Multimodal Brain Radiomics for Predicting Pathological Subtypes of Lung Cancer Brain Metastases: Impact of Imaging Modality, MRI Sequence, and Spatial Extent.
Authors
Affiliations (3)
Affiliations (3)
- Department of Radiology, Harbin Medical University Cancer Hospital, Harbin, Heilongjiang, China (J.G.).
- Department of Radiation Oncology, Harbin Medical University Cancer Hospital, Harbin, Heilongjiang, China (J.X., H.L., Z.Z, C.W., J.L., S.H., S.L., J.M., H.W.).
- Department of Radiation Physics, Harbin Medical University Cancer Hospital, Harbin, Heilongjiang, China (H.W.). Electronic address: [email protected].
Abstract
To develop and validate a brain imaging-based multimodal radiomics model for noninvasive prediction of pathological subtypes in lung cancer patients with brain metastases, and to evaluate the effects of imaging modalities, computed tomography, contrast-enhanced T1-weighted imaging, T2-weighted imaging (CT, CE-T1WI, T2WI), and spatial ranges (tumor vs whole-brain) on model performance for optimizing diagnosis and treatment strategy. A total of 255 lung cancer patients with pathologically confirmed brain metastases were retrospectively enrolled and divided into training and test cohorts at a 7:3 ratio. Radiomic features were extracted from tumor and whole-brain regions on CT, CE-T1WI, and T2WI after image preprocessing. Feature selection was performed via Z-score normalization, t-test, Pearson correlation, minimum redundancy maximum relevance, and least absolute shrinkage and selection operator. Single-modal, spatial-specific, and CT-magnetic resonance imaging (MRI) multimodal models were constructed using multiple machine learning algorithms. Model performance was validated by receiver operating characteristic-area under the curve (AUC), DeLong test, Brier score, calibration curve, decision curve analysis, integrated discrimination improvement, and net reclassification improvement, with SHapley Additive exPlanations adopted for model interpretation. The multimodal fusion model yielded the highest AUC of 0.883 in the test cohort with good calibration and clinical net benefit. The T1WI model outperformed the T2WI, and whole-brain-based models were significantly superior to tumor-region models. Spatial range exerted a greater impact on predictive performance than the imaging sequence. Accordingly, an imaging-clinical integrated model (AUC = 0.890) was established with the objective of further enhancing the model's comprehensive predictive performance and clinical applicability. The CT-MRI multimodal radiomics model accurately predicts pathological subtypes of lung cancer brain metastases. Spatial extent is a critical determinant for model performance.