2024 World Conference on Lung Cancer (WCLC) - Posters-P1.06A.02 Discovery of Morphomolecular Neighbourhoods in Lung Adenocarcinoma Using Self-supervised Learning

Pdf Summary

The study titled "Discovery of Morphomolecular Neighbourhoods in Lung Adenocarcinoma Using Self-supervised Learning" explores the heterogeneity in lung adenocarcinoma (LUAD) through the application of a self-supervised artificial intelligence technique called Histomorphological Phenotype Learning (HPL). The research aims to improve the current understanding and quantification of morphologically diverse growth patterns in LUAD, which could enhance patient risk stratification and inform therapeutic target development.<br /><br />Using a comprehensive dataset from the Leicester Archival Thoracic Tumour Investigatory Cohort – Adenocarcinoma (LATTICe-A), consisting of 4456 whole slide images from 1025 patients, the researchers processed images to extract features using a ResNet backbone algorithm. Through Barlow Twins and Leiden community detection, they identified histomorphological phenotype clusters (HPCs) that encompass morphologically similar regions of the tumor tissues.<br /><br />The study discovered interesting morphomolecular neighborhoods within tumors, some aligning with known growth patterns and others uncovering new phenotypes with potentially significant prognostic implications. These include patterns of immune cell interaction and fibroblast enrichment not typically considered in clinical practice, therefore presenting possible novel therapeutic targets.<br /><br />The survival analysis employed HPC frequency in a Cox proportional hazards model, revealing clustering patterns associated with overall survival. The research demonstrated that HPL-derived risk scores outperform traditional grading methods in survival prediction, suggesting its value in clinical decision-making.<br /><br />Additionally, the study deployed tissue microarrays and high-plex multiplex immunofluorescence panels for further cell phenotyping and bulk-RNA sequencing, facilitating a detailed molecular characterization of the discovered morphomolecular patterns.<br /><br />In summary, the research highlights the potential of self-supervised AI models to reveal insightful and actionable histological features in LUAD that could improve prognostic assessments and therapeutic strategies.

Asset Subtitle

Kai Rakovic

Meta Tag

Speaker Kai Rakovic

Topic Pathology & Biomarkers

Keywords

lung adenocarcinoma

self-supervised learning

Histomorphological Phenotype Learning

morphological diversity

patient risk stratification

ResNet algorithm

histomorphological phenotype clusters

prognostic implications

Cox proportional hazards model

molecular characterization