Artificial Intelligence-Powered Spatial Analysis of the Tumor Microenvironment in Pulmonary Lymphoepithelial Carcinoma.

Lymphoepithelial carcinoma (LEC) can occur in various organs, such as the lung, nasopharynx, and thymus. We investigated the spatial characteristics of the tumor immune microenvironment (TIME) among primary pulmonary LECs (pLECs), pulmonary metastatic nasopharyngeal carcinomas (pmNPCs), and thymic LECs (tLECs). In this retrospective study, a total of 160 surgically resected LEC cases, comprising 116 pLECs, 26 tLECs, and 18 pmNPCs, were included. The TIME features, based on hematoxylin and eosin (H&E)-stained whole-slide images (WSIs) and multiplexed immunofluorescence staining images, were obtained and their association with patient prognosis was analyzed. We developed a semisupervized model for automated tumor segmentation based on H&E WSIs. The performance of the model was robust, with a mean accuracy rate of 0.847 on the testing set. Subsequent TIME analysis revealed different spatial distribution patterns of lymphocytes on H&E WSIs among pLECs, tLECs, and pmNPCs. Lymphocyte count and distribution were prognostically relevant in pLECs, with an increasing trend of lymphocytes from the peripheral normal lung area to the tumor core in patients with a good prognosis. Further TIME analysis based on multiplexed immunofluorescence images uncovered that spatial arrangement and spatial interaction pattern characteristics were dependent on specific tumor types and cell subtypes. Our semisupervized learning model offers an automated and reproducible method for tumor segmentation for the TIME of rare LECs. Our analysis revealed different TIME patterns that distinguish among pLEC, tLEC, and pmNPC and demonstrates that the spatial arrangement and positional interaction patterns of PDL1 tumor cells and FOXP3 regulatory T cells could stratify prognosis in patients with pLEC.