DenseGated-MSNet: A Multi-Sequence Gated Fusion Network for Multiple Sclerosis Classification from Brain MRI

Background/Objectives: Multiple sclerosis (MS) is a chronic demyelinating disease of the central nervous system, and magnetic resonance imaging (MRI) plays a central role in its diagnosis and monitoring. Because different MRI sequences provide complementary information about MS-related pathology, jointly exploiting multiple sequences may improve automated classification. However, many existing automated MS classification studies rely on limited sequence information, are not consistently evaluated under subject-independent protocols, and are constrained by the scarcity of publicly available multi-sequence datasets organized at the subject level. Methods: To address these gaps, we collected and publicly released a new multi-sequence brain MRI dataset and proposed DenseGated-MSNet, a novel end-to-end multi-input deep learning model for MS versus non-MS classification. DenseGated-MSNet processes each co-acquired MRI sequence in parallel through an independent sequence-specific DenseNet-121 branch and integrates the resulting representations through a proposed input-dependent gated feature-level fusion mechanism. Under a strict subject-independent evaluation protocol, the model configuration was determined on a subject-independent development set, and the performance of the final model was internally evaluated using five-fold subject-independent cross-validation on the same cohort at the image and subject levels. Results: In internal five-fold subject-independent cross-validation, DenseGated-MSNet achieved a pooled subject-level accuracy of 90.57% (95% CI: 86.44–93.53%) and an AUC of 0.947 (95% CI: 0.918–0.976). The learned gate values also provided sequence-level modulation information, indicating how strongly each MRI sequence was weighted during multi-sequence integration. Conclusions: Overall, these findings suggest that DenseGated-MSNet can effectively integrate complementary information from multiple MRI sequences for discriminating established MS from selected non-MS controls, while providing sequence-level modulation information. Together with the released dataset, the proposed framework provides a reproducible basis for future subject-independent MS classification studies.