Prediction of brain metastases from nonsmall cell lung cancer using multimodal radiomics and deep learning based on 18FPET/CT images

Brain metastasis (BM) is one of the most common malignant forms of non-small cell lung cancer (NSCLC) and has a poor prognosis. Additionally, there is only 3-5 months of median progressionfree survival after BM. Therefore. Early detection and treatment of NSCLC brain metastasis would help to develop more personalized treatment projects and improve patient survival rates. This study aims to investigate the feasibility of a multimodal radiomics-based method using 3D neural networks to predict the occurrence of BM in NSCLC patients based on
¹⁸
F-FDG PET/CT images. We collected 229 NSCLC patients who underwent both
¹⁸
F-FDG PET/CT scans, including lung and brain parts. Shallow lung tumor features and deep brain features were extracted using the python package PyRadiomics and a 3D neural network, respectively. Least absolute shrinkage and selection operator (LASSO) regression was conducted to select related radiomics signatures, and a support vector machine (SVM) was used to predict whether NSCLC patients would develop BM. Our results showed that the fused lung shallow and brain deep radiomics features demonstrated superior predictive performance with an area under the curve (AUC) of 0.95±0.01. In summary, our findings suggested that fused lung-brain features may significantly enhance predictive performance in recognizing whether NSCLC patients would develop BM, with potential clinical applications.