ESC-Net: A Lightweight EfficientNetV2-Based Framework with Coordinate Attention and Global Context Enhancement for Alzheimer's Disease Auxiliary Diagnosis

Abstract

Alzheimer's disease (AD) is a progressive neurodegenerative disorder, and early identification is crucial for delaying disease progression. Existing MRI-based diagnostic methods still face challenges in computational efficiency, model lightweighting, and multi-stage classification accuracy. This paper proposes a lightweight auxiliary diagnostic model named ESC-Net based on an improved EfficientNetV2 for three-stage classification of cognitively normal (CN), mild cognitive impairment (MCI), and AD. The model adopts a stage-wise heterogeneous convolutional design. Shallow layers use FusedMBConv to enhance training efficiency, while deep layers integrate a Coordinate Attention (CA) mechanism into MBConv to capture both channel relationships and spatial dependencies, improving localization of key pathological regions such as the hippocampus. A progressive stochastic depth regularization strategy is introduced to mitigate overfitting in small-sample medical imaging data. Experimental results on the ADNI dataset show that the proposed model achieves a sensitivity of 99.54% for AD diagnosis, an accuracy of 98.67% for early MCI identification, and a specificity of 98.02% for distinguishing MCI from CN. Compared with traditional deep convolutional networks, this model significantly reduces computational complexity while maintaining excellent classification performance, demonstrating promising potential for clinical application and mobile deployment.