Study on MRI Slices-based Lightweight Neural Network in Alzheimer's Disease Detection

Currently, early detection and intervention are still the key to prevent the aggravation of Alzheimer's disease (AD). Computer assisted diagnosis can provide great help for the early detection of AD. However, compared with natural images, medical images are generally difficult to collect and always have small data sets. To solve this problem, we propose a magnetic resonance imaging (MRI) slices-based lightweight neural network to effectively identify AD. In the proposed method, we use ShuffleNet as the backbone network for feature extraction. Efficient channel attention is added to capture the information of cross-channel interaction information. Triplet loss and cross entropy loss in deep metric learning are combined to ensure the multi-slices under the same subject are classified consistently. The experimental results indicate that, compare with state-of-the-art lightweight methods, our model achieves better classification performance with only 1.52M parameters and 5.56G FLOPs. Experiments verify that our method is helpful to reduce the number of parameters, FLOPs and improve the model performance. Through the construction of the model, it is expected to contribute to the improvement of AD intelligent detection on devices with limited computing power.