Graph perceiver network for lung tumor and premalignant lesion stratification from histopathology

Abstract

AbstractBronchial premalignant lesions (PMLs) precede the development of invasive lung squamous carcinoma (LUSC); however, it is challenging to distinguish between PMLs that will progress to LUSC from those that will regress without intervention. We developed a computational pipeline using H&E whole slide images (WSIs) to stratify PMLs on the spectrum from normal to tumor tissue. Implementing a study across four data cohorts, including lung tumor resection tissue and endobronchial biopsies, our network distinguished carcinoma in situ biopsies by progression status, indicating that WSIs contain important information that could be used in combination with molecular markers to predict progression to cancer.