3D Convolutional Networks Based Automatic Diagnosis of Alzheimer's Disease Using Structural MRI
Early diagnosis of Alzheimer's disease (AD) can help take timely treatment before patients suffer from irreversible brain damage. However, current studies using manually feature extraction and 2D CNN methods cannot make the classification performance optimized. We proposed an approach using 3D convolutional neural network (CNN) to diagnose AD and mild cognitive impairment (MCI), which is the prodromal stage of AD, from structural magnetic resonance images. Subjects of 150 AD, 129 MCI, and 112 normal controls (NC) from the Alzheimer's Disease Neuroimaging Initiative dataset were used in this study. Specifically, weights of a 3-class network were trained to initialize the 2-class network and fine-tuned, which saves time and avoids overfitting. The experimental results show that our 3D-sVoxCNN model significantly improves the classification performance with the accuracy of 95.05% (AD vs MCI vs NC), 99.47% (AD vs. NC), 98.32% (MCI vs. NC), and 98.32% (AD vs. MCI).
Paper
Full text
3D Convolutional Networks Based Automatic Diagnosis of Alzheimer's Disease Using Structural MRI
Semantic Scholar · Medicine · 2019
Abstract
Early diagnosis of Alzheimer's disease (AD) can help take timely treatment before patients suffer from irreversible brain damage. However, current studies using manually feature extraction and 2D CNN methods cannot make the classification performance optimized. We proposed an approach using 3D convolutional neural network (CNN) to diagnose AD and mild cognitive impairment (MCI), which is the prodromal stage of AD, from structural magnetic resonance images. Subjects of 150 AD, 129 MCI, and 112 normal controls (NC) from the Alzheimer's Disease Neuroimaging Initiative dataset were used in this study. Specifically, weights of a 3-class network were trained to initialize the 2-class network and fine-tuned, which saves time and avoids overfitting. The experimental results show that our 3D-sVoxCNN model significantly improves the classification performance with the accuracy of 95.05% (AD vs MCI vs NC), 99.47% (AD vs. NC), 98.32% (MCI vs. NC), and 98.32% (AD vs. MCI).