EEG and MRI Processing for Alzheimer’s Diseases

Abstract

A new method for the diagnosis of Alzheimer’s disease in the mild stage is presented according to combining the characteristics of electroencephalogram (EEG) signal and magnetic resonance imaging (MRI) images. Then, proper features of brain signals are extracted according to the nonlinear and chaotic nature of the brain such as Lyapunov exponent, correlation dimension, and entropy. These features combined with brain MRI images properties include medial temporal lobe atrophy (MTA), cerebrospinal fluid flow (CSF), gray matter (GM), index asymmetry (IA), and white matter (WM) to diagnose the disease. Then two classifiers, the support vector machine and Elman neural network, are used with the optimal combined features extracted by analysis of variance. Results showed that between the three brain signals, and between the four modes of evaluation, the accuracy of the Pz channel and excitation mode was more than the others The accuracy of the results in Elman neural network with the combination of brain signal features and medical images is 94.4% and in the case without combining the signal and image features, the accuracy of the results is 92.2%.

1. Introduction

Alzheimer’s disease is a progressive disease of the mental faculties commonly seen in the elderly. Significant symptoms of this disease are memory loss, judgment, and important behavioral changes in the person [1]. The disease results in the loss of synapses of neurons in some areas of the brain, necrosis of brain cells in different areas of the nervous system, the formation of spherical protein structures called aging plaques outside neurons in some areas of the brain, and fibrous protein structures called neurofibrillary Tangles. A spiral is identified in the cell body of neurons. There is currently no definitive diagnosis or treatment for this disease. The prevalence of Alzheimer’s disease is increasing rapidly [2]. The number of Alzheimer’s patients in Iran has almost doubled in 13 years, according to the Iranian Alzheimer’s Association. On the other hand, the costs of treatment, as well as care and nursing of these patients, are very high and difficult. This disease causes various mental disorders in the patient. It usually takes several years from the first signs of the disease to the acute stages of the disease, when most of the brain cells are destroyed. If this disease is not detected in time, new and up-to-date treatment methods will not work. The solution is to accurately identify the mechanism of this disease and its effect on brain signals, which is very difficult due to the dynamic nature of EEG signals and medical images, which due to the complex nature of this disease as a result, we must determine the best and most effective indicator to identify this disease and how this indicator relates to the characteristics of the brain signal and medical images. Medical image analysis has become very important in the diagnosis of mild Alzheimer’s disease in recent years [3]. The high volume and complexity of medical images make early detection of Alzheimer’s disease difficult for physicians and increase the workload of radiologists, in which case the use of computer-aided diagnostic (CAD), including image processing technologies, can help to increase the accuracy of diagnosis. The use of machine learning systems and deep processing of medical images with proper labeling and feature extraction can be one of the effective methods of diagnosing this disease [4]. Deep learning methods and machine learning techniques can be two effective and accurate methods in the early diagnosis of Alzheimer’s disease [5]. Hippocampal volume analysis is used in medical image processing to diagnose mild Alzheimer’s disease. Because before the atrophy creation, analyzing the volume of hippocampal material in MRI images can be used with deep processing techniques to extract proper features to identify mild-Alzheimer’s disease [6]. 3D segmentation of MRI images further helps researchers diagnose Alzheimer’s disease and obtain important information [7]. Determining the degree of atrophy of MRI images is an effective method for early detection of Alzheimer’s disease. Also, assessing the degree of asymmetry in both the right and left hemispheres and analyzing volumetric mismatch can differentiate from mild to severe Alzheimer’s disease [7]. Using statistical features of signal and obtaining temporal information and using spatial features of MRI images is an effective method for the more accurate evaluation of Alzheimer’s disease [8]. Cortical atrophy means the gradual destruction of the nerve cells that make up the upper regions of the brain, specifically the structures found in the cerebral cortex, mostly due to a reduction or loss of oxygen and nutrients in these areas. There are also different methods for evaluating the medial temporal lobe that has different functional accuracy [9]. Longitudinal T1-weighted MRI studies are another effective way to distinguish mild-Alzheimer’s patients from healthy ones [10]. Also, extracting the appropriate characteristics and deciding on the classification in this field are among the issues to be considered. Currently, there are several methods to diagnose this disease, and it is important to examine two issues in these methods. The cost of these methods and the acceptable accuracy are the points under consideration, so it seems necessary to identify a low-cost method with appropriate accuracy and precision. Therefore, in addition to extracting proper features of EEG signals and MRI images for AD diagnosis, another aim of this study is to identify the proper multimodal combining of these extracted features to increase the accuracy of mild-Alzheimer’s disease detection by using a proper classifier.

2. Materials and methods

In this study, 40 volunteers were used to record brain signals and MRI images in healthy, mild, and severely ill groups. The number of subjects is 19 in the healthy group, 11 in the mild patient group, and 10 in the severely ill group. Forty volunteers with an age range of 60 to 88 years have been used to record brain signals. All participants in all groups were right-handed. Nineteen participants in the Mini-Mental State Exam (MMSE) test scored between 23 and 30 and were included in the group of healthy people. Eleven participants with an MMSE score of 19 to 22 were classified as mild-Alzheimer’s patients, and finally, 10 participants with an MMSE score between 3 and 18 were included in the group of severe Alzheimer’s patients.

The Powerlab SP device with two amplifiers was used to record the brain signal. In this device, three channels for recording the brain signal and one channel for recording the EOG signal, and the other channel for the external audio signal for stimulation have been used so that the stimulation signal and ERP do not occur simultaneously. The signal was recorded in 4-channel mode according to the standard 10–20. The sampling rate of the device is 1 kHz and 16 bits for each sample. Recording the brain signal in the form of subject training, recording the closed eye for 1 minute, recording the open eye for 1 minute, and recording while performing the task assigned to the subject, which includes A. Remembering the displayed shapes; B. The counting of target and non-target sounds in an oddball auditory test. After proper labeling by the physician in segregation of healthy individuals, and mild and severe Alzheimer’s patients by MMSE test in the first part, how to record the brain signal in four steps is explained to people and they are asked to relax during Keep records to prevent the formation of motion artifacts and other unwanted factors, and this method of registration does not cause harm to the person. After preparing the subject, we perform the second step. In the closed eye mode, we record the signal for 1 minute. Then in the third step, the subject will be asked to open the eyes and record the signal for 1 minute. At the end of the third stage, the displayed images have no color so that the color feature does not have a different effect on different subjects. These images are displayed for 1 minute and after this time, the subject will be asked to close the eyes and recall the images (review the images) in the mind. Meanwhile, brain signals will be recorded for 1 minute. Participants will then be asked to open their eyes and express the shapes one by one aloud. In the last part, a sound with a frequency of 1 kHz called non-target sound and 1.5 kHz sound called target sound will be given to the subject. Before playing these two categories of sound, this step has been taught to the subject. In the fourth part of section (b), these sounds are played, and the subject is asked to press the right key as soon as hearing the target sound and the left key as soon as hearing the non-target sound. The interval between stimuli (sound playback) is 2 seconds and the sounds will be played randomly. The only important point is that 75% of the number of stimuli is non-target sound and 25% of the number of stimuli is target sound. If we assume the total number of stimuli to be 120, we will have a total of 30 target stimuli and 90 non-target stimuli, which can be randomly distributed between the non-target stimuli at 2-second intervals. The playing time of each sound (target and non-target) will be 300 milliseconds. This recording section is 276 seconds with the assumption of 120 excitations and the total signal recording time for each subject is approximately 10 minutes. A sample image of recording brain signals is shown in Figure 1.

Figure 1.

A sample research participant during brain signal recording.

The first step in processing brain signals is to eliminate high- and low-frequency noise and interference and to remove motion artifacts. It is clear that the removal of unwanted factors such as motion artifacts, signal deviation from the baseline, high- and low-frequency noise, and reduction of sampling rate is necessary for proper processing of brain signals and extraction of optimal features, and this increases the accuracy of brain signal processing [11]. The motion artifacts in the brain signal are caused by contractions of the muscles of the head and neck as well as the movement of the electrode. On the other hand, transpiration also causes frequency interference. To eliminate these artifacts and noise from the city, a pass filter with cutoff frequencies of 0.5 to 45 Hz was used [12].

Neuroimaging techniques, physiological signs, and genetic analysis are methods used to diagnose Alzheimer’s disease [13]. To detect Alzheimer’s disease in its early stages, neuroimaging methods are used, which include SPECT, PET, and magnetic resonance imaging. The problem with SPECT and PET is the risks of radiation and its very high cost, time-consuming, and inconvenient. Therefore, apart from all these neuroimaging methods, MRI imaging is one of the standard methods used to diagnose Alzheimer’s disease. The advantage of this method is the ease of registration and economic cost over the above methods. MRI images should be at least 3 Tesla and the slices should be 3 mm thick so that acceptable images can be seen to examine the lesions of aging coils and spiral plaques. The MRI image is displayed in three different directions in Figure 2. Then, the appropriate image segmentation, mask, and sharp filter are used for pre-processing.

Figure 2.

MRI image in three different modes.

Various diagnostic tools from the clinical and processing areas for early diagnosis of Alzheimer’s disease have been reviewed. Methods of blood tests, speech therapy, physical function, and hearing status were first examined by a physician and then diagnosed with mild Alzheimer’s disease by recording an electroencephalogram and combining it with medical images [14]. First, the EEG signal is recorded from three channels Fz, Cz, Pz as unipolar and then MRI images from the peritoneal area. Combining MRI images and EEG signals can be a way to diagnose mild Alzheimer’s disease. Medial temporal lobe atrophy, cerebrospinal fluid, white and gray matter volume, and asymmetry between the two hemispheres are effective features in MRI images to diagnose Alzheimer’s disease. Another approach to EEG signal analysis is nonlinear and dynamic signal methods. The parameters that express nonlinear behavior are dual. The first category is parameters that emphasize the dynamics of signal behaviors such as entropy and Lyapunov’s exponent. These parameters describe how the system behaves over time. The second category emphasizes the geometric nature of motion paths in state space, such as the correlation dimension. In this view, the system is allowed to move in the adsorption bed at the appropriate time and then, the geometric dimension of the adsorption bed is obtained. One of the most important tools used to understand the behavior and dynamics of time series of vital signals, which are mainly extracted from nonlinear systems, is the phase diagram. Using this tool, the behavioral characteristics and chaotic nature of the data can be demonstrated appropriately and qualitatively, as well as important parameters such as the path of the system in the state space [15]. In order to draw this diagram using the recorded time series, it is enough to draw each sample at any time in terms of another sample in the previous time. Figure 3 shows the two-dimensional phase curve, and Figure 4 shows the three-dimensional phase curve of the Fz, Cz, and Pz channels of the EEG signal from a healthy person with the eyes closed.

Figure 3.

Two-dimensional phase curve of Fz, Cz, Pz channels EEG signal in closed eye.

Figure 4.

Three-dimensional phase curve of Fz, Cz, Pz channels EEG signal in closed eye.

Lyapanov’s exponent shows the average convergence or divergence of the trajectory path in the phase space. The correlation dimension shows the number of independent variables needed to describe the dynamics of the system and is another way to examine the chaotic signal. If the correlation dimension of a path is zero, it represents a steady state of the system, and if the value is equal to one, it represents prodigal behavior. The value of this variable is incorrect when chaotic behavior occurs. The higher the value of this parameter, the more complex the nonlinear system. Therefore, it can be said that the correlation dimension is the degree of complexity of the distribution of points in the phase space. Figure 5 compares the correlation dimension of Pz channels between 3 groups of healthy people, mild patients, and severe patients. The amount of this feature decreases with the severity of the disease, which is evident in the Pz channel.

Figure 5.

Dimensional comparison of Pz channel correlation for three groups of healthy subjects, mild patient and severe patient.

In patients with a clinical diagnosis of Alzheimer’s disease, atrophy of the inner part of the temporal lobe is evident [16]. In the autosomal dominant form of Alzheimer’s disease, atrophy of the inner part of the temporal lobe in patients, compared with controls, can be detected up to 3 years before the onset of clinical signs of cognitive impairment. In patients with Alzheimer’s disease, hippocampal atrophy was reduced (10–50%), the amygdala was reduced to 40%, and parahippocampus was reduced to 40% compared with the control group, which was standardized for age. There is compelling evidence that atrophy of the internal structures of the temporal lobe, especially the hippocampus and entorhinal cortex, occurs early in the course of the disease and even before the onset of clinical symptoms [17]. The severity of changes in imaging of healthy elderly people makes it difficult to use MRI as a definitive diagnostic method. By the time mild symptoms appear, the volume of the hippocampus may have decreased by more than 25%. Clinically, a reduction in hippocampal volume is associated with the severity of clinical signs and symptoms of memory loss, the patient’s score on cognitive evaluation tests, and pathological findings. Figure 6 shows the determination of spinal atrophy and asymmetry. However, another group believes that there is no clear association between lesions in the course of dementia, including lesions of hyperexcitability of white matter on MRI, and the severity of the symptoms of post-adjustment cognitive impairment for age. They believe that due to the high sensitivity of MRI in the diagnosis of hyperexcitability lesions in T2 view and on the other hand the low specificity of these lesions in the diagnosis of the disease, there is a weak relationship between MRI findings and clinical and neuropathological symptoms. Eq. (1) shows how to determine medial temporal lobe atrophy (MTA):

1. A: Internal temporal loop area.

2. B: Hippocampus and parahippocampus.

3. C: Unilateral lateral ventricle.

Figure 6.

How to determine atrophy in MTA images.

$$\text{MTAi}=\left(\text{A}-\text{B}\right)\times 10/\text{C}$$E1

$$\text{Cleft}:\text{Area}=187.3\text{mm2},Avg=691.1,Dev=\mathrm{128.3.}$$

$$\text{Cright}:\text{Area}=173.1\text{mm2},Avg=648.2,Dev=\mathrm{146.2.}$$

$$\text{Aleft}:\text{Area}=324.7\text{mm2},Avg=323.9,Dev=\mathrm{238.1.}$$

$$\text{Aright}:\text{Area}=325.5\text{mm2},Avg=245.3,Dev=\mathrm{191.3.}$$

$$\text{Bleft}:\text{Area}=200.3\text{mm2},Avg=190.8,Dev=\mathrm{121.6.}$$

$$\text{Bright}:\text{Area}=220.1\text{mm2},Avg=160.4,Dev=\mathrm{118.3.}$$

When MTAi is calculated, two values are determined that each corresponds to a hemisphere. According to MTAi, the asymmetry index is calculated as Eq. (2), and the mean values of MTAi and IA for the three groups are given in Table 1(Figure 7).

Group	Mean MTAi	Mean IA
Healthy	2.3	1.7
Mild AD	4.5	2.5
Severe AD	5.7	3.3

Table 1.

Mean values of MTAi and IA for the three groups are given.

Figure 7.

Cerebrospinal fluid, gray and white matter volume in the image in 18th slice of a participant with mild AD.

$$IA=\left(\text{lMTAi}-\text{dMTAi}\right)/\left(\text{lMTAi}+\text{dMTAi}\right)\times 100$$E2

Measurement of cerebrospinal fluid, gray matter, and white matter volumes from MRI images has been used to diagnose mild Alzheimer’s disease [18].

In this study, nonlinear property that reflects the dynamic nature of the brain signal, including Lyapunov exponent and correlation dimension, is also determined. On the other hand, in order to determine the optimal characteristics in three classes of healthy people, mild patients, and severe patients, the method of analysis of variance has been used. Brain signals from the three channels Fz, Cz, and Pz are recorded in four modes: closed-eye, open-eye, reminder, and stimulus. Forty-five properties are specified in the excitation mode. Table 2 shows the results of analysis of variance for three channels of Fz, Cz, and Pz between the group of healthy individuals, and mild and severe patients [19]. This analysis method is used in the classification of three or more classes to determine the optimal and effective characteristics.

Channel	Benefit features with ANOVA	Number of benefit feature
Fz	A-band power,D3-absolute Mean,D3-average power,D3-Std,Coherence	5
Cz	A-band power,D3-absolute Mean,D3-average power,D3-Std,Colierence,L-ave,D2	7
Pz	T-band power, A-band power,D3-absolute Mean,D3-average power,D3-Std, Coherence,L-ave,D2,ApEn	9

Table 2.

Compare the number and types of optimum features between Fz, Cz, Pz channels.

In this study, the purpose of using the classifier, after extracting the optimal characteristics of brain signals, was to separate the three groups of healthy people, mild and severe patients, and two classifiers such as SVM and the Elman neural network been used, aiming at comparing static and dynamic classifiers [20]. One of the classification methods with the teacher is the backup vector machine method. This view is based on statistical learning theory, but its implementation is similar to the neural network. This method was designed to separate data into two categories. Of course, if you use several SVMs in parallel and with different methods, this method can be used to classify data into more than two categories. SVM claims: It solves the major problem of neural networks, namely overfitting [21]. The results of EEG signal accuracy of different channels in different modes are determined at two levels (mild-severe, mild-healthy, and healthy-severe). Due to the linear separator for three classes containing poor results accuracy, levels are divided into two levels. In this study, a 2-layer Elman neural network was used which has 8 neurons in the latent layer and 1 neuron in the output layer [22]. The number of neural network inputs is equal to the number of features, and the number of hidden layer neurons is equal to the number of optimal features, and to determine the best results, various experiments with different numbers of neurons in the hidden layer have been performed. In the hidden layer and the output, the Sigmoid activation function is used due to its nonlinear property. There are many training functions for teaching the Elman network, wherein in this study the Levenberg-Marquardt error propagation algorithm was used due to higher convergence than other training functions, and the condition for stopping neural network training is an error coefficient of 0.001.

3. Block diagram of proposed method

Figure 8 shows the block diagram of the steps of the proposed method to diagnose Alzheimer’s disease.

Figure 8.

Results of SVM signal separation accuracy with different core functions with optimal characteristics obtained.

from ANOVA.

4. Result

Selected optimal features by ANOVA method in four modes of the closed eye, open eye, Reminder and stimulation are shown in Table 2 to compare the number and types of optimum features between Fz, Cz, Pz channels. The results of the accuracy of the separation of brain signals using SVM with different core functions with optimal characteristics obtained from ANOVA are shown in Figure 8. Figure 9 evaluates closed-eye, open-eye, reminder, and stimulation modes for the desired channels. According to the four modes of closing the eyes, opening the eyes, reminding and stimulating the brain signal, in order to better identify and introduce the features more accurately, each section is considered with a certain index. The closed part is considered with index c, the open eye part is considered with index o, the reminder part is with index r, and the stimulation part is considered with index s. In the excitation section, the target and non-target sound sections are defined with st and ss indices, respectively. In Figure 10, the results of separation of brain signals by SVM classifier with different main functions are compared with optimized ANOVA features with MRI features and without MRI features.

Figure 9.

Evaluation of closed-eye, open-eye, reminder and excitation modes for Fz, Cz, Pz channels in ANOVA mode.

Figure 10.

Compares the results of the resolution of brain signals by the SVM classifier with different core functions with ANOVA-optimized features with MRI features and without MRI features.

Due to the comparison of the results in Figure 11, by the support vector machine with different cores by the optimal brain signal characteristics by ANOVA with the addition of MRI image features, the accuracy of the results is reduced compared to the case where only the brain signal features are used.

Figure 11.

Support vector machine with different cores by the optimal brain signal characteristics by ANOVA with the addition of MRI image.

Finally, by using neural network dynamics because of the nonlinear properties studied and due to the nonlinear dynamics of the EEG signal, Elman neural network is used. The results in Figure 12 are compared by Elman with the optimal features of the brain signal obtained from ANOVA and with the addition of the features of MRI images, and the accuracy of the results is increased compared to the case where only the features of the brain signal are used.

Figure 12.

Compared by Elman with the optimal features of the brain signal obtained from ANOVA and with the addition of the features of MRI images.

5. Discussion

Medical image analysis has become very important in the diagnosis of mild Alzheimer’s disease in recent years [3]. But the important point is that from the way of analyzing medical images, we can determine the most effective channel for recording brain signals. 3D segmentation of MRI images further helps researchers diagnose Alzheimer’s disease and obtain important information [7]. And in 3D images, the most appropriate direction in the image is effective in determining the appropriate features. Determining the degree of atrophy of MRI images is an effective method for early detection of Alzheimer’s disease. Also, assessing the degree of asymmetry in both the right and left hemispheres and analyzing volumetric mismatch can differentiate between mild and severe Alzheimer’s disease [7]. The degree of asymmetry in the left and right hemispheres should be determined by the degree of atrophy and the ratio of the volume of gray matter to the volume of white matter. Using statistical features of signal and obtaining temporal information and using spatial features of MRI images is an effective method for the more accurate evaluation of Alzheimer’s disease [8]. The statistical properties of the signal are temporal in nature and the statistical properties of the image are spatial in nature. Cortical atrophy means the gradual destruction of the nerve cells that make up the upper regions of the brain, specifically the structures found in the cerebral cortex, mostly due to a reduction or loss of oxygen and nutrients in these areas. There are several methods for examining the Medial temporal lobe, the accuracy of which is not clear [9]. However, this condition is more suitable for mild patients. Longitudinal T1-weighted MRI studies are another effective way to distinguish mild Alzheimer’s patients from healthy ones [10]. But this feature has better results in severe patients to differentiate with mild patients. Also, extracting the appropriate characteristics and deciding on the classification in this field are among the issues to be considered.

6. Conclusion

Investigation and analysis of nonlinear dynamics of brain signal show nonlinear and dynamic behavior in different stages of Alzheimer’s disease. Nonlinear dynamics analysis of this signal shows a decrease in the complexity of the brain signal pattern and a decrease in connections due to a decrease in the nonlinear cell dynamics between cortical regions. The next two features are correlation and Lyapanov’s appearance, which indicates the feature space, and the convergence or divergence of this space is slightly reduced in this disease. The courses studied are closed-eyed, open-eyed, reminder, and stimulation, and among these four periods, the stimulation period was the best period for recording brain signals, because to diagnose Alzheimer’s disease, it is more effective to evaluate the speed of stimulus-response. The mean rate of asymmetry and the mean rate of temporal lobe atrophy increase with the progression of Alzheimer’s disease because the amount of damage to the temporal lobe in MRI images of Alzheimer’s disease has increased. The accuracy of the results in Elman neural network with the combination of brain signal features and medical images is 94.4% and in the case without combining the signal and image features, the accuracy of the results is 92.2%. The use of nonlinear classifiers is more appropriate than other classification methods due to the nonlinear dynamics of the brain signal. The accuracy of the results in the support vector machine with RBF core with the combination of brain signal features and medical images is 75.5% and in the case without combining the signal and image features, the accuracy of the results is 76.8%. Due to its nonlinear and normal distribution nature, this nucleus has been able to produce better results. Among the processing methods proposed to classify the three classes of healthy, mild, and severely ill, the method of combining brain signal characteristics and medical images has increased the accuracy of Elman classifier results and decreased the accuracy of SVM results. Because spatial features do not have the same nature as temporal features, and if the classifier divides the groups based on linear and non-return methods by extracting inappropriate features, the correct results will not be created. The main innovation in this research is the extraction of the most appropriate features and the appropriate combination of spatial features of medical images and temporal features of brain signals to diagnose Alzheimer’s disease.