Predicting Radiation-Induced Lymphopenia in Lung Cancer Patients across Institutions after Treatment with Photon- and Proton-based Radiotherapy

Abstract

Purpose: Radiation-Induced Lymphopenia (RIL) has been correlated to inferior survival and hypothesized to adversely impact adjuvant immunotherapy. This study aims to develop and compare RIL prediction models after both Intensity Modulated Radiation Therapy (IMRT) or Proton Therapy (PT) for locallyadvanced lung cancer. Methods and Materials: In total, 178 patients were collected; 45 IMRT patients from Yonsei Cancer Center and 92 IMRT and 41 PT patients from MD Anderson Cancer Center. Clinical variables and Dose Volume Histograms (DVHs) of normal lung and heart were used to predict Grade(G)4 RIL at the end of radiotherapy. Logistic Regression (LR), Random Forests (RF) and a novel Convolutional Neural Network (CNN) that considers the entire DVHs using convolution layers were developed. Models were compared to each other using the area under the Receiver Operating Characteristics curve (AUC). Results: Incidence of G4RIL was 16%/10% after IMRT/PT respectively. For IMRT, predicting G4 RIL using CNN outperformed other methods, increasing the average AUC from 0.69 to 0.74. Model performance significantly decreased when including PT patients for all models to 0.62-0.65, with only small outperformance of CNNs. Standard deviations indicate lower model robustness for mixed IMRT/PT cohorts (0.04) compared to IMRT only (0.02). Conclusion: CNN approaches to include DVH information to predict RIL are successful in IMRT patients, but inhomogeneity in DVHs and baseline patient heterogeneity in combined IMRT/PT cohorts affects model training and robustness. Further research is required to determine if this is due to fundamental differences in the interaction with circulating lymphocytes between IMRT and PT.