ALZHEASE CARE Alzheimer’s Prediction using Deep Learning

Download Full-Text PDF Cite this Publication

Text Only Version

ALZHEASE CARE Alzheimer’s Prediction using Deep Learning

Ansiya Noushad, Austin johns, Bhavya Babu, Joan Vincent, Dr. Remya K Sasi

Computer Science and Engineering

Christ College of Engineering, Irinjalakuda, Kerala India

Abstract – As per reports by the Alzheimer's Disease association, more than 5 million Americans are living with Alzheimer's Disease today, with an anticipated 16 million by 2050. 1 in 3 seniors die because of Alzheimer's disease or different types of dementia. No treatment developed so far can cure a patient who is already in AD. So prevention is better than cure. MR image samples were used to distinguish between healthy individuals, patients with Alzheimers disease, and those with mild cognitive impairment. To understand the changes in the brain we use MRI images as input. Pre-processing methods were performed before images were given as input to the model. After preprocessing the image data, we created CNNs and performed evaluations on these models.

Keywords – 3D Brain image , Image processing, Convolutional Neural Network


    Alzheimer's disease (AD), an irreversible, progressive neurodegenerative disorder, causes problems with memory, thinking, and behavior. The early symptoms of Alzheimers disease are memory loss, mood changes, poor judgement, social withdrawal, changes in vision. This happens because Alzheimer's disease affects the hippocampus, which plays an important role in memory. One in every 3 seconds a new person somewhere is affected by dementia. At first, it typically destroys neurons and their connections in parts of the brain involved in memory, including the entorhinal cortex and hippocampus. Hence, to understand the changes in the brain need to be studied and explored. It is in this context that MRI images are used as input. and a use a deep learning approach is used to study the variations shown by various factors to calculate the transition from Mild Cognitive Impairment (MCI) to AD. So when changes are evident in those factors, one can be aware of such changes and take needed medications.


    The aim of our work is to aid in detecting the different stages of Alzheimers Disease. The system predicts the conversion from Mild Cognitive Impairment (MCI) to probable Alzheimers Disease using Image Processing. A website wherein the patients can access their case files and update their treatment regimen along with their test reports to know the prognosis of the disease has also been created. Here the various stages of Alzheimers disease are detected and to maximize the accuracy of disease prediction.


    Neelaveni and Devasena proposed machine learning techniques along with psychological parameters like Number of visits, Age, MMSE Score, and education of the patient used for early detection of AD. Machine learning algorithms Support vector machine and Decision tree were used to classify AD in patients and distinguish between cognitive impairment.

    Shahbaz et al proposed a six different Machine learning algorithms and Data Mining techniques to classify five different stages of Alzheimers disease and to distinguish different attributes for each stage of Alzheimers Disease among ADNI dataset. The results of this paper revealed that Generalized Linear Model can efficiently classify the stages of Alzheimers Disease with an accuracy of 88.24% on the test data set.

    Alex Fedorov investigates the use of variants of DIM in a setting of progression to Alzheimerss disease comparison with supervised AlexNet and ResNet inspired Convolutional neural networks. Here classification is done between four groups: patients with stable and progressive mild cognitive impairment, with Alzheimerss disease and healthy control. ADNI database is used here.

    Mr Amir Ebrahimi Oshnavieh proposed Alzheimers disease prediction using Deep Learning. Intensity normalization and registration are key preprocessing methods in AD detection. Here for feature extraction patch-based methods are used on disease related regions. Convolutional neural network is used .

    Taeho Jo used Data mining approaches for the early detection and automated classification of AD. Personal information, age, MMSE score is used as data set .Data Mining techniques like CNN and RNN were used. It produces higher accuracy.

    Garam Lee's work proposes a method for prediction of the conversion of Mild cognitive impairment to Alzheimers disease. A deep learning approach called multimodal recurrent neural network was used. There is a transitional stage between cognitively normal adults and AD patients called Mild cognitive impairment (MCI). This method takes the longitudinal and multi-modal nature of available data and it discovers nonlinear patterns associated with MCI progression. The goal of ADNI (Alzheimers Disease Neuroimaging Initiative) is to test whether magnetic resonance imaging (MRI), positron emission tomography (PET), other biological markers, clinical and neuropsychological assessment could be

    used to measure the progression of MCI and AD. The methods use a recurrent neural network and Multi-modal GRU for MCI conversion prediction. The results showed that they achieved the better prediction accuracy of MCI to AD conversion by using longitudinal multi-domain data. A multi-modal deep learning approach has the potential to identify the risk of developing AD who might benefit most from a clinical trial.

    Aunsia Khan r presents a review, analysis and critical evaluation of the recent work done for the early detection of AD using ML techniques. Many other factors such as pre- processing, the number of important attributes for feature selection, class imbalance distinctively affect the assessment of the prediction accuracy. To overcome these limitations, a model is proposed which consists of an initial pre-processing step followed by imperative attributes selection and classification is achieved using association rule mining. The proposed model-based approach gives the right direction for research in early diagnosis of AD and has the potential to distinguish AD from healthy controls.

    Ji Hwan Park in this research, it focuses on developing a progressing framework to diagnose and predict AD at a very early stage with the data collected for AD patients. The collected data is fed into the framework with the input needed to deploy the computational modelling and the machine learning techniques to predict and diagnose AD. The research depends on gathering data from various existing datasets such as ADNI, and it includes collected data related to DNA, dietary, medical history, lifestyle and any other related data linked to the risk factors of AD.

    Fan Zhang proposed the methods accurate diagnosis of MCI is essential for the early diagnosis and treatment of AD. This paper presents a deep learning model for the auxiliary diagnosis of AD, which simulates the clinicians diagnostic process. The proposed model provides a comprehensive analysis about patients pathology and psychology at the same time; therefore, it improves the accuracy of auxiliary diagnosis. The results of multimodal neuro-imaging diagnosis are combined with the results of clinical neuro-psychological diagnosis.

    Ammarah Farooq this work proposes a deep convolutional neural network pipeline based on the diagnosis of Alzheimers disease and its stages using magnetic resonance imaging (MRI) scan. Alzheimers disease causes permanent damage to brain cells associated with memory and memory skills. Diagnosis of Alzheimers in the elderly is very difficult and requires a differentiatin factor due to similar brain patterns and pixel strength. Deep learning strategies are able to learn such representations from the data. In this paper, a 4-way classification is used to differentiate Alzheimers (AD), mild cognitive impairment (MCI), late mild cognitive impairment (LMCI) and healthy persons. Tests were performed using the ADNI database on a high-resolution graphics system and new technological results were available for a variety of disease classifications. This proposed method results in a predictable accuracy of 98.8%, which is a significant increase in inaccuracy compared to previous studies and clearly demonstrates the effectiveness of the proposed method.


    After prepressing the image data, we reted NNs and performed evaluations on these models. NNs are dee artificial neural networks and mmnly used in image related litins such as image lssifitin, clustering, and interpretation. They are inspired by the layered vision mechanism of humans. The improvement f the hardware and the increased ressing ity f grhis ressing units (GUs) has allowed the training of deep networks on muters more efficiently. Big data, which n be lleted from many platforms, are the basis for the implementation of NNs and other deep learning models. Beuse NNs are designed as deep models, they can provide sufficient results in solving mlex problems. The model, which is trained with the bk-rgted algorithm, uld mke predictions n the pixels of the image without feature extrtin. In the following years, many suessful deep models such as lexNet, ZFNet, GgleNet, VGGNet, and ResNet have been reted. The structure f NNs nsists of nvlutinl layer, ling layer, and fully connected layers. Generally, many nvlutinl layers and ling layers are stacked one after the other to rdue feture m, and the generated m is fed into the fully nneted layer. In the nvlutinl layer, filter with a determined size f n×n is lied to the image expressed as pixel matrices.

    The filter is lied by moving over the pixel metrices nd navigating the entire image matrix. Depending on the type of filter lied, some features of the image are revealed. In the ling lyer, the size f the stil dimension is reduced by performing the sub-sampling process. In this way, ling layers provide ease of muttin as well as providing solution for over-fitting. This layer can tinlly be selected as maximum r verge ling. In the maximum ling, the maximum ixel in the window is selected, while the verge f the pixels in the window is obtained in the verge ling. After these ertins, lssifitin rding to the feature m containing the extrted features is made by the fully connected layer, which might be multi-layer eretrns. In the present study, three NN models of similar structure were nstruted. The type of ling layer was determined as maximum ling. rectified linear unit (ReLU), which is widely used in modern deep learning models, was used as the tivtin function. Let x nd f(x) be the number f inputs and an tivtin function, respectively. The ReLU is defined as f(x) = mx(0, x). Moreover, binary cross-entropy function ws employed s loss function.

    This function is also lled log loss due to logarithmic ertins. Different models were obtained by changing the total number and sitins f the layers.


    1. General Framework

      To understand the changes in the brain we use MRI images as input. We use deep learning approach to study the variations shown by these factors to calculate the transition from MCI to AD. So, when changes are happening in those factors, we can be aware about that and take needed medications. Since the number of different prediction images was not the same, the data addition ress was used. Other advancing methods were performed before the images were presented as input t the model. MR images were used t distinguish between healthy individuals, tients with Alzheimers disease, and those with mild dementia. One f the MR data sets was used for the training ress, while the other set of data was used fr testing. Since the number f different prediction images was not the same, the data addition ress was used. Other advancing methods were performed before the images were presented as input to the model. NN models with different layers were created and then applied for the lssifitin of Alzheimers disease types. Performance ws mred in terms f the diagnostic tentil f the rsed model.

    2. Preprocessing of brain MRI data

      In prepressing of brain MRI data, volumetric brain MRI data were studied and imaging ressing techniques used. Alzheimers disease and dementia are neurological disorders that ffet the nerve cells so it is wise to remove unnecessary information. Beuse the skull, eyes, fat, and muscles are unaffected by these diseases, these regions are unnecessary data in the green images. FSL is mlete library of data analysis tls for magnetic resonance imaging (FMRI), MRI, and diffusion tensor imaging (DTI) litins. BET is an automated method of classifying MRI images as res f the brain and nonbrain res. The munt of histogram is lulted from the input image. The triangular tessellation of the spinal rd is initiated internally in the brain and llws for slow rotation of one vertex at time, following the force that keeps the area well serted and smth, while trying to get to the edges of the brain. It is repeated with a high smth bar until len solution is obtained. As result of this ress, nonbrain bjets are removed from the input image. The nonbrain regions found are being analyzed. We selected ten slices from the xil, sagittal, and rnl projections for eh tients volumetric brain data. While selecting these slices, we id attention to hse regions such as the himus, thlmus, hythlmus, mygdl, erebellum, frontal lobe, rietl lobe, iitl lobe, and rus llosum, which are affected in dementia and Alzheimers disease. data augmentation method was lied because of the unbalanced number of tient samples, which might lead to overfitting of the model. For this urse, 10% right and left, up and down shifting, 20% zm, and 20% shear ertins were lied; zoomed and changed sitin images were produced. fter

      an equal number of patient samples were obtained, the images were resized t 150×150×1.

      The nn brain regions obtained are ressed. Image ressing techniques like gray sling and histogram equalization is mainly done on them. Grayscale is range of mnhrmti shades from blk t white. Therefore, grysle image contains only shades f gray and no lr. Gray-scaling is the ress of converting continuous-tone image to an image that muter n mniulte. Histogram equalization is method to ress images in order to adjust the ntrst of an image by modifying the intensity distribution of the histogram. The objective of this technique is to give linear trend to the cumulative probability function ssited with the imge. And then these images are rotated to get the images at various angles. These are also flied to get the mirror image. So, now we have more images in the dataset which can improve the training process and the ury. This will also help our model to classify the images even if they are given rotated or flied.

    3. Implementation of models

    After re-ressing the image data, we created NNs and performed evaluations on these models. NNs are deep artificial neural networks and mmnly used in image related litins such as image lssifitin, clustering, and interpretation. They are inspired by the layered vision mechanism of humans. The improvement f the hardware and the increased ressing ity of grhis ressing units (GUs) has allowed the training of deep ntworks on muters more efficiently. Big data, which n be lleted from many platforms, are the basis for the implementation of NNs and other deep learning models. Beuse NNs are designed as deep models, they can provide sufficient results in solving mlex problems. The obtained desired image is then fed into the nvlutinl Neural Network (NN) Algorithm as input. The structure of NNs nsists of nvlutinl layer, ling layer, and fully nneted layers. Generally, many nvlutinl layers nd ling layers are stked one after the other to rdue feature m, and the generted m is fed into the fully nneted lyer.


    1. Deep Learning

      Deep Learning is n Artificial Intelligence function used to emulate the workings of the human brain in ressing data and also generating tatters for use in decision making. It is also known as Dee Neural Learning or Dee Neural Network and is subset of Mhine Learning in Artificial Intelligence that involves networks ble f training unsupervised from data that is unlabeled or unstructured. Dee Learning is lied rss all industries for number of various tasks. mmeril s that make use of image regnitin, en source platforms with consumer s, and medical research implements that rbe the possibility of drug reusability for new ailments re some of the examples tht come under Deep Learning inrrtin.

    2. Convolutional Neural Network(CNN)

      In deep learning, nvlutinl Neural Network(NN) or nvNet enmsses lss of deep neural networks which is most mmnly lied in order to nlyze visul imgery. In neural networks as whole, NN is one of the most widely used neural networks to erfrm number of litins like image regnitin, image lssifitin, bjet detection, fe regnitin et. The nvlutinl neural Network algorithm which bsilly takes in an input image, assigns learnable weights/bises to various bjets/sets in the image and enables differentiation of one from the other. Hence, in this rjet, the NN algorithm is used for ressing the images before the training. Other lssifitin lgrithms require much more re-ressing than is required in nvNet.

    3. Python

    One of the ulr and consistently ranked programming lnguges is ythn. Its object oriented rh together with its language nstruts aid programmers to write ler, lgil des for both small and lrge-sle rjets. Being dynmilly typed and grbge lleted, it helps to surt multiple programming rdigms mprising structured, bjet oriented and functional programming. Taking all the above fts into consideration, this rjet t detect Alzheimers Disease was designed nd implemented using ythn and its various libraries, including neural network platform librry-Kers running on t of the en source mhine learning platform-TensorFlow and re ythn librry- Numy.

    1. Keras

      A high level neural networks library and the most used deep learning framework written in Python programming language, running on top of the machine learning platform TensorFlow. This makes it an extremely simple and intuitive library to use.

    2. TensorFlow

      An end-to-end open-source Machine Learning platform for Keras. It consists of a comprehensive and flexible ecosystem of libraries, tools and various other resources that aid in providing workflows with high-level APIs.

    3. NumPy

    NumPy, a core Python library that provides the multidimensional array objects and a collection of routines for processing those arrays, is a simple yet powerful data structure used for scientific computing. This method uses NumPy arrays to store the images while it is in training.


    In our rsed model the deep Learning algorithm lled nvlutinl neural Network is used to detect Alzheimers Disease. Feature extraction and lssifitin are the two parts in Dee Neural Networks. Here, we use MRI images of the tient as input to detect the stage of the disease. The input image is then ressed. In image ressing, the images are divided into four stages;

    mild-demented, mdertedemented, nn-demented and very-mild-demented. Before training the mhine, we re- processed the data to make the data set f the same type. To make the rmeters the same, gray sle and dtive equalization are used. Gray scaling is the ress of converting an image from other lr ses (e.g RGB) to shades of gray. dtive histogram equalization is method in image ressing of ntrst adjustment using the images histogram. To train the mhine we want to maintain a minimum range. By maintaining it, the efficiency of the model increases. For that, rotate and flip functions are used to increase the number of images to maintain the minimum unt.

    In reressing the data, we unt the number of images. Images n be rotated to any degree lkwise r nti- lkwise. We just need to define rotation matrix listing rttin int, degree f rttin and the sling ftr. When the range is bve minimum value, flip and rotate functions nt be lied. They re used to increase the number of images. The processed data is given to train the model. In trining, fur lsses nmed tp, tp, tp, and th 4 re used. The ressed data folder ntins the different stages of Alzheimers disease. th is an argument in the load image function where we n give an address. Load image function contains the path frm the lss. Here the images re stred as Numy rrys. We use horizontal stk function to make lbels tgether. Int function is used to specify the data type. Here, we ly normalization t make every value in the sme range. She function is used t get the number of classes of the images. For slitting our image for trining nd testing, the train and test function is lied. nvlutin 2D is the input layer in model function where the input is tken.Thts why input she is specified. Mx ling is used to bring the features of the nvlutin layer together. Dr layer eliminates the unwanted features.

    Batch normalization normalizes the features of eh layer deeply. Flatten is used to make imges in one rry. Next part in the deep neural network is the lssifitin rt. Here the dense layer ntins the number of neurons. Finl dense lyer ntins the total number of neurons. Sft mx tivtin function is used to make the final model layer. Then we mile the model. Validation split function is used to split the 20% f data. Then we save the model. Mtltlib.ylt is used t view the utut of the model trained as grh. Here we lt the grh as two satins. After testing the data 0.96is the ury and prints the data. In the testing rt, it is mbintin of re-ressing. Ld image function in Keras model we load our model and assign it to mdel.Then we normalize data. At final we got the rresnding tients stge of Alzheimers disease as output.

    We implemented this as website. The user can log into the site using their login credentials. Here the tient n input the brain images. After lying the CNN algorithm, the utut determines the stage of Alzheimers disease of the tient.


    During this uncertain erid of vid-19, we me to the realization of how prevention is better than cure. We set out to find suitable way to redit the urrene of Alzheimers, serious disease. Besides the benefit of early diagnosis to the patients, it also ntributes to the national enmy. This enmi contribution is related to the lower st of treatments lied in early stages of dementia mred to in later stages.


    First and foremost, we thank the God Almighty for His blessings showered in us throughout this rjet to successfully complete it. We would like to express our deep sense of gratitude to or rinil Dr. Sajeev John, for his nstnt encouragement and valuable advice throughout the urse. We are extremely grateful to Dr. Remy K. Ssi, Head of Dertment, muter Sin and Engineering, for her valuable guidne thrugh this humble endeavor. Her highly enterrising attitude and tiene fr listening to us and our dubts, timely suggestions and guidance made our rjet reality in stipulated time. We would also like to thank our rjet -rdintr Mrs. Ris Vrghese, ssistnt rfessr, Dertment of muter Siene and Engineering and to our belved guide Dr. Remy K Ssi, Head of Dertment, muter Sin and Engineering, for rer guidance and the valuable assistance. We express our sincere gratitude to all the faculty members of the department of Computer Science and Engineering for their ertin and surt to the rjet. We also express our sincere gratitude to all our friends nd our parents for their ertin and nstnt inspiration.


  1. Neelaveni and Devasana-Alzheimer Disease Prediction using Machine Learning Algorithms, 2020 6th International Conference on Advanced Computing and Communication Systems (ICACCS), Coimbatore, India, 2020, pp. 101-104, doi: 10.1109/ICACCS48705.2020.9074248.J. Clerk Maxwell, A Treatise on Electricity and Magnetism, 3rd ed., vol. 2. Oxford: Clarendon, 1892, pp.68-73.

  2. Shahbaz et al- Classification of Alzheimers Disease using Machine Learning Techniques. 296-303. 10.5220/0007949902960303.

  3. Amir Ebrahimighahnavieh et al- Deep learning to detect Alzheimers disease from neuroimaging: A systematic literature review, Computer Methods and Programs in Biomedicine,Volume 187,2020,105242,ISSN0169-2607.

  4. Garam Lee et al-for Alzheimers Disease Neuroimaging Initiative, Predicting Alzheimers disease progression using multi-modal deep learning approach, Scientific Reports Volume 9, Article number: 1952 (2019).

  5. Altug YIGIT and Zerrin ISIK -Applying deep learning models to structural MRI for stage prediction of Alzheimers disease

  6. Ammarah Farooq et al-A deep CNN based Multi-class Classification of Alzheimers disease using MRI.

  7. Sarraf and Tofighi- Deep learning-based pipeline to recognize Alzheimers disease using fMRI data, 2016 Future Technologies Conference (FTC), San Francisco, CA, 2016, pp. 816-820, doi: 10.1109/FTC.2016.7821697.

  8. Al-Shoukry et al- Alzheimers Diseases Detection by Using Deep Learning Algorithms: A Mini-Review, in IEEE Access, vol. 8, pp. 77131-77141, 2020, doi: 10.1109/ACCESS.2020.2989396.

  9. Tabarestani et al- Longitudinal Prediction Modeling of Alzheimer Disease using Recurrent Neural Networks. 1-4. 10.1109/BHI.2019.8834556.

  10. Tanveer et al-Machine Learning Techniques for the Diagnosis of Alzheimers Disease: A Review. ACM Transactions on Multimedia Computing, Communications and Applications. 16. 35. 10.1145/3344998.

Leave a Reply

Your email address will not be published. Required fields are marked *