Abstract
<title>Abstract</title> <p>Background Preoperative differentiation between glioblastoma (Gb) and solitary brain metastasis (SBM) is clinically important because treatment planning and prognosis differ substantially. This study aimed to develop and externally test habitat-based radiomics models using conventional morphologic MRI and multiple diffusion MRI techniques, and to assess whether multimodal fusion improved diagnostic performance. Methods This retrospective two-center study included 253 patients with histologically confirmed Gb or SBM. Patients from Center A were divided into a training set (n = 142) and an internal test set (n = 60), while patients from Center B formed an independent external test set (n = 51). Radiomics features were extracted from tumor burden volume, peritumoral edema, and abnormal burden volume (ABV). Imaging sources included conventional morphologic MRI, DWI, DTI, DKI, NODDI, and MAP-MRI. Single-modality and multimodal fusion models were developed for each habitat. All data-dependent preprocessing, feature selection, imbalance correction, classifier selection, and hyperparameter tuning were restricted to the training set. Discrimination was assessed using the area under the receiver operating characteristic curve (AUC). External performance was primarily evaluated in the unharmonized external test set, with ComBat harmonization performed separately as a domain-adaptation sensitivity analysis. Calibration, Brier score, and decision curve analysis were assessed in the internal test set. Results Single-modality model performance varied across habitats, imaging sources, and test sets. Multimodal fusion models showed higher observed internal-test AUCs than the corresponding best single-modality models across the three habitats. Among the fusion models, the ABV fusion model achieved the highest observed AUC, reaching 0.940 (95% CI, 0.884–0.996) in the internal test set and 0.839 (95% CI, 0.708–0.969) in the unharmonized external test set. In the ComBat-harmonized domain-adaptation sensitivity analysis, the external-test AUC increased to 0.917 (95% CI, 0.829–1.000). The ABV fusion model also had the lowest Brier score among the evaluated ABV-based models and showed favorable net benefit on decision curve analysis. Conclusions Habitat-based multimodal radiomics integrating conventional morphologic MRI and diffusion MRI showed potential for differentiating Gb from SBM. In this dataset, the ABV fusion model showed the highest observed performance and may provide adjunctive imaging-based support for preoperative diagnostic decision-making.</p>