Diagnosis of Alzheimer’s Disease Based on the Modified Tresnet

• DOI: 10.3390/electronics10161908
• Publication Date: 2021-08-09T09:17:06Z
• Authors: Zelin Xu; Hongmin Deng; Jin Liu; Yang Yang
• Keywords: 0202 electrical engineering, electronic engineering, information engineering; deep learning; 02 engineering and technology; multi-receptive-fields; Alzheimer’s disease; structural MRI

In the medical field, Alzheimer’s disease (AD), as a neurodegenerative brain which is very difficult to diagnose, can cause cognitive impairment and memory decline. Many existing works include variety of clinical neurological psychological examinations, especially computer-aided diagnosis (CAD) methods based on electroencephalographic (EEG) recording or MRI images by using machine learning (ML) combined with different preprocessing steps such hippocampus shape analysis, fusion embedded features, so on, where EEG dataset used for AD usually large complex, requiring extraction series features like entropy spectral feature, etc., it has seldom been applied in detection deep (DL), while were suitable both ML DL. terms structural images, few differences could be found atrophy among three situations: AD, mild (MCI), Normal Control (NC). On other hand, DL have diagnose incorporating recent years, but there not yet many selective models layers. this article, Gray Matter (GM) Magnetic Resonance Imaging (MRI) automatically extracted, better distinguish types situations MCI, NC, compared Cerebro Spinal Fluid (CSF) White (WM). Firstly, FMRIB Software Library (FSL) software utilized batch processing remove skull, cerebellum register heterogeneous SPM + cat12 tool kits MATLAB segment obtaining standard GM images. Next, are trained some new neural networks. The characteristics training process follows: (1) Tresnet, network that achieves best classification effect several networks experiment, selected basic network. (2) A multi-receptive-field mechanism integrated into network, inspired neurons dynamically adjust receptive fields according stimuli. (3) whole realized adding multiple channels convolutional layer, size convolution kernel each channel adjusted. (4) Transfer method train model speeding up optimizing efficiency. Finally, we achieve accuracies 86.9% vs. 63.2% MCI NC respectively, outperform previous approaches. results demonstrate effectiveness our approach.