Abstract
BACKGROUND: This study aims to explore the value of habitat-based magnetic resonance imaging (MRI) radiomics for predicting the origin of brain metastasis (BM). PURPOSE: To investigate whether habitat-based radiomics can identify the metastatic tumor type of BM and whether an imaging-based model that integrates the volume of peritumoral edema (VPE) can enhance predictive performance. METHODS: A primary cohort was developed with 384 patients from two centers, which comprises 734 BM lesions. An independent cohort was developed with 28 patients from a third center, which comprises 70 BM lesions. All patients underwent T1-weighted contrast-enhanced (T1CE) and T2-weighted (T2W) MRI scans before treatment. Radiomics features were extracted from tumor active area (TAA) and peritumoral edema area (PEA) selected using the least absolute shrinkage and selection operator (LASSO) to construct radiomics signatures (Rads). The Rads were further integrated with VPE to build combined models for predicting the metastatic type of BM. Performance of the models were assessed through receiver operating characteristic (ROC) curve analysis. RESULTS: Rads derived from TAA and PEA both showed predictive power for identifying the origin of BM. The developed combined models generated the best performance in the training (AUCs, lung cancer [LC]/non-lung cancer [NLC] vs. small cell lung cancer [SCLC]/non-small cell lung cancer [NSCLC] vs. breast cancer [BC]/gastrointestinal cancer [GIC], 0.870 vs. 0.946 vs. 0.886), internal validation (area under the receiver operating characteristic curves [AUCs], LC/NLC vs. SCLC/NSCLC vs. BC/GIC, 0.786 vs. 0.863 vs. 0.836) and external validation (AUCs, LC /NLC vs. SCLC/NSCLC vs. BC/GIC, 0.805 vs. 0.877 vs. 0.774) cohort. CONCLUSIONS: The developed habitat-based radiomics models can effectively identify the metastatic tumor type of BM and may be considered as a potential preoperative basis for timely treatment planning.