NeuroAlign: Hierarchical Multimodal Fusion of Dynamic and Structural Neuroimaging for MCI Analysis

NeuroAlign presents a hierarchical machine learning framework that fuses functional MRI and diffusion tensor imaging data to improve detection of mild cognitive impairment. The system introduces novel alignment and interaction mechanisms between multimodal neuroimaging datasets, with a new attribution method for interpretability, demonstrating competitive results across multiple medical imaging datasets.

NeuroAlign addresses a critical challenge in neuroimaging analysis: combining complementary but heterogeneous data sources from brain imaging. Functional MRI reveals dynamic neural activity patterns while diffusion tensor imaging maps structural white matter connectivity. Traditional approaches struggle with alignment across these fundamentally different modalities, limiting their diagnostic accuracy for cognitive disorders. This research proposes a structured solution through hierarchical alignment that bridges the representation gap between dynamic and static imaging, then enables fine-grained interaction between connectivity and regional features.

The framework emerges from broader advances in multimodal machine learning and medical AI, where fusion of diverse data types increasingly outperforms single-modality approaches. Early cognitive impairment detection represents a high-value clinical problem, as identifying mild cognitive impairment before progression to dementia enables preventive interventions. The research's validation across three separate datasets (GUTCM, ADNI, OASIS) and preliminary cross-dataset transferability suggests practical clinical utility beyond controlled laboratory conditions.

For the medical AI sector, NeuroAlign demonstrates the continued evolution of interpretable deep learning in healthcare. The Synergistic Activation Mapping method addresses a persistent concern: black-box neural networks provide predictions without clear reasoning. By revealing modality-specific and consistent brain patterns driving classifications, the system builds clinician trust and supports hypothesis generation about disease mechanisms.

Future development will likely focus on clinical validation pipelines and regulatory pathways for such systems. The cross-dataset transferability hints at potential for deployment across hospital networks with varying imaging protocols, though standardization challenges remain significant.

• →Hierarchical multimodal fusion framework improves detection of mild cognitive impairment by aligning functional and structural brain imaging data.
• →Novel Synergistic Activation Mapping method provides interpretable attribution for deep learning predictions in medical imaging without requiring gradients.
• →Validation across three independent datasets demonstrates competitive performance and preliminary cross-dataset generalization capability.
• →Framework reveals modality-specific brain patterns, providing evidence for how different neuroimaging types contribute to cognitive impairment diagnosis.
• →Advances in medical AI interpretability address clinician concerns about black-box decision-making in diagnostic applications.