EGFR Mutation Detection in Whole Slide Images of Non-Small Cell Lung Cancers Using a Two-Stage Deep Transfer Learning...

Area of Science:

• Oncology
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Lung cancer (LC) is a leading cause of cancer mortality globally, with non-small cell lung cancer being the most prevalent subtype.
• Current LC treatment relies on molecular profiling, particularly detecting Epidermal Growth Factor Receptor (EGFR) gene mutations, which are critical for targeted tyrosine kinase inhibitor therapy.

Purpose of the Study:

• To evaluate the efficacy of Convolutional Neural Networks (CNNs) in distinguishing healthy lung tissue from cancerous tissue.
• To assess the potential of CNNs in identifying EGFR-mutated tumor tissue from standard histopathological slides.
• To utilize Explainable AI (Grad-CAM) for visualizing the decision-making process of the CNN models.

Main Methods:

• Two InceptionResNet-V2 based CNNs were applied to Whole Slide Images (WSIs) of lung cancer tissue.
• An Explainable AI technique, Grad-CAM, was integrated to provide visual insights into model predictions.
• The study analyzed 259 lung cancer cases from three distinct Italian centers.

Main Results:

• The CNNs achieved high accuracy (96.17%) in differentiating healthy from cancerous tissue, with an AUC of 0.99.
• For EGFR mutation detection within cancer tissue, the models reached an accuracy of 76.67% and an AUC of 0.77.
• Explainable AI provided visual insights into the model's classification process.

Conclusions:

• The tested CNNs demonstrate significant potential in aiding lung cancer diagnosis, particularly in differentiating tumor from healthy tissue.
• Directly predicting EGFR mutational status from routine H&E slides using these AI models presents challenges.
• The findings suggest that valuable predictive information for LC diagnosis can be extracted from standard histopathological slides via AI analysis.