Early Detection of Alzheimer’s Disease
Dr.Kamatchi Priya L Jayant Harwalkar
Department of Computer Science Department of Computer Science
PES University PES University
Bangalore, India Bangalore, India
priyal@pes.edu jayanthharwalkar@gmail.com
Shravani
Department of Computer Science
PES University
Bangalore, India
shravanim1250@gmail.com
Abstract—Alzheimer’s disease is a progressive neurological
disorder that causes the brain to shrink and brain cells to die.
As of now, the disease is incurable but medication and
management strategies may temporarily improve symptoms.
AD models play an important role in early diagnosis so the
patient can have an opportunity to get the treatment. This
paper contains comprehensive analysis of different machine
learning methods used to detect AD.
Index Terms—component, formatting, style, styling, insert
I. INTRODUCTION
Alzheimer’s disease(AD), a brain disorder, is
considered to be a form of dementia that slowly
destroys memory, thinking, and eventually, the ability to
perform daily tasks. It is caused by the loss and
degeneration of neurons in the brain mostly in the
cortex region. AD is caused by the formation of plagues
in which clumps of abnormal proteins are formed
outside the neuron which block the neuron connections,
disrupting signals, which leads to impairment of the
brain. AD can also be formed by tangles in which a build-
up of protein occurs inside the neuron which affects the
signal transition. In AD, the brain starts to shrink, the
gyri become narrow while the sulci widen. The risk of
getting this disease increases with age and it is mostly
seen in older people.
AD can be diagnosed by doing a brain autopsy and
biopsy and there is no complete cure for the disease.
Having an early detection improves the chances for
effective treatment and the ability of the individual to
participate in a wide variety of clinical trials. Treatment
is effective if given in the early stages. Currently, there
are no treatments to reverse the damage already caused
but proper medication can halt the further progression
of AD and prolong their life.
AD can be detected by performing scans like magnetic
resonance imaging(MRI), computed tomography(CT), or
positron emission tomography(PET)[3]. Researchers use
raw MRI brain scans, demographic images, and clinical
Hemankith Reddy M
Department of Computer Science
PES University
Bangalore, India
hemankith@gmail.com
Neelesh S
Department of Computer Science
PES University
Bangalore, India
neeleshsamptur@gmail.com
data to compare it with normal cognition by using a
developed machine learning and deep learning model
which predicts AD risk.
Despite the currently available bio markers, electronic
healthcare data, health records, and increase in
digitization of data, there is not enough information on
how to use this large-scale health data in the prediction
of AD risk, but there are few studies that demonstrate
that potential AD risk can be predicted when these
resources are combined with data-driven machine
learning models.[5]
The subsequent sections are divided into literature
review, proposed method, implementation details and
conclusion. In proposed method we have described the
architecture and discussed about the model. The
implementation details includes the brief set out of the
data set, prepossessing and winded it with the
comparative analysis with respect to different models.
II. LITERATURE REVIEW
Considering the amount of research that is made,
machine learning techniques, branch of artificial
intelligence are widely used to this. Even with the
existence of all these methods, there are no instruments
for the detection. However, certain, physical,
neurophycological, phycological, neurological tests can
be used for identification of this disease.[12]
SVM was heavily researched for both feature
selection and modelling. There are many variants of
SVM used for classification, for example it can be done
using SPECT images which uses SPECT perfusion imaging
to classify the healthy patients’ images from those
having AD. The approach is based on linear
programming formulation based on linear hyperplane
which performs simultaneous feature selection and
classification. This method has specificity of 90% and
specificity of 84%, this is also proven to be better than
Fisher Linear Discriminant(FLD) and also statistical
parametric mapping (SPM) and also better than human
experts.[6] Using SVM to find Atrophy patterns in AD as
a feature selection is also proven to be one of the best
methods. This method classifies whether the patient has
AD or not based on the anatomical MRI. Even though
this approach provided good results on Cohort 1, the
results weren’t great for inter-cohort as the accuracy
dropped to 74%.
This showed that the selected regions of considered
refined atlas did not have good generalisation ability[7].
SVM was also used for binary classification using LIBSVM
toolbox under MATLAB(the Scikit-learn library can also
be used for python to implement SVM(SVC-Support
Vector Classification)).SVM is less sensitive to the
dimensionality of the problem and hence allows working
with complex problems that involve a large number of
variables. The Radial Basis Function kernel was chosen
as it offers good asymptotic behaviour.However, the
results in some conditions might be nontrival[13].
Decision trees is also one of the most popular
methods as well. The ID3 Decision Tree, along with
measures like Entropy and Information Gain has been
used in this research. At each node in the decision tree,
the attribute with the highest information gain is chosen
as the splitting attribute.
PSO is one of the well known methods for optimising
feature extraction as well as classification. Optimisation
algorithms like GA (Generic programming) for feature
selection, PSO (Particle Swarm Optimisation) for
performance optimisation and ELM (Extreme Learning
Machine) for classification, VBM (Voxal Based
Morphometr for feature extraction, along with ELM and
PSO classified can be used to identify the class of AD
among the three classes. Training and testing accuracy
were 94.57% and 87.23% respectively for GA-ELMPSO
algorithm over 10 random trials.[18] PSO for feature
reduction along with Decision Tree Classifier for
classification achieved an accuracy of 91.24% while the
sensitivity was 91.24% with specificity being 93.10%. In
this method, feature reduction by PSO gave a reduced
set of variables instead of original data and finally
classification is done using decision tree technique there
are many parameters are to be found which takes a lot
of time and energy and process of noise reduction is
difficult as the images were degraded.[19]
CNN was also prominent as prediction and
classification model. The classification is done using two
methods and the first one is building the CNN
architecture from scratch based on MRI scans 2D and 3
D convolutions. The second method is using transfer
learning techniques like VGG19 pre-trained model.
Standard CNN contains feature extraction, feature
reduction and classification so there is no need to do
extraction manually. The weights in initial layers act as
feature extractors and their values can be further
improved by iterative learning[9]. CNN was used as a
feature extractor for Decision tree, SVM, K means
clustering, and Linear Discriminate classifiers are
applied. The classification accuracy was seen to be
improved by including a pre-processing method before
CNN models[15].
A deep learning architecture with a softmax
regression layer and stacked sparse auto-encoders was
also used to develop early diagnosis technique for
Alzheimer’s disease. The autoencoder learns the input
representations. By selecting the highest predicted
probabilities for each label, the softmax regression layer
classifies instances.•The Accuracy, Sensitivity and
Specificity of the model turned out to be 87.76%,
88.57%, and 87.22% for classification of AD vs NC and
showed increase in accuracy compared to convention
methods such as SVM [14]. Another approach
implements feature extraction of brain voxels from grey
matter and classify using the CNN algorithm. Voxels are
enhanced with a Gaussian filter, and unnecessary tissues
are deleted from enhanced voxels. And then CNN
algorithm is used for classification and this method
achived an accuracy of accuracy of 90.47%, 92.59%
precision, and recall of 86.66% in comparison to some
system which uses physician decision.[16]
III. PROPOSED METHODOLOGY
A. Architecture
The images from the database are fed to a pipeline
which consists of a series of pre-processing techniques.
PSO performs feature selection on the pre-processed
images. The resultant images will be stored in a
database and will be used by PSO to get optimal
parameters of the Convolutional Neural Network. This
produces an optimized architecture for CN. The CNN
model is trained, validated and tested.
B. CNN parameter optimisation using PSO
The process of training is repetitive and continued
until the stop criteria is met. The steps to optimize PSO
are:
1) Feed the pre-processed images to as input to the
CNN network. The images should be of the same
size and characteristics. For example, they should of
the same dimensions, scale, color gamma, etc.,
2) Design of PSO parameters. The algorithm’s particle
population is generated. This involves setting the
values of number of particles, number of iterations,
inertial weight, social constant, cognitive constant
 etc., Random values can be set or can also be set X1 coordinate controls the hyper-parameter for
according to some heuristic convolution layer number. If X1 = 4, it means that there
3) With the parameters obtained by the PSO, will be 4 convolution layers. X2 and X3 control the
parameters of CNN are initialised (parameters to be hyper-parameters filter number and size respectively. If
set are given in the table below). The CNN is ready X2 = 32 and X3 = 2, it implies that there will be 32 filters
to be trained now. of size 5x5 (1 is mapped to 3x3, 2 to 5x5, 3 corresponds
4) Training and validation of CNN. The CNN reads, to 7x7, 4 implies 9x9). Similarly, X4 and X5 control filter
processes, validates and tests the input images. This numbers and size for layer 2. The same goes for all the
step produces values for the objective functions. remaining coordinates. X10 represents the batch size for
The objective functions are AIC and Recognition training.
rate. These values are returned to the PSO.
TABLE III: Example particle generated by algorithm
5) Calculate the objective function. The objective
function is calculated by PSO to obtain the optimal 4 100 2 64 2 64 3 96 1 32
values in the search space. IV. IMPLEMENTATION DETAILS
6) PSO parameters are updated. Both, the position of A. Dataset
the particles and the velocity of the particles that The data was obtained from the Alzheimer’s Disease
characterize the particles, are updated by taking Neuroimaging Initiative (ADNI)database
into consideration Pbest and Gbest. They are updated (adni.loni.usc.edu). The ADNI is a long-term study that
based on its own optimal position (Pbest) and the uses indicators such imaging, genetic, clinical, and
optimal position of the entire swarm in the search biochemical markers to follow and detect Alzheimer’s
space (Gbest). disease early. The ADNI data repository has around 2220
7) This process continues until the end criteria is met. patients’ imaging, clinical, and genetic data from four
The end criteria can be the number of iterations or a investigations (ADNI3, ADNI2, ADNI1 and ADNI GO). The
threshold value. image data (MRI scans) was used. ADNI provides
8) It is then determined which architecture is optimal. researchers with as they work on progression of
Here the Gbest particle represents the optimal Alzheimer’s disease. PET images, MRI images, genetics,
architecture. cognitive tests, CSF, and blood data are collected,
To elaborate further on how the algorithm will work, validated and these can be used by researchers as
an example is presented. The particle structure can predictors or the disease. The first goal is to detect AD at
consist of 8 positions as shown below. Each particle has the earliest stage (pre-dementia) and identify
these 8 positions and each of these positions are biomarkers that can be used to track the disease’s
responsible for tuning one hyper-parameter. The hyper- progression. Support breakthroughs in Alzheimer’s
parameters to be optimized are given in the below disease intervention, prevention, and therapy by using
table. innovative diagnostic tools at the earliest possible stage
(when intervention may be most successful).
TABLE I: Structure of Particle ADNI has ground-breaking data-access policy, which
X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 makes data available to all scientists worldwide without
TABLE II: Example Particle Structure restriction. We have acquired 2294 ADNI1 1.5T MRI
Particle Hyper-Parameter Search Space scans which are in the NiFTI format. The images are pre-
Coordinate classified into CN, MCI or AD. Each of the images are of
X1 Convolution layer no. [1, 4] the shape 192 x 192 X 160.
X1 Filter no. (1st layer) [32, 128]
X3 Filter size (1st layer) [1, 3] B. Pre-processing
X4 Filter number (2st layer) [32, 128] The pre-processing steps are:
X5 Filter size (2st layer) [1, 3]
(I) ADNI pipeline
X6 Filter number (3rd layer) [32, 128]
(II) Registration
X7 Filter size (4th layer) [1, 3]
(III) Segmentation and Normalization
X8 Filter number (4th layer) [32, 128]
(IV) Skull Stripping
X9 Filter size (5th layer) [1, 3]
(V) Smoothing
X10 Filter number (5th layer) [32, 128]
(I) ADNI pipeline: The images in the dataset are
obtained from MRI machines, and the machines
use magnetic waves and radio waves to produce
the scans. There are parameters such as radio
frequency, magnetic frequency and uniformity of
the coil which can cause variations in the MRI
scans. To correct such variations in the images,
ADNI pipeline is used. The following is done on the
MRI image as a part of this pipleline
(i) Post-Acquisition Correction: Scanners with
different acquisition parameters provide
considerable hurdles. Small changes in
acquisition parameters for quantitative
sequences have a significant impact on machine
learning models, thus rectifying these
inconsistencies is critical.
(ii) B1 Intensity Variation: B1 errors are one of the
problems in measuring MTR which expands to
magnetization transfer ratio since this MTR
value changes with change in the magnetization
transfer (MT) pulse amplitude. These errors can
also be caused due to nonuniformity in the
radiofrequency and incorrect transmitter output
settings when accounting for changing levels of
RF coil loading. These mistakes need to be
corrected in order to obtain images with no
variations and loss of crucial data.
(iii) Intensity Non-Uniformity: The quality of
acquired data can be affected by intensity non-
uniformity. The term ”intensity non-uniformity”
refers to anatomically unrelated intensity
variance in data. It can be caused by the radio-
frequency coil used, the acquisition pulse
sequence used, and the sample’s composition
and geometry. As a result, it is critical to correct
this variation, and a variety of approaches have
been offered to do so.
(II) Registration: The act of aligning images to be
analyzed is called registration of image, and it is a
critical phase in which data from several images
must be integrated everywhere. They can be taken
at various times, from various perspectives, and
with various sensors. Registration in medical
imaging allows you to merge data from multiple
modalities, such as CT, MR, SPECT, or PET, to get a
full picture of the patient. In our case, since the
MRI scans are taken from different angles, it is the
process of geometrically aligning all the images for
further analysis. It can be used to create
correspondence between features in a set of
images and then infer correspondence away from
those features using a transformation model.
documentclassarticle
(III) Segmentation and Normalization: The division of
brain tissue, which can be split into tissue sections
such as cerebrospinal fluid (CSF) which cushions
the brain, grey matter (GM) where the actual
processing is done and white matter (WM) which
gives communication between different GM areas,
is the major focus of the brain magnetic resonance
imaging (MRI) image segmentation approach (CSF).
Various significant brain regions that could be
useful in identifying Alzheimer’s disease are found
and kept during image segmentation.
Normalization is the process of shifting and scaling
an image so that the pixels have a zero mean and
unit variance. By removing scale invariance, the
model can converge faster.
(IV) Skull Stripping: Skull stripping is a process where in
the skull and the non-brain region of the image is
removed and only the brain portion of the image is
retained as we deal with only this region for
analysis of Alzheimer’s disease. Skull stripping is
one of the first steps in the process of diagnosing
brain disorders. In a brain MRI scan, it is the
method for differentiating brain tissue from non-
brain tissue. Even for experienced radiologists,
separating the brain from the skull is a time-
consuming process, with results that vary widely
from person to person. This is a pipeline that only
needs the input of a raw MRI picture and should
produce a segmented image of the brain after the
necessary preprocessing.
(V) Smoothing: Smoothing involves removing
redundant information and noise from the images.
It helps in easy identification of trends and
patterns in the images. When the image is
produced in an MRI machine, it consists of
different kinds of noise which needs to be removed
in order to obtain clean image without loss of any
crucial information.