2023 World Conference on Lung Cancer (Posters)-P1.20. Mapping the Histopathological Landscape of Lung Adenocarcinoma using Self-Supervised Learning Artificial Intelligence

P1.20. Mapping the Histopathological Landscape of Lung Adenocarcinoma using Self-Supervised Learning Artificial Intelligence - PDF(Abstract)

Back to course

Pdf Summary

In this presentation, the authors introduce a novel self-supervised artificial intelligence (AI) method for mapping the histopathological landscape of lung adenocarcinoma. The method, called Histomorphological Phenotype Learning (HPL), is trained to identify and quantify the underlying tumor morphology without the need for expert annotations.<br /><br />The training pipeline involves the ingestion, normalization, and tiling of whole slide images (WSIs), followed by the Barlow-Twins training of a convolutional neural network that extracts meaningful image features. The vector tile representations are then clustered using Leiden clustering into morphologically similar groups.<br /><br />The HPL method was trained and evaluated on separate collections of lung adenocarcinoma cases from TCGA and NYU. The resulting histomorphological phenotype clusters (HPCs) were examined and classified by expert histopathologists. These HPCs were found to represent consistent and distinct pathological entities, including different varieties of solid pattern growth, normal tissues, and reactive appearances.<br /><br />The authors found that certain HPCs were strongly predictive of patient outcomes. For example, HPCs representing solid pattern growth with few or no tumor infiltrating lymphocytes (TILs) were associated with worse outcomes, while HPCs representing low-grade and/or inflamed patterns had better outcomes. The HPCs associated with worse outcomes were also linked to molecular signatures of proliferation and wound healing, while the HPCs associated with better outcomes were enriched for TIL-specific transcripts.<br /><br />The HPL method has the potential to generate authoritative "foundation models" that can assist in the diagnosis, prognosis, and prediction of lung adenocarcinoma. It can learn the definitive morphological features of the disease from unannotated tissue images and provide valuable insights for further study and research.

Asset Subtitle

John Le Quesne

Meta Tag

Speaker John Le Quesne

Topic Pathology & Biomarkers: Artificial Intelligence in Pathology

Keywords

self-supervised artificial intelligence

histopathological landscape

lung adenocarcinoma

Histomorphological Phenotype Learning

whole slide images

convolutional neural network

histomorphological phenotype clusters

patient outcomes

molecular signatures

unannotated tissue images