Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Diagnosis of Alzheimer Disease Using 2D MRI Slices by Convolutional Neural Network

Diagnosis of Alzheimer Disease Using 2D MRI Slices by Convolutional Neural Network Hindawi Applied Bionics and Biomechanics Volume 2021, Article ID 6690539, 9 pages https://doi.org/10.1155/2021/6690539 Research Article Diagnosis of Alzheimer Disease Using 2D MRI Slices by Convolutional Neural Network 1 1 2 Fanar E. K. Al-Khuzaie , Oguz Bayat , and Adil D. Duru Graduate School of Science and Engineering, Altinbas University, Istanbul, Turkey Department of Physical Education and Sports Teaching, University of Marmara, Istanbul, Turkey Correspondence should be addressed to Fanar E. K. Al-Khuzaie; fanar@itnet.uobabylon.edu.iq Received 26 November 2020; Revised 14 January 2021; Accepted 22 January 2021; Published 2 February 2021 Academic Editor: Mohammed Yahya Alzahrani Copyright © 2021 Fanar E. K. Al-Khuzaie et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. There are many kinds of brain abnormalities that cause changes in different parts of the brain. Alzheimer’s disease is a chronic condition that degenerates the cells of the brain leading to memory asthenia. Cognitive mental troubles such as forgetfulness and confusion are one of the most important features of Alzheimer’s patients. In the literature, several image processing techniques, as well as machine learning strategies, were introduced for the diagnosis of the disease. This study is aimed at recognizing the presence of Alzheimer’s disease based on the magnetic resonance imaging of the brain. We adopted a deep learning methodology for the discrimination between Alzheimer’s patients and healthy patients from 2D anatomical slices collected using magnetic resonance imaging. Most of the previous researches were based on the implementation of a 3D convolutional neural network, whereas we incorporated the usage of 2D slices as input to the convolutional neural network. The data set of this research was obtained from the OASIS website. We trained the convolutional neural network structure using the 2D slices to exhibit the deep network weightings that we named as the Alzheimer Network (AlzNet). The accuracy of our enhanced network was 99.30%. This work investigated the effects of many parameters on AlzNet, such as the number of layers, number of filters, and dropout rate. The results were interesting after using many performance metrics for evaluating the proposed AlzNet. 1. Introduction aspects of our life, for example, genetic algorithm [3–5] and neural networks [6]. Machine learning (ML) is a field of artifi- cial intelligence that usually employs factual procedures to The normal brain of humans consists of mainly three regions, allow PCs to “learn” by utilizing information from saved data namely, white matter (WM), gray matter (GM), and cerebro- sets. At a very basic level, deep learning (DL) is a machine spinal fluid (CSF) [1]. The white matter is called as such learning subset [7]. Deep learning can be defined as a neural because of its white appearance. It contributes about sixty per- cent to the total brain volume. The gray matter is responsible network which uses a huge number of parameters and layers. There are many fundamental network architectures [8] like (i) of the whole processing of the neural signals. It consists of den- convolutional neural networks (CNNs) which are basically a drites and the neuron nuclei. It contributes almost about forty standard neural network that has been extended across space percent of the total brain volume. Cerebrospinal fluid is a col- using shared weights [9]. A convolutional neural network orless fluid that provides protection from mechanical shocks and also emits some important hormones to make the com- (CNN) is designed to recognize images by having convolu- tions inside, which see the edges of a recognized object on munication possible among the white matter, gray matter, the image [10]. (ii) Recurrent neural networks (RNNs) are a and spinal cord of the central nervous system [2]. It is known denomination of artificial neural networks where connections that the family of artificial intelligence (AI) includes many among nodes lay out a directed graph along the temporal algorithms and methods which could be used in different 2 Applied Bionics and Biomechanics Since MRI can provide a lot of invaluable information about sequence. Unlike feedforward neural networks, RNNs have the ability to use their internal state for processing the structures of tissue, such as localization, size, and shape, it is sequences of inputs. RNN is designed to recognize sequences, attracting more of attention for computer-aided diagnosis for example, a speech signal or a text [9]. (iii) Recursive neural and clinical routine [24, 25]. MRI can be divided into func- networks are more like a hierarchical network where there is tional and structural imaging. Functional imaging contains really no time aspect to the input sequence, but the input has tasking state functional MRI (ts-fMRI), resting state functional to be processed hierarchically in a tree fashion [8, 10]. MRI (rs-fMRI), etc., structural imaging contains T1-weighted Generally, different external stimuli match to different brain MRI (T1w), diffusion tensor imaging (DTI), and T2-weighted activities, and the different brain activities display different MRI (T2w) [26]. Medical data systems are diagnostic, and functional brain images [11]. For that, image classification plays analytical systems are applied to help medical centers and phy- asignificant role in identifying different activities of the brain. sicians in disease treatment, and they are critical to improve Recently, many methods of deep learning were proposed to per- treatment and diagnosis. Computer scientists have been inter- form classification of image for different brain activities [12, 13]. ested in this domain given the vital role of medical data in the To identify different activities of the brain including emotions, lives of humans. Physicians may refer to the classification of motor, social, relational and language activities, and working medical data, including medical analyses and symptoms of memory, Koyamada et al. [12] applied a feedforward deep neu- critical diseases, for making the decisions. A data set of disease ral network from images of functional magnetic resonance contains symptoms of patients as attributes besides the num- imaging (fMRI) to implement this mission. The feedforward ber of instances of these symptoms. Health care may use the deep neural network involved a softmax layer and multiple hid- considerable medical data accessible. In the analyses of medi- den layers. Similarly, these hidden layers were used to get high- cal centers, data mining could be used to provide sufficient level latent features, while the softmax layer was applied to cal- sources on illnesses for their prevention and timely detection culate the probability of every subjects in a class. In addition, and to avoid the expensive costs incurred by medical tests dropout, minibatch stochastic gradient descent, [14], and prin- [27]. Representation of features plays a significant role in med- cipal sensitivity analysis [15] were combined into the feedfor- ical image analysis and processing. Deep learning has two ward deep neural network to improve the performance of the obvious advantages in the representation of features: final classification. Recently, to classify different sensorimotor tasks including auditory attention, visual stimulus, right-hand (i) It can be applied to automatically discover features clenching, and left-hand clenching, Jang et al. [13] used fully from a given data set for every specific application. connected feedforward deep neural networks and multiple hid- Usually, methods of traditional feature extraction den layers. In addition to the above classifications, the methods are based on some prior knowledge for extracting of deep learning classification of magnetic resonance imaging features in a certain application. So, these approaches (MRI) images have been used also by other fields of classifica- are semiautomatic learning methods tion, like stroke diagnosis [16], age prediction [17], classification of attention deficit hyperactivity disorder (ADHD) [18], dis- (ii) It can discover new features that are appropriate to spe- crimination of cerebellar ataxia types [19], and emotional cific applications, which have never been discovered by response prediction [20]. Due to the science engineering field, researchers previously. Traditional methods of feature it was doable to create systems of computer-aided diagnosis extraction are often restricted by some a priori knowl- (CAD) that play a critical role in assisting the researchers and edge, which can only extract some features which are physicians when they interpret the medical imaging. Recently, associated with a certain application [28, 29] the use of the machine learning approach, especially DL tech- niques in systems of CAD to diagnose and classify the healthy Medical imaging is the mechanism and process of establishing visual representations of the interior of the body control normal (CN) people, Alzheimer’sdisease (AD), and mild cognitive impairment (MCI) patients, has exponentially for medical intervention and clinical analysis [30]. Machine increased [21, 22]. The Alzheimer’s disease automatic diagno- learning tools and medical image processing can help neurologists in estimating whether a subject is developing sis, especially in its early stage, plays a significant role in human health. Since Alzheimer’s disease is a neurodegenerative illness, Alzheimer’s disease [31]. Alzheimer’s disease is a chronic neurodegenerative disease causing tissue loss throughout it has a long period of incubation. Thus, it is necessary to ana- the brain, and the death of nerve cells usually starts slowly lyze the AD symptoms at different stages. Currently, a lot of researchers have proposed using the classification of images and worsens over time [32]. Alzheimer’s disease is expected to affect more and more people by the year 2050. The cost to perform diagnosis of AD. Moreover, many DL methods have been proposed to implement severity classification of dif- of caring for patients of AD is also expected to rise [33]. Presently, AD is the sixth reason that leads to death in the ferent Alzheimer’s disease patients by using MRI images [22, United States [34]. For this reason, individual computer- 23]. As known in image processing and analyzing, the better the image quality, then the better the results gained. However, aided systems are necessary for accurate and early diagnosis of this disease [33]. There are many approaches for accurate the image quality depends on acquisition of the image, so, when the image acquisition is better, then the image quality is and automatic classification of brain MRI, and one of them is our work. The next part of this article is the related works, higher. Magnetic resonance imaging (MRI) not only keeps then we will talk about our methodology, the results, the the features of noninvasive and good soft tissue contrast, but in addition does not expose humans to high ionizing radiation. discussion, and, at the end, our references. Applied Bionics and Biomechanics 3 matter from the human brain and make the classification by 2. Related Works using the CNN. To enhance the voxels, a Gaussian filter is Researchers have been applying machine learning techniques applied, and to remove the irrelevant tissues, the skull stripping algorithm is used. After that, by applying a hybrid, enhanced, to build classifiers by using clinical measures and imaging data for AD diagnosis. These studies have identified the important and independent component analysis, those voxels are seg- structural differences in the regions such as the entorhinal mented. The input to the CNN was segmented gray matter. cortex and hippocampus entorhinal cortex between the brain Clinical valuation was performed using the provided approach with AD and healthy brain. Different imaging methods like and 90.47 accuracy was achieved. Hamed and Kaabouch [41] proposed a method that yielded good classification accuracy. the functional and structural magnetic resonance imaging (fMRI and sMRI, respectively), single photon emission com- The convolutional neural network with modified architecture puted tomography (SPECT), position emission tomography was used to get the high quality features from the brain MRI (PET), and diffusion tensor imaging (DTI) scans which can to classify people into healthy, early mild cognitive impairment perceive the changes causing AD due to the degeneration of (EMCI), or late mild cognitive impairment (LMCI) groups. The results showed the classification between control normal (CN) cells of the brain. In recent years, deep learning models, especially convolutional neural networks, have demonstrated and LMCI groups in the sagittal view with 94.54 accuracy. outstanding performance for medical image analysis. Payan and Giovanni [35] produced and tested a pattern classification 3. Materials and Methods system which combines convolutional neural network and Inside a CNN, a filter series, with an equivalent size to a small sparse autoencoders. Ehsan et al. [36] adapted a 3D-CNN image patch, automatically searches the entire image to find model for diagnosis of AD. The 3D-CNN is built upon the images of similar spatial features. These filters can be learned 3D convolutional autoencoder, which is pretrained to catch and updated independently; thus, a collection of them can anatomical shape variations in scans of structural brain MRI. detect crucial information of a specific task and data set [42]. Sergey et al. [37] proposed two different kinds of 3D convolu- There are standard steps of CNN. The first step is named tional network architectures to classify the brain MRI which are “convolution”; this step is responsible for finding the features the amendments of residual and plain convolutional neural and applying filters. It is a filter kernel that picks up its weights networks. Applied convolutional neural networks can tackle by convolving the input data tensor with such kernel. There the two problems stated before. These networks can propagate are several variables that effect the convolutional operation local features into the metarepresentation of an object for output such as strides and number of filters. The distance in classification or image recognition. In deep learning for image pixels between two pixels is the stride, while the number of fil- classification, modern advancements like residual network ters states the output feature map number [43]. The operation architectures and batch normalization mechanism alleviate of convolution is just a mathematical operation, which should the issues of having small data sets of training, while providing be treated equally with other operations such as multiplication a frame for automatic feature generation. As a result, these or addition and should not be discussed particularly in the models can be used to 3D MRI images in the absence of literature of machine learning. But, it has still been discussed intermediate handcrafted feature extraction. Karim et al. [38] here for completeness. Convolution is a mathematical opera- adapted three tasks of binary classification which are consid- tion on two functions (e.g., f and g)and producesathird ered for separating the normal control (NC) subject from mild function h; this is an integral that expresses the amount of cognitive impairment (MCI) patients and Alzheimer’sdisease overlap of one function (f ) as it is shifted over the other func- (AD). Two fusion methods on a fully connected (FC) layer tion (g) [44]. Formally, it is described as and on the single-projection CNN output offer better achieve- ment by about 91% accuracy. The outcomes are competitive with the SOA which utilizes a heavier algorithmic chain. Fan ht = ðÞ fTðÞgtðÞ − T dT, ð1Þ and Manhua [39] proposed a classification technique built on −∞ multiple clusters of dense convolutional neural networks (DenseNets) to pick up the various local features for images And denoted as h = f ∗ g. of the MR brain, which are collected for classification of AD. A typically convolutional neural network works with The total brain image is partitioned into different local parts two-dimensional convolution operation that could be sum- and from each region, a number of 3D patches are extracted. marized in Figure 1. As displayed in Figure 1, the input By using theK-means clustering method for grouping the matrix is Figure 1(a), and Figure 1(b) is usually called a kernel patches from each region into different clusters, the DenseNet matrix. So convolution is applied to these matrices, then the had been constructed to pick up the patch features for each result is displayed as in Figure 1(c). The process of convolu- cluster, and the features learned from the characteristic clusters tion can be considered as an element-wise product followed of each part are grouped for classification. The classification by a sum, like what is shown in the example of Figure 1. outputs from different local parts are combined to foster the When the left upper matrix which is 3×3 convoluted with final image classification. This method can progressively learn the kernel, then the result is 29. After that, the target 3×3 the features of MRI from the local patches to the global image matrix slides one column to the right, then is convoluted with level for the task of classification. For preprocessing images of the kernel and gets the result 22. The sliding and recording of MRI, there are no segmentation and rigid registration required. the results have been continued as a matrix. Every target Shaik and Ram [40] provided an approach to extract the gray matrix is 3×3, because the kernel is 3×3; thus, the whole 5 4 Applied Bionics and Biomechanics 10 37 1 5 40 4 2 16 55 20 001 29 22 42 28 17 2 0 47 25 55 39 9 0 19 25 13 10 30 88 830 41 631 (a) (b) (c) Figure 1: A simple illustration of a two-dimension convolution operation: (a) input matrix, (b) kernel matrix, and (c) output matrix after convolution. ×5 matrix is shrunk into a 3×3 matrix when every 3×3 many works have attempted to understand its technique in different perspectives, including [49]. It has also been used to matrix is convoluted to one digit. (Because of 5 − ð3 − 1Þ =3 , the first 3 means the kernel matrix size.) One should realize train other models, such as SVM [50]. that the convolution process is a locally shifted invariant, The CNN architecture which was used in this study is which means that for many different combinations of how composed of five convolutional layers which take an input the nine numbers in the upper matrix 3×3 are placed, the image (the brain’s MRI slice) with a size of 200 ∗ 200. Figure 2 shows some slices of the brain’s MRI; those were convoluted result will be 29. This invariant property plays a crucial role in vision problem because the result of recogni- we used in our research. All five convolutional layers were tion should not be changed due to shift or rotation of features followed by a max-pooling layer. The 64 filters with a kernel in an ideal case. This crucial property is applied to be solved size of 9 ∗ 9 were considered for the first convolutional layer, elegantly by [45], but CNN brought the performance up to a and the max-pooling layer kernel size was set on 2 ∗ 2. The new level. 64 filters with a kernel size of 7 ∗ 7 were considered for the With each convolution layer, there is an activation func- second convolutional layer, and the max-pooling layer kernel tion; the activation is an operation which converts the input size also was set on 2 ∗ 2. The 64 filters with a kernel size of from a linear data tensor to a nonlinear data tensor. In deep 5 ∗ 5 were considered for the third convolutional layer, and learning, many activation functions are popular such as recti- the max-pooling layer kernel size was set on 2 ∗ 2.The 32 fied linear units (ReLU), sigmoid, and tanh [46]. Recently, the filters with a kernel size of 5 ∗ 5 were considered for the fourth rectified linear unit (ReLU) has been used more than the other convolutional layer, and the max-pooling layer kernel size was nonlinear functions, because it does not activate all the set on 2 ∗ 2.The 32 filters with a kernel size of 3 ∗ 3 were con- neurons at the same time [24]. The second step is named sidered for the fifth convolutional layer, and the max-pooling “Max Pooling”; this step is responsible for downsizing the layer kernel size was set on 2 ∗ 2. It is worthwhile to mention image and keeping the important features. Pooling is the that the ReLU (rectified linear unit) function was used as the operation of downsampling which can be performed globally activation functions in all convolutional layers. The ReLU or locally. The function of global pooling returns for every function is used commonly in models of DL; basically, if the 2D feature map a scalar value. The function of local pooling function receives a negative value as input, it returns 0, and downsamples local image parts by a factor [43]. The third step if the function receives a positive value, then the same positive named “flattening” converts to one dimension array (vector). value will return. The fourth step is named “full connection”; this step is respon- The function of ReLU is understood as f ðaÞ = max ð0, aÞ. sible of building all needed connections. The fully connected Figure 3 demonstrates the block diagram of the proposed layer (FC) is typically followed by an activation layer. FC is system (AlzNet). After the convolution layers and the flatten- the layer where the receptive domain is a whole channel of ing layer, there is a dense unit 121, and here, we used a ReLU the former layer [43, 46]. The last step is named “classifier”; as an activation function, then we used a dropout (0.2) to it represents the classification stage to decide if the image is prevent overfitting, then there is a dense unit and sigmoid normal or abnormal [47]. The use of the dropout technique as an activation function; at the last stage, there is a binary is so common in convolutional neural networks. Dropout classifier for displaying the results. was introduced in [14, 48]. This mechanism soon got influen- Table 1 demonstrates the number of MRI slices. There tial, not only because it has good performance but also because are samples for men and women such as a left-handed man of its simplicity of implementation. The idea is very easy: while (L.-handed male), left-handed woman (L.-handed female), training, randomly drop out some of the units. More formally: right-handed man (R.-handed male), and a right-handed for each training case, every hidden unit is randomly omitted woman (R.-handed female); all brain MRIs were in the axial with a probability of p from the network. As suggested in [14], view manner. Keras provides a perfect tool to show a model’s dropout can be seen as an efficient method to perform model summary; Table 2 demonstrates that summary. This displays averaging across a great number of different neural networks, the number of trainable parameters and the output shape for where overfitting can be avoided with less cost of computation each layer. Before starting to fit the model, this is a sanity because of the actual performance which it introduces. Drop- check. So the total params = 414,419, the trainable params out became very popular upon when it was first introduced; = 414,419, and the nontrainable params = 0. Applied Bionics and Biomechanics 5 Left-handed woman Left-handed man Right-handed woman Right-handed man Figure 2: Some brain MRI slices of Alzheimer’s patients. Test Input dataset Output MRI slices Training Abnormal dataset dataset Prediction 2D-CNN Validation Normal dataset ⁎ Dense+ReLU 2 2 Flattening ⁎ ⁎ ⁎ ⁎ 2 2 2 2 2 2 2 2 pooling pooling pooling pooling pooling With Dense+sigmoid dropout ⁎ ⁎ ⁎ ⁎ 7 7 Convolution 5 5 Convolution 5 5 Convolution 3 3 Convolution ⁎ Classifier 9 9 Convolution 64 filters+ReLU 64 filters+ReLU 32 filters+ReLU 32 filters+ReLU 64 filters+ReLU 2D-CNN Figure 3: Block diagram of proposed system (AlzNet). There are five convolutional layers; after each convolutional layer there is a max- pooling layer; the activation function in every convolutional layer was ReLU. After the convolution layers and flattening layer, there is a dense unit 121 with a ReLU as an activation function, with a dropout to prevent overfitting, then there is a dense unit and sigmoid as an activation function; at the last stage, there is a binary classifier. Table 1: Number of MRI slices. 4. Results Number of Number of MRI Subjects Python language 3.6 and Keras have been used for program- subjects slices ming this work. Keras is a high-level library; it is an open R.-handed male (AD patient) 60 3050 source machine learning library that is written in Python. L.-handed male (AD patient) 40 2600 Keras is used for numerical computation purposes; it is used L.-handed female (AD patient) 30 2150 to perform the computations more easily and efficiently in R.-handed female (AD patient) 40 2600 practice. The training data set was 75% and the validation Female (NC) 25 2000 data set was 25%. There were many practical experiments Male (NC) 45 2800 that had been done in this research for trying to find the best parameters of this convolutional neural network. So we try to find the best number of dense units for hidden layers depend- ing on the result of accuracy, whereas the other researches used different numbers at each one. Another parameter we tested many times is the rate of dropout to find the fit rate 6 Applied Bionics and Biomechanics Table 2: Summary of the proposed model. Table 3: Accuracy of AlzNet depending on dropout rate and dense unit. Layer (type) Output shape Param # Dropout rate Conv2d_1 (Conv2D) (None, 192, 192, 64) 15,616 0.5 0.4 0.3 0.2 0.1 Max_pooling2d_1 (MaxPooling2) (None, 96, 96, 64) 0 130 94.90 94.21 94.78 94.02 94.12 Conv2d_2 (Conv2D) (None, 90, 90, 64) 200,768 129 95.18 95.72 95.61 95.31 95.48 Max_pooling2d_2 (MaxPooling2) (None, 45, 45, 64) 0 128 94.85 94.56 94.71 94.65 94.58 Conv2d_3 (Conv2D) (None, 41, 41, 64) 102,464 127 95.09 95.12 95.01 95.23 95.31 Max_pooling2d_3 (MaxPooling2) (None, 20, 20, 64) 0 126 96.61 96.72 96.40 95.97 96.01 Conv2d_4 (Conv2D) (None, 16, 16, 32) 51,232 Dense unit 125 96.15 96.05 96.11 96.44 96.21 Max_pooling2d_4 (MaxPooling2) (None, 8, 8, 32) 0 124 96.09 96.12 96.02 96.30 96.05 Conv2d_5 (Conv2D) (None, 6, 6, 32) 9248 123 95.91 95.72 95.82 95.01 95.11 Max_pooling2d_5 (MaxPooling2) (None, 3, 3, 32) 0 122 96.85 96.56 96.90 95.77 95.83 Flatten_1 (flatten) (None, 288) 0 121 97.32 97.16 97.30 97.88 97.06 Dense_1 (dense) (None, 121) 34,969 120 95.43 95.66 94.94 94.88 94.90 Dropout_1 (dropout) (None, 121) 0 Dense_2 (dense) (None, 1) 122 True Positive + True Negative Accuracy = : for our convolutional neural network by observing the results True positive + True negative + False positive + False negative of accuracy. That is figured out in Table 3. ð5Þ So it is obvious from Table 3 that the highest accuracy value was when we utilized 0.2 for the dropout rate and 121 for the dense unit. In fact, the range of the dropout rate which In this work, the measure metrics have been applied on we tested was (from 0.1 to 0.5) increasing by 0.1, when the the training data set and test data set (see Table 4). number of dense units was 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, and 130. Figure 4 shows the accuracy 5. Discussion value depending on the relationship between the dense unit number and the dropout rate. For our current work, we develop an efficient deep convolu- There are several metrics for measuring the performance tional neural network based on a classifier and demonstrate of binary classification [51], such as recall, precision, specific- very good performance by using the OASIS data set [54]. ity, and Accuracy. Precision is very helpful because we want The OASIS-3 data set has been saved in the XNAT central to be confident of our forecast, since it tells us how many of repository [55]. In our work, a total of 15,200 MRI axial slices the values expected as positive are actually positive [50], as were used. The data was used to include the MRI scans of follows: about 170 AD patients and 70 NC. They are all from different subjects that make the test of recognition performance more True positive reliable. The age of each patient is between the range of 65-90 Precision = : ð2Þ years old, both male and female in this work. At the first stage True positive + False positive of data preprocessing, we obtained 2D slices from each MRI image. Then, the every last 20 dark slices with each time Recall (sensitivity) is another very valuable measure that course was discarded because they included no functional helps one, for instance, to know the proportion of the num- ber of values accurately labeled as positive on the overall information. Preprocessed images are augmented by rotating the slices to see whether or not the model can recognize the values which are actually positive, as follows: images; this increased the samples size and made a good True positive training of the CNN model. The scans are T1-weighted Recall = : ð3Þ whereas those of [40] were T2-weighted. Whereas [35, 41] True positive + False negative used three convolution layers, our proposed system (CNN model) has 5 convolution layers; each convolution layer has Using the F1 score is a safe way to get a complete impres- ReLU as an activation function. Each convolution layer is sion of recall and precision. The F1 score provides us the har- followed by max pooling layers. Our proposed approach per- monic mean of recall and precision [52, 53], as follows: forms binary classification to fit the model in a batch size of Precision ∗ Recall 64 in 150 epochs. Table 2 summarizes the total architecture F1 score = 2 ∗ : ð4Þ of the proposed system. When [40, 41] used Adam optimiza- Precision + Recall tion, we trained the model by Adadelta optimization with a Accuracy is the proportion of accurate predictions (both dropout rate of 0.2 for the dropout layer which had been true negative and true positive) among the entire number of utilized to prevent the overfitting like in [40], but [41] used cases examined [52], as follows: a 0.5 dropout rate. The number of dense units was 121 when Applied Bionics and Biomechanics 7 Accuracy of AlzNet depending on dropout rate and dense unit Table 5: Literature reviews. 98.5 Methods Year Accuracy 97.5 Sergey et al. [37] 2017 80.00 96.5 96 Fan and Manhua [39] 2018 89.5 95.5 Shaik and Ram [40] 2019 90.47 94.5 Ehsan et al. [36] 2016 97.6 Karim et al. [38] 2017 91.41 93.5 Payan and Giovanni [35] 2015 95.39 130 129 128 127 126 125 124 123 122 121 120 Hamed and Kaabouch [41] 2019 94.54 0.5 0.2 0.4 0.1 Proposed method 2020 99.30 0.3 Figure 4: Accuracy of AlzNet depending on dropout rate and dense 6. Conclusions unit. In order to diagnose Alzheimer’s disease, deep neural networks, especially CNNs, can provide meaningful informa- Table 4: Measure metrics of training data set and test data set. tion. A CNN-based method for extracting discriminatory Dataset Precision Recall F1 score Accuracy features from structural MRI was proposed in this paper, with the goal of classifying Alzheimer’s disease and healthy Training 97.06 97.99 97.52 97.88 subjects using 2D MRI slices. For potential AD individuals, Test 98.92 99.53 99.22 99.30 the suggested approach can lead to many advantages and can also lead to an early diagnosis of AD. The experimental results of the OASIS database for 240 subjects demonstrated Measure metrics of training data set that our proposed method of extraction and classification of and test data set features provided high accuracy for AD and CN. The best 99.53 99.3 99.22 results have been obtained for the classification between the 98.92 CN and the AD axial view of the MRI. The proposed method 97.99 97.88 98 yielded a classification accuracy of 99.30 percent. The above 97.52 97.06 results indicate higher reliability, recall, precision, and F1 score of our proposed method for the diagnosis of AD and the classification between CN and AD. Training data set Test data set Precision F1 score Data Availability Recall Accuracy Our data set was from the OASIS database; the website is Figure 5: Measure metrics of training data set and test data set. https://www.oasis-brains.org. The OASIS-3 data set has been saved in the XNAT central repository; the website is https:// central.xnat.org. the number of dense units of [37] was 128; actually, we made many experiments to decide which is the best number of dense units we should take, and Table 3 shows that. In this Conflicts of Interest work, we tried to put different values of the neural network The authors declare that there is no conflict of interest parameters by trial and error, by relying on the accuracy regarding the publication of this paper. value, and comparison with previous researches. During binary classification, we trained the classifier for AD and CN images, and the model resulted in 97.88% training Acknowledgments accuracy and 99.30 test accuracy. It is required to mention that our proposed framework had been trained, and the Data were provided in MRI by OASIS-3. The principal prediction was made with utmost accuracy. Figure 5 demon- investigators are T. Benzinger, D. Marcus, and J. Morris. The strates that. The accuracy of the proposed system has been study was funded by NIH (P50AG00561, P30NS09857781, compared with different models discussed in literature P01AG026276, P01AG003991, R01AG043434, UL1TR000448, reviews as shown in Table 5. and R01EB009352). It is observed that the proposed model achieves remark- able performance. The last thing we have to say is that neural References networks have plenty of parameters, and any change in one of them will make the value of results different, and also, [1] M. Nazir, F. Wahid, and S. Ali Khan, “A simple and intelligent there is a big important reason for making a variation of approach for brain MRI classification,” Journal of Intelligent & results—it is the data set and its type. Fuzzy Systems, vol. 28, no. 3, pp. 1127–1135, 2015. 8 Applied Bionics and Biomechanics [17] J. H. Cole, R. P. K. Poudel, D. Tsagkrasoulis et al., “Predicting [2] M. H. Fazel Zarandi, M. Zarinbal, and M. Izadi, “Systematic image processing for diagnosing brain tumors: a type-II fuzzy brain age with deep learning from raw imaging data results in a expert system approach,” Applied Soft Computing, vol. 11, reliable and heritable biomarker,” NeuroImage, vol. 163, no. 1, pp. 285–294, 2011. pp. 115–124, 2017. [18] D. P. Kuang and L. H. He, “Classification on ADHD with deep [3] H. Nabeel Kaghed, S. Eman al–Shamery, and F. E. K. al-Khu- zaie, “Multiple sequence alignment based on developed genetic learning,” in Proc., Int. Conf. Cloud Computing and Big Data (CCBD), pp. 27–32, Wuhan, China, 2014. algorithm,” Indian Journal of Science and Technology, vol. 9, no. 2, pp. 1–7, 2016. [19] Z. Yang, S. H. Zhong, A. Carass, S. H. Ying, and J. L. Prince, “Deep learning for cerebellar ataxia classification and func- [4] N. H. Kaghed, E. S. Al, and F. E. K. Al-Khuzaie, “Comparative study of genetic algorithm and dynamic programming of DNA tional score regression,” in Proc. 5th Int. Workshop on Machine Learning in Medical Imaging, pp. 68–76, Boston, multiple sequence alignment,” Journal of University of Baby- lon, vol. 25, no. 2, pp. 403–414, 2017. MA, USA, 2014. [20] H. C. Kim and J. H. Lee, “Deep learning to predict the emo- [5] F. Al-Khuzaie, “Multiple sequence alignment based on devel- tional response using functional MRI data,” NeuroImage, oped genetic algorithm,” (Master thesis), University of Baby- vol. 186, pp. 607–627, 2019. lon, Iraq, 2016, https://www.researchgate.net/publication/ 347511906_Multiple_Sequence_Alignment_Based_on_ [21] H. T. Gorji and J. Haddadnia, “A novel method for early diag- Developed_Genetic_Algorithm?channel=doi&linkId= nosis of Alzheimer's disease based on pseudo Zernike moment 5fdf736a45851553a0db2ee7&showFulltext=true. from structural MRI,” Neuroscience, vol. 305, pp. 361–371, [6] V. G. Gudise and G. K. Venayagamoorthy, “Comparison of par- ticle swarm optimization and backpropagation as training algo- [22] S. Liu, S. Liu, W. Cai, S. Pujol, R. Kikinis, and D. Feng, “Early rithms for neural networks,” in Proceedings of the 2003 IEEE diagnosis of Alzheimer’s disease with deep learning,” in Pro. Swarm Intelligence Symposium. SIS'03 (Cat. No.03EX706), IEEE 11th International Symposium on Biomedical Imaging pp. 110–117, Indianapolis, IN, USA, 2003. (ISBI), pp. 1015–1018, Beijing, China, 2014. [7] T. H. H. Aldhyani, M. Alrasheedi, A. A. Alqarni, M. Y. Alzah- [23] F. Li, L. Tran, K. H. Thung, S. W. Ji, D. G. Shen, and J. Li, “A robust deep model for improved classification of AD/MCI rani, and A. M. Bamhdi, “Intelligent hybrid model to enhance time series models for predicting network traffic,” IEEE Access, patients,” IEEE J Biomed Health Inform, vol. 19, no. 5, pp. 1610–1616, 2015. vol. 8, pp. 130431–130451, 2020. [8] G. Litjens, T. Kooi, B. E. Bejnordi et al., “A survey on deep [24] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet clas- sification with deep convolutional neural networks,” in Proc. of learning in medical image analysis,” Medical Image Analysis, vol. 42, pp. 60–88, 2017. the Advances in Neural Information Processing Systems, pp. 1097–1105, Lake Tahoe, CA, USA, 2012. [9] M. Z. Alom, T. M. Taha, C. Yakopcic et al., “A state-of-the-art [25] J. Liu, M. Li, Y. Pan et al., “Complex brain network analysis survey on deep learning theory and architectures,” Electronics, and its applications to brain disorders: a survey,” Complexity, vol. 8, no. 3, p. 292, 2019. vol. 2017, Article ID 8362741, 27 pages, 2017. [10] D. Ravi, C. Wong, F. Deligianni et al., “Deep learning for [26] L. Jin, P. Yi, L. Min et al., “Applications of deep learning to health informatics,” IEEE Journal of Biomedical and Health MRI images: a survey,” Big Data Mining and Analytics, Informatics, vol. 21, no. 1, pp. 4–21, 2017. vol. 1, no. 1, pp. 1–18, 2018. [11] T. Horikawa, M. Tamaki, Y. Miyawaki, and Y. Kamitani, [27] O. S. Soliman and E. A. El Hamd, “Classification of breast can- “Neural decoding of visual imagery during sleep,” Science, cer using differential evolution and least squares support vec- vol. 340, no. 6132, pp. 639–642, 2013. tor machine,” Int J Emerg Trends Technol Comput Sci, vol. 3, [12] S.Koyamada,Y.Shikauchi,K. Nakae,M.Koyama,and no. 2, pp. 155–161, 2014. S. Ishii, “Deep learning of fMRI big data: a novel approach [28] J. Liu, M. Li, J. Wang, F. Wu, T. Liu, and Y. Pan, “A survey of to subject-transfer decoding,” 2015, https://arxiv.org/abs/ MRI-based brain tumor segmentation methods,” Tsinghua 1502.00093. Science and Technology, vol. 19, no. 6, pp. 578–595, 2014. [13] H. Jang, S. M. Plis, V. D. Calhoun, and J. H. Lee, “Task-specific [29] J. Schmidhuber, “Deep learning in neural networks: an over- feature extraction and classification of fMRI volumes using a view,” Neural Networks, vol. 61, pp. 85–117, 2015. deep neural network initialized with a deep belief network: [30] M. Natarajan and S. Sathiamoorthy, “A novel pre-processing evaluation using sensorimotor tasks,” Neuro Image, vol. 145, approach for the denoising of Alzheimer disease image data- Part B, pp. 314–328, 2017. set,” Asian Journal of Computer Science and Technology, [14] G. E. Hinton, N. Srivastava, A. Krizhevsky, I. Sutskever, and vol. 7, no. 2, pp. 107–112, 2018. R. R. Salakhutdinov, “Improving neural networks by prevent- [31] E. Moradi, A. Pepe, C. Gaser, H. Huttunen, J. Tohka, and Alz- ing co-adaptation of feature detectors,” 2012, https://arxiv.org/ heimer's Disease Neuroimaging Initiative, “Machine learning abs/1207.0580. framework for early MRI-based Alzheimer's conversion predic- [15] S. Koyamada, M. Koyama, K. Nakae, and S. Ishii, “Principal tion in MCI subjects,” NeuroImage, vol. 104, pp. 398–412, 2015. sensitivity analysis,” in Proc. 19th Pacific-Asia Conf. Knowledge [32] C. Vaishnavi and P. Anna, “Alzheimer’s disease: a comprehen- Discovery and Data Mining, pp. 621–632, Ho Chi Minh City, sive review,” EC Neurology, vol. 8, no. 1, pp. 16–32, 2017. Vietnam, 2015. [33] R. Brookmeyer, E. Johnson, K. Ziegler-Graham, and H. M. [16] T. Schmah, G. E. Hinton, R. S. Zemel, S. L. Small, and Arrighi, “Forecasting the global burden of Alzheimer’s dis- S. Strother, “Generative versus discriminative training of rbms ease,” Alzheimer's & Dementia, vol. 3, no. 3, pp. 186–191, 2007. for classification of fMRI images,” in Proc. 21st Int. Conf. Neu- ral Information Processing Systems, pp. 1409–1416, Vancou- [34] S. L. Murphy, J. Xu, and K. D. Kochanek, “Deaths: final data ver, Canada, 2008. for 2010,” Natl Vital Stat Rep, vol. 61, no. 4, pp. 1–117, 2013. Applied Bionics and Biomechanics 9 [51] T. H. Aldhyani, A. S. Alshebami, and M. Y. Alzahrani, “Soft [35] A. Payan and M. Giovanni, “Predicting Alzheimer's disease: a neuroimaging study with 3D convolutional neural networks,” Computing Model to Predict Chronic Diseases,” Journal of in Proc. Int. Conf. Pattern Recog. Appl. Methods, 2015. Information Science and Engineering, vol. 36, no. 2, pp. 365– 376, 2020. [36] E. Hosseini-Asl, R. Keynton, and A. El-Baz, “Alzheimer's dis- ease diagnostics by adaptation of 3D convolutional network,” [52] T. H. H. Aldhyani, A. S. Alshebami, and M. Y. Alzahrani, “Soft in 2016 IEEE International Conference on Image Processing clustering for enhancing the diagnosis of chronic diseases over (ICIP), pp. 126–130, Phoenix, AZ, USA, September 2016. machine learning algorithms,” Journal of Healthcare Engineer- ing, vol. 2020, Article ID 4984967, 16 pages, 2020. [37] S. Korolev, A. Safiullin, M. Belyaev, and Y. Dodonova, “Resid- ual and plain convolutional neural networks for 3D brain MRI [53] A. M. A. Mohamed, O. N. Uçan, O. Bayat, and A. D. Duru, classification,” in 2017 IEEE 14th International Symposium on “Classification of resting-state status based on sample entropy Biomedical Imaging (ISBI 2017), pp. 835–838, Melbourne, and power spectrum of electroencephalography (EEG),” VIC, Australia, April 2017. Applied Bionics and Biomechanics, vol. 2020, Article ID 8853238, 10 pages, 2020. [38] A. Karim, B. P. Jenny, and A. Karim, “Classification of sMRI for Alzheimer's disease diagnosis with CNN: single Siamese [54] P. J. LaMontagne, T. L. S. Benzinger, J. C. Morris et al., networks with 2D+? Approach and fusion on ADNI,” in Proc. “OASIS-3: longitudinal neuroimaging, clinical, and cognitive of the 2017 ACM on International Conference on Multimedia dataset for normal aging and Alzheimer disease,” med Rxiv, Retrieval, pp. 494–498, 2017. [39] F. Li, M. Liu, and Alzheimer's Disease Neuroimaging Initiative, [55] R. Herrick, W. Horton, T. Olsen, M. McKay, K. A. Archie, and “Alzheimer's disease diagnosis based on multiple cluster dense D. S. Marcus, “XNAT Central: open sourcing imaging research convolutional networks,” Computerized Medical Imaging and data,” NeuroImage, vol. 124, Part B, pp. 1093–1096, 2016. Graphics, vol. 70, pp. 101–110, 2018. [40] S. Basheera and M. S. Sai Ram, “Convolution neural network– based Alzheimer's disease classification using hybrid enhanced independent component analysis based segmented gray matter of T2 weighted magnetic resonance imaging with clinical val- uation,” Alzheimer's & Dementia: Translational Research & Clinical Interventions, vol. 5, no. 1, pp. 974–986, 2019. [41] H. T. Gorji and N. Kaabouch, “A deep learning approach for diagnosis of mild cognitive impairment based on MRI images,” Brain Sciences, vol. 9, no. 9, pp. 217–231, 2019. [42] S. Qiu, G. H. Chang, M. Panagia, D. M. Gopal, R. Au, and V. B. Kolachalama, “Fusion of deep learning models of MRI scans, mini–mental state examination, and logical memory test enhances diagnosis of mild cognitive impairment,” Alzhei- mer's & Dementia: Diagnosis, Assessment & Disease Monitor- ing, vol. 10, no. 1, pp. 737–749, 2018. [43] J. Gu, Z. Wang, J. Kuen et al., “Recent advances in convolu- tional neural networks,” Pattern Recognition, vol. 77, pp. 354–377, 2018. [44] W. Haohan and R. Bhiksha, “On the origin of deep learning,” 2017, https://arxiv.org/abs/1702.07800. [45] B. Herbert, T. Tinne, and V. G. Luc, “Surf: speeded up robust features,” in European Conference on Computer Vision, pp. 404–417, Springer, Berlin, Heidelberg, 2006. [46] A. M. Mahdi, “Visual saliency analysis, prediction, and visual- ization: a deep learning perspective,” Ph. D. dissertation, Dept. ECE, Southern Illinois Univ., Carbondale, USA, 2019. [47] H. Samer, R. Kumar, and C. Rowen, Using convolutional neu- ral networks for image recognition, Cadence Design Systems Inc., San Jose, CA, USA, 2015. [48] S. Nitish, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov, “Dropout: a simple way to prevent neural networks from overfitting,” The Journal of Machine Learning Research, vol. 15, no. 1, pp. 1929–1958, 2014. [49] B. Pierre and J. S. Peter, “Understanding dropout,” Advances in neural information processing systems, pp. 2814–2822, 2013. [50] H. I. Alsaadi, A. M. Rafah, O. Bayat, and O. N. Ucani, “Com- putational intelligence algorithms to handle dimensionality reduction for enhancing intrusion detection system,” Journal of Information Science and Engineering, vol. 36, no. 2, pp. 293–308, 2020. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Applied Bionics and Biomechanics Hindawi Publishing Corporation

Diagnosis of Alzheimer Disease Using 2D MRI Slices by Convolutional Neural Network

Download PDF
9 pages

Loading next page...
 
/lp/hindawi-publishing-corporation/diagnosis-of-alzheimer-disease-using-2d-mri-slices-by-convolutional-tmFDMrFlmK

References (58)

Publisher
Hindawi Publishing Corporation
Copyright
Copyright © 2021 Fanar E. K. Al-Khuzaie et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
ISSN
1176-2322
eISSN
1754-2103
DOI
10.1155/2021/6690539
Publisher site
See Article on Publisher Site

Abstract

Hindawi Applied Bionics and Biomechanics Volume 2021, Article ID 6690539, 9 pages https://doi.org/10.1155/2021/6690539 Research Article Diagnosis of Alzheimer Disease Using 2D MRI Slices by Convolutional Neural Network 1 1 2 Fanar E. K. Al-Khuzaie , Oguz Bayat , and Adil D. Duru Graduate School of Science and Engineering, Altinbas University, Istanbul, Turkey Department of Physical Education and Sports Teaching, University of Marmara, Istanbul, Turkey Correspondence should be addressed to Fanar E. K. Al-Khuzaie; fanar@itnet.uobabylon.edu.iq Received 26 November 2020; Revised 14 January 2021; Accepted 22 January 2021; Published 2 February 2021 Academic Editor: Mohammed Yahya Alzahrani Copyright © 2021 Fanar E. K. Al-Khuzaie et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. There are many kinds of brain abnormalities that cause changes in different parts of the brain. Alzheimer’s disease is a chronic condition that degenerates the cells of the brain leading to memory asthenia. Cognitive mental troubles such as forgetfulness and confusion are one of the most important features of Alzheimer’s patients. In the literature, several image processing techniques, as well as machine learning strategies, were introduced for the diagnosis of the disease. This study is aimed at recognizing the presence of Alzheimer’s disease based on the magnetic resonance imaging of the brain. We adopted a deep learning methodology for the discrimination between Alzheimer’s patients and healthy patients from 2D anatomical slices collected using magnetic resonance imaging. Most of the previous researches were based on the implementation of a 3D convolutional neural network, whereas we incorporated the usage of 2D slices as input to the convolutional neural network. The data set of this research was obtained from the OASIS website. We trained the convolutional neural network structure using the 2D slices to exhibit the deep network weightings that we named as the Alzheimer Network (AlzNet). The accuracy of our enhanced network was 99.30%. This work investigated the effects of many parameters on AlzNet, such as the number of layers, number of filters, and dropout rate. The results were interesting after using many performance metrics for evaluating the proposed AlzNet. 1. Introduction aspects of our life, for example, genetic algorithm [3–5] and neural networks [6]. Machine learning (ML) is a field of artifi- cial intelligence that usually employs factual procedures to The normal brain of humans consists of mainly three regions, allow PCs to “learn” by utilizing information from saved data namely, white matter (WM), gray matter (GM), and cerebro- sets. At a very basic level, deep learning (DL) is a machine spinal fluid (CSF) [1]. The white matter is called as such learning subset [7]. Deep learning can be defined as a neural because of its white appearance. It contributes about sixty per- cent to the total brain volume. The gray matter is responsible network which uses a huge number of parameters and layers. There are many fundamental network architectures [8] like (i) of the whole processing of the neural signals. It consists of den- convolutional neural networks (CNNs) which are basically a drites and the neuron nuclei. It contributes almost about forty standard neural network that has been extended across space percent of the total brain volume. Cerebrospinal fluid is a col- using shared weights [9]. A convolutional neural network orless fluid that provides protection from mechanical shocks and also emits some important hormones to make the com- (CNN) is designed to recognize images by having convolu- tions inside, which see the edges of a recognized object on munication possible among the white matter, gray matter, the image [10]. (ii) Recurrent neural networks (RNNs) are a and spinal cord of the central nervous system [2]. It is known denomination of artificial neural networks where connections that the family of artificial intelligence (AI) includes many among nodes lay out a directed graph along the temporal algorithms and methods which could be used in different 2 Applied Bionics and Biomechanics Since MRI can provide a lot of invaluable information about sequence. Unlike feedforward neural networks, RNNs have the ability to use their internal state for processing the structures of tissue, such as localization, size, and shape, it is sequences of inputs. RNN is designed to recognize sequences, attracting more of attention for computer-aided diagnosis for example, a speech signal or a text [9]. (iii) Recursive neural and clinical routine [24, 25]. MRI can be divided into func- networks are more like a hierarchical network where there is tional and structural imaging. Functional imaging contains really no time aspect to the input sequence, but the input has tasking state functional MRI (ts-fMRI), resting state functional to be processed hierarchically in a tree fashion [8, 10]. MRI (rs-fMRI), etc., structural imaging contains T1-weighted Generally, different external stimuli match to different brain MRI (T1w), diffusion tensor imaging (DTI), and T2-weighted activities, and the different brain activities display different MRI (T2w) [26]. Medical data systems are diagnostic, and functional brain images [11]. For that, image classification plays analytical systems are applied to help medical centers and phy- asignificant role in identifying different activities of the brain. sicians in disease treatment, and they are critical to improve Recently, many methods of deep learning were proposed to per- treatment and diagnosis. Computer scientists have been inter- form classification of image for different brain activities [12, 13]. ested in this domain given the vital role of medical data in the To identify different activities of the brain including emotions, lives of humans. Physicians may refer to the classification of motor, social, relational and language activities, and working medical data, including medical analyses and symptoms of memory, Koyamada et al. [12] applied a feedforward deep neu- critical diseases, for making the decisions. A data set of disease ral network from images of functional magnetic resonance contains symptoms of patients as attributes besides the num- imaging (fMRI) to implement this mission. The feedforward ber of instances of these symptoms. Health care may use the deep neural network involved a softmax layer and multiple hid- considerable medical data accessible. In the analyses of medi- den layers. Similarly, these hidden layers were used to get high- cal centers, data mining could be used to provide sufficient level latent features, while the softmax layer was applied to cal- sources on illnesses for their prevention and timely detection culate the probability of every subjects in a class. In addition, and to avoid the expensive costs incurred by medical tests dropout, minibatch stochastic gradient descent, [14], and prin- [27]. Representation of features plays a significant role in med- cipal sensitivity analysis [15] were combined into the feedfor- ical image analysis and processing. Deep learning has two ward deep neural network to improve the performance of the obvious advantages in the representation of features: final classification. Recently, to classify different sensorimotor tasks including auditory attention, visual stimulus, right-hand (i) It can be applied to automatically discover features clenching, and left-hand clenching, Jang et al. [13] used fully from a given data set for every specific application. connected feedforward deep neural networks and multiple hid- Usually, methods of traditional feature extraction den layers. In addition to the above classifications, the methods are based on some prior knowledge for extracting of deep learning classification of magnetic resonance imaging features in a certain application. So, these approaches (MRI) images have been used also by other fields of classifica- are semiautomatic learning methods tion, like stroke diagnosis [16], age prediction [17], classification of attention deficit hyperactivity disorder (ADHD) [18], dis- (ii) It can discover new features that are appropriate to spe- crimination of cerebellar ataxia types [19], and emotional cific applications, which have never been discovered by response prediction [20]. Due to the science engineering field, researchers previously. Traditional methods of feature it was doable to create systems of computer-aided diagnosis extraction are often restricted by some a priori knowl- (CAD) that play a critical role in assisting the researchers and edge, which can only extract some features which are physicians when they interpret the medical imaging. Recently, associated with a certain application [28, 29] the use of the machine learning approach, especially DL tech- niques in systems of CAD to diagnose and classify the healthy Medical imaging is the mechanism and process of establishing visual representations of the interior of the body control normal (CN) people, Alzheimer’sdisease (AD), and mild cognitive impairment (MCI) patients, has exponentially for medical intervention and clinical analysis [30]. Machine increased [21, 22]. The Alzheimer’s disease automatic diagno- learning tools and medical image processing can help neurologists in estimating whether a subject is developing sis, especially in its early stage, plays a significant role in human health. Since Alzheimer’s disease is a neurodegenerative illness, Alzheimer’s disease [31]. Alzheimer’s disease is a chronic neurodegenerative disease causing tissue loss throughout it has a long period of incubation. Thus, it is necessary to ana- the brain, and the death of nerve cells usually starts slowly lyze the AD symptoms at different stages. Currently, a lot of researchers have proposed using the classification of images and worsens over time [32]. Alzheimer’s disease is expected to affect more and more people by the year 2050. The cost to perform diagnosis of AD. Moreover, many DL methods have been proposed to implement severity classification of dif- of caring for patients of AD is also expected to rise [33]. Presently, AD is the sixth reason that leads to death in the ferent Alzheimer’s disease patients by using MRI images [22, United States [34]. For this reason, individual computer- 23]. As known in image processing and analyzing, the better the image quality, then the better the results gained. However, aided systems are necessary for accurate and early diagnosis of this disease [33]. There are many approaches for accurate the image quality depends on acquisition of the image, so, when the image acquisition is better, then the image quality is and automatic classification of brain MRI, and one of them is our work. The next part of this article is the related works, higher. Magnetic resonance imaging (MRI) not only keeps then we will talk about our methodology, the results, the the features of noninvasive and good soft tissue contrast, but in addition does not expose humans to high ionizing radiation. discussion, and, at the end, our references. Applied Bionics and Biomechanics 3 matter from the human brain and make the classification by 2. Related Works using the CNN. To enhance the voxels, a Gaussian filter is Researchers have been applying machine learning techniques applied, and to remove the irrelevant tissues, the skull stripping algorithm is used. After that, by applying a hybrid, enhanced, to build classifiers by using clinical measures and imaging data for AD diagnosis. These studies have identified the important and independent component analysis, those voxels are seg- structural differences in the regions such as the entorhinal mented. The input to the CNN was segmented gray matter. cortex and hippocampus entorhinal cortex between the brain Clinical valuation was performed using the provided approach with AD and healthy brain. Different imaging methods like and 90.47 accuracy was achieved. Hamed and Kaabouch [41] proposed a method that yielded good classification accuracy. the functional and structural magnetic resonance imaging (fMRI and sMRI, respectively), single photon emission com- The convolutional neural network with modified architecture puted tomography (SPECT), position emission tomography was used to get the high quality features from the brain MRI (PET), and diffusion tensor imaging (DTI) scans which can to classify people into healthy, early mild cognitive impairment perceive the changes causing AD due to the degeneration of (EMCI), or late mild cognitive impairment (LMCI) groups. The results showed the classification between control normal (CN) cells of the brain. In recent years, deep learning models, especially convolutional neural networks, have demonstrated and LMCI groups in the sagittal view with 94.54 accuracy. outstanding performance for medical image analysis. Payan and Giovanni [35] produced and tested a pattern classification 3. Materials and Methods system which combines convolutional neural network and Inside a CNN, a filter series, with an equivalent size to a small sparse autoencoders. Ehsan et al. [36] adapted a 3D-CNN image patch, automatically searches the entire image to find model for diagnosis of AD. The 3D-CNN is built upon the images of similar spatial features. These filters can be learned 3D convolutional autoencoder, which is pretrained to catch and updated independently; thus, a collection of them can anatomical shape variations in scans of structural brain MRI. detect crucial information of a specific task and data set [42]. Sergey et al. [37] proposed two different kinds of 3D convolu- There are standard steps of CNN. The first step is named tional network architectures to classify the brain MRI which are “convolution”; this step is responsible for finding the features the amendments of residual and plain convolutional neural and applying filters. It is a filter kernel that picks up its weights networks. Applied convolutional neural networks can tackle by convolving the input data tensor with such kernel. There the two problems stated before. These networks can propagate are several variables that effect the convolutional operation local features into the metarepresentation of an object for output such as strides and number of filters. The distance in classification or image recognition. In deep learning for image pixels between two pixels is the stride, while the number of fil- classification, modern advancements like residual network ters states the output feature map number [43]. The operation architectures and batch normalization mechanism alleviate of convolution is just a mathematical operation, which should the issues of having small data sets of training, while providing be treated equally with other operations such as multiplication a frame for automatic feature generation. As a result, these or addition and should not be discussed particularly in the models can be used to 3D MRI images in the absence of literature of machine learning. But, it has still been discussed intermediate handcrafted feature extraction. Karim et al. [38] here for completeness. Convolution is a mathematical opera- adapted three tasks of binary classification which are consid- tion on two functions (e.g., f and g)and producesathird ered for separating the normal control (NC) subject from mild function h; this is an integral that expresses the amount of cognitive impairment (MCI) patients and Alzheimer’sdisease overlap of one function (f ) as it is shifted over the other func- (AD). Two fusion methods on a fully connected (FC) layer tion (g) [44]. Formally, it is described as and on the single-projection CNN output offer better achieve- ment by about 91% accuracy. The outcomes are competitive with the SOA which utilizes a heavier algorithmic chain. Fan ht = ðÞ fTðÞgtðÞ − T dT, ð1Þ and Manhua [39] proposed a classification technique built on −∞ multiple clusters of dense convolutional neural networks (DenseNets) to pick up the various local features for images And denoted as h = f ∗ g. of the MR brain, which are collected for classification of AD. A typically convolutional neural network works with The total brain image is partitioned into different local parts two-dimensional convolution operation that could be sum- and from each region, a number of 3D patches are extracted. marized in Figure 1. As displayed in Figure 1, the input By using theK-means clustering method for grouping the matrix is Figure 1(a), and Figure 1(b) is usually called a kernel patches from each region into different clusters, the DenseNet matrix. So convolution is applied to these matrices, then the had been constructed to pick up the patch features for each result is displayed as in Figure 1(c). The process of convolu- cluster, and the features learned from the characteristic clusters tion can be considered as an element-wise product followed of each part are grouped for classification. The classification by a sum, like what is shown in the example of Figure 1. outputs from different local parts are combined to foster the When the left upper matrix which is 3×3 convoluted with final image classification. This method can progressively learn the kernel, then the result is 29. After that, the target 3×3 the features of MRI from the local patches to the global image matrix slides one column to the right, then is convoluted with level for the task of classification. For preprocessing images of the kernel and gets the result 22. The sliding and recording of MRI, there are no segmentation and rigid registration required. the results have been continued as a matrix. Every target Shaik and Ram [40] provided an approach to extract the gray matrix is 3×3, because the kernel is 3×3; thus, the whole 5 4 Applied Bionics and Biomechanics 10 37 1 5 40 4 2 16 55 20 001 29 22 42 28 17 2 0 47 25 55 39 9 0 19 25 13 10 30 88 830 41 631 (a) (b) (c) Figure 1: A simple illustration of a two-dimension convolution operation: (a) input matrix, (b) kernel matrix, and (c) output matrix after convolution. ×5 matrix is shrunk into a 3×3 matrix when every 3×3 many works have attempted to understand its technique in different perspectives, including [49]. It has also been used to matrix is convoluted to one digit. (Because of 5 − ð3 − 1Þ =3 , the first 3 means the kernel matrix size.) One should realize train other models, such as SVM [50]. that the convolution process is a locally shifted invariant, The CNN architecture which was used in this study is which means that for many different combinations of how composed of five convolutional layers which take an input the nine numbers in the upper matrix 3×3 are placed, the image (the brain’s MRI slice) with a size of 200 ∗ 200. Figure 2 shows some slices of the brain’s MRI; those were convoluted result will be 29. This invariant property plays a crucial role in vision problem because the result of recogni- we used in our research. All five convolutional layers were tion should not be changed due to shift or rotation of features followed by a max-pooling layer. The 64 filters with a kernel in an ideal case. This crucial property is applied to be solved size of 9 ∗ 9 were considered for the first convolutional layer, elegantly by [45], but CNN brought the performance up to a and the max-pooling layer kernel size was set on 2 ∗ 2. The new level. 64 filters with a kernel size of 7 ∗ 7 were considered for the With each convolution layer, there is an activation func- second convolutional layer, and the max-pooling layer kernel tion; the activation is an operation which converts the input size also was set on 2 ∗ 2. The 64 filters with a kernel size of from a linear data tensor to a nonlinear data tensor. In deep 5 ∗ 5 were considered for the third convolutional layer, and learning, many activation functions are popular such as recti- the max-pooling layer kernel size was set on 2 ∗ 2.The 32 fied linear units (ReLU), sigmoid, and tanh [46]. Recently, the filters with a kernel size of 5 ∗ 5 were considered for the fourth rectified linear unit (ReLU) has been used more than the other convolutional layer, and the max-pooling layer kernel size was nonlinear functions, because it does not activate all the set on 2 ∗ 2.The 32 filters with a kernel size of 3 ∗ 3 were con- neurons at the same time [24]. The second step is named sidered for the fifth convolutional layer, and the max-pooling “Max Pooling”; this step is responsible for downsizing the layer kernel size was set on 2 ∗ 2. It is worthwhile to mention image and keeping the important features. Pooling is the that the ReLU (rectified linear unit) function was used as the operation of downsampling which can be performed globally activation functions in all convolutional layers. The ReLU or locally. The function of global pooling returns for every function is used commonly in models of DL; basically, if the 2D feature map a scalar value. The function of local pooling function receives a negative value as input, it returns 0, and downsamples local image parts by a factor [43]. The third step if the function receives a positive value, then the same positive named “flattening” converts to one dimension array (vector). value will return. The fourth step is named “full connection”; this step is respon- The function of ReLU is understood as f ðaÞ = max ð0, aÞ. sible of building all needed connections. The fully connected Figure 3 demonstrates the block diagram of the proposed layer (FC) is typically followed by an activation layer. FC is system (AlzNet). After the convolution layers and the flatten- the layer where the receptive domain is a whole channel of ing layer, there is a dense unit 121, and here, we used a ReLU the former layer [43, 46]. The last step is named “classifier”; as an activation function, then we used a dropout (0.2) to it represents the classification stage to decide if the image is prevent overfitting, then there is a dense unit and sigmoid normal or abnormal [47]. The use of the dropout technique as an activation function; at the last stage, there is a binary is so common in convolutional neural networks. Dropout classifier for displaying the results. was introduced in [14, 48]. This mechanism soon got influen- Table 1 demonstrates the number of MRI slices. There tial, not only because it has good performance but also because are samples for men and women such as a left-handed man of its simplicity of implementation. The idea is very easy: while (L.-handed male), left-handed woman (L.-handed female), training, randomly drop out some of the units. More formally: right-handed man (R.-handed male), and a right-handed for each training case, every hidden unit is randomly omitted woman (R.-handed female); all brain MRIs were in the axial with a probability of p from the network. As suggested in [14], view manner. Keras provides a perfect tool to show a model’s dropout can be seen as an efficient method to perform model summary; Table 2 demonstrates that summary. This displays averaging across a great number of different neural networks, the number of trainable parameters and the output shape for where overfitting can be avoided with less cost of computation each layer. Before starting to fit the model, this is a sanity because of the actual performance which it introduces. Drop- check. So the total params = 414,419, the trainable params out became very popular upon when it was first introduced; = 414,419, and the nontrainable params = 0. Applied Bionics and Biomechanics 5 Left-handed woman Left-handed man Right-handed woman Right-handed man Figure 2: Some brain MRI slices of Alzheimer’s patients. Test Input dataset Output MRI slices Training Abnormal dataset dataset Prediction 2D-CNN Validation Normal dataset ⁎ Dense+ReLU 2 2 Flattening ⁎ ⁎ ⁎ ⁎ 2 2 2 2 2 2 2 2 pooling pooling pooling pooling pooling With Dense+sigmoid dropout ⁎ ⁎ ⁎ ⁎ 7 7 Convolution 5 5 Convolution 5 5 Convolution 3 3 Convolution ⁎ Classifier 9 9 Convolution 64 filters+ReLU 64 filters+ReLU 32 filters+ReLU 32 filters+ReLU 64 filters+ReLU 2D-CNN Figure 3: Block diagram of proposed system (AlzNet). There are five convolutional layers; after each convolutional layer there is a max- pooling layer; the activation function in every convolutional layer was ReLU. After the convolution layers and flattening layer, there is a dense unit 121 with a ReLU as an activation function, with a dropout to prevent overfitting, then there is a dense unit and sigmoid as an activation function; at the last stage, there is a binary classifier. Table 1: Number of MRI slices. 4. Results Number of Number of MRI Subjects Python language 3.6 and Keras have been used for program- subjects slices ming this work. Keras is a high-level library; it is an open R.-handed male (AD patient) 60 3050 source machine learning library that is written in Python. L.-handed male (AD patient) 40 2600 Keras is used for numerical computation purposes; it is used L.-handed female (AD patient) 30 2150 to perform the computations more easily and efficiently in R.-handed female (AD patient) 40 2600 practice. The training data set was 75% and the validation Female (NC) 25 2000 data set was 25%. There were many practical experiments Male (NC) 45 2800 that had been done in this research for trying to find the best parameters of this convolutional neural network. So we try to find the best number of dense units for hidden layers depend- ing on the result of accuracy, whereas the other researches used different numbers at each one. Another parameter we tested many times is the rate of dropout to find the fit rate 6 Applied Bionics and Biomechanics Table 2: Summary of the proposed model. Table 3: Accuracy of AlzNet depending on dropout rate and dense unit. Layer (type) Output shape Param # Dropout rate Conv2d_1 (Conv2D) (None, 192, 192, 64) 15,616 0.5 0.4 0.3 0.2 0.1 Max_pooling2d_1 (MaxPooling2) (None, 96, 96, 64) 0 130 94.90 94.21 94.78 94.02 94.12 Conv2d_2 (Conv2D) (None, 90, 90, 64) 200,768 129 95.18 95.72 95.61 95.31 95.48 Max_pooling2d_2 (MaxPooling2) (None, 45, 45, 64) 0 128 94.85 94.56 94.71 94.65 94.58 Conv2d_3 (Conv2D) (None, 41, 41, 64) 102,464 127 95.09 95.12 95.01 95.23 95.31 Max_pooling2d_3 (MaxPooling2) (None, 20, 20, 64) 0 126 96.61 96.72 96.40 95.97 96.01 Conv2d_4 (Conv2D) (None, 16, 16, 32) 51,232 Dense unit 125 96.15 96.05 96.11 96.44 96.21 Max_pooling2d_4 (MaxPooling2) (None, 8, 8, 32) 0 124 96.09 96.12 96.02 96.30 96.05 Conv2d_5 (Conv2D) (None, 6, 6, 32) 9248 123 95.91 95.72 95.82 95.01 95.11 Max_pooling2d_5 (MaxPooling2) (None, 3, 3, 32) 0 122 96.85 96.56 96.90 95.77 95.83 Flatten_1 (flatten) (None, 288) 0 121 97.32 97.16 97.30 97.88 97.06 Dense_1 (dense) (None, 121) 34,969 120 95.43 95.66 94.94 94.88 94.90 Dropout_1 (dropout) (None, 121) 0 Dense_2 (dense) (None, 1) 122 True Positive + True Negative Accuracy = : for our convolutional neural network by observing the results True positive + True negative + False positive + False negative of accuracy. That is figured out in Table 3. ð5Þ So it is obvious from Table 3 that the highest accuracy value was when we utilized 0.2 for the dropout rate and 121 for the dense unit. In fact, the range of the dropout rate which In this work, the measure metrics have been applied on we tested was (from 0.1 to 0.5) increasing by 0.1, when the the training data set and test data set (see Table 4). number of dense units was 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, and 130. Figure 4 shows the accuracy 5. Discussion value depending on the relationship between the dense unit number and the dropout rate. For our current work, we develop an efficient deep convolu- There are several metrics for measuring the performance tional neural network based on a classifier and demonstrate of binary classification [51], such as recall, precision, specific- very good performance by using the OASIS data set [54]. ity, and Accuracy. Precision is very helpful because we want The OASIS-3 data set has been saved in the XNAT central to be confident of our forecast, since it tells us how many of repository [55]. In our work, a total of 15,200 MRI axial slices the values expected as positive are actually positive [50], as were used. The data was used to include the MRI scans of follows: about 170 AD patients and 70 NC. They are all from different subjects that make the test of recognition performance more True positive reliable. The age of each patient is between the range of 65-90 Precision = : ð2Þ years old, both male and female in this work. At the first stage True positive + False positive of data preprocessing, we obtained 2D slices from each MRI image. Then, the every last 20 dark slices with each time Recall (sensitivity) is another very valuable measure that course was discarded because they included no functional helps one, for instance, to know the proportion of the num- ber of values accurately labeled as positive on the overall information. Preprocessed images are augmented by rotating the slices to see whether or not the model can recognize the values which are actually positive, as follows: images; this increased the samples size and made a good True positive training of the CNN model. The scans are T1-weighted Recall = : ð3Þ whereas those of [40] were T2-weighted. Whereas [35, 41] True positive + False negative used three convolution layers, our proposed system (CNN model) has 5 convolution layers; each convolution layer has Using the F1 score is a safe way to get a complete impres- ReLU as an activation function. Each convolution layer is sion of recall and precision. The F1 score provides us the har- followed by max pooling layers. Our proposed approach per- monic mean of recall and precision [52, 53], as follows: forms binary classification to fit the model in a batch size of Precision ∗ Recall 64 in 150 epochs. Table 2 summarizes the total architecture F1 score = 2 ∗ : ð4Þ of the proposed system. When [40, 41] used Adam optimiza- Precision + Recall tion, we trained the model by Adadelta optimization with a Accuracy is the proportion of accurate predictions (both dropout rate of 0.2 for the dropout layer which had been true negative and true positive) among the entire number of utilized to prevent the overfitting like in [40], but [41] used cases examined [52], as follows: a 0.5 dropout rate. The number of dense units was 121 when Applied Bionics and Biomechanics 7 Accuracy of AlzNet depending on dropout rate and dense unit Table 5: Literature reviews. 98.5 Methods Year Accuracy 97.5 Sergey et al. [37] 2017 80.00 96.5 96 Fan and Manhua [39] 2018 89.5 95.5 Shaik and Ram [40] 2019 90.47 94.5 Ehsan et al. [36] 2016 97.6 Karim et al. [38] 2017 91.41 93.5 Payan and Giovanni [35] 2015 95.39 130 129 128 127 126 125 124 123 122 121 120 Hamed and Kaabouch [41] 2019 94.54 0.5 0.2 0.4 0.1 Proposed method 2020 99.30 0.3 Figure 4: Accuracy of AlzNet depending on dropout rate and dense 6. Conclusions unit. In order to diagnose Alzheimer’s disease, deep neural networks, especially CNNs, can provide meaningful informa- Table 4: Measure metrics of training data set and test data set. tion. A CNN-based method for extracting discriminatory Dataset Precision Recall F1 score Accuracy features from structural MRI was proposed in this paper, with the goal of classifying Alzheimer’s disease and healthy Training 97.06 97.99 97.52 97.88 subjects using 2D MRI slices. For potential AD individuals, Test 98.92 99.53 99.22 99.30 the suggested approach can lead to many advantages and can also lead to an early diagnosis of AD. The experimental results of the OASIS database for 240 subjects demonstrated Measure metrics of training data set that our proposed method of extraction and classification of and test data set features provided high accuracy for AD and CN. The best 99.53 99.3 99.22 results have been obtained for the classification between the 98.92 CN and the AD axial view of the MRI. The proposed method 97.99 97.88 98 yielded a classification accuracy of 99.30 percent. The above 97.52 97.06 results indicate higher reliability, recall, precision, and F1 score of our proposed method for the diagnosis of AD and the classification between CN and AD. Training data set Test data set Precision F1 score Data Availability Recall Accuracy Our data set was from the OASIS database; the website is Figure 5: Measure metrics of training data set and test data set. https://www.oasis-brains.org. The OASIS-3 data set has been saved in the XNAT central repository; the website is https:// central.xnat.org. the number of dense units of [37] was 128; actually, we made many experiments to decide which is the best number of dense units we should take, and Table 3 shows that. In this Conflicts of Interest work, we tried to put different values of the neural network The authors declare that there is no conflict of interest parameters by trial and error, by relying on the accuracy regarding the publication of this paper. value, and comparison with previous researches. During binary classification, we trained the classifier for AD and CN images, and the model resulted in 97.88% training Acknowledgments accuracy and 99.30 test accuracy. It is required to mention that our proposed framework had been trained, and the Data were provided in MRI by OASIS-3. The principal prediction was made with utmost accuracy. Figure 5 demon- investigators are T. Benzinger, D. Marcus, and J. Morris. The strates that. The accuracy of the proposed system has been study was funded by NIH (P50AG00561, P30NS09857781, compared with different models discussed in literature P01AG026276, P01AG003991, R01AG043434, UL1TR000448, reviews as shown in Table 5. and R01EB009352). It is observed that the proposed model achieves remark- able performance. The last thing we have to say is that neural References networks have plenty of parameters, and any change in one of them will make the value of results different, and also, [1] M. Nazir, F. Wahid, and S. Ali Khan, “A simple and intelligent there is a big important reason for making a variation of approach for brain MRI classification,” Journal of Intelligent & results—it is the data set and its type. Fuzzy Systems, vol. 28, no. 3, pp. 1127–1135, 2015. 8 Applied Bionics and Biomechanics [17] J. H. Cole, R. P. K. Poudel, D. Tsagkrasoulis et al., “Predicting [2] M. H. Fazel Zarandi, M. Zarinbal, and M. Izadi, “Systematic image processing for diagnosing brain tumors: a type-II fuzzy brain age with deep learning from raw imaging data results in a expert system approach,” Applied Soft Computing, vol. 11, reliable and heritable biomarker,” NeuroImage, vol. 163, no. 1, pp. 285–294, 2011. pp. 115–124, 2017. [18] D. P. Kuang and L. H. He, “Classification on ADHD with deep [3] H. Nabeel Kaghed, S. Eman al–Shamery, and F. E. K. al-Khu- zaie, “Multiple sequence alignment based on developed genetic learning,” in Proc., Int. Conf. Cloud Computing and Big Data (CCBD), pp. 27–32, Wuhan, China, 2014. algorithm,” Indian Journal of Science and Technology, vol. 9, no. 2, pp. 1–7, 2016. [19] Z. Yang, S. H. Zhong, A. Carass, S. H. Ying, and J. L. Prince, “Deep learning for cerebellar ataxia classification and func- [4] N. H. Kaghed, E. S. Al, and F. E. K. Al-Khuzaie, “Comparative study of genetic algorithm and dynamic programming of DNA tional score regression,” in Proc. 5th Int. Workshop on Machine Learning in Medical Imaging, pp. 68–76, Boston, multiple sequence alignment,” Journal of University of Baby- lon, vol. 25, no. 2, pp. 403–414, 2017. MA, USA, 2014. [20] H. C. Kim and J. H. Lee, “Deep learning to predict the emo- [5] F. Al-Khuzaie, “Multiple sequence alignment based on devel- tional response using functional MRI data,” NeuroImage, oped genetic algorithm,” (Master thesis), University of Baby- vol. 186, pp. 607–627, 2019. lon, Iraq, 2016, https://www.researchgate.net/publication/ 347511906_Multiple_Sequence_Alignment_Based_on_ [21] H. T. Gorji and J. Haddadnia, “A novel method for early diag- Developed_Genetic_Algorithm?channel=doi&linkId= nosis of Alzheimer's disease based on pseudo Zernike moment 5fdf736a45851553a0db2ee7&showFulltext=true. from structural MRI,” Neuroscience, vol. 305, pp. 361–371, [6] V. G. Gudise and G. K. Venayagamoorthy, “Comparison of par- ticle swarm optimization and backpropagation as training algo- [22] S. Liu, S. Liu, W. Cai, S. Pujol, R. Kikinis, and D. Feng, “Early rithms for neural networks,” in Proceedings of the 2003 IEEE diagnosis of Alzheimer’s disease with deep learning,” in Pro. Swarm Intelligence Symposium. SIS'03 (Cat. No.03EX706), IEEE 11th International Symposium on Biomedical Imaging pp. 110–117, Indianapolis, IN, USA, 2003. (ISBI), pp. 1015–1018, Beijing, China, 2014. [7] T. H. H. Aldhyani, M. Alrasheedi, A. A. Alqarni, M. Y. Alzah- [23] F. Li, L. Tran, K. H. Thung, S. W. Ji, D. G. Shen, and J. Li, “A robust deep model for improved classification of AD/MCI rani, and A. M. Bamhdi, “Intelligent hybrid model to enhance time series models for predicting network traffic,” IEEE Access, patients,” IEEE J Biomed Health Inform, vol. 19, no. 5, pp. 1610–1616, 2015. vol. 8, pp. 130431–130451, 2020. [8] G. Litjens, T. Kooi, B. E. Bejnordi et al., “A survey on deep [24] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet clas- sification with deep convolutional neural networks,” in Proc. of learning in medical image analysis,” Medical Image Analysis, vol. 42, pp. 60–88, 2017. the Advances in Neural Information Processing Systems, pp. 1097–1105, Lake Tahoe, CA, USA, 2012. [9] M. Z. Alom, T. M. Taha, C. Yakopcic et al., “A state-of-the-art [25] J. Liu, M. Li, Y. Pan et al., “Complex brain network analysis survey on deep learning theory and architectures,” Electronics, and its applications to brain disorders: a survey,” Complexity, vol. 8, no. 3, p. 292, 2019. vol. 2017, Article ID 8362741, 27 pages, 2017. [10] D. Ravi, C. Wong, F. Deligianni et al., “Deep learning for [26] L. Jin, P. Yi, L. Min et al., “Applications of deep learning to health informatics,” IEEE Journal of Biomedical and Health MRI images: a survey,” Big Data Mining and Analytics, Informatics, vol. 21, no. 1, pp. 4–21, 2017. vol. 1, no. 1, pp. 1–18, 2018. [11] T. Horikawa, M. Tamaki, Y. Miyawaki, and Y. Kamitani, [27] O. S. Soliman and E. A. El Hamd, “Classification of breast can- “Neural decoding of visual imagery during sleep,” Science, cer using differential evolution and least squares support vec- vol. 340, no. 6132, pp. 639–642, 2013. tor machine,” Int J Emerg Trends Technol Comput Sci, vol. 3, [12] S.Koyamada,Y.Shikauchi,K. Nakae,M.Koyama,and no. 2, pp. 155–161, 2014. S. Ishii, “Deep learning of fMRI big data: a novel approach [28] J. Liu, M. Li, J. Wang, F. Wu, T. Liu, and Y. Pan, “A survey of to subject-transfer decoding,” 2015, https://arxiv.org/abs/ MRI-based brain tumor segmentation methods,” Tsinghua 1502.00093. Science and Technology, vol. 19, no. 6, pp. 578–595, 2014. [13] H. Jang, S. M. Plis, V. D. Calhoun, and J. H. Lee, “Task-specific [29] J. Schmidhuber, “Deep learning in neural networks: an over- feature extraction and classification of fMRI volumes using a view,” Neural Networks, vol. 61, pp. 85–117, 2015. deep neural network initialized with a deep belief network: [30] M. Natarajan and S. Sathiamoorthy, “A novel pre-processing evaluation using sensorimotor tasks,” Neuro Image, vol. 145, approach for the denoising of Alzheimer disease image data- Part B, pp. 314–328, 2017. set,” Asian Journal of Computer Science and Technology, [14] G. E. Hinton, N. Srivastava, A. Krizhevsky, I. Sutskever, and vol. 7, no. 2, pp. 107–112, 2018. R. R. Salakhutdinov, “Improving neural networks by prevent- [31] E. Moradi, A. Pepe, C. Gaser, H. Huttunen, J. Tohka, and Alz- ing co-adaptation of feature detectors,” 2012, https://arxiv.org/ heimer's Disease Neuroimaging Initiative, “Machine learning abs/1207.0580. framework for early MRI-based Alzheimer's conversion predic- [15] S. Koyamada, M. Koyama, K. Nakae, and S. Ishii, “Principal tion in MCI subjects,” NeuroImage, vol. 104, pp. 398–412, 2015. sensitivity analysis,” in Proc. 19th Pacific-Asia Conf. Knowledge [32] C. Vaishnavi and P. Anna, “Alzheimer’s disease: a comprehen- Discovery and Data Mining, pp. 621–632, Ho Chi Minh City, sive review,” EC Neurology, vol. 8, no. 1, pp. 16–32, 2017. Vietnam, 2015. [33] R. Brookmeyer, E. Johnson, K. Ziegler-Graham, and H. M. [16] T. Schmah, G. E. Hinton, R. S. Zemel, S. L. Small, and Arrighi, “Forecasting the global burden of Alzheimer’s dis- S. Strother, “Generative versus discriminative training of rbms ease,” Alzheimer's & Dementia, vol. 3, no. 3, pp. 186–191, 2007. for classification of fMRI images,” in Proc. 21st Int. Conf. Neu- ral Information Processing Systems, pp. 1409–1416, Vancou- [34] S. L. Murphy, J. Xu, and K. D. Kochanek, “Deaths: final data ver, Canada, 2008. for 2010,” Natl Vital Stat Rep, vol. 61, no. 4, pp. 1–117, 2013. Applied Bionics and Biomechanics 9 [51] T. H. Aldhyani, A. S. Alshebami, and M. Y. Alzahrani, “Soft [35] A. Payan and M. Giovanni, “Predicting Alzheimer's disease: a neuroimaging study with 3D convolutional neural networks,” Computing Model to Predict Chronic Diseases,” Journal of in Proc. Int. Conf. Pattern Recog. Appl. Methods, 2015. Information Science and Engineering, vol. 36, no. 2, pp. 365– 376, 2020. [36] E. Hosseini-Asl, R. Keynton, and A. El-Baz, “Alzheimer's dis- ease diagnostics by adaptation of 3D convolutional network,” [52] T. H. H. Aldhyani, A. S. Alshebami, and M. Y. Alzahrani, “Soft in 2016 IEEE International Conference on Image Processing clustering for enhancing the diagnosis of chronic diseases over (ICIP), pp. 126–130, Phoenix, AZ, USA, September 2016. machine learning algorithms,” Journal of Healthcare Engineer- ing, vol. 2020, Article ID 4984967, 16 pages, 2020. [37] S. Korolev, A. Safiullin, M. Belyaev, and Y. Dodonova, “Resid- ual and plain convolutional neural networks for 3D brain MRI [53] A. M. A. Mohamed, O. N. Uçan, O. Bayat, and A. D. Duru, classification,” in 2017 IEEE 14th International Symposium on “Classification of resting-state status based on sample entropy Biomedical Imaging (ISBI 2017), pp. 835–838, Melbourne, and power spectrum of electroencephalography (EEG),” VIC, Australia, April 2017. Applied Bionics and Biomechanics, vol. 2020, Article ID 8853238, 10 pages, 2020. [38] A. Karim, B. P. Jenny, and A. Karim, “Classification of sMRI for Alzheimer's disease diagnosis with CNN: single Siamese [54] P. J. LaMontagne, T. L. S. Benzinger, J. C. Morris et al., networks with 2D+? Approach and fusion on ADNI,” in Proc. “OASIS-3: longitudinal neuroimaging, clinical, and cognitive of the 2017 ACM on International Conference on Multimedia dataset for normal aging and Alzheimer disease,” med Rxiv, Retrieval, pp. 494–498, 2017. [39] F. Li, M. Liu, and Alzheimer's Disease Neuroimaging Initiative, [55] R. Herrick, W. Horton, T. Olsen, M. McKay, K. A. Archie, and “Alzheimer's disease diagnosis based on multiple cluster dense D. S. Marcus, “XNAT Central: open sourcing imaging research convolutional networks,” Computerized Medical Imaging and data,” NeuroImage, vol. 124, Part B, pp. 1093–1096, 2016. Graphics, vol. 70, pp. 101–110, 2018. [40] S. Basheera and M. S. Sai Ram, “Convolution neural network– based Alzheimer's disease classification using hybrid enhanced independent component analysis based segmented gray matter of T2 weighted magnetic resonance imaging with clinical val- uation,” Alzheimer's & Dementia: Translational Research & Clinical Interventions, vol. 5, no. 1, pp. 974–986, 2019. [41] H. T. Gorji and N. Kaabouch, “A deep learning approach for diagnosis of mild cognitive impairment based on MRI images,” Brain Sciences, vol. 9, no. 9, pp. 217–231, 2019. [42] S. Qiu, G. H. Chang, M. Panagia, D. M. Gopal, R. Au, and V. B. Kolachalama, “Fusion of deep learning models of MRI scans, mini–mental state examination, and logical memory test enhances diagnosis of mild cognitive impairment,” Alzhei- mer's & Dementia: Diagnosis, Assessment & Disease Monitor- ing, vol. 10, no. 1, pp. 737–749, 2018. [43] J. Gu, Z. Wang, J. Kuen et al., “Recent advances in convolu- tional neural networks,” Pattern Recognition, vol. 77, pp. 354–377, 2018. [44] W. Haohan and R. Bhiksha, “On the origin of deep learning,” 2017, https://arxiv.org/abs/1702.07800. [45] B. Herbert, T. Tinne, and V. G. Luc, “Surf: speeded up robust features,” in European Conference on Computer Vision, pp. 404–417, Springer, Berlin, Heidelberg, 2006. [46] A. M. Mahdi, “Visual saliency analysis, prediction, and visual- ization: a deep learning perspective,” Ph. D. dissertation, Dept. ECE, Southern Illinois Univ., Carbondale, USA, 2019. [47] H. Samer, R. Kumar, and C. Rowen, Using convolutional neu- ral networks for image recognition, Cadence Design Systems Inc., San Jose, CA, USA, 2015. [48] S. Nitish, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov, “Dropout: a simple way to prevent neural networks from overfitting,” The Journal of Machine Learning Research, vol. 15, no. 1, pp. 1929–1958, 2014. [49] B. Pierre and J. S. Peter, “Understanding dropout,” Advances in neural information processing systems, pp. 2814–2822, 2013. [50] H. I. Alsaadi, A. M. Rafah, O. Bayat, and O. N. Ucani, “Com- putational intelligence algorithms to handle dimensionality reduction for enhancing intrusion detection system,” Journal of Information Science and Engineering, vol. 36, no. 2, pp. 293–308, 2020.

Journal

Applied Bionics and Biomechanics – Hindawi Publishing Corporation

Published: Feb 2, 2021

There are no references for this article.