2023 World Conference on Lung Cancer (Posters)-P1.20. Predictability of Actionable Mutations in NSCLC using Attention-Based Multiple-Instance Learning on H&E Images

P1.20. Predictability of Actionable Mutations in NSCLC using Attention-Based Multiple-Instance Learning on H&E Images - PDF(Abstract)

Back to course

Pdf Summary

This document discusses a study that explores the use of artificial intelligence (AI) in predicting actionable mutations in non-small cell lung cancer (NSCLC) using H&E histology images. The study aims to overcome limitations in acquiring sufficient tumor tissue and the prohibitive costs of genomic testing by predicting cancer biomarkers from readily available histology images. The researchers used the Cancer Genome Atlas (TCGA) lung adenocarcinoma dataset as a discovery cohort for predicting KRAS, KRAS-G12C, and EGFR mutations. They also used an in-house dataset from St. James's Hospital for external validation. The models were trained using attention-based multiple-instance learning (MIL) and evaluated using the area under the receiver operating characteristic (AUROC) curve.<br /><br />The results show that KRAS-MUT mutation status could be predicted with a mean AUROC of 0.607 across 5 folds in the TCGA dataset, with the best-performing fold achieving an AUROC of 0.71. The external validation on the St. James's Hospital cohort achieved an AUROC of 0.575. However, KRAS-G12C mutations could not be distinguished from non-G12C KRAS mutations in the TCGA dataset. EGFR mutation status, on the other hand, could be predicted with a mean AUROC of 0.707 in the TCGA training set, with the best-performing fold achieving an AUROC of 0.780. The validation on a held-out set of samples achieved an AUROC of 0.839.<br /><br />The study concludes that the small number of oncogene-positive samples in existing datasets limits the ability to reliably train machine learning models in digital pathology. Out of the three mutations studied, only EGFR could be predicted with an AUROC greater than 0.75 using attention-based MIL. However, further validation in an external cohort is necessary. The study highlights the importance of developing larger datasets to improve the predictability of mutations using AI in digital pathology. The keywords associated with the study include digital pathology, molecular diagnostics, and deep learning.

Asset Subtitle

Pierre Murchan

Meta Tag

Speaker Pierre Murchan

Topic Pathology & Biomarkers: Artificial Intelligence in Pathology

Keywords

artificial intelligence

predicting mutations

lung cancer

H&E histology

cancer biomarkers

TCGA dataset

KRAS mutations

EGFR mutations

multiple-instance learning

digital pathology