A Radiogenomics Ensemble to Predict EGFR and KRAS Mutations in NSCLC.

Area of Science:

• Oncology
• Bioinformatics
• Machine Learning

Background:

• Lung cancer is a leading global cause of cancer death.
• Accurate detection of EGFR and KRAS mutations is crucial for effective lung cancer treatment.
• Non-invasive methods for mutation detection and limited public datasets hinder classifier performance.

Purpose of the Study:

• To develop an ensemble approach for enhanced prediction of EGFR and KRAS mutations using limited data.
• To introduce and evaluate a novel voting scheme, Selective Class Average Voting (SCAV).
• To assess the performance improvement of machine learning and Convolutional Neural Networks (CNNs) with SCAV.

Main Methods:

• Application of an ensemble learning strategy to small datasets.
• Development and implementation of the Selective Class Average Voting (SCAV) scheme.
• Evaluation of SCAV with both traditional machine learning models and CNNs for mutation prediction.

Main Results:

• For EGFR mutations, machine learning sensitivity increased from 0.66 to 0.75, and AUC from 0.68 to 0.70. Deep learning achieved an AUC of 0.846, with SCAV boosting accuracy from 0.80 to 0.857.
• For KRAS mutations, AUC increased significantly in machine learning models (0.65 to 0.71) and deep learning models (0.739 to 0.778).
• The ensemble approach with SCAV demonstrated improved performance for both EGFR and KRAS mutation prediction.

Conclusions:

• Ensemble methods, particularly with SCAV, effectively enhance the predictive performance of machine learning classifiers and CNNs on small datasets.
• This approach offers a viable strategy for predicting EGFR and KRAS mutations, augmenting clinical capabilities.
• The findings support the potential of these tools as large datasets become available.