Advertisement

Novel Methodology to Investigate the Effect of Radiation Dose to Heart Substructures on Overall Survival

  • Alan McWilliam
    Correspondence
    Corresponding author: Alan McWilliam, PhD
    Affiliations
    Division of Clinical Cancer Science, School of Medical Sciences, Faculty of Biology, Medicine and Health, University of Manchester, Manchester, United Kingdom

    Department of Radiotherapy Related Research, The Christie NHS Foundation Trust, Manchester, United Kingdom
    Search for articles by this author
  • Jonathan Khalifa
    Affiliations
    Department of Radiation Oncology, Institut Universitaire du Cancer de Toulouse, Toulouse, France
    Search for articles by this author
  • Eliana Vasquez Osorio
    Affiliations
    Division of Clinical Cancer Science, School of Medical Sciences, Faculty of Biology, Medicine and Health, University of Manchester, Manchester, United Kingdom

    Department of Radiotherapy Related Research, The Christie NHS Foundation Trust, Manchester, United Kingdom
    Search for articles by this author
  • Kathryn Banfill
    Affiliations
    Division of Clinical Cancer Science, School of Medical Sciences, Faculty of Biology, Medicine and Health, University of Manchester, Manchester, United Kingdom

    Department of Radiotherapy Related Research, The Christie NHS Foundation Trust, Manchester, United Kingdom
    Search for articles by this author
  • Azadeh Abravan
    Affiliations
    Division of Clinical Cancer Science, School of Medical Sciences, Faculty of Biology, Medicine and Health, University of Manchester, Manchester, United Kingdom

    Department of Radiotherapy Related Research, The Christie NHS Foundation Trust, Manchester, United Kingdom
    Search for articles by this author
  • Corinne Faivre-Finn
    Affiliations
    Division of Clinical Cancer Science, School of Medical Sciences, Faculty of Biology, Medicine and Health, University of Manchester, Manchester, United Kingdom

    Department of Radiotherapy Related Research, The Christie NHS Foundation Trust, Manchester, United Kingdom
    Search for articles by this author
  • Marcel van Herk
    Affiliations
    Division of Clinical Cancer Science, School of Medical Sciences, Faculty of Biology, Medicine and Health, University of Manchester, Manchester, United Kingdom

    Department of Radiotherapy Related Research, The Christie NHS Foundation Trust, Manchester, United Kingdom
    Search for articles by this author
Open AccessPublished:June 22, 2020DOI:https://doi.org/10.1016/j.ijrobp.2020.06.031

      Purpose

      For patients with lung cancer treated with radiation therapy, a dose to the heart is associated with excess mortality; however, it is often not feasible to spare the whole heart. Our aim is to define cardiac substructures and dose thresholds that optimally reduce early mortality.

      Methods and Materials

      Fourteen cardiac substructures were delineated on 5 template patients with representative anatomies. One thousand one hundred sixty-one patients with non-small cell lung cancer were registered nonrigidly to these 5 template anatomies, and their radiation therapy doses were mapped. Mean and maximum dose to each substructure were extracted, and the means were evaluated as input to prediction models. The cohort was bootstrapped into 2 variable reduction techniques: elastic net least absolute shrinkage and selection operator and the random survival forest model. Each method was optimized to extract variables contributing most to overall survival, and model coefficients were evaluated to select these substructures. The most important variables common to both models were selected and evaluated in multivariable Cox-proportional hazard models. A threshold dose was defined, and Kaplan-Meier survival curves plotted.

      Results

      Nine hundred seventy-eight patients remained after visual quality assurance of the registration. Ranking the model coefficients across the bootstraps selected the maximum dose to the right atrium, right coronary artery, and ascending aorta as the most important factors associated with survival. The maximum dose to the combined cardiac region showed significance in the multivariable model, a hazard ratio of 1.01/Gy, and P = .03 after accounting for tumor volume (P < .001), N stage (P < .01), and performance status (P = .01). The optimal threshold for the maximum dose, equivalent dose in 2-Gy fractions, was 23 Gy. Kaplan-Meier survival curves showed a significant split (log-rank P = .008).

      Conclusions

      The maximum dose to the combined cardiac region encompassing the right atrium, right coronary artery, and ascending aorta was found to have the greatest effect on patient survival. A maximum equivalent dose in 2-Gy fractions of 23 Gy was identified for consideration as a dose limit in future studies.
      Summary
      This work identifies the cardiac substructures where excess dose is most associated with early mortality. The right atrium, origin of the right coronary artery, and the ascending aorta are identified with a maximum equivalent dose in 2-Gy fractions of 23 Gy presented as a dose limit for future studies.

      Introduction

      For patients with non-small cell lung cancer (NSCLC) treated with radiation therapy, cardiac dose is associated with a decrease in overall survival. The Radiation Therapy Oncology Group (RTOG) 0617 found worse overall survival in the high-dose arm. Patients with NSCLC treated in the high-dose 74-Gy arm had a median survival of 20.3 months compared with 28.7 months in the standard 60-Gy arm. In the analysis, higher dose to the heart (v5 and v30) was identified as associated with excess mortality.
      • Bradley J.D.
      • Paulus R.
      • Komaki R.
      • et al.
      Standard-dose versus high-dose conformal radiotherapy with concurrent and consolidation carboplatin plus paclitaxel with or without cetuximab for patients with stage IIIA or IIIB non-small-cell lung cancer (RTOG 0617): A randomised, two-by-two factorial phase 3 study.
      Additional publications that followed the reporting of RTOG 0617 have explored retrospective and clinical trial data to develop a better understanding of the influence of cardiac dose on overall survival after radiation therapy.
      • McWilliam A.
      • Kennedy J.
      • Hodgson C.
      • Vasquez Osorio E.
      • Faivre-Finn C.
      • van Herk M.
      Radiation dose to heart base linked with poorer survival in lung cancer patients.
      • Stam B.
      • van der Bijl E.
      • van Diessen J.
      • et al.
      Heart dose associated with overall survival in locally advanced NSCLC patients treated with hypofractionated chemoradiotherapy.
      • Stam B.
      • Peulen H.
      • Guckenberger M.
      • et al.
      Dose to heart substructures is associated with non-cancer death after SBRT in stage I-II NSCLC patients.
      • Dess R.T.
      • Sun Y.
      • Matuszak M.M.
      • et al.
      Cardiac events after radiation therapy: Combined analysis of prospective multicenter trials for locally advanced non-small-cell lung cancer.
      • Vivekanandan S.
      • Landau D.B.
      • Counsell N.
      • et al.
      The impact of cardiac radiation dosimetry on survival after radiation therapy for non-small cell lung cancer.
      • Chun S.G.
      • Hu C.
      • Choy H.
      • et al.
      Impact of intensity-modulated radiation therapy technique for locally advanced non-small-cell lung cancer: a secondary analysis of the NRG Oncology RTOG 0617 randomized clinical trial.
      • Guberina M.
      • Eberhardt W.
      • Stuschke M.
      • et al.
      Heart dose exposure as prognostic marker after radiotherapy for resectable stage IIIA/B non-small-cell lung cancer: Secondary analysis of a randomized trial.
      • Ma J.T.
      • Sun L.
      • Sun X.
      • et al.
      Is pulmonary artery a dose-limiting organ at risk in non-small cell lung cancer patients treated with definitive radiotherapy?.
      • Ning M.S.
      • Tang L.
      • Gomez D.R.
      • et al.
      Incidence and predictors of pericardial effusion after chemoradiation therapy for locally advanced non-small cell lung cancer.
      • Schytte T.
      • Hansen O.
      • Stolberg-Rohr T.
      • Brink C.
      Cardiac toxicity and radiation dose to the heart in definitive treated non-small cell lung cancer.
      • Speirs C.K.
      • DeWees T.A.
      • Rehman S.
      • et al.
      Heart dose is an independent dosimetric predictor of overall survival in locally advanced non-small cell lung cancer.
      • Wang K.
      • Eblan M.J.
      • Deal A.M.
      • et al.
      Cardiac toxicity after radiotherapy for stage III non-small-cell lung cancer: Pooled analysis of dose-escalation trials delivering 70 to 90 Gy.
      A pooled analysis of 112 patients with NSCLC in clinical trials identified heart mean dose as being independently associated with cardiac events after treatment. Cardiac events were observed in 23% of the patients for whom the mean dose to the heart was identified as a predictor in multivariable analysis, including correcting for baseline cardiac risk. Risk of a cardiac event for patients with a mean heart dose of <10 Gy, 10 to 20 Gy, or >20 Gy was 4%, 7%, and 21%, respectively, adjusted for competing risk.
      • Wang K.
      • Eblan M.J.
      • Deal A.M.
      • et al.
      Cardiac toxicity after radiotherapy for stage III non-small-cell lung cancer: Pooled analysis of dose-escalation trials delivering 70 to 90 Gy.
      Retrospective analyses have highlighted the V40-55 dose to pulmonary artery
      • Ma J.T.
      • Sun L.
      • Sun X.
      • et al.
      Is pulmonary artery a dose-limiting organ at risk in non-small cell lung cancer patients treated with definitive radiotherapy?.
      and high dose to the left atrium and superior vena cava
      • Stam B.
      • van der Bijl E.
      • van Diessen J.
      • et al.
      Heart dose associated with overall survival in locally advanced NSCLC patients treated with hypofractionated chemoradiotherapy.
      as being associated with early mortality. In addition, changes in electrocardiograms (ECGs) were observed owing to the large dose to the left atrial wall.
      • Vivekanandan S.
      • Landau D.B.
      • Counsell N.
      • et al.
      The impact of cardiac radiation dosimetry on survival after radiation therapy for non-small cell lung cancer.
      Retrospective analyses are limited by the lack of availability of segmentations of cardiac substructures. This limitation can be overcome to an extent by using analysis methodologies that do not require a cardiac segmentation in the analysis. Our previous work used a technique called image-based data mining to analyze the plan dosimetry for 1101 patients with NSCLC without the need for segmentations. This work highlighted the base of the heart as the most dose-sensitive region,
      • McWilliam A.
      • Kennedy J.
      • Hodgson C.
      • Vasquez Osorio E.
      • Faivre-Finn C.
      • van Herk M.
      Radiation dose to heart base linked with poorer survival in lung cancer patients.
      with a hazard ratio of 1.25 and a median survival of 20 months for patients receiving >16.3 Gy in that region, versus 28 months in the low-dose group. The region identified includes multiple substructures in the base of the heart comprising the origin of the coronary arteries, pulmonary arteries, aortic valve, and sinoatrial and atrioventricular nodes. It is likely that each of these structures will display an independent dose-response curve, and one or more could be the primary driver of this observed early mortality. To optimize the radiation therapy plan to minimize the risk of a cardiac event, we need to better define the cardiac substructures to be spared. Furthermore, it is important not to compromise the coverage of the target volume, and therefore sparing a smaller and more targeted region is more likely to be achievable than sparing the entire heart in patients with lung cancer.
      In this work, we aim to define those cardiac substructures where excess dose results in worse patient survival. We perform a retrospective dosimetric analysis of the dose received by 14 cardiac substructures. We use a template patient methodology in which the dose from the radiation therapy plan for each patient is evaluated against multiple template anatomies where cardiac substructures have been segmented manually. Dosimetric parameters will be highly correlated across cardiac substructures; therefore, we apply variable reduction techniques to rank the variables that primarily drive excess mortality. This approach will identify the most important cardiac substructures and allow a threshold dose to be found that can be brought forward into planning studies and, ultimately, translated in clinical practice.

      Methods and Materials

      Data for 1161 patients with NSCLC treated with curative-intent radiation therapy (55 Gy in 20 fractions) were extracted from the treatment planning archive from a single institution. Institutional approval was granted to use the data (research ethics committee reference: 17/NW/0060). Five patients were selected to act as template patients with representative anatomy. Images of the cardiac substructure delineation for the 5 template anatomies are included in Figure E1. These template anatomies were selected as follows: (1) the tumor was not in proximity to the heart (ie, the planning target volume should not overlap the heart) and (2) the heart volume was representative of distribution of heart volume in the study patients (template 1, 919 cm3; template 2, 641 cm3l template 3, 738 cm3; template 4, 632 cm3; template 5, 735 cm3). These heart volumes fall within one standard deviation (SD) of the mean heart volume of the population of patients (837 cm3; SD ± 213 cm3). For these 5 template anatomies, a single clinician who specialized in thoracic radiation oncology (J.K.) delineated the following 14 cardiac substructures: aortic valve, ascending aorta, circumflex, left anterior descending coronary artery, left atrium, left ventricle, left coronary artery, mitral valve, pulmonary artery, pulmonary valve, right coronary artery, right atrium, right ventricle, tricuspid valve, and the whole heart. Contours were verified and amended after discussion with a consultant radiologist with expertise in cardiac anatomy.
      All patients in the cohort were registered nonrigidly to each of the 5 template anatomies in turn using the open source niftyReg code (http://sourceforge.net/projects/niftyreg/). Registrations were inspected manually to remove any gross errors, such as those caused by previous lung surgery or full lung collapse. Seven hundred seventy-eight patients had the origin of the left and right coronary arteries identified manually and a point of interest placed by a clinical oncologist. The systematic error in these points was evaluated as a quantitative measure of the nonrigid registration uncertainty for the base of the heart. These points were chosen because the base of the heart contains many substructures of interest, the lack of image contrast makes registration difficult, and previous work has identified the base of the heart as the region most strongly associated with early mortality. In addition, a subset of 386 patients had a heart segmentation available for comparison with the delineated heart on each template anatomy. To quantify the global uncertainty in the nonrigid registration, the SD of the difference in the center of mass (CoM) coordinates between each patient heart and the template heart segmentations were calculated for each cardinal direction. To account for this uncertainty, the planned dose distribution for each patient was blurred using a 3D-gaussian function with width set to the SD for each direction before extracting dose statistics.
      By nonrigidly registering each patient onto each template patient, the dose distribution for that patient can be evaluated on the template cardiac anatomy (ie, as if the patient’s radiation therapy plan had been created for the template anatomy). Therefore, for each patient, the mean and maximal doses were calculated for the 14 cardiac substructures on each of the 5 template anatomies. If the nonrigid registration was perfect, then the mean and maximal dose for a given patient would be identical for each cardiac substructure across the 5 template anatomies. However, there will be uncertainties in the nonrigid registration, and these uncertainties will interact with the dose gradients present within the dose distribution. For each patient and cardiac substructure, the mean was calculated for the mean and the maximal doses across the 5 template anatomies. The corresponding standard error of the mean was calculated to assess consistency of the extracted values across template patients.
      Cardiac substructures exist in close proximity within the heart boundary, as any such dosimetric parameters extracted will be highly correlated. Univariable and simple multivariable Cox-proportional hazard models for each substructure were investigated. However, because of the collinearity of the dosimetric parameters, assessment of the significance on overall survival directly by Cox-proportional hazard models was not deemed appropriate. Therefore, variable reduction techniques were applied to identify which dosimetric parameter and cardiac substructure has the largest influence in worse overall survival. Two variable reduction approached were investigated: (1) an elastic net least absolute shrinkage and selection operator (LASSO) predicting overall survival at 1 year and (2) a time-to-event random survival forest model. The data was bootstrapped 500 times to access the stability of the model in selecting the substructures mostly strongly influencing overall survival.
      Briefly, elastic net is a method of implementing a general linear model, particularly suited to high dimensional and correlated variables allowing for variable selection; it is a balance between a ridge penalty and a LASSO technique. The elastic-net balances between these 2 approaches by optimizing a hyper-parameter in the model, alpha, using cross-validation to find the optimal model.
      • Friedman J.H.
      • Hastie T.
      • Tibshirani R.
      Regularization paths for generalized linear models via coordinate descent.
      For completeness, a ridge general linear model, a full LASSO model and an elastic-net model were created using the mean and maximal dose for all cardiac substructures as variables in the model. Coefficients were calculated for all variables selected by each approach, for each bootstrap. Two approaches to select the most important variables were used, adapted from the work by Abram et al
      • Abram S.V.
      • Helwig N.E.
      • Moodie C.A.
      • DeYoung C.G.
      • MacDonald A.W.
      • Waller H.G.
      Bootstrap enhanced penalized regression for variable selection with neuroimaging data.
      analyzing neuroimaging data: (1) simply counting the times each variable was selected by each bootstrapped model
      • Bunea F.
      • She Y.
      • Ombao H.
      • Gongvatana A.
      • Devlin K.
      • Cohen R.
      Penalized least squares regression methods and applications to neuroimaging.
      ; however, this approach ignores the magnitude of the coefficients in the model, therefore (2) the quantiles (10%, 25%, 50%, 75%, 90%) of the bootstrapped model coefficients were calculated, and variables where the lowest quantiles were greater than zero were deemed most important. Histograms of the coefficients from each bootstrap were also plotted for visualization.
      Second, a random survival forest model was implemented as described by Ishwaren et al.
      • Ishwaran H.
      • Kogalur U.B.
      • Blackstone H.
      • Lauear M.S.
      Random survival forests.
      This approach is commonly used for analyzing genetic data and is, therefore, designed to handle high dimensional and correlated data sets. This design makes it an ideal model for analyzing dosimetric data from a large amount of neighboring structures where the doses are likely to be highly correlated. This implementation of the random survival model also allows data to be censored where patients have been lost to follow-up. A limit of 500 trees was set with each node allowed to split 5 variables (square root of the number of variables). Permutation testing, with 100 permutations, was used to rank the variables (cardiac substructures) that contributed most in the model and therefore patient survival. (In the permutation testing, the association between a variable of interest and the outcome is broken by permuting values for all individuals.) The difference in the prediction accuracy of the model versus the permuted model gives the importance for a single tree. The average of all permutations gives the importance for the variable. This process is repeated for each variable in the model.
      • Wright M.N.
      • Ziegler A.
      • Konig I.R.
      Do little interactions get lost in dark random forests?.
      ,
      • Breiman L.
      Random forests.
      Similar to the assessment of the bootstrapped models described here, 2 methods were used: (1) counting the number of occasions that a variable was selected as most important and (2) assessing the magnitude of the calculated importance values. Histograms of the importance values were plotted for each substructure for visualization.
      The common and most highly ranked dosimetric variables for the selected cardiac substructures were taken forward into a survival analysis for a combined cardiac region. For the survival analysis, the cohort was randomly split into two-thirds training and one-third validation. Clinical and dosimetric parameters were compared with a Wilcoxon test to ensure balance of these cohorts. Univariable and multivariable Cox-proportional hazards models were created to calculate the hazard ratios of the selected dosimetric parameters. Finally, to identify a potential threshold dose-level, Kaplan-Meier curves were plotted using an optimal split technique to identify the threshold with the largest separation between groups survival. The optimal split was determined on the training data set and evaluated on the test data set. Log-rank was calculated for to assess any significance in the survival of the 2 groups. All statistical analyses were performed in R version 3.6.

      Results

      Visual checks for gross registration errors showed that the nonrigid registration performed well for 978 patients for all 5 template anatomies. Patients were mainly excluded where atelectasis and previous surgery caused the registration to fail. Patient characteristics for the 978 patients included in the analysis are shown in Table 1. Mapping of the origin of the left and right coronary arteries showed acceptable systematic errors (1.1 mm right-left, 1.5 mm anterior-posterior and 4.5 mm cranial-caudal), and full results included in Table E1. Comparison of the CoM coordinates for the registered hearts on the template anatomy compared with the CoM of the template patients resulted in the SD in the left-right direction of 6.5 mm, anterior-posterior direction of 7.1 mm, and the cranial-caudal direction of 11 mm. These values were applied to a 3D-Gaussian filter to blur the dose distribution of each patient before extracting the dosimetric statistics for each cardiac substructure.
      Table 1Patient demographics from the study population
      VariableSubvariableSubtotalPercentage complete
      SexMale527100%
      Female451
      Median age, years (range)73 (38-95)100%
      Median tumor volume, cm3 (range)35.4 (7-528)100%
      Smoking history, nCurrent

      ex-smoker
      145 18034%
      Life-long nonsmoker7
      Comorbidly score04136%
      (ACE27)1114
      2127
      373
      T StageT113193%
      T2397
      T3225
      T4162
      N stageN047093%
      N1130
      N2245
      N365
      Performance0140100%
      Status1438
      2339
      361
      Induction chemotherapyYes

      No
      260 718100%
      LateralityRight

      Left
      540 438100%
      The mean and maximal dose to each cardiac substructure was extracted for each patient evaluated on the 5 template anatomies. For each substructure, the mean across the 5 mean doses and maximal doses were calculated. Summary statistics for each substructure are included in Table E2. The SD and the standard error of the mean across the doses extracted from each template patient is also included in Table E2. These results demonstrate an acceptable range, with larger substructures generally showing better stability than small substructures and that the mean dose was more stable than the maximal dose. As expected, because of the proximity of these structures, all dosimetric variables showed a strong positive correlation Figure E2. In addition, the Cox-proportional hazard models are included in Table E3 for all substructures and mean and maximal doses. In univariable analysis, dosing to all substructures showed an association with overall survival. Because of the importance of tumor size, a simple multivariable model is also included, and again the majority of structures show significance. This finding illustrates the collinearity of the data.
      The full summary statistics and histograms of coefficients from the bootstrapped models are included in Table E4 for the ridge, Table E5 for LASSO, and Table E6 for elastic net (S8) models. These tables included the 10%, 25%, 50%, 75%, and 90% quantiles for the coefficients, and they include the total number of times each variable was selected by the model. The LASSO model typically underfits data; across the majority of bootstraps, few variables were selected, and the only variable with a 75% quantile greater than zero was the maximal dose to the right atrium (Table E5). This variable was selected the most into the models (306/500). Ridge models tend to overfit data, and as seen in Table E4, every variable has been selected into the models, 13 selected greater than 400/500 times. This makes the results from the ridge model difficult to interpret. The histograms of variable coefficients from the bootstrapped models illustrate these issues. Finally, the elastic net model showed the max dose to the right atrium selected into 464 of 500 models and the only variable with the 10% and 25% quantiles greater than zero (Table E6). The mean dose to right ventricle was selected for 332 of 500 models with the greatest 75% and 90% quantile values. The maximal dose to the right coronary artery and mean dose to the ascending aorta were selected for 307, 261, and 232 models, respectively. The summary of these results is included in Table 2.
      Table 2Results from the variable selection techniques.
      Cardiac substructureSelected into model (500 bootstraps)
      Elastic netRandom survival forest
      Right atrium (max dose)464397
      Right coronary (max dose)30712
      Ascending aorta (max dose)23247
      Right ventricle (mean dose)332
      Right ventricle (max dose)1991
      Ascending aorta (mean dose)26113
      The full results from the bootstrapped random survival forest models are included in the Table E6 and summarized in Table 2. The maximal dose to the right atrium was ranked most important in 397 of 500 models, and it displayed the highest importance values on permutation testing. The next highest ranked variables were the maximal dose to the ascending aorta (47 of 500 models), the mean dose to ascending aorta (13 of 500 models), and maximal dose to right coronary artery (12 of 500 models). Table 2 shows that the most common, highest ranked variables between the elastic net and random survival forest models were the maximal dose to the right atrium, right coronary artery, and the ascending aorta. These substructures were combined into one cardiac region and taken forward into the final survival analysis.
      Next, the patient cohort was split randomly into a training data set (647 patients) and a test data set (331 patients). Variables were well balanced between groups (Table E7); tumor volume, age, sex, mean lung dose, received induction chemotherapy, T stage, N stage, and performance status were tested using a Wilcoxon test. Table 3 shows the results for the univariable and multivariable models including this combined region. In the multivariable model, the region showed a hazard ratio (HR) of 1.01 Gy–1 (95% confidence interval [CI], 1.01-1.02; P = .03, evaluated as a continuous variable). The multivariable model showed that (log) tumor volume was significant with an HR of 1.25 cm–3 (95% CI, 1.12-1.40; P < .001, continuous variable); age HR of 1.02 per year (95% CI, 1.01-1.03; P < .001, continuous variable); N-stage P < .01; performance status P = .01; and sex P = .05. Note, the logarithm of tumor volume was used to remove the skew in the distribution in which a small number of large volume tumors might dominate the multivariable model (Table E8). Interestingly, mean lung dose was not significant (P = .85) for lung mean dose across the cohort of 12.5 Gy (minimum mean dose, 2.7 Gy; maximum mean dose, 23.4 Gy). The lung V5 and V20 showed a positive correlation with mean lung dose, on inclusion into the multivariable model in place of the mean lung dose, a similar model performance was found. Tumor laterality was significant in univariable analysis (P = .02) but not in multivariable analysis (P = .28), where previous work, using our institutional data, showed that patients with right-sided tumors perform worse than those with left-sided tumors.
      • McWilliam A.
      • Vasquez Osorio E.
      • Faivre-Finn C.
      • van Herk M.
      Influence of tumour laterality on patient survival in non-small cell lung cancer after radiotherapy.
      The identified cardiac substructures are located on the right side of the heart; for right-sided tumors, the maximum dose received to the combined region is 45.2 Gy (range, 0-56 Gy) and for left-sided tumors is 29.8 Gy (range, 0-55 Gy).
      Table 3Univariable and multivariable models including the maximum dose to the defined cardiac substructures, right atrium, and right coronary artery
      Tumor volume has been included and other clinical factors that may affect outcomes.
      UnivariateMultivariate
      HR (95% CI)P valueHR (95% CI)P value
      Combined region maximum dose (continuous)1.01 (1.01-1.02)<.0011.01 (1.01-1.02).03
      Tumor volume (log) (continuous)1.34 (1.24-1.45)<.0011.25 (1.12-1.40)<.001
      Age (continuous)1.01 (1.00-1.01).321.02 (1.01-1.03)<.001
      Laterality (right vs left)1.24 (1.04-1.48).021.12 (0.91-1.38).28
      Mean lung dose (continuous)1.06 (1.04-1.08)<.0011.00 (0.97-1.04).85
      Gender (male vs female)1.14 (0.96-1.35).131.22 (1.00-1.49).05
      Induction chemotherapy (yes vs no)1.08 (0.89-1.31).430.90 (0.71-1.15).40
      T-Stage
       (T1 ref)
       T21.32 (0.98-1.76).060.96 (0.70-1.31).79
       T31.67 (1.23-2.27)>.011.27 (0.88-1.82).20
       T41.66 (1.19-2.32)>.011.01 (0.68-1.49).95
      N-stage
       (N0 ref)
       N10.99 (0.75-1.31).950.82 (0.60-1.11).19
       N21.45 (1.18-1.79)>.011.47 (1.13-1.90)<.01
       N31.60 (1.15-2.24)>.011.51 (1.00-2.25).05
      Performance status
       (PS0 ref)
       11.30 (0.98-1.71).061.28 (0.95-1.74).11
       21.55 (1.17-2.06)>.011.43 (1.04-1.96).03
       31.82 (1.21-2.73)>.011.85 (1.16-2.96).01
      Abbreviations: CI = confidence interval; HR = hazard ratio.
      Tumor volume has been included and other clinical factors that may affect outcomes.
      The optimal cut was determined for the combined region from the training data to be 19.5 Gy. Figure 1 shows the Kaplan-Meier curves for the test data set (323 patients) with patients grouped on those receiving a maximum dose to the combined region greater or less than 19.5 Gy threshold dose (log rank P = .008). Mean survival times for the patients were: >19.5 Gy for 12 months (95% CI, 10-14 months) and <19.5 Gy for 21 months (95% CI, 17-23 months).
      Figure thumbnail gr1
      Fig. 1Kaplan-Meier curves for patients in the test data set receiving a maximum dose greater than or less than 19.5 Gy to the combined region: right atrium, right coronary artery, and ascending aorta. Patients receiving a higher dose do worse with the curves showing a significant split on the log-rank test. Numbers at risk are included for each curve.
      All patients were planned for treatment with 20 fractions, and a threshold maximum dose of 19.5 Gy was used. Converting to an equivalent dose in 2-Gy fractions (EQD2), assuming an alpha/beta ratio of 2 Gy,
      • Lauk S.
      • Ruth S.
      • Trott K.R.
      The effects of dose-fractionation on radiation-induced heart disease in rats.
      ,
      • Schultz-Hector S.
      • Sund M.
      • Thames H.D.
      Fractionation sensitivity and repair kinetics of radiation-induced heart failure in the rat.
      results in a threshold EQD2 dose of 23 Gy.
      • Fowler J.F.
      Sensitivity analysis of parameters in linear-quadratic radiobiologic modeling.

      Discussion

      We have applied novel variable reduction methodologies to identify cardiac substructures where excess dose is most associated with early mortality. Dosimetric parameters across various cardiac substructures will inherently be highly correlated; therefore, we used variable reduction tools, an elastic net model, and a random survival forest model to rank the importance of the dosimetric parameters for 14 cardiac substructures. Such techniques are commonly used in bioinformatic approaches to analysis genetic data, which is also highly correlated, and are perfectly suited to performing this analysis. However, because of this collinearity of the data, these results should be interpreted with care. Both techniques identified the right atrium, right coronary artery, and ascending aorta as the substructures highly associated with early mortality. The agreement between the 2 different approaches used in the study provides more confidence in the results. A threshold maximum EQD2 dose of 23 Gy to these substructures was identified and proposed as a dose limit to investigate in future studies. Importantly, the patient cohort represents the heterogeneity of the ‘real world’ with patients older than 70 years and performance status 2+, patients who are often underrepresented in clinical trials.
      Manual or automatic segmentation of cardiac substructures in routine radiation therapy planning computed tomography (CT) scans is challenging and time consuming.
      • Morris E.D.
      • Ghanem A.I.
      • Dong M.
      • et al.
      Cardiac substructure segmentation with deep learning for improved cardiac sparing.
      • Zhu J.
      • Zhang J.
      • Qiu B.
      • Liu Y.
      • Liu X.
      • Chen L.
      Comparison of the automatic segmentation of multiple organs at risk in CT images of lung cancer between deep convolutional neural network-based and atlas-based techniques.
      • Finnegan R.
      • Dowling J.
      • Koh E.S.
      • et al.
      Feasibility of multi-atlas cardiac segmentation from thoracic planning CT in a probabilistic framework.
      • Zhou R.
      • Liao Z.
      • Pan T.
      • et al.
      Cardiac atlas development and validation for automatic segmentation of cardiac substructures.
      Delineation is difficult, particularly for small substructures such as the coronary arteries or the valves. In this work, we reversed this methodology. Instead of manually contouring substructures across many patients or attempting an auto-segmentation method, we nonrigidly registered all patients onto 5 template anatomies in which the cardiac substructures had been contoured. This method allows the radiation therapy plan from each patient to be evaluated on the template anatomy. The nonrigid registration will contain uncertainties; to quantify this uncertainty, we compared the CoM of each patient’s heart contour evaluated on the template anatomy. The SD of the CoM coordinates in each cardinal direction was used to blur the dose distribution accounting for the uncertainty in mapping the cardiac dose into the template patients’ spatial frame of reference. In addition, we used 5 template patients calculating the mean of the mean and maximal doses for each substructure. This method will minimize the potential effects of the outliers in driving the variable selection, and it provides a measure of uncertainty in the evaluated structure doses. Table E1 includes the SD and standard error of the mean across the 5 template patients, showing that good agreement was found. The maximal doses showed the greatest variation, as would be expected from the potential interaction of the dose gradients with these structures.
      The right atrium, right coronary artery, and ascending aorta were combined into one region, and the maximum dose was included in a multivariable survival model hazard ratio of 1.01 Gy-1 (P = .03). Interestingly, for the 386 patients with heart contours available, whole-heart dosimetric parameters were not significantly associated with overall survival on univariable analysis (mean heart dose, P = .3; heart v30; P = .4; and heart v5, P = .6). The substructures that we identified are in agreement with work presented in the literature.
      • McWilliam A.
      • Kennedy J.
      • Hodgson C.
      • Vasquez Osorio E.
      • Faivre-Finn C.
      • van Herk M.
      Radiation dose to heart base linked with poorer survival in lung cancer patients.
      ,
      • Stam B.
      • Peulen H.
      • Guckenberger M.
      • et al.
      Dose to heart substructures is associated with non-cancer death after SBRT in stage I-II NSCLC patients.
      Previous work in our institution identified dose to base of the heart to be associated with higher mortality. We used an image-based data mining technique in 1101 patients treated with 55 Gy in 20 fractions and validated in 89 patients undergoing stereotactic ablative body radiation therapy (SABR) with 60 Gy in 5 fractions.
      • McWilliam A.
      • Kennedy J.
      • Hodgson C.
      • Vasquez Osorio E.
      • Faivre-Finn C.
      • van Herk M.
      Radiation dose to heart base linked with poorer survival in lung cancer patients.
      The identified region included multiple cardiac substructures. The work presented here builds on this analysis by identifying the substructures and dosimetric parameters within this region that are most strongly associated with early mortality.
      The Gaussian blur technique used to account for the registration uncertainty and calculating a mean dose across 5 template anatomies introduces a risk that the maximal dose values used in the models are lower than the actual, planned, maximal dose. A Gaussian blur will act to reduce the maximal dose in the plan. Conversely, if the highest dose is in a neighboring substructure or neighboring in the lung it might act to raise the dose in any given substructure. By calculating the mean dose across 5 template anatomies, we aimed to remove the effect of outliers that might have driven the models. These effects become more important because high-dose gradients are likely to exist across the mediastinum. Small registration uncertainties can lead to a large difference in reported dosimetric values. Therefore, an attempt to account and minimize these uncertainties, although imperfect, is required. Although not a direct comparison, the article by McKenzie et al
      • McKenzie A.
      • van Herk M.
      • Mijnheer B.
      Margins for geometric uncertainty around organs at risk in radiotherapy.
      describes the effects of daily treatment uncertainties on the dose to organs at risk, describing how the blurring can cause both an increase or a decrease in dose. These factors need to be considered when interpreting our conclusion.
      Further work has identified structures in the base of the heart. Stam et al
      • Stam B.
      • Peulen H.
      • Guckenberger M.
      • et al.
      Dose to heart substructures is associated with non-cancer death after SBRT in stage I-II NSCLC patients.
      analyzed 803 lung cancer patients treated with SABR. This work nonrigidly registered to a single average patient anatomy and identified the maximum dose on the left atrium and the near minimum dose on the superior vena cava as associated with noncancer death in multivariable analysis.
      • Stam B.
      • Peulen H.
      • Guckenberger M.
      • et al.
      Dose to heart substructures is associated with non-cancer death after SBRT in stage I-II NSCLC patients.
      The population of patients with early-stage lung cancer treated with SABR will show marked differences compared with the population of stage 3 patients in our study population. Treatment fields will be smaller, and there will not be lymph node involvement. Importantly, patients treated with SABR tend to show worse cardiac comorbidities, hence treatment with SABR and not surgery. This latter reason might explain the difference in substructures identified in each study; however, without better prospective data on cardiac health before radiation therapy, we cannot confirm this hypothesis. Complementary work by Johnson-Hart et al
      • Johnson-Hart C.N.
      • Price G.J.
      • Faivre-Finn C.
      • et al.
      Residual setup errors towards the heart after image guidance linked with poorer survival in lung cancer patients: Do we need stricter IGRT protocols?.
      investigated the effect of small residual errors after image guided radiation therapy for patients with lung cancer.
      • Johnson-Hart C.N.
      • Price G.J.
      • Faivre-Finn C.
      • et al.
      Residual setup errors towards the heart after image guidance linked with poorer survival in lung cancer patients: Do we need stricter IGRT protocols?.
      The work showed that when the (∼1 mm) residual setup errors point toward the base of the heart, then patients have worse overall survival. The effect of these small residual errors suggests that there is a substructure in the base of the heart that displays a steep dose-response curve.
      One cardiac structure that is difficult to define on CT is the electrical conduction system. The electrical conduction system originates at the sinoatrial node, which is located in the myocardium superior in the right atrium. The electrical impulse is transmitted through the right atrium to the atrioventricular (AV) node at the junction of the cardiac atria and ventricles. The impulse is stalled at the AV node before proceeding through the bundle of His and Purkinje fibers in the myocardium to ensure coordinated cardiac contraction. The maximal dose to the right atrium may therefore be a surrogate to the dose delivered to the electrical conduction system. The identification of a maximal dose threshold suggests that the structure has a serial behavior (ie, nerves). The analysis by Vivekanandan et al identified associations between changes on ECG and radiation therapy dose to the left atrial wall.
      • Vivekanandan S.
      • Landau D.B.
      • Counsell N.
      • et al.
      The impact of cardiac radiation dosimetry on survival after radiation therapy for non-small cell lung cancer.
      However, this work analyzed only 78 patients, and ECG changes were reported for 38% of patients after treatment. Importantly, patients included in this study were performance status 0 to 1, in comparison with our patient population in which 40% were performance status 2 and higher. It is unknown whether these patients with poorer performance status had more significant cardiac comorbidities. In addition, the results indicate that a dose >63 Gy is associated with these changes (patients isotoxically escalated to 63-73 Gy in 30 fractions) compared with the EQD2 23 Gy maximum dose suggested in our study. ECGs are not routinely collected for lung cancer radiation therapy patients before and after treatment; therefore, it is difficult to validate these findings further in a larger cohort.
      There is evidence that radiation therapy dose to the normal lung tissue is also associated with cardiac events post treatment. Research in mice models suggests an interaction between radiation therapy dose to the lung and elevated pulmonary hypertension through damage to the microvasculature structure.
      • Van Luijk P.
      • Gorter T.
      • Willems T.
      • et al.
      Induction of pulmonary hypertension may explain early mortality after thoracic radiotherapy.
      In our multivariable analysis, the mean lung dose was not significantly associated with survival. However, it may be that lower doses to the lung are important, and further retrospective and prospective studies are needed to further understand this interplay. In addition, our previous work has shown that patients with right-sided lung tumors have worse overall survival than patients with left-sided tumors do (HR, 1.26; P = .004). Mean dose to the right lung volume was significantly associated with overall survival in univariable analysis (P = .01), but mean dose to the left lung was not.
      • McWilliam A.
      • Vasquez Osorio E.
      • Faivre-Finn C.
      • van Herk M.
      Influence of tumour laterality on patient survival in non-small cell lung cancer after radiotherapy.
      Laterality was not significant in the multivariable analysis presented in this work, suggesting that the laterality effect may, in part, be driven by excess dose to the structures in the right side of the heart. This interplay of radiation dose to the normal lung tissue and cardiac substructures needs further investigation.
      Interestingly, previous studies have also shown that dose to the whole heart affects patient outcomes.
      • Dess R.T.
      • Sun Y.
      • Matuszak M.M.
      • et al.
      Cardiac events after radiation therapy: Combined analysis of prospective multicenter trials for locally advanced non-small-cell lung cancer.
      ,
      • Wang K.
      • Eblan M.J.
      • Deal A.M.
      • et al.
      Cardiac toxicity after radiotherapy for stage III non-small-cell lung cancer: Pooled analysis of dose-escalation trials delivering 70 to 90 Gy.
      However, in the cohort analyzed here, the whole-heart dosimetric parameters were not significant on univariable survival analysis. Recent research suggests that whole-heart dose has an immunosuppressive effect that results in early mortality, where radiation dose affects the circulating immune cells in the blood.
      • Ladbury C.J.
      • Rusthoven C.G.
      • Camidge D.R.
      • et al.
      Impact of radiation dose to the host immune system on tumor control and survival for stage III non-small cell lung cancer treated with definitive radiation therapy.
      • Contreras J.A.
      • Lin A.J.
      • Weiner A.
      • et al.
      Cardiac dose is associated with immunosuppression and poor survival in locally advanced non-small cell lung cancer.
      • Jin J.-Y.
      • Mereniuk T.
      • Yalamanchali A.
      • et al.
      A framework for modeling radiation induced lymphopenia in radiotherapy.
      Conversely, higher doses to the certain cardiac substructures can directly affect cardiac function. Such results highlight the multifaceted nature of cardiac toxicity, where it is likely there are multiple mechanisms at play.
      There are limitations of such retrospective data analysis, in particular regarding incomplete data. We do not have access to cause of death data for these patients; therefore, we cannot distinguish between cancer and noncancer death. There is also incomplete information regarding patient comorbidities. For a subset of these patients, we have an ACE-27 score; however, at the time these patients were treated this was not routine. ACE-27 captures severe comorbidities, and we do not have detailed data on cardiac comorbidities. Prospective data collection is required to better understand the interaction between cardiac dose and a patient’s underlying cardiac comorbidities. This understanding will allow personalized cardiac risks to be calculated, individualized radiation dose constraints, and targeted follow-up of patients. To further this understanding, we are recruiting patients to a prospective cardiac biomarker study funded by Yorkshire Cancer Research (Research ethics committee: 18/NW/0706). Detailed comorbidity data will be collected as well as prospective ultrasound CT angiogram alongside circulating cardiac biomarkers. Additional prospective studies, such as the lungART trial (NCT00410683), the CLARIFY study in Groningen, and the HALO study in Amsterdam, are collecting detailed cardiac imaging before and during follow-up of patients with lung cancer to assess cardiac function. These studies will inform the mechanistic understanding of cardiac toxicities. In addition, detailed prospective studies may inform how the cardiac regions target constraint should be individualized for each patient—that is, patients with a higher burden of untreated cardiac comorbidities might need a lower dose constraint than a healthier patient would. This approach will best balance cure against toxicity.
      This work has identified cardiac substructures associated with excess mortality and will inform the implementation of cardiac sparing in radiation therapy. An EQD2 dose of 23 Gy is proposed as a maximum dose to a cardiac subregion defined by the right atrium and right coronary arteries. Prospective studies are required to investigate the proposed constraint.

      Acknowledgments

      The authors thank Dr Sean Brown, Dr Anthea Cree, and Dr David Cobben for providing the anatomic identification points.

      Supplementary Data

      References

        • Bradley J.D.
        • Paulus R.
        • Komaki R.
        • et al.
        Standard-dose versus high-dose conformal radiotherapy with concurrent and consolidation carboplatin plus paclitaxel with or without cetuximab for patients with stage IIIA or IIIB non-small-cell lung cancer (RTOG 0617): A randomised, two-by-two factorial phase 3 study.
        Lancet Oncol. 2015; 16: 187-199
        • McWilliam A.
        • Kennedy J.
        • Hodgson C.
        • Vasquez Osorio E.
        • Faivre-Finn C.
        • van Herk M.
        Radiation dose to heart base linked with poorer survival in lung cancer patients.
        Eur J Cancer. 2017; 85: 106-113
        • Stam B.
        • van der Bijl E.
        • van Diessen J.
        • et al.
        Heart dose associated with overall survival in locally advanced NSCLC patients treated with hypofractionated chemoradiotherapy.
        Radiother Oncol. 2017; 125: 62-65
        • Stam B.
        • Peulen H.
        • Guckenberger M.
        • et al.
        Dose to heart substructures is associated with non-cancer death after SBRT in stage I-II NSCLC patients.
        Radiother Oncol. 2017; 123: 370-375
        • Dess R.T.
        • Sun Y.
        • Matuszak M.M.
        • et al.
        Cardiac events after radiation therapy: Combined analysis of prospective multicenter trials for locally advanced non-small-cell lung cancer.
        J Clin Oncol. 2017; 35: 1395-1402
        • Vivekanandan S.
        • Landau D.B.
        • Counsell N.
        • et al.
        The impact of cardiac radiation dosimetry on survival after radiation therapy for non-small cell lung cancer.
        Int J Radiat Oncol Biol Phys. 2017; 99: 51-60
        • Chun S.G.
        • Hu C.
        • Choy H.
        • et al.
        Impact of intensity-modulated radiation therapy technique for locally advanced non-small-cell lung cancer: a secondary analysis of the NRG Oncology RTOG 0617 randomized clinical trial.
        J Clin Oncol. 2017; 35: 56-62
        • Guberina M.
        • Eberhardt W.
        • Stuschke M.
        • et al.
        Heart dose exposure as prognostic marker after radiotherapy for resectable stage IIIA/B non-small-cell lung cancer: Secondary analysis of a randomized trial.
        Ann Oncol. 2017; 28: 1084-1089
        • Ma J.T.
        • Sun L.
        • Sun X.
        • et al.
        Is pulmonary artery a dose-limiting organ at risk in non-small cell lung cancer patients treated with definitive radiotherapy?.
        Radiat Oncol. 2017; 12: 34
        • Ning M.S.
        • Tang L.
        • Gomez D.R.
        • et al.
        Incidence and predictors of pericardial effusion after chemoradiation therapy for locally advanced non-small cell lung cancer.
        Int J Radiat Oncol Biol Phys. 2017; 99: 70-79
        • Schytte T.
        • Hansen O.
        • Stolberg-Rohr T.
        • Brink C.
        Cardiac toxicity and radiation dose to the heart in definitive treated non-small cell lung cancer.
        Acta Oncol. 2010; 49: 1058-1060
        • Speirs C.K.
        • DeWees T.A.
        • Rehman S.
        • et al.
        Heart dose is an independent dosimetric predictor of overall survival in locally advanced non-small cell lung cancer.
        J Thorac Oncol. 2017; 12: 293-301
        • Wang K.
        • Eblan M.J.
        • Deal A.M.
        • et al.
        Cardiac toxicity after radiotherapy for stage III non-small-cell lung cancer: Pooled analysis of dose-escalation trials delivering 70 to 90 Gy.
        J Clin Oncol. 2017; 35: 1387-1394
        • Friedman J.H.
        • Hastie T.
        • Tibshirani R.
        Regularization paths for generalized linear models via coordinate descent.
        J Stat Softw. 2010; 33: 1-22
        • Abram S.V.
        • Helwig N.E.
        • Moodie C.A.
        • DeYoung C.G.
        • MacDonald A.W.
        • Waller H.G.
        Bootstrap enhanced penalized regression for variable selection with neuroimaging data.
        Front Neurosci. 2016; 10: 344
        • Bunea F.
        • She Y.
        • Ombao H.
        • Gongvatana A.
        • Devlin K.
        • Cohen R.
        Penalized least squares regression methods and applications to neuroimaging.
        Neuroimage. 2011; 55: 1519-1527
        • Ishwaran H.
        • Kogalur U.B.
        • Blackstone H.
        • Lauear M.S.
        Random survival forests.
        Ann Appl Stat. 2008; 2: 841-860
        • Wright M.N.
        • Ziegler A.
        • Konig I.R.
        Do little interactions get lost in dark random forests?.
        BMC Bioinformatics. 2016; 17: 145
        • Breiman L.
        Random forests.
        Mach Learn. 2001; 45: 5-32
        • McWilliam A.
        • Vasquez Osorio E.
        • Faivre-Finn C.
        • van Herk M.
        Influence of tumour laterality on patient survival in non-small cell lung cancer after radiotherapy.
        Radiother Oncol. 2019; 137: 71-76
        • Lauk S.
        • Ruth S.
        • Trott K.R.
        The effects of dose-fractionation on radiation-induced heart disease in rats.
        Radiother Oncol. 1987; 8: 363-367
        • Schultz-Hector S.
        • Sund M.
        • Thames H.D.
        Fractionation sensitivity and repair kinetics of radiation-induced heart failure in the rat.
        Radiother Oncol. 1992; 23: 33-34
        • Fowler J.F.
        Sensitivity analysis of parameters in linear-quadratic radiobiologic modeling.
        Int J Radiat Oncol Biol Phys. 2009; 73: 1532-1537
        • Morris E.D.
        • Ghanem A.I.
        • Dong M.
        • et al.
        Cardiac substructure segmentation with deep learning for improved cardiac sparing.
        Med Phys. 2020; 47: 576-586
        • Zhu J.
        • Zhang J.
        • Qiu B.
        • Liu Y.
        • Liu X.
        • Chen L.
        Comparison of the automatic segmentation of multiple organs at risk in CT images of lung cancer between deep convolutional neural network-based and atlas-based techniques.
        Acta Oncol. 2019; 58: 257-264
        • Finnegan R.
        • Dowling J.
        • Koh E.S.
        • et al.
        Feasibility of multi-atlas cardiac segmentation from thoracic planning CT in a probabilistic framework.
        Phys Med Biol. 2019; 64085006
        • Zhou R.
        • Liao Z.
        • Pan T.
        • et al.
        Cardiac atlas development and validation for automatic segmentation of cardiac substructures.
        Radiother Oncol. 2017; 122: 66-71
        • McKenzie A.
        • van Herk M.
        • Mijnheer B.
        Margins for geometric uncertainty around organs at risk in radiotherapy.
        Radiother Oncol. 2002; 62: 299-307
        • Johnson-Hart C.N.
        • Price G.J.
        • Faivre-Finn C.
        • et al.
        Residual setup errors towards the heart after image guidance linked with poorer survival in lung cancer patients: Do we need stricter IGRT protocols?.
        Int J Radiat Oncol Biol Phys. 2018; 102: 434-442
        • Van Luijk P.
        • Gorter T.
        • Willems T.
        • et al.
        Induction of pulmonary hypertension may explain early mortality after thoracic radiotherapy.
        Radiother Oncol. 2017; 123: s67-s68
        • Ladbury C.J.
        • Rusthoven C.G.
        • Camidge D.R.
        • et al.
        Impact of radiation dose to the host immune system on tumor control and survival for stage III non-small cell lung cancer treated with definitive radiation therapy.
        Int J Radiat Oncol Biol Phys. 2019; 105: 346-355
        • Contreras J.A.
        • Lin A.J.
        • Weiner A.
        • et al.
        Cardiac dose is associated with immunosuppression and poor survival in locally advanced non-small cell lung cancer.
        Radiother Oncol. 2018; 128: 498-504
        • Jin J.-Y.
        • Mereniuk T.
        • Yalamanchali A.
        • et al.
        A framework for modeling radiation induced lymphopenia in radiotherapy.
        Radiother Oncol. 2020; 144: 105-113

      Comments

      Commenting Guidelines

      To submit a comment for a journal article, please use the space above and note the following:

      • We will review submitted comments as soon as possible, striving for within two business days.
      • This forum is intended for constructive dialogue. Comments that are commercial or promotional in nature, pertain to specific medical cases, are not relevant to the article for which they have been submitted, or are otherwise inappropriate will not be posted.
      • We require that commenters identify themselves with names and affiliations.
      • Comments must be in compliance with our Terms & Conditions.
      • Comments are not peer-reviewed.