Predicting EGFR mutational status from pathology images using a real-world dataset.

Area of Science:

• Computational pathology
• Oncology
• Artificial intelligence in medicine

Background:

• Non-small cell lung cancer (NSCLC) treatment is increasingly driven by biomarkers, such as epidermal growth factor receptor (EGFR) mutations.
• Targeted therapies offer significant benefits for patients with specific genomic alterations.

Purpose of the Study:

• To develop and validate deep learning algorithms for assessing EGFR status from hematoxylin and eosin (H&E) stained images of lung adenocarcinoma.
• To evaluate the performance of an attention-based algorithm in a real-world cohort with diverse morphology and low tumor content.

Main Methods:

• Development of a set of algorithms, including an attention-based model, to analyze EGFR status and morphology.
• Utilized a cohort of 2099 patients with advanced lung adenocarcinoma and H&E images.
• Validated the best-performing algorithm on a separate cohort with known EGFR mutation prevalence.

Main Results:

• The attention-based EGFR algorithm achieved an AUC of 0.870, NPV of 0.954, and PPV of 0.410.
• The attention model outperformed a heuristic-based model, utilizing both tumor and non-tumor regions.
• High-attention regions correlated with solid growth patterns and increased peritumoral immune presence for predicted EGFR negativity.

Conclusions:

• Deep learning algorithms show potential for rapid, image-based screening of biomarker alterations in NSCLC.
• This approach may help prioritize tissue usage for essential biomarker testing, optimizing resource allocation.
• The algorithm's ability to extract signal from non-tumor regions suggests a broader utility in pathological assessment.