High accuracy epidermal growth factor receptor mutation prediction via histopathological deep learning

Area of Science:

• Oncology
• Pathology
• Artificial Intelligence

Background:

• Accurate detection of epidermal growth factor receptor (EGFR) mutations is crucial for guiding tyrosine kinase inhibitor therapy in non-small cell lung cancer (NSCLC).
• Traditional EGFR mutation detection relies on tissue samples, which are often difficult to obtain, potentially delaying or preventing treatment.
• There is a need for non-invasive or less invasive methods for EGFR mutation status prediction.

Purpose of the Study:

• To develop a high-accuracy deep learning model for predicting EGFR mutation status using only routine haematoxylin and eosin (H&E)-stained slides.
• To assess the model's performance in classifying EGFR mutation status and its potential as a pre-screening tool.

Main Methods:

• A convolutional neural network based on ResNet-50 was trained on 226 H&E-stained NSCLC slides (88 with EGFR mutations).
• The model was tested on 100 independent H&E-stained NSCLC slides (50 with EGFR mutations).
• Slide-level classification of EGFR mutation status was performed.

Main Results:

• The model achieved a sensitivity of 76% and specificity of 74% (AUC 0.82) for EGFR mutation prediction.
• A double-threshold approach allowed 33% of patients to be classified with 100% sensitivity and 87.5% specificity.
• Incorporating adenocarcinoma subtype information improved sensitivity to 100% for 37.3% of adenocarcinoma patients.

Conclusions:

• Deep learning models utilizing H&E slides can serve as a rapid, cost-effective pre-screening tool for EGFR mutations in NSCLC.
• This approach can complement molecular detection methods, especially for patients with limited tissue availability.
• The model has the potential to expand treatment opportunities for NSCLC patients by facilitating timely mutation status assessment.