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

Background: Lung cancer (LC) is the leading cause of cancer death worldwide. Non-small cell lung cancer is the most frequent and includes adenocarcinoma and squamous cell carcinoma. Currently, LC treatment is based on tumor molecular profiling. LC may display Epidermal Growth Factor Receptor (EGFR) gene mutation. Detecting mutations in the EGFR gene is crucial for the tyrosine kinase inhibitory therapy. Methods: This study used a computer-based methodology with two Convolutional Neural Networks (CNNs) based on InceptionResNet-V2, applied to Whole Slide Images, to distinguish healthy from cancerous tissue and then EGFR mutated tumor tissue samples. We also integrated an Explainable AI technique (Grad-CAM) to clearly visualize insights into the model's decision-making process. The analysis was conducted on 259 LC cases collected from three different centers (Florence, Rome, and Sassari). Results: This methodology achieved an accuracy of 96.17% in distinguishing healthy from cancerous tissue, with specificity of 87.89%, sensitivity of 98.43%, an F1 score of 97.59% and an AUC of 0.99. Additionally, Cohen's Kappa indicated a consistency of 0.7982, and the confusion matrix showed a correct classification rate of 96.2%. For EGFR mutation detection in cancer tissue, slide-level performance after aggregation reached an accuracy of 76.67% with specificity of 80.77%, sensitivity of 73.53%, an F1 score of 78.12%, a consistency of 0.5583 of Cohen's Kappa and an AUC of 0.77. The confusion matrix showed 76.7% as a correct classification rate. Conclusion: The two tested CNNs showed potential for assisting LC diagnosis, especially in distinguishing healthy from tumor tissue. While the direct detection of EGFR mutational status remains challenging, the results suggest that relevant predictive signals can still be extracted from routine H&E slides.