Leveraging Deep Learning for Enhanced Detection of Alzheimer’s Disease Through Morphometric Analysis of Brain Images

Prevalent as the most common form of neurodegenerative disorder, Alzheimer’s Disease (AD) significantly strains global socio-economic structures [1]. Projections indicate that the incidence of dementia attributable to AD is set to rise rapidly in forthcoming years, engendering an amplified demand for healthcare services and resources [1]. The resultant pressure is expected to not only burden the infrastructure of national health systems but also impose additional fiscal and ethical responsibilities on families tasked with patient care [2]. According to the World Health Organization (WHO), over fifty-five million individuals globally grapple with AD, with societal costs estimated at US\$1.3 trillion in 2019 [2]. Predictions suggest that by 2030, these costs could exceed US\$2.8 trillion, driven by an increase in dementia prevalence and escalating healthcare expenses [3].

AD primarily manifests as a deterioration of memory, followed by a progressive decline in walking ability, language, emotions, and numerous cognitive and behavioral functions [4]. Characterized by the accumulation of abnormal proteins in the form of amyloid-beta plaques and neurofibrillary tangles, AD induces volumetric deformations in brain tissue over time, rendering neurons dysfunctional [4]. Regrettably, these deformations are irreversible, rendering the prevention of AD progression before extensive neurodegeneration a crucial area of research [5].

AD is typically categorized into three stages: cognitively normal (CN), mild cognitive impairment (MCI), and full-blown AD [6-9]. The MCI stage, considered an intermediary between physiological aging and AD, is marked by a mild but measurable cognitive decline. Crucially, early diagnosis during the MCI stage can potentially delay or prevent disease progression to AD, hence its recognition as a critical research focus [10].

The deployment of computer-based AD diagnosis can assist clinicians in identifying high-risk groups, thereby preventing disease progression and enabling early diagnosis [11]. Magnetic Resonance Imaging (MRI), a common medical imaging method, offers detailed visualization of brain structures for AD detection [12]. However, it is challenging to identify patterns of physiological changes in the brain structure during the early stages of AD using traditional radiological readings or quantitative analysis [13].

Thus, the development of reliable auxiliary systems is necessary to improve the accuracy of early disease detection, supplementing the observations and decisions of physicians [13]. Utilizing MRI-based methods, machine learning (ML)-based models can reveal fine-scale anatomical changes in the brain's internal structure associated with cognitive decline, thereby bolstering early AD diagnosis [14].

The increasing popularity of research on automatic classification of AD progression from MRI images using machine learning methods underlines the significant role of these methods in developing computer-based decision support systems [15]. Structural deformations in the brain of an AD patient, most often observed in the medial temporal lobes and hippocampus, can be perceived with the naked eye from an MRI scan [16, 17]. Figure 1 illustrates the MRI images of subjects with CN, MCI, and AD, showing the reduction in hippocampus tissue size and increase in ventricle size as the disease progresses.

This study aims to develop an efficient early diagnosis system for patients in the MCI stage before progressing to AD. Limited studies in the literature focus on Alzheimer's diagnosis based on deep learning (DL)-based feature extraction and analysis of morphometric methods, particularly tensor-based morphometry (TBM), a medical imaging method that allows analysis of morphological brain changes [18]. However, the visual interpretation of these images can be challenging, highlighting the need for DL-based methods for automatic extraction of disease-specific features.

This research aims to enhance AD detection by integrating DL and TBM methods by comparing model performances trained with medial slices (in the x, y, and z direction) or axial slices covering the hippocampus and temporal lobe. The effect of using small kernel sizes for detailed feature extraction from local volume changes in the brain on model success was examined. While binary classification (CN, AD) studies are prevalent in the literature, multiple classification (AD, MCI, CN) studies are limited. Therefore, a multiple classification approach was adopted in this research.

The structure of this paper includes "Related Work" in chapter 2, "Methods" in chapter 3, "Result and Discussion" in chapter 4, and "Conclusion" in chapter 5.

1.png

Figure 1. MRI scans of a) CN, b) MCI, and c) AD [17]

3.1 Data

The use of a freely available common dataset among researchers increases the credibility of the studies. For this reason, the ADNI database was used as part of this research. The ADNI is a global project dedicated to discovering, analyzing, and curing disease with the aim of slowing and stopping the progression of AD. The ADNI provides a range of datasets, offering a valuable resource for researchers aiming to detect AD in its early stages. By supplying these standardized datasets, ADNI facilitates consistent research practices and promotes the global sharing of compatible data among researchers.

The ADNI was launched in 2003 as a public-private partnership, led by Principal Investigator Michael W. Weiner, MD. The primary goal of ADNI has been to test whether serial MRI, positron emission tomography (PET), other biological markers, and clinical and neuropsychological assessment can be combined to measure the progression of MCI and early AD. As such, the investigators within the ADNI contributed to the design and implementation of ADNI and/or provided data but did not participate in analysis or writing of this report. A complete listing of ADNI investigators can be found online. (GE Healthcare, Philips Medical Systems, or Siemens) [37].

3.2 Tensor-based morphometry

For large, multi-site neuroimaging analyses and clinical trials of AD and MCI, TBM is an objective, dependable, high-throughput imaging measure [38]. TBM aims to determine local volume differences among groups of brains using deformation fields that map points in a template (x₁; x₂; x₃) to corresponding locations in individual source pictures (y₁; y₂; y₃) using a Jacobian matrix. The local volume differences contain information about the local stress, shear, and rotation involved in the deformation. Eq. (1) shows how to compute a Jacobian matrix [39].

$\mathbf{J}=\left[\begin{array}{lll}\partial y_1 / \partial x_1 & \partial y_1 / \partial x_2 & \partial y_1 / \partial x_3 \\ \partial y_2 / \partial x_1 & \partial y_2 / \partial x_2 & \partial y_2 / \partial x_3 \\ \partial y_3 / \partial x_1 & \partial y_3 / \partial x_2 & \partial y_3 / \partial x_3\end{array}\right]$                    (1)

T1 MRI images were pre-processed by Hua et al. [40] and prepared for TBM analysis. In brief, all images were resampled to 1mm voxel sizes in x, y, and z dimensions, then non-linearly aligned to a group-average template created with 40 randomly selected control subjects, and individual Jacobian maps were estimated from the warp-fields. Deformation fields calculated for deformation-based morphometry and TBM are shown in Figure 2 [39].

The DBM field calculates the large-scale difference between the subject image and template image, whereas the TBM field calculates local differences. TBM is a mapping method applied to MRI images to visualize brain tissue loss and enlargement. Figure 3 shows MRI image of an AD subject in the ADNI database.

In Figure 4, the TBM analysis image of the AD subject, whose MRI image is shown in Figure 3, is presented. As can be seen in Figure 4, compared to minimally processed MRI images, TBM images are blurry and difficult to interpret by eye. This is due to the fact that TBM images present the local morphological differences of each subject compared to the group mean. They are meant to be used for group-level statistical analysis rather than visual inspection [41].

In this study, we used different variations of ADNI dataset for three different models. In the first model, the network was trained with only medial slices of three axes (axial, sagittal, and coronal) meaning that there were 3 MRI images for each subject. Then, the performance values of the network trained with those images were analyzed. 156 AD subjects [mean age: 75.4±7.5 year, 81 males (M) and 75 females (F)]; 330 patients with MCI [mean age: 74.8±7.4 year, 172 (M) and 158 (F)]; and 190 CN subjects [mean age: 76.0±5.0 years, 100 (M)/90 (F)] are used. The demographic characteristics of a total of 676 subjects used in the first model are presented in Table 1.

2.png

Figure 2. Differences between a) DBM and b) TBM [39]

3.png

Figure 3. Original MRI scan image of a single subject: a) Coronal MRI slice, b) Sagittal MRI slice and c) Axial MRI slice

4.png

Figure 4. TBM image of the subject: a) Coronal slice; b) Sagittal slice; c) Axial slice

The second and third models were trained with the same database using only slices of axial planes that cover the temporal lobes and hippocampi. Then, the performance values of these networks were analyzed. 28 AD subjects [mean age: 75.0±5.0 year, 16 males (M)/12 females (F)]; 88 patients with MCI [mean age: 73.8±5.4 years, 47 (M)/41 (F)]; and 54 CN subjects [mean age: 74.4±5.5 years, 30 (M)/24 (F)] are used. The demographic characteristics of a total of 170 subjects used in the second and third model are presented in Table 2.

Table 1. The demographic characteristics of the first model

Table 2. The demographic characteristics of the second and third models

5.png

Figure 5. Obtaining the dataset used in feeding the models

3.3 Deep learning

The images, after being downloaded to our local servers, were converted to 2D png images from 3D nifti files by using med2image software [42]. 3D TBM images are downloaded from ADNI. First, the height and width of each nifti file are saved. Witht this, axial images (z axis) were sliced at 5-pixel intervals starting from the 48th pixel to the 115th pixel, covering only the hippocampus and temporal lobe. As a result, 12 2D axial png brain slices were obtained for each subject. Considering that the hippocampus and temporal lobe are the regions most affected by AD, slices with these regions in the range of 48-115 pixels were taken [6]. The steps for converting TBM images from 3D medical image format to 2D png format are presented in Figure 5.

Within the scope of this investigation, we adapted a 2D CNN-based DL architecture to diagnose AD early using TBM images. We used the AlexNet architecture, which is proven to be successful in image classification processes [43, 44]. AlexNet has won the 2012 “Imagenet Large-Scale Visual Recognition Competition” with an extraordinary difference [45]. This architecture is a deep CNN consisting of eight layers, five of which are convolutional layers with rectified linear unit (ReLU) functions and three of which are batch normalization layers. Three of the batch normalization layers are followed by max pooling layers, A learnable filter that extracts features from an input image is represented by the convolutional layer. For a 3D image of dimensions H, W, and C, where H stands for height, W for width, and C for channel count. When a three-dimensional filter is used, it can be represented as FC (number of filter channels), FH (filter height), and FW (filter width). Since AH stands for activation height and AW for activation width, the output activation map size must be AHxAW. Using the Eq. (2) and Eq. (3), activation height and width values can be calculated [46].

$A_H=1+\frac{H-F_H+2 P}{S}$            (2)

$A_W=1+\frac{W-F_W+2 P}{S}$            (3)

P stands for padding, S for stride, and since there are n filters, the dimensions of the activation map should be AWx AWxn.

The pooling layer's key purpose is to reduce the size of the feature maps. Thereby, there will be fewer parameters to learn and fewer computations to be made by the network. By applying a non-linear conversion to the given inputs, an activation function addresses non-linearity in the network [47] Eq. (4).

$\sigma(\vec{z})_i=\frac{e^{z_i}}{\sum_{j=1}^K \, e^{z_j}}$             (4)

where, $\vec{z}$ is the softmax function's input vector, and each z_i value is an input vector component, which can have any real value. The normalizing term, which appears in the formula below, makes sure that the function's output values add up to 1, producing a legitimate probability distribution. K is the number of classes in the classifier with multiple classes. If the input is positive, the rectified linear activation function, or ReLU for short, will output the input directly; if it is negative, it will output zero Eq. (5). Because a model that utilizes ReLU is simpler to train and frequently performs better, it has evolved into the default activation function for many diverse types of neural networks [48].

$f_{\mathrm{ReIU}}=\max (0, x)$           (5)

In this study, we used three models. The first model that was trained using medial brain slices (in x, y, and z directions) belongs to 676 subjects (train: 1622; test: 406). In the second and third models, twelve axial brain slices obtained at 5-pixel intervals covering the hippocampus and temporal lobe from 170 subjects were used (train: 1632; test: 406). Since the number of medical images used for the training of all three models was intended to be close to each other, fewer subjects were used in the second and third models. 20% of the resulting dataset was separated as test datasets for each clinical group (AD, CN, MCI). 80% of the data was allocated for training and 20% for testing. For the training of the models, 80% of the data was used for training and 20% for testing. During model training, 20% of the training data was selected as validation data.

In our study, models are trained using AlexNet or fine-tuned AlexNet architecture. The architecture ends with two dense layers and a dropout layer with 0.5 rates between them.

In the first model, we observed the performance of original AlexNet by feeding it medial slices (in the x, y, and z directions) of TBM images from each subject. The model produced an accuracy of less than 50%, failing to perform a successful diagnosis.

For the second and third model of the study, 5-pixel-spaced axial images covering only the hippocampus and temporal lobe were extracted, yielding twelve images per subject. Hippocampus and temporal lobe were selected as the regions of interest because they are the regions that suffer the most from AD [6]. The second model that fed these images achieved over 93% accuracy using the original AlexNet network [44].

In the third model, we fine-tuned the parameters of the second model to see if we could improve the performance of the diagnosis. In the third model, in addition to implementing the original AlexNet model, an adaptation of it with a smaller kernel size and stride length was used. The idea was to make the model more consistent with the nature of the TBM images. Since they show differences in more minor scales, the kernel size and stride length were fine-tuned to catch those differences. Specifically, kernel size and stride length in the first convolutional layer were set to 3×3 and 2×2, respectively. And the strides in the first max pooling layer were changed to 2×2, the second convolutional layer was changed to 3×3 and 2×2, the second max pooling layer was changed to 2×2, the third convolutional layer was set to 2×2 and 1×1, the fourth was changed to 1×1 and 1×1, and the last max pooling layer was formed at 1×1. Naturally, the decrease in resolution caused an increase in the total learning parameters. The details of the second model can be seen in Table 3.

The accuracy of our third fine-tuned model, trained using axial slices covering the temporal lobe, was over 92%.

Table 3. Details of the CNN architecture used for the third model

This study presents a new method to distinguish between AD, MCI, and CN stages using TBM and DL techniques. The study proposed three different models to differentiate between subjects. In addition, TBM-based analyses are compared with VBM-based analyses used in previous studies. Local morphological changes are helpful in identifying AD and MCI. As a result, TBM-based analyses achieve similar success levels compared to VBM-based analyses. The CNN-based methodology employed Within the confines of this study effectively discriminates between individuals with AD and those with MCI based on localized morphological alterations quantified through TBM. Consistent with its concept, TBM was useful only when the brain regions that are subject to the most severe local atrophy (namely, the hippocampi and temporal lobes) were used to feed the CNN algorithm. When medial slices from each axis were used, although they also contained meaningful information such as larger ventricles in AD subjects, the CNN algorithm was not able to capture the differences. However, this was not the case for the studies that used VBM or sMRI as input. A high level of predictive accuracy was reached for different 2D ROIs (covering the hippocampus). TBM is more suitable to diagnose diseases or abnormalities that are associated with a fairly small region in the brain, such as tumors, and the regions that are susceptible to atrophy should be chosen as the region of interest to run the prediction algorithm. The results show that TBM analysis using DL-based methods can be successfully used as an effective and usable method in the early diagnosis of Alzheimer's. Although the TBM contained information about minor regional differences, extracting more details using small kernels from the images did not significantly alter the model results. According to the results of all models in the study, DL-based systems can successfully perform early detection of Alzheimer's using TBM covering the hippocampus. In addition, TBM can be more successful than DBM-based morphometric methods in the early diagnosis of Alzheimer's.

This study brings a new perspective to research in this area. In future studies, researchers should explore other neural networks such as Xception, MobileNet, and newer state-of-the-art networks to construct the classifier. Additionally, they can experiment with preprocessing steps like skull stripping and density normalization to achieve similar or better results. Furthermore, by leveraging the feature extraction ability of pre-trained models, they can enhance the overall performance of the different classifiers they propose.