LM3DFN: An end-to-end model for non-invasive prediction of EGFR mutation in non-small cell lung cancer

Objective

To explore the feasibility of constructing a deep learning model based on chest CT images to predict epidermal growth factor receptor (EGFR) mutation in non-small cell lung cancer (NSCLC), providing an innovative solution for non-invasive molecular typing.

Methods

This study retrospectively included 623 pathologically confirmed NSCLC patients admitted to our hospital from January 2020 to December 2024 (EGFR mutant: 326 cases, 52.3 %; EGFR non-mutant: 297 cases, 47.7%). All cases had complete CT images and EGFR test results. A Lightweight Multimodal 3D Fusion Network (LM3DFN) deep learning framework was developed, incorporating an attention mechanism to enhance key regional image features and integrate critical imaging information. The dataset was randomly divided into a training set (467 cases) and a test set (156 cases) in a 3:1 ratio. Model performance was evaluated using multi-dimensional metrics, including accuracy (ACC), precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC).

Results

The LM3DFN model demonstrated excellent predictive performance in the test set (ACC = 0.836[0.818–0.854], Precision = 0.825[0.813–0.863], Recall = 0.779[0.751–0.802], F1 = 0.801[0.772–0.834], AUC = 0.889[0.885–0.923]). Visualization of attention analysis indicated a correlation between EGFR mutations and tumor texture and grayscale.

Conclusion

This study confirmed that the LM3DFN model can effectively mine phenotypic features in CT images related to EGFR mutations, providing a non-invasive and reproducible alternative for molecular typing in clinical practice. This model is particularly suitable for dynamic monitoring of gene status evolution during targeted therapy, offering important technical support for the optimization and translational application of precision diagnosis and treatment systems for lung cancer.