MNA-net: Multimodal Neuroimaging Attention-Based Architecture for Cognitive Decline Prediction

Jamie Vo, Naeha Sharif, Ghulam Mubashar Hassan

Abstract: Early detection of Alzheimer’s Disease (AD) is imperative for timely intervention and improving patient prognosis. Predicting whether a Cognitively Normal (CN) patient will develop Mild Cognitive Impairment (MCI) or AD is particularly beneficial. Recently, several studies have utilized Convolutional Neural Networks (CNN) which integrate Magnetic Resonance Imaging (MRI) and Positron Emission Tomography (PET) for classifying MCI and AD. However, in these studies, the integration of MRI and PET features is primarily accomplished by simple concatenation, which fails to facilitate meaningful cross-modal interactions. This paper proposes MNA-net, a multimodal neuroimaging attention-based model designed to predict the conversion of CN individuals to MCI or AD within next 10 years. Addressing a gap in current research, MNA-net leverages attention mechanisms to integrate MRI and PET images, enhancing prediction accuracy. Tested on the OASIS-3 dataset, MNA-net achieved 82.9% accuracy, an 80% true negative rate, and an 85.7% true positive rate. The results highlight the effectiveness of attention mechanisms in fusing neuroimaging modalities and demonstrate the advantages of patch-based feature extraction to enhance the prediction of cognitive impairment.