3D-CNN HadNet classification of MRI for Alzheimer’s Disease diagnosis

Abstract

The human brain consists of billions of neurons and their loss caused by neurodegenerative disorder dramatically affects cognitive brain functions. Such a state is known as Alzheimer's disease (AD). At this moment, AD remains without a decent cure, and its diagnosis mostly depends on the experience of clinicians; therefore, early diagnostics is critical because AD is progressive and at the beginning, its' development can be slowed down. In this paper, we use a deep learning advances to developed a classification system based on a 3D convolutional neural network for analyzing Magnetic Resonance Imaging (MRI) data collected for healthy individuals, patients with mild cognitive impairment (MCI) and with AD. The dataset of MR images was collected from Alzheimer's Disease Neuroimaging Initiative (ADNI); spatially normalized with statistical parametric mapping (SPM) toolbox and the skull-stripped for better 3D-CNN (HadNet) training. The backbone of the HadNet architecture is to use stacked convolutions (inception approach) which allows accessing more internal features of the MR image related to the AD. The hyperparameters of the HadNet were fine-tuned through the Bayesian optimization process. The developed classifier does not use segmented brain regions and can automatically process the whole MR image and based on learned features, during training, detect to which class input belongs. In our work, we select three classes: Healthy, MCI and AD; the final 3D-CNN architecture consists of three inception blocks. The HadNet was end-to-end trained using MR brain scans of 530 subjects including 185 AD patients, 185 MCI patients and 160 healthy individuals (HC). Evaluation results show that the trained classifier can distinguish between AD, MCI and HC with accuracy of 88.31% what is a promising classification results.