Dementia Alzheimer's Disease Diagnosis Using Convolutional Neural Networks on Two-Dimensional MRI Slices

Neurodegenerative diseases, such as Alzheimer's disease (AD), have profited enormously from recent improvements in medical imaging and diagnostics, which are rapidly developing. The use of modern computational approaches, such as convolutional neural networks (CNNs), has greatly enhanced diagnosis and contributed considerably to this success. One of the most significant developments in medical research has been the use of convolutional neural networks, or CNNs, to read two-dimensional magnetic resonance imaging slices to identify Alzheimer's disease. This initiative uses artificial intelligence to improve the precision and efficacy of illness diagnosis and treatment. White matter, grey matter, and cerebrospinal fluid are components of the human brain, a complex organ made up of multiple independent elements. Despite accounting for the bulk of the brain's volume, these parts are critical for proper brain function. To completely comprehend brain disorders, one must first understand both white matter (so-called because of its colour) and grey matter (both of which are important in the processing of cerebral impulses). The health of the brain and spinal cord is dependent on the maintenance of key hormonal balances as well as the defense of the central nervous system. Each of these processes relies heavily on cerebrospinal fluid [1]. A full understanding of these elements is critical for investigating brain illnesses, particularly neurodegenerative diseases such as Alzheimer's. Alzheimer's disease is a neurological condition that gradually deteriorates memory and cognitive function. People affected by this ailment have a considerably worse quality of life. A precise and fast diagnosis is critical for effective treatment of this illness. This is because it allows for rapid reactions, which may slow the disease's development. It is worth noting that using traditional diagnostic approaches for early-stage diagnosis has various obstacles, emphasizing the necessity for more modern therapies.

Convolutional neural networks (CNNs) are a form of deep learning algorithm that has gained popularity because of their exceptional image processing and recognition capabilities. Because of their layered construction, CNNs can identify and extract complicated patterns from photos. As a result, medical imaging, especially MRI scans, becomes incredibly important in their investigation [2]. The use of convolutional neural networks (CNNs) to identify Alzheimer's disease based on magnetic resonance imaging (MRI) pictures has greatly increased medical diagnosis accuracy. CNNs have frequently proven their capacity to detect minor morphological and functional changes in the brain that are indicative of Alzheimer's disease in its early stages. This capacity goes beyond what is currently possible using current approaches. CNNs achieve this by analysing MRI scan slices. This strategy also addresses the growing demand for a range of effective diagnostic tools considering Alzheimer's disease's increasing global incidence. A decrease in the amount of labour required by medical workers, a faster diagnostic procedure, and consistent and trustworthy diagnostic findings are only a few of the advantages of adopting CNNs [3]. Building a system that can adapt to new diagnostic data and raise its accuracy and efficiency in accordance with the new data is a promising topic since convolutional neural networks (CNNs) are learning and adaptive systems.

1.png

Figure 1. Architecture of convolution neural network

This study on convolutional neural networks (CNNs) for Alzheimer's disease identification in MRI images exemplifies the multidisciplinary collaboration of neurology, radiology, and artificial intelligence. This study not only demonstrates medical technological breakthroughs but also the multidisciplinary teamwork that has happened. The convergence of technical innovation and human knowledge opens new options for innovative solutions to long-standing medical problems [4]. It is difficult to overestimate the significance of these advances, especially in view of the worldwide catastrophe posed by the increasing prevalence of Alzheimer's disease. There is a ray of hope for the millions of families impacted by this dreadful disease because of the improvement of CNN-based diagnostic tools. It is a critical area of study and development in the realm of medical technology since it is a significant step towards improving patient outcomes and boosting the efficacy of sickness treatment [5]. The convolutional process depicted in Figure 1 exemplifies how a typical CNN achieves its goal.

These approaches are capable of autonomous decision-making and learning. This category includes numerous approaches, two of which are genetic algorithms [5] and neural networks [6]. To provide only one example: It is a subfield of artificial intelligence (AI) that focuses on teaching computers to "learn" by extrapolating patterns and insights from large data sets. Machine learning, or ML for short, is a discipline that focuses on training computers to "learn" via the use of empirical approaches. In other words, machine learning trains computers to "learn" by searching for patterns and insights in data. When simplified to its most basic components, the concept of deep learning, abbreviated as deep learning for short, may be considered a subfield within the larger topic of machine learning [7].

2.png

Figure 2. Alzheimer diseases identification with CNN

Recurrent neural networks (RNNs), unlike feedforward neural networks, can process input sequences using their internal state. Alzheimer Diseases Identification with CNN is shown in Figure 2.

Scientific engineering enabled the development of computer-aided diagnostic (CAD) systems, a critical tool for assisting academics and medical practitioners in making sense of imagery related to medical problems. CAD systems are critical for improving medical care. These systems were developed because of scientific and technological advancements. Current studies are focused on scientific engineering, the discipline of science in charge of the development of these systems. Since a few years ago, the usage of DL techniques in CAD systems, which are part of the machine learning approach, has expanded dramatically [8-10]. The use of machine learning has grown in unison with the introduction of deep learning techniques.

3.png

Figure 3. MRI image skull removal, grey matter cutting process

CAD software has advanced significantly over the years, and customers now have access to a wide selection of options. Engineers and scientists must use a wide range of approaches to build this complex system. This notion was first described using the phrase "deep learning," although it is now more generally used as shorthand. The abbreviation DL stands for "deep learning." These methods can be used to detect and classify people with dementia such as vascular dementia, Alzheimer's disease, and moderate cognitive impairment. The abbreviation CN stands for "control normal," and it relates to a medically accepted concept of "healthy." Although the initial stages of Alzheimer's disease may be the most successfully treated, an automated diagnosis is critical for sustaining human health throughout the illness's course [11, 12]. When it comes to disease in its early stages, this argument is much stronger, Figure 3 shows MRI Image Skull Removal, Grey Matter Cutting Process. The current study focuses on employing convolutional neural networks, or CNNs, to evaluate two-dimensional convolutional magnetic resonance imaging (MRI) slices to identify Alzheimer's disease early. The research intends to enhance patient care, improve treatment results for this neurological condition, and increase diagnostic efficiency and accuracy.

A CNN is a technology that utilizes multiple filters, each of which is the size of a tiny patch of visual information and does an automatic search across the full picture to identify additional images with identical spatial properties. Each of these filters may be learned and developed alone, but when combined, they can provide critical information about a certain activity and data collection. CNN has established standards. The first stage, called convolution, is responsible for locating features and applying filters. This sort of kernel is known as a filter kernel, and the method it uses to calculate its weights is known as convolving [26]. When anything is convolved, the input data tensor is convolved with itself. Aside from the number of filters and stages used in the process, the results of using a convolutional method are affected by a variety of other factors. Convolution should be approached in the same way as addition and multiplication are. This is how all other simple mathematical operations should be performed. Alternatively, the convolution process should be addressed in the same way as addition and multiplication. However, to be as thorough as possible, we will continue to investigate it throughout the remainder of this article. When two functions are combined, the mathematical procedure known as "convolution" yields a third function. Convolution is described using the phrase "convolution." It is properly referred to in formal documents [27].

This section will go over the notion of convolution in the context of convolutional neural networks (CNNs) briefly. Each filter kernel in a CNN convolves the data to discover features in the input data. The convolution procedure is comparable to operations in elementary mathematics, and the outcome is a function formed from the input data.

This paper investigates the usage of three activation functions in convolutional neural networks (CNNs): sigmoid, tanh, and rectified linear units (ReLU). The importance of these activation functions is also investigated. Given that it provides personalised neural stimulation, this illustrates ReLU's growing popularity.

The article also discusses max pooling, which minimises the size of an image without surrendering any of its key parts, and flattening, which condenses an array into a single plane.

This article will discuss the dropout approach, which helps lower the chance of overfitting in convolutional neural networks (CNNs). This approach selects neurons at random from the training set to eliminate from the dataset. This increases the model's durability as well as the complementarity of ensemble research methodologies.

To evaluate brain pictures generated from magnetic resonance imaging (MRI), a convolutional neural network (CNN) was used. As a result, this page contains information on a specific CNN architecture. Several convolutional layers, max-pooling layers, and 64-filter kernels are used, among other things.

The chapter is concluded with a description of DenseNet, a network that can incorporate features from many clusters to achieve flawless classification. This network substantially simplifies the classification of MRI scan pictures because pre-segmentation or registration are not necessary prior to the operation.

The target matrix is then convolved with the kernel by shifting one column to the right, yielding the desired output which is completed. The desired outcome is then achieved. The results of the sliding approach have been given in a matrix for your convenience. Even if its size has been reduced from its original form in the kernel, the value remains present in all target matrices after being simplified to a single digit [28]. This immediately reduces the size of the initial matrix to that of a matrix. (The first three indicate the kernel matrix's dimensions; when combined, they provide the matrix's size.) It is critical to note that the convolution process is unaffected when applied to locally shifting situations. This is since the process of convolving impacts the outcome. Even though shows that this important part is already being taken care of well, CNN improved its performance to meet its standards.

This function is an operation that generates a nonlinear data tensor from a linear data tensor as input. The function's input is a linear data tensor. The convolutional representation is made up of many layers, each with its own activation function. The Sigmoid, Tanh and Rectified Linear Units are only three of the most important activation functions used in deep learning [29-31]. The rectified linear unit (ReLU) has lately surpassed the other nonlinear functions in popularity. This is because, unlike other nonlinear functions, the ReLU does not simultaneously stimulate all neurons. The second stage, known as "Max Pooling," aims to reduce the size of the image while retaining all the important properties that define it. Depending on the circumstances, the down sampling operation, also known as pooling, can be performed on a global or regional scale. In response to your use of the global pooling function on a two-dimensional feature map, it will only provide a single scalar result [32-34]. The local pooling function can be used to acquire smaller sample sizes for each component of the local image. This is feasible since the function pools the data locally. In the last phase, the array is consolidated into a single plane, a process known as "flattening". The final phase, often known as the "full connection" phase, is responsible for the installation of all connections and cables necessary for the phase to be considered complete.

5.png

Figure 5. The proposed model

An activation layer is frequently placed after a fully connected (FC) layer shown in Figure 5. However, this is not always the case. The receptive zone of an FC layer surrounds the channel of the layer above it. The "classifier" is the element in charge of finally assessing whether the image is normal. Dropout is utilized as a technique in convolutional neural networks an amazingly high proportion of the time. This strategy quickly gained popularity not just because of the numerous benefits it provides, but also because it is so simple to implement. Its widespread acceptance favored both characteristics. The strategy is simple: we'll choose a few folks at random and kick them out of the program. There is a random chance that one or more of the network's covert units will be taken out during each training scenario. This might happen to any of the units at any time. The genuine performance of dropouts may aid in minimizing overfitting at a reduced cost to computer resources. A dropout is a tool that may be used to do this. The dropout strategy may work well when putting together the results of several neural networks into a single prediction.

6.png

Figure 6. The proposed CNN architecture

As a result of Dropout's swift climb to renown following its initial release, several studies, including having since attempted to study the game's gameplay from numerous angles. One of these is "Dropout: A Game Design Analysis." Also, one of the many ways it has been used is to train other models, like support vector machines [35-37]. The CNN architecture used in this study required a picture of the same size and resolution as the MRI brain slice under consideration. Each part was evaluated independently see Figure 6. The max-pooling layer of the neural network was the final layer to be completed, following the completion of five convolutional layers. The final layer in the series has the most pooling throughout it. The kernel size of the most complex pooling layer is 64 filters. We looked at convolutional layers with kernel sizes up to 64 filters. The model's second convolutional layer had 64 filters with varying kernel sizes. This kernel size was also employed for the maximum-pooling layer. For the third convolutional layer, we attempted 64 filters with a kernel size of. However, when it came to the maximum pooling layer, the kernel size that we used was It was concluded that this would be the largest kernel size that could be used in the pooling layer after evaluating a total of 32 different filters with that kernel size to see whether they could be utilized in the final convolutional layer [38, 39]. Trials were carried out to determine if the filters could be used. This was done to improve the model's accuracy. Because of its track record of efficacy, the therapy was chosen to be a component of the treatment plan. Deep learning modeling heavily relies on the ReLU function. To put it another way, if a negative number is passed into the method, it will always return 0. If, on the other hand, a positive value is supplied, the result will remain the same regardless of what is done.

7.png

Figure 7. Proposed CNN network with dense layers

DenseNet was designed to take up patch characteristics for each cluster and learn those properties. The characteristics learnt from each component's normal clusters are being aggregated for categorization shown in Figure 7. The Dense Net was created to learn such traits. The mean clustering technique was employed to attain this purpose and categorize the patches based on their distinct areas. Following the creation of many local scale classifications in the picture, the findings of these classifications are combined to provide a more precise global label. This approach can learn the MRI properties for classification at every level, from discrete, microscopic bits up to the complete picture. No segmentation or accurate registration is necessary before using the images acquired by the MRI scan. These operations, however, can be performed beforehand.

A deep convolutional neural network based on an OASIS data set classifier is being created now. We are currently conducting research on this. We have attained unrivaled performance levels owing to our network. The OASIS-3 dataset is the primary archive maintained by the XNAT network. For the purposes of this investigation, 15,200 axial slices from an MRI were employed in total. Magnetic resonance imaging (MRI) scans from 170 Alzheimer's sufferers and 70 healthy people serving as controls were evaluated. Each one is picked from a different demography to ensure that the sample for the recognition test is as accurate, and representative as is practically practicable. Adults, adolescents, and children of all ages are taking part in this study. As a result, a wide age range is represented, ranging from 65 to 90 years old. Before processing the data, we "flattened" each 3-D MRI image into a 2-D representation. The initial part of the larger project is finished. Finally, it was concluded that the additional 20 dark slices were unnecessary for any of the time periods. They had no option but to surrender because they provided no evidence to back up their claims. To improve preprocessed pictures, the model's recognition ability may be checked and rotated. As a result, there was more data available for training a CNN model. The activation function for each convolution layer in our proposed system (the CNN model) is ReLU, as opposed to the systems reported, which only employed three convolution layers. This is due to the CNN model used by our system. The suggested convolutional neural network model has five layers in total. Following the completion of the convolution layers, the pooling layers begin. We noticed that increasing the batch size to 64 or 150 binary classification iterations yielded the best results. To train the model, we utilized Adadelta optimization and set the dropout rate on the dropout layer to 0.2. This procedure is quite like the methods used when Adam optimization was utilized. This has been implemented to avoid overfitting, as was done using a dropout rate of 0.5 rather than the planned dropout rate of 1.0. Finally, keep in mind that neural networks are made up of multiple components and that changing even one of them might affect the validity of the findings. The main and most important reasons for the different results are how the data was collected and what kind of information it had. In the comments section of your research project, you can go into detail about the specifics of the Convolutional Neural Network (CNN) technique you used for your question. CNNs are good for processing visual data because they use a plethora of filters to scan an entire image for elements that are similar in their spatial connection. While each of these filters must be understood and built independently, when coupled, they provide critical insights for certain activities and data sets. Each filter extracts a particular aspect of the data.

Convolution, one of the first steps of your process, is critical for applying filters and finding characteristics. Convolution of the input data tensor using filter kernels is one way; this works similarly to adding and multiplying in elementary arithmetic. Using filter kernels, this technique comprises addition and multiplication. The number of filters and stages is simply one of the variables that might impact the output of convolution. The CNNs utilised in your study differ because they use different activation functions, such as Tanh, Sigmoid, and Rectified Linear Units (ReLU). The ability of ReLU to selectively excite neurons contributes to its rising popularity. The "max pooling" phase accomplishes this by employing both local and global pooling processes. This phase is critical for lowering the file size of the image without compromising its essential components. Figure 6 depicts the CNN architecture used in your study, which includes numerous convolution and pooling layers. The last layer is the most difficult since it has the greatest pooling over 64 filters. The network is designed to analyse each component individually and is capable of processing images with the same resolution and size as MRI brain slices. The CNN design is well-known for its versatility to include findings from other networks, as well as its simplicity and efficiency in minimising overfitting. Another essential aspect of dropout approaches is their ability to decrease overfitting. DenseNet was used in your study to learn patch attributes, as seen in Figure 7. Clustering and global label development were particularly addressed. This method eliminates the need for exact registration or segmentation prior to using MRI images, enabling the learning of MRI characteristics at any size.

The results make use of the OASIS data collection, which works effectively, as seen in Figure 9. Table 2 shows that the CNN approach you gave had the highest accuracy, precision, area under the curve (AUC), and sensitivity when compared to more traditional techniques such as KNN, SVM, and LDA. Finally, as shown in Table 3, your CNN architecture is meticulously built. It starts with the input layer and progresses through a series of dropout and dense layers. The output layer includes a softmax activation function designed for speedy classification and reliable feature extraction. Figure 10 shows how critical it is to establish this equilibrium for the proposed hybrid model to correctly categorise Alzheimer's illnesses. When dealing with multi-class classification issues, your technique emphasises the significance of accurate training and parameter modifications to achieve the greatest possible balance of efficiency and performance. In this approach, tools such as the Adam optimizer and the categorical cross-entropy loss function are important. The study's vast patient group reinforces and makes the findings more practical, emphasising the importance of memory and accuracy in outcome prediction.