End-to-end Prediction of EGFR Mutation Status with Denseformer

Area of Science:

• Radiology
• Oncology
• Artificial Intelligence

Background:

• Accurate epidermal growth factor receptor (EGFR) genotyping is crucial for lung adenocarcinoma treatment.
• Current methods like biopsy and sequence testing are invasive and complex.
• Deep learning with CT imagery offers a non-invasive alternative, but limitations exist.

Purpose of the Study:

• To propose a Denseformer framework for end-to-end, non-invasive EGFR mutation status identification directly from 3D lung CT images.
• To overcome limitations of manual tumor annotation and insufficient feature exploitation in existing methods.
• To leverage whole-lung CT data, considering the association of EGFR status with the broader lung microenvironment.

Main Methods:

• Developed a Denseformer framework integrating Convolutional Neural Network (CNN) and Transformer architectures.
• Utilized 3D whole-lung CT images as direct input, eliminating the need for manual nodule annotation.
• Incorporated a combined Transformer module for global integration of multi-level and multi-layer features.

Main Results:

• The Denseformer framework demonstrated effective feature extraction from 3D CT images for accurate EGFR mutation status prediction.
• The model achieved superior performance compared to state-of-the-art methods using only CT imaging.
• The approach successfully predicted EGFR mutation status without invasive procedures or manual annotations.

Conclusions:

• The proposed Denseformer framework offers an accurate, non-invasive method for EGFR genotyping in lung adenocarcinoma using 3D CT scans.
• This deep learning approach eliminates the need for manual tumor segmentation, reducing time and subjectivity.
• Denseformer represents a significant advancement in predicting EGFR mutation status from medical imaging.