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(Slides)

Back to course

Pdf Summary

Researchers have developed a self-supervised machine learning pipeline called Histomorphological Phenotype Learning (HPL) to analyze histopathological images of lung adenocarcinoma. The pipeline uses unannotated whole slide images (WSIs) to generate a phenotypic dictionary consisting of histomorphological phenotype clusters (HPCs). The researchers tiled 4,472 WSIs from a cohort of 1,007 patients and trained the model to extract features from each tile. These features were then clustered using Leiden clustering to group similar tiles. Representative tiles from each HPC were histopathologically reviewed and assigned descriptions.<br /><br />The researchers found that the HPCs were histologically coherent, had prognostic power, and aligned with molecular modules. They observed differences between prognostically favorable and lethal HPCs in the tumor microenvironment, with lethal HPCs featuring dense and immune "cold" stroma, while favorable HPCs were rich in lymphocytes.<br /><br />To further understand the mechanistic underpinnings of the HPCs, the researchers projected the HPCs onto tissue microarrays (TMAs) to capture regional transcriptomic profiles. They conducted gene set enrichment analysis against hallmark pathways and found correlations between HPC composition and pathways related to proliferation, myogenesis, epithelial-mesenchymal transition, and inflammation.<br /><br />The HPCs can be used for downstream tasks such as survival analysis and patient stratification for adjuvant treatment. The researchers demonstrated the potential utility of their approach for overall prognostication and patient stratification.<br /><br />In conclusion, the researchers developed a self-supervised learning approach using HPL to analyze histological datasets of lung adenocarcinoma. The HPCs generated through this approach were found to have prognostic power and align with molecular phenotypes. These findings have the potential to advance our understanding of the tumor microenvironment and help guide treatment decisions for patients with lung adenocarcinoma.

Asset Subtitle

John Le Quesne

Meta Tag

Speaker John Le Quesne

Topic Pathology & Biomarkers: Artificial Intelligence in Pathology

Keywords

Histomorphological Phenotype Learning

lung adenocarcinoma

histopathological images

whole slide images

phenotypic dictionary

histomorphological phenotype clusters

Leiden clustering

prognostic power

tumor microenvironment

patient stratification