Fused multimodal radiomics with deep learning for predicting EGFR-sensitizing mutations based on ¹⁸F- PET/CT images

Epidermal growth factor receptor (EGFR) is one of the most common driver genes in non-small cell lung cancer (NSCLC), including EGFR-resistance mutations and EGFR-sensitizing mutations. Meanwhile, only NSCLC patients with EGFR-sensitizing mutations exhibit a positive response to tyrosine kinase inhibitors (TKIs). Radiomics-based deep learning methods provide noninvasive technologies for predicting EGFR-sensitizing mutations in NSCLC patients. In our work, we fused multimodal radiomics data from 18F-FDG PET/CT images with a deep learning algorithm to predict EGFR-sensitizing mutations in NSCLC patients. Specifically, 202 NSCLC patients scanned with 18F-FDG PET/CT were collected. Traditional radiomics features were extracted by PyRadiomics software, while deep features were extracted by a residual neural network from PET/CT images. We used the least absolute shrinkage and selection operator (LASSO) and support vector machine (SVM) to reduce redundant features and predict EGFR-sensitizing mutations. Additionally, we introduced different deep models and imaging modalities to compare the predictive performance. Our results showed that the fused ResNet-based deep and traditional features demonstrated excellent prediction performance. In addition, multimodal PET/CT radiomics-based methods might predict mutations better than CT-only and PET-only methods.