A convolutional neural network based deep learning model was constructed to distinguish between Alzheimer disease (AD) and behavioral variant frontotemporal dementia (bvFTD) in individual patients, and to improve the accuracy of their differential diagnosis.

High-resolution 3D structural brain images of 47 probable AD and 39 probable bvFTD were acquired by a 3.0 Tesla MRI scanner. According to the AAL2 template and Harvard-Oxford template, the ROI (region of interest) brain gray matter volume features were extracted from processed structural MRI data. A convolutional neural network based classification model was constructed, and its performance was compared to that of traditional classification algorithms in machine learning. Experiments on main different brain regions of AD and bvFTD were implemented, and the influence of subjects' ages on the proposed model was also explored.

Using AAL2 template and Harvard-Oxford template, the highest classification accuracies of the proposed convolutional neural network model were 82.6% and 83.7%, respectively. They were higher than classification accuracies of traditional machine learning algorithms, 75.6% and 76.7%. Using data of the most different ROIs between AD and bvFTD (26 ROIs in AAL2 template and 31 ROIs in Harvard-Oxford template), the highest classification accuracies generated by the proposed model were 79.1%和80.2%.

The computational model based on convolutional neural network can be used to distinguish disease-specific gray matter patterns in patients with AD and those with bvFTD, and can help the clinician to arrive at a better differential diagnosis of AD and bvFTD.

Alzheimer disease; Dementia; Magnetic resonance imaging; Neural networks (computer)