Abstract

e20043 Background: Tumor spread through air space (STAS) is a novel histologic pattern of invasion related to high recurrence and poor prognosis. Meanwhile, there is a potential relationship between STAS status and the tumor immune microenvironment. The aim of this study was to predict STAS status in lung adenocarcinoma (LUAD) using a CT-based 3D-CNN deep learning model and explore the tumor immune microenvironment in the GSE58661 cohort. Methods: We retrospectively collected data from 368 patients with lung adenocarcinoma and STAS status from Guangdong Province General Hospital between 2021 and 2023. We performed propensity score matching (PSM) based on gender, year, BMI, and smoking status to obtain a cohort of 368 patients without STAS status. A 3D-Resnet-Resnetearning model was developed to identify the STAS status of lung adenocarcinoma, and the performance of the models was evaluated using the area under the curve (AUC), accuracy, sensitivity, and specificity. Meanwhile, bule RNA sequencing and baseline CT of LUAD from 41 patients in the GSE58661 cohort and the TCIA dataset were used to investigate the potential biological mechanism in the patients with STAS status. Results: We included 584 and 152 patients in the training and testing sets, respectively. The 3D-Resnet algorithm was used in the 3D patch with voxel sizes of 128*128*128, 96*96*96, and 64*64*64 to construct the DL-STAS model. With the training set's AUC of 0.899, the results showed that the model achieved an AUC, sensitivity, specificity, and accuracy of 0.884, 0.831, 0.838, and 0.835 in the testing cohort, respectively. According to our model, we divided the patients in GSE58661 into STAS and non-STAS. By identifying and analyzing the differentially expressed genes between the STAS and non-STAS groups, we found that several immune-related pathways, such as the T cell receptor signaling pathway, the PD-1/PD-L1 expression pathway, the natural killer cell-mediated cytotoxicity pathway, and the macrophage activation pathway, were enriched in the group with STAS status. Conclusions: The DL-STAS model we developed has great potential for accurately predicting STAS status in lung adenocarcinoma. In the meanwhile, the prognosis of patients with STAS status may be impacted by their particular tumor immune microenvironment.