Abstract:

Alzheimer’s disease (AD) is one of the common medical issues that the world is facing today. This disease has a high prevalence of memory loss and cognitive decline primarily in the elderly. At present, there is no specific treatment for this disease, but it is thought that identification of it at an early stage can help to manage it in a better way. Several studies used machine learning (ML) approaches for AD diagnosis and classification. In this study, we considered the Open Access Series of Imaging Studies-3 (OASIS-3) dataset with 2,168 Magnetic Resonance Imaging (MRI) images of patients with very mild to different stages of cognitive decline. We applied deep learning-based convolution neural networks (CNN) which are well-known approaches for diagnosis-based studies. The model training was done by 70% of images and applied 10-fold cross-validation to validate the model. The developed architecture model has successfully classified the different stages of dementia images and achieved 83.3% accuracy which is higher than other traditional classification techniques like support vectors and logistic regression.

Keyword:dataset of Alzhemiers,Machine Learning (CNN)

Introduction:

Alzheimer's Disease (AD) is the most well-known and largely diffused neurodegenerative disorder occurring in the elderly. AD negatively affects patients' everyday lives, causing an advanced decline of cognitive capabilities such as memory, language, behaviour, and critical thinking (Alzheimer’s Disease International. Changes in cognitive impairment of AD patients start slowly and evolve rapidly over the long run. Similar to other body parts, brain can change as people get older. Some people lost thinking and incidental issues with recollecting certain things. Excessive cognitive decline, and other significant changes in the manner in which brain function is impaired. The first symptoms of AD are trouble recalling recently learned data because Alzheimer's progressions regularly start in the brain areas involved in learning and memory.

As Alzheimer's progresses progressively severe symptoms like confusion, mood changes, disorientation, unwarranted doubts about family and companions, and trouble talking appear. Individuals with cognitive decline or other potential indications 1of AD may think that it’s difficult to remember they have an issue. AD is a type of dementia with several implications on the cognitive domain, affecting primarily thinking and memory. Specialists and different parental figures screen the movement of AD in patients by assessing the level of decrease in the patients' psychological capacities that are often classified into three stages: very mild (normal cognitive), mild cognitive impairment (MCI), and demented (Gaugler et al. 2016). Figure 1 presents the magnetic resonance image (MRI) images of different AD conditions. Although the MCI and dementia patients both are experiencing a reduction of cognitive abilities, dementia patients would suffer from more pronounced difficulties with thinking or hampered judgment.

While reviewing the literature, Structural Magnetic Resonance Imaging (SMRI) was found to be one of the most effective biomarkers in the diagnosis of Alzheimer’s Disease(AD) as it is a rich source of brain anatomical information MRI is a non-invasive neuro imaging modality utilized to image the internal structures of the brain. An MRI usespowerful magnetic fields and radiofrequency pulses to form the image of bones, organs, and especially soft tissues. At present, MRI is considered the most sensitive imaging modality of the brain. MRI gives the morphology of the white matter, gray matter, and cerebrospinal fluid (CSF). It helps to understand the brain atrophy regions and anatomical changes corresponding to Alzheimer’s disease. Thus they are recognized as an important biomarker for detecting AD and its progression. Nowadays, deep learning networks have become a robust and efficient tool for classification, especially when there are many datasets . Deep networks learn the weight parameters more efficiently to give higher classification accuracy from test data . This paper uses a cascaded 3DCNN text ract the features independently from 27 partitioned patches of the brain. The features are concatenated to form the global

Alzheimer’s Disease (AD) is a neurological, chronic, and progressive brain disorder that leads to memory loss, less thinking ability, and inability to perform simple tasks. AD usually affects older people of more than 65 age. The statistics tell that around 90 million people are affected by AD, and it is anticipated to rise towards 300 million by 2050 . Mild cognitive impairment (MCI) is a prodromal stage, or it is considered as a transitional state from Normal Control (NC) to AD dementia . At present, there is no treatment to cure AD; thus, it is critical to diagnose the disease accurately at its early stage for providing patient care and for developing future treatments. The family and friends are the ones who notice the behavioral and cognitive changes an AD patient undergoes. The gravity of symptoms is not easily noticed as the AD symptoms confuse the normal aging process, especially in rural areas. Thus, doctors are not consulted until being too late, resulting in a late diagnosis. A survey was conducted in which 64% caretakers of AD patients affirm that the behavioral changes they notice are due to the normal aging process, and 67% of them agree that it is due to this reason they delay consult doctor. Even after consulting physicians, many are not able to diagnose the disease. It is said that even the most experienced specialist fails by 10 -15 % cases to diagnose the disease correctly. This situation raises the importance of an accurate computer aided detection device for assisting doctors in diagnosing AD and especially early diagnosis of the disease. Early diagnosis increases AD patients’ survival rate, but the diagnosis is harder as it needs to detect the pathology’s most subtle symptoms. Definite diagnosis is only possible after the post-mortem of the brain. Still, neuroimaging technology is increasingly used as it offers the possibility to examine the patients when they are alive.feature representing the whole brain. The fused feature vector is further trained and classified using Multilayered Perceptron to attain a good accuracy for three-level classification of AD. In , the authors used a nonlinear graph to perform three level and binary level classifications. They used information from different modalities like PET, CSF, and
MRI to perform three-level classifications and achieved an accuracy of 56.6%, which is very less.

They extracted around two lakh features from PET, CSF, and MRI from 83 ROI. One of the

Advantages of this method is that it is a non-parametric method. In the work the authors used ANN and initialized weights using the Centered Kernel Alignment (CKA) method to increase the accuracy of three-level classification. They used 324 features extracted from MRI to train and test their algorithm and achieved an accuracy of 70.3%. The architecture complexity is less, but it cannot be implemented for diagnosis due to less accuracy. The authors of applied the wavelet decomposition method. They used MRI images as input and then followed the discrete wavelet transform method for feature extraction. To reduce the complexity, they applied principal component analysis and then classified data with an accuracy of 83.33% using the Support Vector Machine (SVM) method.
In , the authors tried many machine learning methods and ANN techniques and came up with a maximum accuracy of77.1% for three classifications. They extracted 81 features:cortical volume feature, cortical thickness feature, and hippocampus volume features. The SVM classifier gave only 58.4% accuracy.Authors in used a Multilayer autoencoder to train
and classify AD, MCInc, MCIc, and NC. The algorithm achieved a classification accuracy of 47.2%. They segmented MRI and PET images to get 83 regions of interest areas and extracted the classification features.

The method lacked accuracy and used conventional segmentation methods and handcrafted features that are time-consuming and need domain experts. Authors in introduced zero maskings in stacked autoencoder to learn better feature representation from MRI and PET features and achieved an accuracy of 53.7%. They throw away some features to reduce dimensionality, and due to which less accuracy is gained. Two models were created
by authors in using stacked autoencoders to achieve an accuracy of 56.6% and 58% for AD, MCI, and NC classification. Texture features were extracted from the segment hippocampus and other ROI regions. The authors throw light into the extraction of texture features using different methods. But the accuracy is still less and needs the hectic process of segmentation to be done for feature extraction. In the author utilized Three Dimensional Convolutional Neural Network (3DCNN) architecture to learn features directly from the images without any need of human expertise and classified AD, MCI, and NC successfully to achieve an accuracy of 9.4%. The 3DCNN learned the hierarchal features from low level to high-level feature, which enabled the algorithm to classify well. In their work, they resized the whole image to a smaller dimension that is not appreciable in medical imaging. Authors in , and utilized 3DCNN architecture and tried to increase the performance by initializing the weights
using pre-trained Three Dimensional (3D) autoencoders. The work is good enough, but they considered the whole image as input that demands heavy computational resources and is costlier to train such architecture. Inspired by ResNet, a 39 layered 3DCNN architecture was proposed in classify AD, MCI, and NC. In 3D VGG based architecture was proposed to classify the same. The number of parameters is huge, requires high computational resources, and is difficult to perform hyperparameter tuning. To overcome all such problems and increase accuracy, we have proposed a patch based method utilizing a cascaded 3DCNN to extract the features and a multilayer neural network for classification.