L. Yang, B. Li, and L. Wang; Department of Radiation Oncology, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Jinan, China
Purpose/Objective(s): The combination of thoracic radiotherapy (TRT) and immune checkpoint inhibitors (ICIs) increases the risk of pneumonitis. This study aimed to construct a deep pathomics model based on H&E-stained pathological slides to accurately predict the radioimmune-associated pneumonitis in non-small cell lung cancer (NSCLC). Materials/
Methods: NSCLC patients treated with TRT combined with ICIs in Shandong Cancer Hospital were included. Whole slide images (WSIs) images of H&E-stained pathological slides were collected. Included patients were randomly divided into training set, validation set and test set in a ratio of 3:1:1. The labels of the images were defined as the occurrence of grade 2 or higher radioimmune-associated pneumonitis. The Vision Transformer-Recursive Neural Network model (ViT-RNN) network was used to train the model. The patch-level model was trained to identify predictive patches and obtain patch-level probability distribution based on ViT, which is mapped to the patient-level model to get the final prediction. Results: A total of 234 NSCLC patients with 344 H&E WSIs were included, including 59 patients (25.2%) with grade 2 or higher pneumonitis. There were 141 patients with 190 images in training set, 47 patients with 83 images in validation set and 46 patients with 71 images in test set. The deep pathomics model achieved the area under curve (AUC), sensitivity, and specificity of 0.935, 0.888, and 0.981 in validation set, and 0.907, 0,869 and 0.945 in test set. Conclusion: The pathomics signature based on deep learning has the potential to accurately predict the radioimmune-associated pneumonitis in NSCLC patients receiving combined therapy, which may provide an opportunity to explore the related microenvironment and biological mechanism.
