Deep Learning Predicts EGFR Mutation Status from Histology Images in Non-Small Cell Lung Cancer.

Area of Science:

• Digital Pathology
• Artificial Intelligence in Oncology
• Biomarker Discovery

Background:

• EGFR mutation screening in non-small cell lung cancer (NSCLC) has global variability, creating a significant care gap.
• Deep learning (DL) shows promise for extracting actionable features from histology images, with regulatory approvals in other areas.
• Integrating predictive DL could enhance EGFR mutation screening rates in NSCLC.

Purpose of the Study:

• To develop and validate a DL model for predicting EGFR mutation status from routine hematoxylin and eosin (H&E) stained histology images in NSCLC.
• To assess the model's performance across diverse datasets, including various histologic subtypes, specimen types, and imaging platforms.

Main Methods:

• A DL model, Lunit SCOPE Genotype Predictor, was trained on over 12,000 whole-slide images.
• The model was validated on diverse datasets (n=1,461 and n=599) and a multi-scanner test set (n=2,261).
• Performance was evaluated using the area under the receiver operating characteristic curve (AUROC) across different subgroups.

Main Results:

• The DL model achieved an overall AUROC of 0.905 on the initial validation set.
• Robust performance was observed across specimen types (biopsies: 0.804, resections: 0.912) and histologic subtypes (adenocarcinoma: 0.880).
• The model achieved an AUROC of 0.860 on an independent international test set and demonstrated high concordance across multiple slide scanners.

Conclusions:

• The Lunit SCOPE Genotype Predictor effectively predicts EGFR mutation status from routine histology images in NSCLC.
• The model's validated performance across diverse settings supports its potential application in routine clinical practice.
• This AI-driven approach may help augment molecular EGFR mutation screening and improve biomarker testing rates.