A Deep Learning Framework with Attention Mechanism for the Classification of Mild Cognitive Impairment using Brain MRI

Mild Cognitive Impairment (MCI) refers to a neurological disorder marked by observable changes in cognitive decline. MCI signifies an intermediate phase linking normal aging and Alzheimer's Disease (AD). It is typically categorized into two types: Early MCI (EMCI) and Late MCI (LMCI), representing varying levels of cognitive impairment. An effective classification between EMCI and LMCI is vital, as it allows timely interventions that can potentially slow the progression of AD, improving patient outcomes and quality of life. Identifying the subtle features and patterns in EMCI and LMCI is essential for early diagnosis and effective intervention. This work proposes a novel Deep Learning (DL) network integrating EfficientNet-B0, a Global Context Network (GCN), and Neighborhood Attention blocks. The EfficientNet-B0 acts as the foundation for feature extraction in the proposed model. It uses its efficient architecture and pre-trained capabilities to capture meaningful patterns. To further enhance feature representation, a Global Context Network (GCN) is employed to capture long-range dependencies and global semantic information. Additionally, a Neighborhood Attention Module is used to refine local features by focusing on important regions in the MRI scans. The proposed framework balances global semantic understanding and localized feature extraction, resulting in improved classification performance. With systematic training and evaluation, the model achieves an accuracy of 95.3%, outperforming existing approaches and providing a better solution for neuroimaging-based MCI subtype classification.