Early Detection of Alzheimer's Disease Using Magnetic Resonance Imaging: A Novel Approach Combining Convolutional Neural Networks and Ensemble Learning

Abstract

Early detection is critical for effective management of Alzheimer's disease (AD) and screening for mild cognitive impairment (MCI) is common practice. Among several deep-learning techniques that have been applied to assessing structural brain changes on magnetic resonance imaging (MRI), convolutional neural network (CNN) has gained popularity due to its superb efficiency in automated feature learning with the use of a variety of multilayer perceptrons. Meanwhile, ensemble learning (EL) has shown to be beneficial in the robustness of learning-system performance via integrating multiple models. Here, we proposed a classifier ensemble developed by combining CNN and EL, i.e., the CNN-EL approach, to identify subjects with MCI or AD using MRI: i.e., classification between (1) AD and healthy cognition (HC), (2) MCIc (MCI patients who will convert to AD) and HC, and (3) MCIc and MCInc (MCI patients who will not convert to AD). For each binary classification task, a large number of CNN models were trained applying a set of sagittal, coronal, or transverse MRI slices; these CNN models were then integrated into a single ensemble. Performance of the ensemble was evaluated using stratified fivefold cross-validation method for 10 times. The number of the intersection points determined by the most discriminable slices separating two classes in a binary classification task among the sagittal, coronal, and transverse slice sets, transformed into the standard Montreal Neurological Institute (MNI) space, acted as an indicator to assess the ability of a brain region in which the points were located to classify AD. Thus, the brain regions with most intersection points were considered as those mostly contributing to the early diagnosis of AD. The result revealed an accuracy rate of 0.84 ± 0.05, 0.79 ± 0.04, and 0.62 ± 0.06, respectively, for classifying AD vs. HC, MCIc vs. HC, and MCIc vs. MCInc, comparable to previous reports and a 3D deep learning approach (3D-SENet) based on a more state-of-the-art and popular Squeeze-and-Excitation Networks model using channel attention mechanism. Notably, the intersection points accurately located the medial temporal lobe and several other structures of the limbic system, i.e., brain regions known to be struck early in AD. More interestingly, the classifiers disclosed multiple patterned MRI changes in the brain in AD and MCIc, involving these key regions. These results suggest that as a data-driven method, the combined CNN and EL approach can locate the most discriminable brain regions indicated by the trained ensemble model while the generalization ability of the ensemble model was maximized to successfully capture AD-related brain variations early in the disease process; it can also provide new insights into understanding the complex heterogeneity of whole-brain MRI changes in AD. Further research is needed to examine the clinical implication of the finding, capability of the advocated CNN-EL approach to help understand and evaluate an individual subject's disease status, symptom burden and progress, and the generalizability of the advocated CNN-EL approach to locate the most discriminable brain regions in the detection of other brain disorders such as schizophrenia, autism, and severe depression, in a data-driven way.

Keywords: Alzheimer’s Disease Neuroimaging Initiative; Alzheimer’s disease; MCI-to-AD conversion; MRI biomarkers; convolutional neural networks; ensemble learning; magnetic resonance imaging; mild cognitive impairment.

FIGURE 1

Preprocessing pipeline—an example showing the formation of a transverse slice used in the learning. (A) Original image. (B) Skull-stripping and spatial normalization. (C) Smoothing. (D) Gray normalization. (E) Slicing and resizing.

FIGURE 2

The cropping range (inside the blue rectangle) of the slices used to train the model on (A) a sagittal plane and (B) a coronal plane, respectively. (A) Sagittal Plane. (B) Coronal plane.

FIGURE 3

Basic structures of the CNN convolutional layer and pooling layer. (A) Convolutional layer. (B) Pooling layer.

FIGURE 4

The architecture of the classifier ensemble based on the three sets of 2D slices (from left to right: sagittal, coronal, and transverse).

FIGURE 5

Experimental flow chart. (A) Training phase. (B) Testing phase.

FIGURE 6

Base-classifier architecture used in the CNN-EL approach proposed here.

FIGURE 7

The architecture of the 3D-SENet model. (A) Convolution block (Conv), (B) Squeeze-and-Excitation block (Se_block), (C) 3D-SENet model.

FIGURE 8

The list of brain regions with the classification capacity in each classification task. (A) Discriminable brain regions in the AD vs. HC. (B) Discriminable brain regions in the MCIc vs. HC. (C) Discriminable brain regions in the MCIc vs. MCInc.

FIGURE 9

Top 10 most discriminable brain regions in each binary classification task: (A) AD vs. HC; (B) MCIc vs. HC; (C) MCIc vs. MCInc.

FIGURE 10

Distributions of the identified brain regions on the relevant behavioral domains in each binary classification task: (A) AD vs. HC; (B) MCIc vs. HC; (C) MCIc vs. MCInc.