A fast and accurate 3D fine-tuning convolutional neural network for alzheimer’s disease diagnosis

Abstract

The fast and accurate diagnosis of Alzheimer’s Disease (AD) plays a significant part in patient care, especially at the early stage. The main difficulty lies in the three-class classification problem with AD, Mild Cognitive Impairment (MCI) and Normal Cohort (NC) subjects, due to the high similarity on brain patterns and image intensities between AD and MCI’s Magnetic Resonance Imaging (MRI). So far, many studies have explored and applied various techniques, including static analysis methods and machine learning algorithms for Computer Aided Diagnosis (CAD) of AD. But there is still lack of a balance between the speed and accuracy of existing techniques, i.e., fast methods are not accurate while accurate algorithms are not fast enough. This paper proposes a new deep learning architecture to achieve the tradeoff between the speed and accuracy of AD diagnosis, which predicts three binary and one three-class classification in a unified architecture named 3D fine-tuning convolutional neural network (3D-FCNN). Experiments on the standard Alzheimer’s disease Neuroimaging Initiative (ADNI) MRI dataset indicated that the proposed 3D-FCNN model is superior to conventional classifiers both in accuracy and robustness. In particular, the achieved binary classification accuracies are 96.81% and AUC of 0.98 for AD/NC, 88.43% and AUC of 0.91 for AD/MCI, 92.62% and AUC of 0.94 for MCI/NC. More importantly, the three-class classification for AD/MCI/NC achieves the accuracy of 91.32%, outperforming several state-of-the-art approaches.