MRA-Net: Multiscale Residual Attention Network for Multiclass Alzheimer Disease Classification

Alzheimer’s Disease (AD) is a prevalent neurodegenerative disorder due to degeneration of brain cells resulting in progressive decline of an individual’s ability to perform activities independently. It has also been established that this cognitive impairment can only be effectively prevented or delayed, by early detection. Magnetic Resonance Imaging is also a valuable noninvasive tool for repeated assessments and is very important as a marker of disease progression. Our proposed methodology is particularly intended for identifying the characteristics of Alzheimer’s disease at the initial stage of its development, as well as for the classification of the stages of the disease. Thus, we propose MRA-Net - an attention-based residual multiscale convolution neural network intended for AD classification. More specifically, here we present a lightweight residual-Net based on the multiscale CNN which represents mainly high-level features of a given input from different receptive fields. Moreover, we place a lightweight attention module (LAM) on the top of Residual-Net to learn and focus on more discriminative features while revealing significant feature dependencies. The performance of the proposed approach is investigated using the OASIS dataset: a relatively high level of accuracy is achieved. The results of experiments show that the proposed MRA-Net is superior to other models, likewise, providing 97.37% accuracy.