MssNet: An Efficient Spatial Attention Model for Early Recognition of Alzheimer's Disease

Abstract

Deep learning models are widely used in medical image-guided disease recognition and have achieved outstanding performance. Voxel-based models are typically the default choice for deep learning-based MRI analysis, which require high computational resources and large data volumes, making them inefficient for rapid disease screening. Simultaneously, the existing Alzheimer's disease (AD) recognition model is primarily comprised of Convolutional Neural Network (CNN) structures. With the increasing of the network depth, the fine-grained details of global features tend to be partially lost. Therefore, we propose a Multi-scale spatial self-attention Network (MssNet) that effectively captures both coarse-grained and fine-grained features. We design to select the target slice based on image entropy to achieve efficient slice-based AD recognition. To capture multi-level spatial information, a novel spatial attention mechanism and spatial self-attention attention are designed. The former is utilized to collect critical spatial information and identify areas that are likely to be lesions, the latter investigates the relationship between features in different image regions through spatial interaction by pure convolutional blocks. MssNet fully utilizes multi-scale information at different granularities for spatial feature interaction, providing it with strong modeling and information understanding capabilities. It has achieved excellent performance in the recognition tasks of Alzheimer's Disease Neuroimaging Initiative (ADNI) and Open Access Series of Imaging Studies (OASIS) datasets. Moreover, MssNet is a lightweight model involving lower scale parameters against the Voxel-based ones, while demonstrating strong generalization capability.