Handcrafted Radiomics, Deep Learning Radiomics in the Prediction of Radiation Pneumonitis for NSCLC Patients Treated with Immunotherapy Followed with Thoracic Radiotherapy

Abstract

Our previous study has shown that NSCLC patients previously received immune checkpoint inhibitors (ICIs) underwent thoracic intensity modulated radiotherapy have a higher risk of acute radiation pneumonitis (RP). This study aimed to establish machine learning models using handcrafted radiomics (HCR), deep learning-based radiomics (DLR) and clinical characteristics to improve the prediction of symptomatic radiation pneumonitis (RP) (grade ≥ 2) status for NSCLC patients treated with immunotherapy followed with thoracic radiotherapy. This study retrospectively collected data of 61 NSCLC patients meeting the requirements of study enrollment. Of these 61 patients, 35 developed symptomatic graded ≥ 2 RP. We defined 3 regions of interest (ROIs) in planning CT images including gross tumor volume (GTV), planning tumor volume (PTV), PTV-GTV. We calculated the mean dose, V5, V10, V20, and V30 within TL-GTV, and the volume of GTV, PTV and total lung. A total of 516 handcrafted radiomics features and 512 deep features were extracted from each 3 ROIs. Person Correlation Analysis and Least Absolute Shrinkage and Selection Operator (LASSO) were used to reduce the dimension of features. The HCR models, DLR models and the fusion models across different ROIs with machine learning classifiers were built and compared. In multi-classifier modeling, models with PTV under logistic regression (LR) classifiers showed better prediction than other ROIs under different machine learning algorithms. Based on PTV with LR, HCR+ DLR model had better performance, with an area under the curve (AUC) of 0.95 (95% confidence interval (CI): 0.893-1) in the training cohort and 0.87 (95% CI: 0.698-1) in the test cohort, which was higher than that of HCR model, with an AUC of 0.86 (95% CI: 0.755-0.9) in the training cohort and 0.82 (95% CI: 0.624-1) in the test cohort, the results of fusion model with HCR, DLR and 7 clinical characteristics including T, N, clinical stage, age, smoking, radiotherapy alone/combined and V30, demonstrated the best distinguishing performance, with an AUC of 0.99 (95% CI: 0.970-1) in the training cohort and 0.91 (95% CI: 0.784-1) in the test cohort. The combination of HCR, DLR and clinical characteristic underwent machine learning algorithms can improve the prediction of symptomatic RP in NSCLC patients treated with ICIs followed with thoracic radiotherapy.