In two groups of patients with different radiosensitivity of satisfactory positive net benefits of the nomogram at the lung tissue, we cannot judge the probability of RP by standard threshold probabilities. clinical features and lung dosimetric parameters. However, Frontiers in Oncology | www.frontiersin.org 8 December 2020 | Volume 10 | Article 596013 Du et al. A Novel Nomogram Predicting RP for ESCC BC DE Decision curve FIGURE 6 | (A) The comprehensive nomogram incorporates V5, MLD, tumor stage, and rad-score (the first period) to predict the risk of RP in EC patients. V5: relative volume of total lung treated with ≥5 Gy; MLD: mean lung dose. (B, C) ROC curves of the comprehensive nomogram in the primary and validation cohorts. (D) Calibration curves of the comprehensive nomogram with the addition of V5, MLD, tumor stage, and radiomics features. The diagonal dotted line represents an ideal evaluation, and the other two lines next to it represent the performance of the nomogram. A closer fit to the diagonal dotted line indicates a better evaluation. (E) Decision curves of the radiomics features model and the combination model (comprehensive nomogram) predicting RP. The y-axis represents the net benefit. The red curve represents the comprehensive nomogram, and the green line represents the radiomics features model. The horizontal black line indicates that the assumption is valid. The oblique gray line indicates that the assumption is invalid. Frontiers in Oncology | www.frontiersin.org 9 December 2020 | Volume 10 | Article 596013 Du et al. A Novel Nomogram Predicting RP for ESCC radiomics can analyze the data by extracting features from CT reports not being sufficient to provide specific and safe standard images of the lung, thus providing a powerful method for the doses (34). Chargari et al. (35) find that V5 is a risk factor for noninvasive description of lung tissue radiosensitivity. This may acute or chronic lung toxicity. Cho et al. (6) find that MLD is the be why the radiomics features are superior to the clinical features most related factor that predicts RP rather than V5, V10, or V20. and dosimetric parameters in current studies. In this study, this Some clinical features have emerged as important risk factors advantage in AUC value, sensitivity, and specificity performance contributing to RP progression. Some studies show that is not particularly good, but through our research analysis, smoking is related to the severity of RP (36, 37). Takeda et al. radiomics features of RP risk prediction cannot be ignored. (38) and Kimura et al. (39) report that COPD is a significant risk The successful establishment of the prediction model is based factor for RP in patients with EC after RT. In this study, we find on the standardization of data collection and the rationalization that smoking status, COPD, and concurrent chemoradiotherapy of data processing. First, we should consider that the feature are not correlated with the incidence of RP, and so these factors extraction data are affected by CT parameters (31) because the are not included in our combined model, but this does not mean CT features may be different under different image-acquisition that they are not important. After LASSO logistic regression conditions. In this study, based on the CBCT of the Varian analysis, several significant variables, including V5, MLD, and accelerators in our center, these devices have the same tube tumor stage, were integrated into the nomogram to predict PR. voltage, tube current, exposure time, exposure amount, and pixel The results were as follows: clinical-dose characteristic model matrix size, which can help control for the differences between (AUC values: V5 = 0.698, MLD = 0.685, tumor stage = 0.662), the scanners and acquisition parameters. Second, to develop the radiomics model (primary cohort AUC 0.700, validation cohort radiomics signature, a total of 851 candidate features were AUC 0.765), and nomogram (primary cohort AUC 0.836, reduced to a set of only a few potential descriptors by using validation cohort AUC 0.905). The nomogram demonstrates a the LASSO logistic regression model to realize feature selection better ability to predict RP than the other models. by constantly adjusting the regularization parameter l to make How to use this information in the treatment plan or the weight coefficient of the feature approach zero. The LASSO alternative program to help clinicians is our greatest concern. (20) logistic regression model is suitable for analyzing large sets Fortunately, the goal of radiomics is to develop a decision- of radiomics features with a relatively small sample size, and it is making tool that meets the needs of clinicians. This is because such a tool could combine radiomics features with other patient designed to avoid overfitting high-dimensional data (21, 32). At the same time, the LASSO logistic regression model allows the characteristics to improve the capability of the decision support radiomics signature to be constructed by combining the selected model (15, 40). We show that radiomics features complement features, so it allows the model to more easily identify the most clinical features and lung dosimetric parameters, helping to closely related features in patients with RP. Finally, the nested provide better predictive ability for RP. The clinical decision cross-validation method (25) was used for internal validation to curve of this nomogram shows that the effectiveness of the improve the accuracy of the model. nomogram in predicting RP is higher than that of using It should be noted that the difference in the irradiation mode radiomic characteristics or other parameters alone. In short, (3-D conformal radiation therapy and IMRT) affects the under the threshold probability, the decision curve exhibits a potential dose distribution of the lung, which may affect the satisfactory positive net benefit of the nomogram. selection of clinical features and dosimetric parameters as risk Our results demonstrate the potential value of radiomics characteristics of RP. This can be quickly confirmed by techniques in the risk prediction of RP patients. If more comparing Tucker et al. (33) and Shane et al. (13) where, in clinical variables are included in the nomogram, there will be the former, 75% received 3-D conformal radiation therapy, and more room for future development of this model, and the the latter 83% received IMRT. Therefore, it seems complicated to resulting prediction effect will be better. A recent study (41)by establish a general model with good discriminant performance another of our teams found that subjective global assessment under different technical conditions. score (SGA), pulmonary fibrosis score (PFS), planning target The clinical factors (age, tumor stage, KPS score, chronic volume/total lung volume (PTV/LV), MLD, and ratio of change lung disease, diabetes, chemotherapy lack thereof) and lung regarding systemic immune inflammation index at 4 weeks (4w dosimetric parameters (V5, V10, V20, MLD) related to RP are SII) were potential valuable markers in predicting severe acute reported in previous studies. To provide better help for the radiation pneumonitis (SARP). Subsequently, the team oncologist, we designed a clinical nomogram to combine the developed a nomogram and corresponding risk classification above available RP risk factors with radiomics features. system with superior prediction ability for SARP. In the next Therefore, we aim to establish a combined model, maximizing step, we will consider combining the research results of this team clinical utility and accuracy of prediction ability, and so the with radiomics to establish a new RP prediction model for better initial experimental design was not expected to rely solely on the clinical application. radiomics model as the final prediction model. Of course, Although our study has many strengths, several limitations judging from the AUC value, sensitivity, and specificity of the should be addressed here. First, the sample size is small, which radiomics model in each period of RT, these characteristics can lead to the inability to apply nonlinear classifiers, such as alone are not perfect in predicting RP. Dose-volume histogram neural networks (42, 43). Because a nonlinear classifier uses a (DVH) metrics have been extensively observed and reported to more extensive data set, it is beneficial to improve the accuracy of be correlated with RP despite the current data and research the RP model. Second, our analysis does not account for two-way Frontiers in Oncology | www.frontiersin.org 10 December 2020 | Volume 10 | Article 596013 Du et al. A Novel Nomogram Predicting RP for ESCC or higher-order interactions between features. If interactions diagnosis and treatment of diseases and the prediction between features had been identified, the interaction terms that of complications. were most strongly associated with the outcome interactions would have been selected when we constructed the radiomics signature, and this could have improved performance. However, CONCLUSIONS uncovering the interactions of multiple attributes is a challenging problem. Third, we used a validation cohort that was drawn from CT radiomics has powerful data-processing and analysis abilities. the same institution as the primary cohort, which prevented us In this context, we explored a method to predict RP based on the from investigating the generalizability of the results to other lung CBCT radiomics features for EC patients. More institutions and settings. In addition, there is a lack of sufficient importantly, we used this method to successfully build and external data validation. Fourth, selection bias occurred when validate a novel nomogram with good predictive value, which strict criteria were used, and this may affect the model training. can help clinicians identify high-risk RP patients early and guide For instance, all patients are middle thoracic EC patients, which personalized treatment and clinical decisions. limits the application of this method to patients with cervical, upper, and lower thoracic segment EC radiotherapy. Also, all patients experienced uniform CBCT imaging scanners and DATA AVAILABILITY STATEMENT parameters, which does not guarantee the reproducibility and stability of radiomics features under other conditions. In the The raw data supporting the conclusions of this article will be future, we should conduct a prospective, multicenter, large- made available by the authors, without undue reservation. cohort study to further develop and validate nomograms in terms of predicting RP. As a future study, we will add different types of patients, including those with different EC locations (cervical, upper AUTHOR CONTRIBUTIONS thoracic, lower thoracic segments) and different RT techniques FD and NT are responsible for analyzing data and writing (3DCRT, TOMO, VMAT). We will also include more laboratory papers. JL designed experiments to guide the writing and indicators that may reflect RP, such as inflammatory indexes and revision of papers. WW directed the writing and revision of immune inflammatory indexes. In terms of basic research, we papers. YC, YZ, and ZL were responsible for radiomics diagnosis should also improve the model of radiomics, especially the and radiomics data processing of patients. All authors combination of radiomics and genomics. The former focuses contributed to the article and approved the submitted version. on medical imaging of the normal tissues or tumors and performs diagnosis and prognosis based on quantitative imaging features, and the latter discovers and notes the gene sequences to study the function and structure of genomes of the FUNDING diseases. Besides this, if we can combine available radiation metabolomics (44) with functional CT (45, 46)radiomics Funding was obtained from the National Key Research Program features, it may help us understand the differences in radiation of China (No. 2016YFC0904700), the National Natural sensitivity and tissue cell metabolism in order to establish a more Science Foundation of China (No. 81773287), and The Key robust prediction model. Therefore, it can be predicted that the Research Development Program of Shandong Province combination of multiple omics will be the best plan for the future (No. 2016GSF201093). 6. Cho WK, Oh D, Kim HK, Ahn YC, Noh JM, Shim YM, et al. Dosimetric REFERENCES predictors for postoperative pulmonary complications in esophageal cancer 1. Bray F, Ferlay J, Soerjomataram I, Siegel RL, Torre LA, Jemal A. Global cancer following neoadjuvant chemoradiotherapy and surgery. Radiother Oncol statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide (2019) 133:87–92. doi: 10.1016/j.radonc.2019.01.005 for 36 cancers in 185 countries. CA Cancer J Clin (2018) 68(6):394–424. 7. Rodrigues G, Lock M, D’Souza D, Yu E, Dyk JV. Prediction of radiation doi: 10.3322/caac.21492 pneumonitis by dose-volume histogram parameters in lung cancer-A systematic 2. Ajani JA, D'Amico TA, Bentrem DJ, Chao J, Corvera C, Das P, et al. review. Radiother Oncol (2004) 71(2):127–38. doi: 10.1016/j.radonc.2004.02.015 Esophageal and Esophagogastric Junction Cancers, Version 2.2019,NCCN 8. SenanS,CardenalF,VansteenkisteJ,Stigt J, AkyolF, Neve WD. A Clinical Practice Guidelines in Oncology. J Natl Compr Canc Netw (2019) 17 randomized phase II study comparing induction or consolidation (7)855–83. doi: 10.6004/jnccn.2019.0033 chemotherapy with cisplatindocetaxel,plus radical concurrent 3. Deng W, Lin SH. Advances in radiotherapy for esophageal cancer. Ann Transl chemoradiotherapy with cisplatin-docetaxel,in patients with unresectable Med (2018) 6(4):79. doi: 10.21037/atm.2017.11.28 locally advanced non-small-cell lung cancer. Ann Oncol (2011) 22:553–8. 4. Tonison JJ, Fischer SG, Viehrig M, Welz S, Boeke S, Zwirner K, et al. doi: 10.1093/annonc/mdq388 Radiation pneumonitis after intensity-modulated radiotherapy for 9. Graham M V, Purdy JA, Emami B, Harms W, Bosch W, Lockett MA. Clinical esophageal cancer: institutional data and a systematic review. Sci Rep dose-volume histogram analysis for pneumonitis after 3D treament for non- (2019) 9(1):2255. doi: 10.1038/s41598-018-38414-5 small cell lung cancer(NSCLC). Int J Radiat Oncol Biol Phys (1999) 45(2):323– 5. Zhang Y, Wang H, Huang X, Zhang Q, Ren R, Sun R, et al. Dosimetric 9. doi: 10.1016/s0360-3016(99)00183-2 comparison of TomoDirect, helical tomotherapy, VMAT,and ff-IMRT for 10. Bashir U, Siddique MM, Mclean E, Goh V, Cook GJ. Imaging heterogeneity in upper thoracic esophageal carcinoma. Med Dosim (2019) 44(2):167–72. lung cancer:techniques,applications,and challenges. Am J Roentgenol (2016) doi: 10.1016/j.meddos.2018.05.001 207(3):534–43. doi: 10.2214/AJR.15.15864 Frontiers in Oncology | www.frontiersin.org 11 December 2020 | Volume 10 | Article 596013 Du et al. A Novel Nomogram Predicting RP for ESCC ́ ́ 11. Yip SS, Aerts HJ. Applications and limitations of radiomics. Phys Med Biol 32. Daghir-Wojtkowiak E, Wiczling P, Bocian S, Kubik Ł, Koslinski P, Bocian S, (2016) 61(13):R150–66. doi: 10.1088/0031-9155/61/13/R150 et al. Least absolute shrinkage and selection operator and dimensionality 12. Cunliffe A, Armato SG, Castillo R, Pham N, Guerrero T, Al-Hallaq HA. Lung reduction techniques in quantitative structure retention relationship modeling texture in serial thoracic computed tomography scans:correlation of of retention in hydrophilic interaction liquid chromatography. J Chromatogr radiomics-based features with radiation therapy dose and radiation A (2015) 1403:54–62. doi: 10.1016/j.chroma.2015.05.025 pneumonitis development. Int J Radiat Oncol Biol Phys (2015) 91(5):1048– 33. Tucker S L, Liao Z, Dinh J, Bian SX, Mohan R, Martel MK, et al. Is there an 56. doi: 10.1016/j.ijrobp.2014.11.030 impact of heart exposure on the incidence of radiation pneumonitis? Analysis 13. Krafft SP, Rao A, Stingo F, Briere TM, Court LE, Liao Z, et al. The utility of of data from a large clinical cohort. Acta Oncol (2014) 53(5):590–6. quantitative CT radiomics features for improved prediction of radiation doi: 10.3109/0284186X.2013.831185 pneumonitis. Med Phys (2018) 45(11):5317–24. doi: 10.1002/mp.13150 34. Zhao J, Yorke ED, Li L, Kavanagh BD, Li XA, Das S, et al. Simpe factors 14. Lambin P, Rios-Velazquez E, Leijenaar R, Carvalho S, van Stiphout RG, associated with radiatio-reduced lung toxicity after stereotactic body radiation Granton P, et al. Radiomics: Extracting more information from medical therapy of the thorax:a pooled analysis of 88 studies. Int J Radiat Oncol Biol images using advanced feature analysis. Eur J Cancer (2012) 48(4):441–6. Phys (2016) 95(5):1357–66. doi: 10.1016/j.ijrobp.2016.03.024 doi: 10.1016/j.ejca.2011.11.036 35. Chargari C, Riet F, Mazevet M, Morel E, Lepechoux C, Deutsch E. 15. Gillies RJ, Kinahan PE, Hricak H. Radiomics:images are more than pictures,they Complications of thoracic radiotherapy. Presse Med (2013) 42(9 pt 2):e342– are data. Radiology (2016) 278(2):563–77. doi: 10.1148/radiol.2015151169 51. doi: 10.1016/j.lpm.2013.06.012 16. Parmar C, Rios Velazquez E, Leijenaar R, Jermoumi M, Carvalho S, Mak RH, 36. Tucker S L, Liu HH, Liao Z, Wei X, Wang S, Jin H, et al. Analysis of radiation et al. Robust radiomics feature quantification using semiautomatic volumetric pneumonitis risk using a generalized lyman model. Int J Radiat Oncol Biol segmentation. PloS One (2014) 9(7):e102107. doi: 10.1371/journal.pone.0102107 Phys (2008) 72(2):568–74. doi: 10.1016/j.ijrobp.2008.04.053 17. Doumou G, Siddiques M, Tsoumpas C, Goh V, Cook GJ. The precision of 37. Jin H, Tucker SL, Liu HH, Wei X, Yom SS, Wang S, et al. Dose-volume textural analysis in (18)F-FDG-PET scans of oesophageal cancer. Eur Radiol thresholds and smoking status for the risk of treatment-related pneumonitis (2015) 25(9):2805–12. doi: 10.1007/s00330-015-3681-8 in inoperable non-small cell lung cancer treated with definitive radiotherapy. 18. Coroller TP, GrossmannP,Hou Y, Velazquez ER, Leijenaar RTH, HermannG, et al. Radiother Oncol (2009) 919(3):427–32. doi: 10.1016/j.radonc.2008.09.009 CT-based radiomic signature predicts distant metastasis in lung adenocarcinoma. 38. Takeda A, Kunieda E, Ohashi T, Aoki Y, Oku Y, Enomoto T, et al. Severe Radiother Oncol (2015) 114(3):345–50. doi: 10.1016/j.radonc.2015.02.015 COPD is correlated with mild radiation pneumonitis following stereotactic 19. Tibshirani R. Regression Shrinkage and Selection via the Lasso:a retrospective. body radiotherapy. Chest (2012) 141(4):858–66. doi: 10.1378/chest.11-1193 J Roy Statist Soc Ser B (1996) 58(1):267–88. doi: 10.1111/j.1467-9868. 39. Kimura T, Togami T, Takashima H, Nishiyama Y, Ohkawa M, Nagata Y. 2011.00771.x Radiation pneumonitis in patients with lung and mediastinal tumours:a 20. Gui J, Li H. Penalized Cox regression analysis in the high-dimensional and retrospective study of risk factors focused on pulmonary emphysema. Br J low-sample size settings, with applications to microarray gene expression Radiol (2012) 85(1010):135–41. doi: 10.1259/bjr/32629867 data. Bioinformatics (2005) 21(13):3001–8. doi: 10.1093/bioinformatics/bti422 40. Permuth JB, Choi J, Balarunathan Y, Kim J, Chen DT, Chen L, et al. 21. Hepp T, Schmid M, Gefeller O, Waldmann E, Mayr A. Approaches to Combining radiomic features with a miRNA classififier may improve regularized regression–a comparison between gradient boosting and the prediction of malignant pathology for pancreatic intraductal papillary Lasso. Methods Inf Med (2016) 55(5):422–30. doi: 10.3414/ME16-01-0033 mucinous neoplasms. Oncotarget (2016) 7:85785–97. doi: 10.18632/ 22. Ndhlovu ZM, Chibnik LB, Proudfoot J, Vine S, McMullen A, Cesa K, et al. oncotarget.11768 High-dimensionai immunomonitonitoring models of HIV-1-specific CD8 T- 41. Wang L, Liang S, Li C, Sun X, Pang L, Meng X, et al. A novel nomogram and cell responses accurately identify subjects achieving spontaneous viral control. risk classification system predicting radiation pneumonitis in patients with Blood (2013) 121(5):801–11. doi: 10.1182/blood-2012-06-436295 esophageal cancer receiving radiotherapy. Int J Radiat Oncol Biol Phys (2019) 23. Kumamaru KK, Saboo SS, Aghayev A, Cai P, Quesada CG, George E, et al. CT 105(5):1074–85. doi: 10.1016/j.ijrobp.2019.08.024 pulmonary angiography-based scoring system to predict the prognosis of 42. Dreiseitl S, Ohno-Machado L. Logistic regression and artificial neural network acute pulmonary embolism. J Cardiovasc Comput Tomogr (2016) 10(6):473–9. classifications models: A methodology review. J BioMed Inform (2002) 35(5- doi: 10.1016/j.jcct.2016.08.007 6):352–9. doi: 10.1016/s1532-0464(03)00034-0 24. Vasquez MM, Hu C, Roe DJ, Chen Z, Halonen M, Guerra S. Least absolute 43. Ayer T, Chhatwal J, Alagoz O, Jr CEK, Woods RW, Burnside ES. Informatics shrinkage and selection operator type methods for the identification of serum in radiology:Comparison of logistic regression and artificial neural network biomarkers of overweight and obesity: simulation and application. BMC Med models in breast cancer risk estimation. Radiographics (2010) 30(1):13–22. Res Methodol (2016) 16(1):154. doi: 10.1186/s12874-016-0254-8 doi: 10.1148/rg.301095057 25. Xu C-J, van der Schaaf A, van’t Veld AA, Langendijk JA, Schilstra C. Statistical 44. Coy S L, Cheema AK, Tyburski JB, Laiakis EC, Collins SP, Fornace A Jr. validation of normal tissue complication probability models. Int J Radiat Radiation metabolomics and its potential in biodosimetry. Int J Radiat Biol Oncol Biol Phys (2012) 84(1):e123–9. doi: 10.1016/j.ijrobp.2012.02.022 (2011) 87(8):802–23. doi: 10.3109/09553002.2011.556177 26. Friedman J, Hastie T, Tibshirani R. Regularization Paths for Generalized 45. Hoover DA, Reid RH, Wong E, Stitt L, Sabondjian E, Rodrigues GB, et al. Linear Models via Coordinate Descent. J Stat software (2010) 33(1):1–22. SPECT-based functional lung imaging for the prediction of radiation doi: 10.1016/j.jspi.2009.07.020 pneumonitis: A clinical and dosimetric correlation. J Med Imaging Radiat 27. Aerts HJ, Velazquez ER, Leijenaar RT, Parmar C, Grossmann P, Carvalho S, Oncol (2014) 58(2):214–22. doi: 10.1111/1754-9485.12145 et al. Decoding tumour phenotype by noninvasive imaging using aquantitative 46. Castillo R, Pham N, Ansari S, Meshkov D, Castillo S, Li M, et al. Pre- radiomics approach. Nat Commun (2014) 5:4006. doi: 10.1038/ncomms5006 radiotherapy FDG PET predicts radiation pneumonitis in lung cancer. Radiat 28. Jenkins P, Welsch A. Computed tomography appearance of early radiation Oncol (2014) 9:74. doi: 10.1186/1748-717X-9-74 injury to the lung: Correlation with clinical and dosimetric factors. Int J Radiat Oncol Biol Phys (2011) 81(1):97–103. doi: 10.1016/j.ijrobp.2010.05.017 29. Xu L, Jiang J, Li Y, Zhang L, Li Z, Xian J, et al. Genetic variants of SP-D confer Conflict of Interest: The authors declare that the research was conducted in the susceptibility to radiation pneumonitis in lung cancer patients undergoing thoracic absence of any commercial or financial relationships that could be construed as a radiation therapy. Cancer Med (2019) 8(5):2599–611. doi: 10.1002/cam4.2088 potential conflict of interest. 30. Yin M, Liao Z, Liu Z, Wang L, O'Reilly M, Gomez D, et al. Genetic variants of the nonhomologous end joining gene LIG4 and severe radiation pneumonitis Copyright © 2020 Du, Tang, Cui, Wang, Zhang, Li and Li. This is an open-access in nonsmall cell lung cancer patients treated with definitive radiotherapy. article distributed under the terms of the Creative Commons Attribution License Cancer (2012) 118:528–35. doi: 10.1002/cncr.26214 (CC BY). The use, distribution or reproduction in other forums is permitted, provided 31. Mackin D, Fave XB, Zhang L, Fried D, Yang J, Taylor B, et al. Measuring the original author(s) and the copyright owner(s) are credited and that the original computed tomography scanner variability of radiomics features. Invest Radiol publication in this journal is cited, in accordance with accepted academic practice. No (2015) 50:(11):757–65. doi: 10.1097/RLI.0000000000000180 use, distribution or reproduction is permitted which does not comply with these terms. Frontiers in Oncology | www.frontiersin.org 12 December 2020 | Volume 10 | Article 596013

Journal

Frontiers in OncologyPubmed Central

Published: Dec 17, 2020

There are no references for this article.