Journal of University of Shanghai for Science and Technology ISSN: 1007-6735

A DEEP LEARNING-BASED APPROACH FOR

DETECTION OF DEMENTIA FROM BRAIN MRI
Reshma L1, Sai Priya Nalluri2, Priya R Sankpal3

Student, BNM Institute of Technology-Bengaluru, Karnataka, India

1

Student, BNM Institute of Technology-Bengaluru, Karnataka, India

2

Professor, BNM Institute of Technology-Bengaluru, Karnataka, India

3

1
[email protected], [email protected], [email protected]

Abstract:In this paper, a user-friendly system has been developed which will provide the result
of medical analysis of digital images like magnetization resonance of image scan of brain for
detection and classification of dementia. The small structural differences in the brain can slowly
and gradually become a major disease like dementia. Progression of dementia can be slowed
when identified early. Hence, this paper aims at developing a robust system for classification
and identifying dementia at the earliest. The method used in this paper for initial disclosure and
diagnosis of dementia is deep learning since it can give important results in a shorter period of
time. Deep Learning methods such as K-means clustering, Pattern Recognition and Multi-class
Support Vector Machine (SVM) have been used to classify different stages of dementia. The goal
of this study is to provide a user interface for deep learning-based dementia classification using
brain magnetic resonance imaging data. The results show that the created method has an
accuracy of 96% and may be utilized to detect people who have dementia or are in the early
stages of dementia.

Keywords: Dementia, MRI, K-means clustering, Pattern Recognition, Multi-class SVM.

1. INTRODUCTION

Dementia is a collection of conditions determined by deterioration of at least two brain functions,

such as memory loss and judgement. Dementia is characterized by a group of conditions and as
a result, it is not a particular illness. It is an all-inclusive term for a slew of symptoms linked to a
decrease in memory or other cognitive abilities severe enough to impair a person's ability to
carry out everyday tasks [1]. The various causes of dementia are as follows:

 Alzheimer’s Disease: 60-80% of the total cases.

 Lewy body Dementia: 5-10% of the total cases.
 Vascular Dementia: 5-10% of the total cases.
 Frontotemporal Dementia: 5-10% of the total cases.
 Others: Parkinson’s Disease, Huntington’s Disease account for <5% of the cases.

However, there are a variety of other disorders that might be mistaken for dementia symptoms,
including those that are curable but are often overlooked, such as vitamin insufficiency and
thyroid issues.
Dementia can be broadly classified into four stages:
 Mild cognitive impairment
 Mild dementia
 Moderate dementia
 Severe Dementia

Volume 23, Issue 7, July - 2021 Page -516

Journal of University of Shanghai for Science and Technology ISSN: 1007-6735

These stages are explained in Table 1.

Table 1. Characteristics and Symptoms of various stages of Dementia.

Sl. No. Stages of Dementia Characteristics Symptoms

1. Stage-I Mild cognitive Subtle decline in
impairment thinking
2. Stage-II Mild dementia Memory changes,
confusion
3. Stage-III Moderate dementia Inability to eat,
bathe or dress
without help
4. Stage-IV Severe Dementia Loss of ability to
communicate and
recognize loved
ones

These stages are detected using Magnetic Resonance Imaging (MRI) which is a diagnostic
imaging technique that presently offers the most accurate non-invasive way of imaging the
spinal cord, brain, and other areas of the human body. It is the most commonly used imaging
method to help confirm the diagnosis of dementia and to survey the course of the disease [3,4].
Visualization and understanding about the underline pathology of the disease has been made
easier due to MRI scans.
The number of individuals living with dementia is expected to rise quickly as average life
expectancy rises. By 2050 it is predicted that 131.5 million people will be affected by dementia
as shown in Figure 1 [4,5].
In light of this information, a strong system to track demented people and their illness
development is required, reducing the load on those liable for their care.

Figure 1. Estimation of the number of people that will be affected by Dementia.

Volume 23, Issue 7, July - 2021 Page -517

Journal of University of Shanghai for Science and Technology ISSN: 1007-6735

2. RELATED WORKS
For Alzheimer's disease detection, Shrikant Patro and Prof. Nisha V M employed Bicubic
Interpolation and Watershed Segmentation [6]. They suggested an image processing technique
for processing coronal, axial, and sagittal plane magnetic resonance imaging (MRI) of the brain.
In a brain MRI, image segmentation is utilized to identify the afflicted region. The regions of the
brain that have been identified using MRI include hippocampus and brain volume It is
determined who has Alzheimer's disease, who has mild cognitive impairment, and who belongs
to the healthy cohort.

L. Jaba Sheela and Dr. V. Shanthi later employed the CBA algorithm to achieve the same goal
[7]. Harlick proposes five textural characteristics based on the MRI images that define the
disease. An updated CBA technique is utilized to identify the pictures as Normal or Abnormal,
which is supported by the rule set established during the training phase. The tests were
conducted on the OASIS dataset, and the results demonstrate that the proposed system classifies
with 98% accuracy, showing the application of association rules in classification.

For the detection of Alzheimer's disease, Rigel Mahmood and Bishad Ghimire employed
supervised learning [8]. They developed a new approach for better classification of Alzheimer's
disease based on mathematical and image processing approaches. The most frequent method
these days is to analyse MRI scans using diffeomorphism principles to create a mapping from
one MRI to another.

Hiroki Fuse, Kota Oishi, Norihide Maikusa, and Tadanori Fukami investigated the application of
Shape Analysis of MRI Images using Contour extraction and Feature extraction by P-type
Fourier descriptor to identify Alzheimer's disease. The lateral ventricle, including the septum
lucidum, was examined using a P-type Fourier Descriptor as shape information in this paper.
They used a Support Vector Machine to do classification utilising a mixture of multiple
descriptors as features (SVM). The classification accuracy was 87.5%, which was superior than
the accuracy achieved using the volume ratio to intracranial volume (81.5%), which is
frequently used to evaluate morphological change [9].

P. Padilla, M. López, J. M. Górriz, J. Ramrez, D. Salas-González, and I. lvarez created an NMF-

SVM based CAD tool that used Intensity Normalization and Fisher Discriminant Ratio for
feature selection and achieved an accuracy of 85-91%. The functional brain scans are examined
and divided using a computer-aided scripting language. A single photon emission computed
tomography (SPECT) database and positron emission tomography (PET) images were used for
this purpose, both of which comprised data for both Alzheimer's Disease patients and healthy
subjects as a benchmark [10]. The Fisher discriminant ratio (FDR) and nonnegative matrix
factorization are used to evaluate these databases.

However, these papers failed to classify the various stages of dementia. In this paper, a deep
learning-based network is used to train the different types of dementia features along with
the label data set. This trained network is used to classify the test case images into different
stages and classes.

Volume 23, Issue 7, July - 2021 Page -518

Journal of University of Shanghai for Science and Technology ISSN: 1007-6735

3. METHODOLOGY
Multiple image processing techniques were used to pre-process the MRI images of the brain
such as contrast enhancement, filtering, segmentation, morphological processing etc. These
images are then fed into pattern recognition neural network where they are trained [11]. The
block diagram of the method used in the study is shown in Figure 2. These steps are explained in
the successive sections.
In this study, 104 T2-weighted MRI images of various stages of dementia in .jpg format have
been considered. The size of the images range between 250-800kB.

3.1. Image Acquisition

The database images used contain structural MRI images taken by a means of a medical camera.
The database images are collected from OASIS-2 and the size of the images is not standard.

Figure 2. Block Diagram of the proposed system.

3.2. Image Pre-processing

The use of a computer to manipulate digital images is known as digital processing. Because raw
data from imaging sensors contains faults, it must go through several stages of processing to
eliminate these flaws and ensure information originality. Contrast enhancement is a technique
for making visual characteristics stand out more clearly by making good use of the display's
colors or output device. Later noise is removed using median filter [11,12]. The steps involved
are shown in Figure 3.

Volume 23, Issue 7, July - 2021 Page -519

Journal of University of Shanghai for Science and Technology ISSN: 1007-6735

Figure 3. Pre-processing Block Diagram.

3.3. Segmentation

These images are segmented after they have been pre-processed. Segmentation is the process of
dividing an image into distinct zones using pixels with the same axis. To be relevant and useful
for image analysis and evaluation, the reasoning must have a strong connection to the exhibited
item or feature of interest. From low-level image processing to high-level image representation
in terms of features, objects, and scenes, meaningful segmentation is the first step in
transforming a grayscale image into one or more other images. The quality of segmentation
determines the effectiveness of image analysis, yet precise image partitioning is a tough
challenge to tackle [13]. Segmentation outcomes that are commonly obtained for an MRI image
are shown in Figure 4. In this study we have used K-means Clustering which will be discussed
in further sections.

Figure 4. Segmentation of an MRI.

3.4. Morphological Operations

Morphological operation aims at removing imperfections caused due to noise and texture caused
by distortion in the binary regions of the image. Hence a non-linear set of operations associated
to the morphology or shape of features is used [13]. Various types of morphological operations
are illustrated in Figure 5.

Volume 23, Issue 7, July - 2021 Page -520

Journal of University of Shanghai for Science and Technology ISSN: 1007-6735

Figure 5. Morphological operations.

3.5. Feature Extraction

Feature extraction is a form of dimensionality reduction that condenses a huge amount of raw
data into a smaller group that can be processed. The new set of features should then be able to
highlight the old set's array of data. In this approach, a condensed version of the original feature
can be made by combining the original set of images [14]. The common types of features
extracted are shown in Figure 6.

Figure 6. Various Features.

3.6. Pattern Recognition Neural Network

Data categorization and grouping are aided by neural networks. They help in grouping unlabeled
data based on similarities among the exampled inputs when they have a labelled data set to train
on, and they classify data when they have a labelled data set to train on. Deep neural networks
are parts of larger machine learning systems that include classification, reinforcement learning,
and regression techniques. Neural networks can extract characteristics that are then input into
clustering and classification algorithms. During this study we have used Pattern recognition
neural network. These networks are used for training of data based on knowledge already gained
from statistical information [15]. Their applications include speech recognition, speaker
identification and automatic medical diagnosis. The steps involved in mapping features using a
pattern recognition network is shown in Figure 7.

Volume 23, Issue 7, July - 2021 Page -521

Journal of University of Shanghai for Science and Technology ISSN: 1007-6735

Figure 7. Block diagram of pattern recognition neural network.

3.7. Multi-class SVM

The formal definition of a support vector machine (SVM) is a discriminative classifier with a
separating hyper plane. In other words, given labelled training data, the algorithm generates an
optimum hyper plane that categorizes new samples (supervised learning).
Multi class SVM works on the principal of one-to-rest. Here the classifier can use m-SVMs.
Each SVM would predict membership in one of the ‘m’ classes [16]. Classified results obtained
using an SVM is shown in Figure 8.

Figure 8. Classification using SVM.

4. EXPERIMENTAL RESULTS

The implemented GUI contains a total of 9 push buttons each corresponding to the set to be
executed. MATLAB's graphical user interface design environment (guide) is used to create the
GUI.

4.1. Push button 1: Load Image

When the load image button is clicked, MATLAB will direct us to a folder in our local system
where the data set for our study is stored. The selected image is then uploaded in the GUI.

4.2. Push button 2: Contrast enhancement

Before performing contrast enhancement, the chosen image is first converted into a gray scale
image. Then, contrast enhancement is performed using adaptive histogram equalization as it
provides better clarity. This method works on small areas of an image rather than the entire
image. Contrast of each pixel is enhanced hence the features of the entire image stands out.

Volume 23, Issue 7, July - 2021 Page -522

Journal of University of Shanghai for Science and Technology ISSN: 1007-6735

4.3. Push button 3: Filtering

Processes implemented along with filtering include smoothing, sharping, and edge enhancement.
In this study median filtering has been performed using medfilt2 which is a 2-D filter.

4.4. Push button 4: Segmentation

Initially a colorspace function is used to convert a color image to LAB colorspace. Lab
colorspace stands for L for lightness and a and b for the color dimension. It is a colorspace with
3-axis color system. The obtained output is segmented using K-means clustering algorithm.

K-means clustering
Clustering is one of the most common exploratory data analysis technique used to get an
intuition about the structure of the data. It can be defined as the task of identifying
subgroups in the data such that data points in the same subgroup (cluster) are very similar
while data points in different cluster are very different. The decision of the similarity
measure used is application-specific. K-means is one of the simplest unsupervised learning
algorithms that solves the clustering problem [13]. The procedure follows a simple and easy
way to classify a given data set through a certain number of clusters (k) fixed apriori.

Algorithm of K-means Clustering

 Determine the values of ‘k’.
 Initialize the centroid.
 Measure the distance.
 Assign to the nearest cluster. If there is no convergence go back to step 3
 If there is convergence measure the variance. Repeat step 2-5 until you get the lowest
sum of variance.
 Clustering results are obtained.
Clustering results obtained after segmentation are shown in Figure 9.

Figure 9. Segmentation using K-means algorithm.

4.5. Push button 5: Morphological operation

Firstly, light structures which are connected to the image border are suppressed. Dilation is
performed next. It adds pixels to the boundary of objects in an image.

Volume 23, Issue 7, July - 2021 Page -523

Journal of University of Shanghai for Science and Technology ISSN: 1007-6735

4.6. Push button 6: Feature Extraction

Firstly, a GLCM matrix is created. Then the following properties are obtained.
 Contrast: the intensity contrast between a pixel and its neighbor over the entire image.
 Correlation: It is the statistical measure of how closely a pixel is connected to its
neighbors throughout the full image.
 Energy: summation of squared elements in the GLCM.
 Homogeneity: refers to how near the distribution of elements in the GLCM is to the
diagonal of the GLCM.
 Entropy: Entropy is a randomness measure that takes low values for smooth images.
 Mean: It is the average value.
 Standard Deviation: This is extracted using std2 i.e. B = std2(A) computes the standard
deviation of the values in A.
 RMS: It stands for Root mean squared value. For vectors, rms(X) is the root mean
squared value in X.
 Variance: It is extracted using var. For vectors, Y = var(X) returns the variance of the
values in X.
 Kurtosis: K = kurtosis(X) returns the sample kurtosis of the values in X.
 Skewness: S = skewness(X) returns the sample skewness of the values in X.
 IDM: It stands for Inverse Difference Movement. Inverse difference moment is the
measure of local homogeneity.
The extracted features are stored in a .mat file for further steps as shown in Figure 10.

Figure 10. Feature extraction table.

4.7. Push button 7: Training

In the recent years, Magnetic Resonance Texture Analysis has been picking up pace to
determine the level of damage in neuroimaging data based on GLCM and histogram features.
This is where pattern recognition network comes into the picture.
The extracted GLCM features from the previous step are trained for classification using a simple
feed-forward pattern recognition neural network [16]. The following steps are followed in
training this model:
 Create a Pattern Recognition Network using MATLAB in-built function patternet [16].
 The hidden layer is taken as 10.
 The next process is to set up data division for testing, validation, and training.
 Train the network: This is accomplished using the [NET,TR] = train(NET,X,T) function,
which takes a network NET, input data X, and target data T and returns the network as
well as a training record TR after training. This is shown in the Figure 11.
 The network is then tested using perform(NET,T,Y,EW), which receives a network,
targets T and outputs Y, and optionally error weights EW, and returns performance
using the network's default performance function. This is shown in the Figure 12.
 Finally, the accuracy of the model is calculated.

Volume 23, Issue 7, July - 2021 Page -524

Journal of University of Shanghai for Science and Technology ISSN: 1007-6735

Figure 11. Pattern recognition network.

Figure 12. Neural network training.

4.8. Push button 8: Classification

In this study we want to classify the level of dementia into four classes i.e. Stage-I (Mild
cognitive impairment), Stage-II (mild dementia), Stage-III (moderate dementia) and Stage-IV
(severe dementia). For this to happen we first need to teach the classifier on what basis the levels
are supposed to be classified. The image is classified using the SVM after comparison of
features obtained from the selected image and the pre extracted features which are labeled using
pattern recognition network [17]. In more technical terms, an SVM produces a hyperplane or a
family of hyperplanes for classification in a high or infinite-dimensional space.

Volume 23, Issue 7, July - 2021 Page -525

Journal of University of Shanghai for Science and Technology ISSN: 1007-6735

Steps: SVM
 Import the data set.
 Explore the data and pre-process them.
 Split the data into attributes and label.
 Train the SVM algorithm and make predictions.
 Evaluate the results.

The classification of various stages of dementia is shown in Figure 13.

Figure 13. Classification using SVM.

The final outline of GUI is shown in Figure 14. The obtained results are shown in Table 2.

Figure 14. Outline of the GUI.

Volume 23, Issue 7, July - 2021 Page -526

Journal of University of Shanghai for Science and Technology ISSN: 1007-6735

Accuracy is calculated using the formula,

(TP+TN)
Accuracy = (TP+TN+FP+FN) (1)

where TP = True Positives, TN = True Negatives, FP = False Positives, and FN = False

Negatives. These values are obtained from the confusion matrix.

Table 2. Obtained accuracy values

PercentErrors Accuracy
0.1 96.07%
0.2 88.23%
0.3 86.27%

5. CONCLUSION
There is no way to tell if someone has dementia without doing a test. A thorough medical
history, physical examination, lab examination, and the typical changes in thought and behavior
associated with each type of dementia help doctors diagnose dementia and its type. Doctors can
detect whether or not a person has dementia with a high degree of confidence, but determining
the kind of dementia is more difficult since the brain alterations of distinct dementias might
overlap. Hence in this paper Magnetic resonance texture analysis has been used to differentiate
between various stages. The pre-extracted features are trained and then assigned to each class
using a pattern recognition neural network and based on this the chosen image is classified into a
particular stage using multi class SVM. The model accuracy was found to be about 96%.
Majority of the papers which were referred during the study used supervised learning methods
such neural networks (CNN, ANN). However, in the study Multi-Class SVM has been used to
classify stages of dementia.
To determine whether we can figure out which portions of the brain are damaged by dementia,
future enhancements to this research will look at particular regions of the MR image rather than
the full image. Accuracy can be improved further by including hybrid optimization techniques
with the present mechanisms. Further investigations could also include optimization of
classifiers. Investigating the efficiency of the proposed method for a larger data set could be an
area of research potential.

Acknowledgments
We express our heartfelt gratitude to the college management of BNM Institute of Technology,
Bengaluru for enabling us to carry out the research work with all the resources and subsequent
publication. Also, we express our sincere gratitude to Visvesvaraya Technological Institute
(VTU), Belagavi for providing us the research platform.

REFERENCES
[1] H. M. Tarek Ullah, Zishan Ahmed Onik, Riashat Islam, and Dip Nandi, “Alzheimer's
Disease and Dementia Detection from 3D Brain MRI Data Using Deep Convolutional Neural
Networks,” 2018 3rd International Conference for Convergence in Technology (I2CT), Pune,
India, doi: 10.1109/I2CT.2018.8529808.

[2] Doaa Ebrahim, Amr M. T. Ali-Eldin, Hossam E. Moustafa, and Hesham Arafat, “Alzheimer
Disease Early Detection Using Convolutional Neural Networks,” 2020 15th International
Conference on Computer Engineering and Systems (ICCES), Cairo, Egypt, doi:
10.1109/ICCES51560.2020.9334594.

Volume 23, Issue 7, July - 2021 Page -527

Journal of University of Shanghai for Science and Technology ISSN: 1007-6735

[3] J A Akhila, Christin Markose and R P Aneesh, “Feature extraction and classification of
Dementia with neural network”, 2017 International Conference on Intelligent Computing,
Instrumentation and Control Technologies (ICICICT), Kerala, India, doi:
10.1109/ICICICT1.2017.8342782.

[4] Maxim Mizotin, Jenny Benois-Pineau, Michèle Allard, and Gwenaelle Catheline, “Feature-
based brain MRI retrieval for Alzheimer disease diagnosis”, 2012 19th IEEE International
Conference on Image Processing, Orlando, FL, USA, doi: 10.1109/ICIP.2012.6467091.

[5] Luis Javier Herrera, Ignacio Rojas, H. Pomares, A. Guillén, O. Valenzuela and O. Baños,
“Classification of MRI Images for Alzheimer's Disease Detection”,2013 International
Conference on Social Computing, Alexandria, VA, USA, doi:10.1109/SocialCom.2013.127.

[6] Shrikant Patro, Prof. Nisha V M, 2019, Early Detection of Alzheimer’s Disease using Image
Processing, International Journal of Engineering Research & Technology (IJERT) Volume 08,
Issue 05 (May 2019).

[7] L. J. Sheela and V. Shanthi, "An Image Analysis and Classification Protocol for
Characterization of Normal and Abnormal Memory Loss in Aging from Structural MRI," 2009
International Conference on Signal Processing Systems, Singapore, 2009, pp. 639- 643, doi:
10.1109/ICSPS.2009.199.

[8] R. Mahmood and B. Ghimire, "Automatic detection and classification of Alzheimer's

Disease from MRI scans using principal component analysis and artificial neural networks,"
2013 20th International Conference on Systems, Signals and Image Processing (IWSSIP),
Bucharest,2013, pp. 133-137, doi: 10.1109/IWSSIP.2013.6623471.

[9] Hiroki Fuse, Kota Oishi, Norihide Maikusa, and Tadanori Fukami, “Detection of
Alzheimer’s Disease with Shape Analysis of MRI Images”, 2018 Joint 10th International
Conference on Soft Computing and Intelligent Systems and 19th International Symposium on
Advanced Intelligent Systems, pp. 978-1-5386-2633-7, doi: 10.1109/SCISISIS.2018.00171.

[10]P. Padilla, M. Lopez, J. M. Gorriz, J. Ramirez, D. Salas-Gonzalez, and I. Alvarez,

“NMFSVM Based CAD Tool Applied to Functional Brain Images for the Diagnosis of
Alzheimer’s Disease”, 2012 IEEE Transactions on Medical Imaging, pp. 0278-0062.

[11] A. W. Salehi, P. Baglat, B. B. Sharma, G. Gupta and A. Upadhya, “A CNN Model: Earlier
Diagnosis and Classification of Alzheimer Disease using MRI”, 2020 International Conference
on Smart Electronics and Communication (ICOSEC), Trichy, India, 2020, pp. 156-161, doi:
10.1109/ICOSEC49089.2020.9215402.

[12] Sagar Adatrao and Mayank Mittal, “An analysis of different image preprocessing
techniques for determining the centroids of circular marks using hough transform”, 2016 2nd
International Conference on Frontiers of Signal Processing (ICFSP), Warsaw, Poland,
doi:10.1109/ICFSP.2016.7802966.

[13] Xin Zheng, Qinyi Lei, Run Yao, Yifei Gong & Qian Yin,” Image segmentation based on
adaptive K-means algorithm”, EURASIP Journal on Image and Video Processing volume 2018,
doi: 10.1186/s13640-018-0309-3.

[14] P. Mohanaiah, P. Sathyanarayana, and L. GuruKumar, “Image Texture Feature

Extraction Using GLCM Approach”, 2013 International Journal of Scientific and Research
Publications, Volume 3, Issue 5, ISSN 2250-3153.

[15] M. Yu. Mikheev, Yu. S. Gusynina, and T. A. Shornikova, “Building Neural Network for
Pattern Recognition”, 2020 International Russian Automation Conference, Sochi, Russia, doi:
10.1109/RusAutoCon49822.2020.9208207.

Volume 23, Issue 7, July - 2021 Page -528

Journal of University of Shanghai for Science and Technology ISSN: 1007-6735

[16] Sai Yeshwanth Chaganti, Ipseeta Nanda, Koteswara Rao Pandi, Tavva G.N.R.S.N.
Prudhvith, and Niraj Kumar, “Image Classification using SVM and CNN”, 2020 International
Conference on Computer Science, Engineering and Applications (ICCSEA), Gunupur, India, doi:
10.1109/ICCSEA49143.2020.9132851.

[17] Guang-Bin Huang; Hongming Zhou; Xiaojian Ding; Rui Zhang, “Extreme Learning
Machine for Regression and Multiclass Classification”, IEEE Transactions on Systems, Man,
and Cybernetics, Part B (Cybernetics), doi:10.1109/TSMCB.2011.2168604

Volume 23, Issue 7, July - 2021 Page -529