Deep learning for preoperative prediction of the EGFR mutation and subtypes based on the MRI image of spinal metastasis from primary NSCLC

Abstract

Preoperative prediction of EGFR mutation status and subtypes is essential to choose appropriate treatment strategies and increase the survival of NSCLC patients. However, existing computer-aided predicting methods mainly focused on the primary tumour and rely on traditional handcrafted-based radiomics features and machine learning classifiers, which have inherent limitations due to huge computation complexity and low accuracy. This study made the first attempt to investigate an end-to-end deep learning technique for detecting EGFR mutations and subtypes based on bone metastasis originated from primary NSCLC. A CM-EfNet was proposed by integrating the convolutional block attention module (CBAM) and multi-resolution feature fusion mechanism (MFM) with EfficientNet v2. For detecting EGFR mutations, the proposed CM-EfNet achieved the highest AUCs of 0.851 and 0.764 in primary and external validation cohorts, respectively. For predicting EGFR mutation sites in exon 19 versus exon 21, the CM-EfNet also generated the highest AUCs of 0.711 and 0.687 in primary and external validation cohorts, respectively. Our research offers a potential non-invasive end-to-end imaging tool for preoperative prediction of EGFR mutations and mutation subtypes for metastatic NSCLCs.