LungRad-BM: Accurately Predicting Brain Metastasis in Non-small Cell Lung Cancer via a Multimodal Deep Learning Framework on ¹⁸F-FDG PET/CT Radiomics.

Authors

Abstract

Brain metastasis (BM) poses a major threat to patients with non-small cell lung cancer (NSCLC), making early risk evaluation a pressing need. To address this challenge, we developed a novel framework, namely LungRad-BM, to achieve non-invasive diagnosis of BM in NSCLC patients by integrating multiregional and multimodal features derived from ¹⁸F-fluorodeoxyglucose positron emission tomography/computed tomography (¹⁸F-FDG PET/CT) radiomics. We retrospectively enrolled 61 treatment-naive NSCLC patients, including 21 with BM and 40 without BM, and manually delineated intra-tumoral region (ITR) and peri-tumoral region (PTR) on PET/CT radiomics to subsequently extract 2,264 radiomic features. Using a combined strategy of feature selection via the Least Absolute Shrinkage and Selection Operator (LASSO) followed by modeling with the TabNet deep learning architecture, we developed models to predict BM with multiregional and multimodal features, respectively, and demonstrated excellent accuracy across different imaging regions and modalities in the testing dataset. Furthermore, by integrating multimodal features from PET and CT, we individually developed two LungRad-BM models for ITR and PTR, with area under the curve values exceeding 0.90. With validation from five-fold cross-validation and an additional independent dataset, and systematic comparison against multiple baseline algorithms, the LungRad-BM models exhibited superior performance. We anticipate that the LungRad-BM integrated models, decoded the ¹⁸F-FDG PET/CT radiomics, will enable non-invasive prediction of BM risk in NSCLC patients and provide feature-level transparency to support individualized screening and early intervention planning.

Tags

Topics