A Multiparametric MRI‑Based Deep Learning Model for Binary Triage of Lung Versus Non‑lung Brain Metastases: A...

Identifying the origin of brain metastases (BM) is essential for personalized treatment, particularly in patients with carcinoma of unknown primary or multiple primary neoplasms. We aimed to develop a multiparametric MRI-based three-dimensional (3D) deep learning (DL) fusion model for noninvasive differentiation of lung cancer (LC) versus non-lung cancer (NLC) origin BM. A total of 458 patients with 585 BM lesions from three centers were retrospectively included. A 3D DL model was developed using paired CE-T1 and T2-FLAIR sequences. The model employs a shared-weight Siamese encoder to extract cross-modal spatial features, combined with spatial attention mechanisms that highlight diagnostically relevant tumor regions. To capture intratumoral heterogeneity, Local Moran's I spatial autocorrelation analysis was used to partition each lesion into four biologically interpretable habitat subregions, generating subregion-level feature tokens. Cross-modal attention fusion was then applied to integrate complementary information from both sequences at both whole-lesion and subregion levels. Model robustness was evaluated using sensitivity analyses across histologic subtypes and at the patient level. Ablation studies were conducted to compare multimodal and single-modality performance and a multi-reader study to assess the model's clinical utility in assisting radiologists. The fusion model achieved an area under the receiver operating characteristic curve (AUC) of 0.925, significantly outperforming single-modality models using CE-T1 (AUC = 0.817) and T2-FLAIR (AUC = 0.802). Sensitivity analyses demonstrated robust performance across all subtypes (AUCs, 0.895-0.956). The patient-level analysis yielded an AUC of 0.942. Radiologists' performance improved with model assistance (accuracy trainee, 0.500-0.583; experienced, 0.595-0.679; expert, 0.690-0.845). The proposed 3D DL fusion model demonstrated strong performance in differentiating LC from NLC BM and improved radiologists' accuracy, supporting its potential as a noninvasive decision-support tool for BM origin assessment.