Diagnosis of Alzheimer’s Disease using Structural MRI and Convolution Neural Network

Background: Alzheimer’s disease (AD) is a prevalent, neurological disease without effective treatment. However, if diagnosed early, the progression of the disease could be delayed through medication. Currently, one method to effectively diagnose AD early is to use Alternate Covering Neural Network (ACNN) network to discern various non-invasive Magnetic Resonance Imaging (MRI) images. This research aims to create an approach better than the current one and thus increase the accuracy of classifying MRI images, thereby diagnosing AD earlier and more perfectly. Methods: Alzheimer’s Disease Neuroimaging Initiative (ADNI; National Institutes of Health Grant U19 AG024904) database provided 3013 different sets of 3D MRI images labeled as cognitively normal (CN), mild cognitive impairment (MCI), and AD. A newly-proposed, modified Residual Network (ResNet) and an ACNN network were then constructed. Their common goal was to learn how to classify these labeled MRI images. After training, the two models got unique parameters for using the updated network to diagnose new images. Finally, inference, or testing the diagnostic accuracy of the two models, were performed based on another 469 different 3D MRI image sets. The accuracy of classification for two separate models were compared. Results: Compared with the ACNN network with a weighted classification accuracy of 80.17%, the newly proposed ResNet network enhances the weighted accuracy to 85.07% and showed statistical significance (p0.9196; AD 0.9389>0.9146). Conclusions: The research proposed a new deep learning convolutional network to classify 3D structural MRI images. The new ResNet is better in that it showed increased accuracy with statistical significance and had fewer false-positive results compare with the traditional ACNN network, thereby promising to help doctors diagnose AD more quickly and more accurately