Radiotherapy-immunotherapy related pneumonitis prediction from pre-treatment CT using a deep graph-based integrative model.

2580 Background: Early prediction of symptomatic pneumonitis (grade ≥ 2) could potentially assist the comprehensive management for non-small cell lung cancer (NSCLC) patients who received radiotherapy combined with immunotherapy. A more accurate integrative model for symptomatic pneumonia prediction using CT images is needed. Methods: This retrospective study contains 243 NSCLC patients (62 symptomatic pneumonitis) who underwent radiotherapy combined with immunotherapy between January 2015 and June 2021. Five-fold cross-validation was performed for training and testing. There were 195 cases in the training set (50 symptomatic pneumonitis) and 48 cases in the validation set (12 symptomatic pneumonitis) in each fold. The deep graph integrative model (DG) was composed of two pre-trained 3D UNet encoders to extract deep features from tumor and lung volumes of pre-treatment CT images, respectively, and a graph attention layer (GAT) for integration and classification. The encoders were fine-tuned using manually segmented tumor and lung CT volumetric patches from the training set. Evaluation measures include area under ROC curve (auc), sensitivity (sen), and specificity (spe). Results: Our new DG achieved auc, sen, and spe of 0.823, 0.767, and 0.810, which outperformed conventional CT radiomics model (auc 0.743, sen 0.620, spe 0.752), 3D UNet based deep radiomics model (auc 0.761, sen 0.746, spe 0.737), and our model without GAT (auc 0.796, sen 0.762, spe 0.782). The improvement was statistically significant (p < 0.001). Conclusions: Our DG model improved symptomatic pneumonia prediction using CT images, which can be used as a tool to effectively improve the safety and personalized treatment of combined radiotherapy with immunotherapy for NSCLC patients.