Automated classification of Alzheimer's disease and mild cognitive impairment using a single MRI and deep neural networks

Abstract

We built and validated a deep learning algorithm predicting the individual diagnosis of Alzheimer's disease (AD) and mild cognitive impairment who will convert to AD (c-MCI) based on a single cross-sectional brain structural MRI scan. Convolutional neural networks (CNNs) were applied on 3D T1-weighted images from ADNI and subjects recruited at our Institute (407 healthy controls [HC], 418 AD, 280 c-MCI, 533 stable MCI [s-MCI]). CNN performance was tested in distinguishing AD, c-MCI and s-MCI. High levels of accuracy were achieved in all the classifications, with the highest rates achieved in the AD vs HC classification tests using both the ADNI dataset only (99%) and the combined ADNI + non-ADNI dataset (98%). CNNs discriminated c-MCI from s-MCI patients with an accuracy up to 75% and no difference between ADNI and non-ADNI images. CNNs provide a powerful tool for the automatic individual patient diagnosis along the AD continuum. Our method performed well without any prior feature engineering and regardless the variability of imaging protocols and scanners, demonstrating that it is exploitable by not-trained operators and likely to be generalizable to unseen patient data. CNNs may accelerate the adoption of structural MRI in routine practice to help assessment and management of patients.

Keywords: Alzheimer's disease; Convolutional neural networks; Deep learning; Diagnosis; Mild cognitive impairment; Prediction.

Fig. 1

Architecture of a typical convolutional neural network. a) Input layer: the data is given to the network. b) Convolutional layer: neurons identify the main features that characterize the images, storing the information into a ‘feature map’ (e.g., red, blue and yellow blocks). c) Pooling layer: the size of each feature map is reduced with a downsampling operation along the spatial dimension (e.g., red, blue and yellow blocks). d) Fully-connected layer: the neurons are connected to all neurons from the previous layer. e) Output layer: the step that returns the probability of the input data to belong to each class.

Fig. 2

Flowchart of the main steps of the experiments performed. MRI data of each classification dataset (AD vs HC, c-MCI vs HC, s-MCI vs HC, AD vs c-MCI, AD vs s-MCI, c-MCI vs s-MCI) were randomly split into a large training and validation set (90% of images) and a testing set (10% of images). Data augmentation was applied on images selected for training and validation. See text for further details.

Fig. 3

Examples of images after data augmentation, i.e., deformation, cropping, rotation, flipping, and scaling. Axial and coronal images are shown. A = anterior; L = left; P = posterior; R = right.