Dementia is an umbrella term for a group of neurological diseases in which cognitive capabilities deteriorate over time. Alzheimer's disease (AD), the most common type of dementia, includes 60 to 80 percent of all dementia cases (Jain, Jain, Aggarwal, & Hemanth, 2019; Ramzan et al., 2020). It is a progressive and irreversible neurodegenerative disease associated with symptoms such as a decline in cognitive functionality, deficiency of memory, and disturbance of daily activities (Jin et al., 2020). Albeit there is no compelling evidence for the leading cause of AD, it has been reported that some pathophysiological changes in the brain, beginning several years before the final stage, are responsible for the occurrence of AD. These changes are comprised of the emersion of neurofibrillary tangles within the neurons, which leads to the death of neurons and the accumulation of amyloid plaques among nerve cells, disturbing the usual path of neurotransmitters (Janghel & Rathore, 2021; S. Liu et al., 2015; Menikdiwela, Nguyen, & Shaw, 2018).
Mild Cognitive Impairment (MCI), an intermediate stage between AD and normal control (NC), refers to a detectable decline in cognitive abilities with no disruption in the patient's daily life (Sarraf, Desouza, Anderson, & Saverino, 2019). Although not all individuals with MCI progress to AD and dementia, there is a high rate of probability for the conversion from MCI to AD. Therefore, in many studies, MCI has been recognized as the prodromal stage of AD (Abrol et al., 2020; Gorji & Kaabouch, 2019).
Due to the growing demands of global life, the prevalence rate of age-related diseases such as AD has increased in recent years (Lu, Popuri, Ding, Balachandar, & Beg, 2018). The death rate of heart disease and prostate cancer has decreased in the last two decades; meantime, the death rate of AD has increased by 145 percent, making AD the sixth leading cause of death in the US (Association, 2019; Basheer, Bhatia, & Sakri, 2021; Nawaz et al., 2021). Although some recent studies have shown promising results for new drugs against AD, there is still no approved treatment for the disease (Hu, Ju, Shen, Zhou, & Li, 2016). As mentioned above, finding a way to diagnose AD accurately in the early stages can have many benefits, including stopping or decreasing the progression of the disease, reducing healthcare costs, and improving people's quality of life.
To the authors' best knowledge, there are three different approaches to diagnosing AD. In the first one, which is the most popular due to its ease of use and low cost, specialists utilize clinical information, symptoms, and other criteria like cognitive assessment scales and questionnaires to diagnose AD. However, this approach has some major drawbacks, such as being influenced by subjective factors and having undesirable performance results (Sun, Wang, Wang, & Liu, 2021). In the second approach, the clinical biomarkers, including the level of tau and amyloid-beta proteins, are measured through the cerebrospinal fluid (CSF) or brain autopsy. Despite its acceptable performance, this approach usually requires invasive procedures for measurement, making it unpopular as a routine method for early diagnosis of AD (Basheera & Ram, 2021; Sun et al., 2021). In the third approach, neuroimaging modalities such as MRI, fMRI, and PET are used to show the structure and functionality of the brain. This method can provide large amounts of information in a short period of time; however, interpreting all the detailed information in images is relatively challenging for physicians (Basheera & Ram, 2021; Sun et al., 2021).
Advances in computing power and the availability of open-access AD-related datasets, have led to the use of machine learning (ML) approaches in the context of early diagnosis of AD (Pellegrini et al., 2018). Deep learning (DL) has recently received much attention due to its brilliant results in different fields and medical image analysis as well (Basaia et al., 2019). DL approaches can extract high-level features, shown to be more efficient than other traditional approaches in a number of studies (S. Liu et al., 2015). In parallel with the widespread use of DL in various fields, especially medicine, its application in AD diagnosis has emerged. In this regard, Suk et al. (Suk & Shen, 2013) conducted the first study on the use of DL in AD diagnosis in 2013. They used the stacked auto-encoder (SAE) method and support vector machine (SVM) classifier in the feature extraction and classification steps, respectively (Suk & Shen, 2013). The following paragraphs summarize some similar studies. You can also see our recent systematic review on the current status of using DL in the early diagnosis of AD for a more comprehensive overview (Fathi, Ahmadi, & Dehnad, 2022).
Li et al. (A. Li et al., 2021) aimed to diagnose AD through a hippocampal shape and asymmetry analysis by cascaded convolutional neural networks (CNN). Compared to their previous study (Cui & Liu, 2019), which used only hippocampal shape features for classification, their performance was slightly lower this time. Both Mehmood et al. (Mehmood et al., 2021) and Kang et al. (Kang, Jiang, Huang, & Zhang, 2020) used a 2D-CNN-based architecture called VGG and transfer learning for early diagnosis of AD; however, Kang et al. utilized a multi-modal (MRI/DTI) approach.
ResNet, as the most popular CNN architecture in the literature, was used in a number of studies such as Abrol et al. (Abrol et al., 2020), Ramzan et al. (Ramzan et al., 2020), Odusami et al. (Odusami, Maskeliūnas, Damaševičius, & Krilavičius, 2021), Shanmugam et al. (Shanmugam, Duraisamy, Simon, & Bhaskaran, 2022), Li Y et al. (Y. Li, Ding, Wang, Li, & Tang, 2021), Ji et al. (Huanhuan Ji, Liu, Yan, & Klette, 2019) and Jabason et al. (Jabason, Ahmad, & Swamy, 2019). Some of these studies used the ResNet and other DL methods as an ensemble method (Jabason et al., 2019; Huanhuan Ji et al., 2019) or comparative analysis (Y. Li et al., 2021; Odusami et al., 2021; Shanmugam et al., 2022). In the study of Zhang et al. (X. Zhang, Han, Zhu, Sun, & Zhang, 2021), a 3D-ResNet with the attention mechanism, was proposed to create an explainable model for early AD diagnosis. A few other studies have also utilized the attention mechanism mainly aimed at adding explainability to the black box nature of CNN-based models (Guan, Wang, Cheng, Jing, & Liu, 2022; H. Ji, Liu, Yan, & Klette, 2020; Z. Liu et al., 2022; J. Zhang et al., 2021; Y. Zhang, Teng, Liu, Liu, & He, 2022). According to the literature, VGG and DenseNet were the second and third most popular CNN architectures. Most studies, using VGG, employed its standard versions, VGG16 or VGG19; however, some studies proposed customized versions, including the studies of Zhang P et al. (P. Zhang, Lin, Qiao, & Tu, 2021) and Yu et al. (Yu, Peng, Shi, Zhu, & Dai, 2019).
DenseNet, one of the most successful architectures for early AD diagnosis, was used in some recent studies. Li and Liu (F. Li & Liu, 2018, 2019; M. Liu et al., 2020) applied 3D-DenseNet in their studies in order to extract high-level features and classify different stages of AD. The authors in (F. Li & Liu, 2018) extracted features from various parts of the brain using patch-based strategies, whereas in (F. Li & Liu, 2019) and (M. Liu et al., 2020), only the hippocampus region was used to extract features. Several DenseNet architectures are combined in ensemble approaches reported in the studies of Wang et al. (Wang et al., 2019), Ruiz et al. (Ruiz, Mahmud, Modasshir, & Shamim Kaiser, 2020), and Islam et al. (Islam & Zhang, 2018). Some reviewed studies have proposed customized CNN architectures to diagnose AD and its prodromal stages, MCI or its subcategories, namely the early MCI (EMCI) and late MCI (LMCI). Basaia et al. (Basaia et al., 2019) have proposed a 3D-CNN with 12 convolutional blocks, a rectified linear unit (ReLu) as the activation layer, a fully connected layer, and a logistic regression layer as the classifier for automatic classification of AD and subcategories of MCI. Gorji et al. (Gorji & Kaabouch, 2019) developed a simple 2D-CNN architecture with three convolution layers, each followed by a max-pooling, a fully-connected layer, and a sigmoid classifier for binary classification of prodromal stages of AD. The study focused on gray matter (GM) due to its proven effect on the early onset of AD. Pan et al. (Pan et al., 2020) employed an ensemble scheme based on different 2D-CNN classifiers for early detection of AD.They built various base CNN classifiers on single-axis slices of MR images and created an ensemble model based on the five best classifiers for each axis.
While many reviewed studies reported promising results, most did not address all clinically valuable classification groups. In addition, there is no comprehensive comparative analysis among various individual base classifiers and ensemble approaches. In the current study, we aimed to propose an ensemble method based on deep learning for the early diagnosis of AD using MRI images. The proposed method was comprised of six well-known convolutional neural networks (CNN) based on a novel approach called the weighted probability-based ensemble method (WPBEM). The main contribution of the current study are described as follows:
-
A novel ensemble method called WPBEM was used to enhance the performance of the individual CNN models for early diagnosis of AD.
-
A comparative analysis was utilized to find hyperparameters and the optimal scenario for combining the individual CNNs.
-
In order to enhance the performance of the models, we used a domain adaptation transfer learning approach producing superior results to any other parameter initialization methods.
-
The current study has addressed all the valuable binary and multiclass classification groups.