Multimodal Convolutional Mixer for Mild Cognitive Impairment Detection

Brain imaging is important in detecting Mild Cognitive Impairment (MCI) and related dementias. Magnetic Resonance Imaging (MRI) provides structural insights, while Positron Emission Tomography (PET) evaluates metabolic activity, aiding in the identification of dementia-related pathologies. This study integrates multiple data modalities—T1-weighted MRI, Pittsburgh Compound B (PiB) PET scans, cognitive assessments such as Mini-Mental State Examination (MMSE), Clinical Dementia Rating (CDR) and Functional Activities Questionnaire (FAQ), blood pressure parameters, and demographic data—to improve MCI detection. The proposed improved Convolutional Mixer architecture, incorporating B-cos modules, multi-head self-attention, and a custom classifier, achieves a classification accuracy of 96.3% on the Mayo Clinic Study of Aging (MCSA) dataset (sagittal plane), outperforming state-of-the-art models by 5%–20%. On the full dataset, the model maintains a high accuracy of 94.9%, with sensitivity and specificity reaching 89.1% and 98.3%, respectively. Extensive evaluations across different imaging planes confirm that the sagittal plane offers the highest diagnostic performance, followed by axial and coronal planes. Feature visualization highlights contributions from central brain structures and lateral ventricles in differentiating MCI from cognitively normal subjects. These results demonstrate that the proposed multimodal deep learning approach improves accuracy and interpretability in MCI detection.