Classification of Mild Cognitive Impairment via Improved Multimodal Convolutional Mixer Model

This study presents a novel multimodal deep learning approach for the early detection of Mild Cognitive Impairment (MCI) by integrating neuroimaging and clinical data. Leveraging T1-weighted Magnetic Resonanse Imaging (MRI) and Pitsburgh Compound B Positron Emission Tomography (PiBPET) images alongside cognitive tests, clinical parameters and demographic information, we propose a Multimodal Convolutional Mixer (MCM) architecture that incorporates multi-head self-attention, a custom classifier, and specialized B-cos modules for enhanced interpretability. Data from three major datasets - ADNI, MCSA, and OASIS3 - were preprocessed, extracted slices were fused across sagittal, coronal, and axial planes, and used to train our model. Our MCM model achieved classification accuracies as high as $\mathbf{9 4. 9 2 \%}$ in the sagittal plane, significantly outperforming state-of-the-art models such as ResNet and Vision Transformer (ViT). Ablation studies confirmed that the multihead self-attention module is critical, with its removal resulting in a $\mathbf{1 5 - 2 0 \%}$ performance drop. Feature contribution maps further revealed that the model focuses on central brain regions (e.g., corpus callosum, thalamus, lateral ventricles) in MCI cases, consistent with existing literature linking these areas to cognitive decline. Despite some variability across imaging planes, our results demonstrate that the proposed multimodal approach not only enhances diagnostic accuracy but also provides interpretable insights, offering more effective clinical decision-making in the early detection of MCI.