Local-global attention network (LGA-Net): An efficient hybrid CNN-transformer model with multi-attention for non-invasive gene mutation prediction in NSCLC

Abstract

To address the challenge of low accuracy in gene mutation prediction for non-small cell lung cancer, we propose the LGA-Net model to integrate a lightweight Transformer with a convolutional neural network (CNN) module. The model predicts Epidermal Growth Factor Receptor (EGFR) and Kirsten Rat Sarcoma Virus Oncogene (KRAS) gene mutations in CT images using the Reinforced Local Perception Module (RLPM) and incorporates an Attention-based lightweight multi-head mechanism transformer (ATLM). Experimental results indicate that the Local-Global Attention Network (LGA-Net) model achieves accuracy rates of 89.56% and 88.29% for EGFR and KRAS mutation predictions, respectively. Ablation experiments validate the effectiveness of each module. Our experiments show that the model significantly enhances prediction performance and presents a promising approach for the non-invasive detection of gene mutations in non-small cell lung cancer.