Alzheimer's Disease Auxiliary Diagnosis System with Enhanced Feature Fusion Based on RGT and LSTM

Alzheimer's disease (AD) has the highest prevalence of all neurological brain disorders, and timely detection can be helpful in significantly ameliorating the symptoms of patients. In this work, we use brain PET images of patients in the OASIS and ADNI databases, and apply deep learning-based computer-aided diagnosis of AD, proposing a model called R-Le-LSTM. This model utilizes pretrained super-resolution reconstruction network (RGT) output to obtain the image reconstruction, applying LSTM networks to strengthen feature extraction in convolutional networks (CNNs). The introduced model allow high-level multi-scale feature fusion at fully connected stage after convolution layers. This work performs brain PET image analysis and AD disease prediction based on this model. Experimental data showing performance of the multi-scale feature fusion method that has high accuracy in the proposed solution. The experimental results show that the proposed model can learn faster than the traditional convolutional networks and LSTM networks, and finally obtain better recognition performance. The methodology established can serve as a good reference for future medical imaging studies of other brain diseases.