Self-supervised AI reveals a lethal discohesive phenotype in lung adenocarcinoma

Abstract

Applications of artificial intelligence (AI) to histopathology are now common, but most require supervision which inherently limits their scope. By using self-supervised learning (SSL), we discover and quantify the full range of histopathological appearances in a disease, and associate them with clinicopathological ground truths such as prognosis. We used this approach to discover under-appreciated morphologies of lung adenocarcinoma (LUAD), using a highly characterised resected tumour cohort of over 4000 slides from over 1000 patients. By constructing an authoritative lexicon of recurrent LUAD appearances, we ab initio discovered several stromal morphologies strongly predictive of outcome. With multimodal data integration and external dataset validation, we propose that epithelial discohesion is lethal, but only in the context of immunologically cold stroma. Both these morphological features are independent of current prognostic schema. Crucially, we describe these features in the context of real-world diagnostic histopathology, giving them immediate clinical translatability. Statement of SignificanceHistopathology relies heavily on epithelial morphology, often neglecting the stroma. We use self-supervised AI to identify under-appreciated morphologies in LUAD linked to poor outcome and validate these observations in an external cohort, demonstrating the utility of self-supervised AI as a powerful biological discovery tool.