Tackling Tumor Heterogeneity Issue: Transformer-Based Multiple Instance Enhancement Learning for Predicting EGFR...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate prediction of epidermal growth factor receptor (EGFR) mutations is vital for non-small cell lung cancer (NSCLC) diagnosis and treatment.
• Computed tomography (CT) imaging is a promising tool, but current fully supervised methods struggle with tumor heterogeneity, limiting prediction accuracy.

Purpose of the Study:

• To develop a novel weakly supervised method, TransMIEL, to accurately predict EGFR mutations in NSCLC by addressing tumor heterogeneity.
• To improve the discriminative power of image features and enhance tumor representation for better diagnostic accuracy.

Main Methods:

• Introduced TransMIEL, a weakly supervised method utilizing multiple instance learning for EGFR mutation prediction.
• Developed an instance enhancement learning (IEL) strategy with self-derived pseudo-labels to strengthen instance features.
• Designed a spatial-aware transformer (SAT) to capture inter-instance relationships and an instance adaptive gating (IAG) module for dynamic feature aggregation.

Main Results:

• TransMIEL significantly outperformed existing fully and weakly supervised methods on both public and in-house NSCLC datasets.
• The method demonstrated effectiveness in highlighting intra-tumor and peri-tumor areas relevant to EGFR mutation status through visualization.
• TransMIEL showed improved model generalization performance in predicting EGFR mutations.

Conclusions:

• TransMIEL offers a powerful and accurate approach for non-invasive EGFR mutation prediction in NSCLC, effectively handling tumor heterogeneity.
• The developed method provides a novel perspective for future research on tumor heterogeneity in medical image analysis.
• TransMIEL holds significant potential as an effective tool for clinical application in NSCLC management.