Using deep learning to predict radiation pneumonitis in patients treated with stereotactic body radiotherapy (SBRT) for pulmonary nodules: preliminary results

Abstract

This study aimed to develop a predictive model using clinical, dosimetric, and radiomic features for radiation-induced pneumonitis (RP) after lung stereotactic body radiation therapy (SBRT). We retrospectively analyzed the clinical data of 153 patients who underwent SBRT for lung nodules between 2010 and 2019. A total of 3,350 radiomic computed tomography (CT) features of radiotherapy simulation (shape, intensity, texture, and log filters) were extracted. Among them, 30 factors were selected through Pearson's correlation analysis and subjected to analysis. A proposed lung toxicity prediction model was developed using a deep neural network algorithm. The programming language used was Python. The data were divided into a training set (70% of data) and a test/validation set (30% of data). We adjusted the original data by oversampling to correct the uneven sample distribution to balance the data set. The Talos library was used in this study for hyperparameter determination and the model with the highest accuracy was selected. The Talos library provided the RP prediction model with the highest accuracy of 94.8%. The area under the curve of the receiver operating characteristics curve was 0.90, which was relatively fair. It showed relatively high accuracy in the RP prediction model based on the clinical, dosimetric, and radiomic factors of patients who received SBRT for lung nodules. A further study using more cases from other medical centers is being planned for external validation.