Artificial Intelligence-Based Medical Image Classification using a
Multilayer Fuzzy Approach
KISHORE KUMAR AKULA1, ALEXANDER GEGOV2,3, FARZAD ARABIKHAN2
1Statistics eTeachers Group,
Royal Statistical Society
2110, 100 Leeward Glenway, Toronto, Ontario, M3C 2Z1,
CANADA
2School of Computing,
University of Portsmouth,
Winston Churchill Ave, South Sea, Portsmouth PO1 2UP Portsmouth,
UNITED KINGDOM
3English Faculty of Engineering,
Technical University of Sofia,
Sofia 1756,
BULGARIA
Abstract: - A Convolutional Neural Network (CNN) is an effective Artificial Intelligence (AI) technique for the
automation of image analysis. However, to achieve a high level of accuracy, a CNN usually requires a large
amount of data and a long training time. The current study addresses the above problem by proposing a novel
AI technique. The latter can detect and classify abnormalities in images using a small amount of available data
and a short training time. The proposed technique, Artificial Intelligence Based Medical Image Classification
Using a Multilayer Fuzzy Approach (MFA), was validated using open access medical image data, where an
image with a particular type of abnormal object contained in it was compared with a normal image with the
same object in it. The similarity was then computed in percentages and subtracted from the hundred, which is
the abnormality in the first image. The results showed that the novel MFA outperforms significantly better than
the benchmark, CNN, and is a useful tool for automated analysis of medical image data sets.
Key-Words: - Multilayer fuzzy, convolutional neural network, image, training, classification, CT scan.
Received: July 2, 2022. Revised: August 19, 2023. Accepted: September 22, 2023. Published: October 25, 2023.
1 Introduction
To find abnormalities within objects in images,
normally the experts in a particular field visually
examine them. Medical diagnosis often requires
visual examination of diagnostic images to find any
abnormalities within the images. For instance,
doctors examine CT scans to determine if cancer
exists or not. Automation of the image analysis
process would greatly cut down the time required to
arrive at a diagnosis. Several methods have been
developed since the rise of the field of Artificial
Intelligence (AI) that rely on concepts from
cognitive science.
1.1 Cognitive Science
Cognitive science is the study of the mind and its
processes, such as learning, which contains sub-
domains like knowledge, understanding, analysis,
synthesis, and evaluation. The sub-domain of
evaluation includes the process of comparing, [1].
Comparing two images to find their similarity is the
key approach of this study. For example, when two
objects in a CT scan like two images of lungs are
compared, then the similarity or dissimilarity can be
determined. When one normal image of the lung is
compared with an abnormal lung image, then the
difference is the abnormality. However, many
precautions must be taken to ensure accurate
comparisons and to extract the true abnormality,
which will be discussed in the forthcoming sections.
1.2 Convolutional Neural Networks
The contemporary method used to find
abnormalities in the objects of images utilizes one of
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.24
Kishore Kumar Akula, Alexander Gegov,
Farzad Arabikhan
E-ISSN: 2224-2872
206
Volume 22, 2023
the architectures of deep learning, which is a
convolution neural network (CNN). CNN is used as
a method for automatic, accurate, and reliable image
analysis, [2].
The CNN finds patterns in images and then helps
classify the images. In most of the literature where
CNN has been used, abnormalities in objects within
images cannot be found directly, but the model has
to be trained with large data sets containing images
in which the abnormality has already been identified
and classified by humans. However, to get
categorized data to train and test is a laborious
process. Furthermore, if the data and categories are
changed, then the software code has to be changed
again.
The data size required for CNN is typically large.
However, literature has shown that there are CNN
techniques capable of utilizing small datasets, as
demonstrated in, [3], [4], [5]. Nevertheless, it should
be noted that even these small datasets often consist
of hundreds or thousands of samples. In contrast, the
MFA approach utilizes a mere 21 images for all
steps, including the training, testing, validation,
detection, and classification of abnormalities.
For example, if cancer needs to be identified and
classified into one of several types of tumors using
CT scan images, a CNN does not directly identify
the cancer in the images to classify the cancer types.
To do so, the CNN has to be trained with data in
which cancer has already been identified, as well as
images from healthy individuals without cancer.
This research proposes an AI-based method as an
alternative that can bypass some of these challenges
by using a Multilayer Fuzzy Approach.
1.3 Computer Vision
Computer vision is an interdisciplinary science to
gain micro and macroscopic understanding of an
image. It provides a straightforward and meaningful
description of real-world objects from images and
involves digital images or medical scans. Computer
vision deals with images in terms of pixels, and its
features include image processing, feature
extraction, and cropping images. Additionally,
computer vision techniques like structural similarity
(SS) help with comparing two similar images like
CT scans of lungs from two patients. Using this
technique, it is possible to find the similarity or
differences among the images in terms of a
percentage or a fraction.
1.3.1 Image Comparison
An image has a fuzzy nature since it is a two-
dimensional projection of a three-dimensional real
object, [6]. Images are used in many scientific fields
like geography, medicine, social sciences, and
psychometrics. While it might be used to find water
resources in the field of geography, it can be used to
analyze CT scans, X-rays, and MRI scans, to study a
pathological problem, a cancer tumor, or an
infection in the field of medicine.
Images for medical sciences are typically
standardized with a specific setup, angle, and other
technical arrangements. For instance, a CT scan of
the chest to detect a lung cancer tumor is acquired
using the same settings. All images are taken from
approximately the same side of the object in the
image, like for instance, the front view of lungs. In
addition, images of the same kind of object are
acquired, such as the lungs or the brain. All images
are also taken with the same lighting and intensity,
covering approximately the same area. Furthermore,
the images can be microscopic or macroscopic but
not mixed.
1.3.2 Pixels
Each image is represented by units known as pixels.
Additionally, an image can be represented as an
image function, which, in turn, is the mathematical
representation of an image. The image function is a
vector-valued function consisting of a small number
of arguments. The digital image is a special case of
an image function, [6]. Once an image is
represented as an image function, many of the
statistical and mathematical concepts can be
applied. These pixels are also considered to be fuzzy
numbers.
1.3.3 Grey Scale Image
Generating a grey-scale image is a process in which
each pixel of the image is represented by only the
intensity information of the light. The colors of the
image are black, white, and grey.
1.3.4 Structural Similarity Index (SSI)
The SSI is obtained when two similar images are
compared, and this parameter reveals the percentage
similarity between the two images, [7]. When more
than 2 similar images are compared, then we get a
structural similarity score. Several software like
MATLAB, Python, and CRAN-R can be used to
determine the SSI. The software will divide the
parts of the images into sub-frames and compare
them to find the similarity between one image and
the next. Python was used in this study.
1.4 Fuzziness
In many real-world situations, information is not
always clear-cut or binary. Instead, it allows for the
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.24
Kishore Kumar Akula, Alexander Gegov,
Farzad Arabikhan
E-ISSN: 2224-2872
207
Volume 22, 2023
representation and manipulation of vague or
ambiguous data.
1.4.1 Fuzzy Set
The fuzzy set that will be used in this study is
'{ID, Similarity percent},' where ID is the
identification of the image and similarity percent is
the similarity obtained when two images are
compared. Along with fuzzy sets, both cognitive
science and computer vision can be used to form an
AI technique that can be used to detect an
abnormality in the objects of an image by
comparing images with normal objects in it and a
similar image with abnormal objects in it.
The images must be acquired using a uniform
set of specifications mentioned in the following
sections, and the images can be from any of the
sciences or social sciences. Using this technique,
abnormalities in a massive number of images can be
found. Furthermore, there is no need for data with
identified and categorized abnormalities to train the
process.
1.4.2 Multilayer Fuzzy Notion
The multilayer fuzzy system is based on some of the
theoretical concepts from, [8], [9]. In particular, two
layers are used which implement each of the two
consecutive stages in image manipulation described
below. The notion of a multilayer fuzzy set is being
introduced in this study. An image is a two-
dimensional projection of a three-dimensional
object. A three-dimensional object has length,
width, and thickness or height, whereas a two-
dimensional object only has length and width. When
an image is acquired, whether it is a CT scan or a
normal photograph, the three-dimensional nature
will be converted to a two-dimensional nature. That
is, the projections assumed from the edges of a
three-dimensional object will create a two-
dimensional object. This creates a fuzzy notion in an
image in computer vision, [6]. In addition to this
fuzzy quality, a second fuzzy quality can be
attributed to the pixel, which is the conversion of a
continuous picture into a digital format, which
involves fuzziness. Thus, digital images consist of
pixels that are viewed as fuzzy numbers.
Additionally, to run some computer vision
commands, the conversion of a color image to a
grey scale is needed, and this process involves
fuzziness. Consequently, when a study deals with
images, these images can be described by multilayer
fuzziness, forming a multilayer fuzzy input, which
will herein be referred to as the multilayer fuzzy
notion.
1.4.3 Multilayer Fuzzy Set
As mentioned above, digital images fall under the
multilayer fuzzy notion. Additionally, when two
images are compared, the SSI will be obtained,
which also possesses a multilayer fuzzy nature. In
this study, the SSI will be equal to the membership
value and the multilayer fuzzy set will be of the
form, '{Identification of the image, SSI}.' For
example, the multilayer fuzzy set when one CT scan
of the lungs of a normal or healthy individual is
compared with a CT scan of the lungs of a patient
with some disease can be, '{patient ID = 111, SSI =
0.089}.'
1.5 Rationale for the study
Many hours are spent by medical personnel on
disease diagnoses that could be spent on other
patient care activities. The use of cognitive science
concepts can be applied to aid in such diagnoses
while also improving the test-retest reliability.
However, the caveat is that existing methods, such
as CNNs, require a large data set and a separate data
set that has already been categorized as having the
abnormality being studied, which makes this
procedure less effective for analyzing diagnostic
images of new or rare diseases that have less data.
Furthermore, even with established technology, it
may be difficult to obtain a data set large enough to
train the program. Fuzzy systems can bypass some
of these issues since they do not require a training
data set, and the minimum number of images
required to make a comparison is two, making it
easier to make comparisons with a limited number
of images.
As an application of the method developed in the
current MFA study, the medical field was chosen.
Specifically, the application will focus on the
identification of COVID-19 infection and
categorization based on the spread of the disease as
seen on the CT scan. Using the MFA method, a
normal image is sufficient to compare other images
to detect whether abnormalities exist. Moreover, this
is the first time that abnormalities will be found by
comparing an image with ‘good’ objects, such as
healthy tissue, and images with objects with
abnormalities, such as tissue infected with a virus.
2 Problem Formulation
2.1 Primary Aim
The primary aim of this research is to develop an
AI-based MFA using computer vision and a small
data set that can detect and classify accurately
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.24
Kishore Kumar Akula, Alexander Gegov,
Farzad Arabikhan
E-ISSN: 2224-2872
208
Volume 22, 2023
abnormalities in medical images within a fairly
short training time.
2.2 Secondary Aim
The secondary aim is to apply MFA to a small data
set of 22 images of healthy and COVID-19-positive
lung CT scans, [10].
3 Problem Solution
3.1 Approach for Primary Aim
An image with a normal object in it and several
other images with the same type of object, but with
abnormalities, were considered. The normal image
was compared with the abnormal images one by
one, and each time, the structural similarity index
(SSI), [8], score was calculated. In this case, the SSI
score is a percentage-based fraction representing the
level of abnormality of the objects in the images.
After finding the SSI for the images, several
techniques, including a multilayer fuzzy system,
other computational intelligence methods, and
software code testing tools, were used to detect and
classify the images.
3.2 Approach for Secondary Aim
MFA was used on a data set of CT scans of normal
right lung images andCOVID–19–confirmed
images.
3.3 Images
MFA can be implemented for any image, such as,
for example, pathological images, a geographic
region, or an astronomical image, as long as they are
acquired from the same angle and the same frame
dimensions. However, in MFA, the focus will be on
medical images for the application part of the
primary aim.
3.4 The Data Set
The data set for the application of the study method,
MFA, was acquired from the National Cancer
Institute, [10], and the data sets were downloaded by
checking the option titled, ‘COVID.’ The images
were a mix of COVID-19-positive, as well as
normal images, the latter of which were CT scans
without any abnormalities. A total of 21 images
were used, of which normal and COVID-19-positive
images were present. Among these 21 images, a few
CT scans had high contrast. Out of all the images,
one normal CT scan was taken as the standard
image. As there is no training or testing process, the
number of images with and without COVID-19 is
independent of finding the abnormality in the
images. The minimum number of images required to
test abnormality using the MFA method is two: one
normal image which is the standard image and one
image being tested for an abnormality. To classify
images, any number of images can be considered
other than the normal image, without an upper or
lower limit for the number of images.
3.4.1 The Prediction Data Set
The prediction data set consisted of 21 randomly
selected CT scans, [10], that were acquired to
investigate COVID-19.
3.4.2 Standard Image
A standard or normal image is a reference image
with which the other images under study will be
compared. Both the standard image and other
images under study are of the same file type (e.g.,
.dcm, .jpg), and have the same contrast, brightness,
and cropping as the image in Figure 1(1.1). Even
with different brightness or contrast (like the images
in Figure 1(1.1) and Figure 1(1.2) the MFA method
works, but it takes a greater amount of time to find
the similarity values.
3.4.3 The Images under Study
The standard image and the images being studied
will be of the same kind. This is because to study
diagnostic problems in the lungs, the standard
image, as well as the images under study, should be
images of lungs acquired with the same
specifications, such as medical CT scans. Some
examples of images with different contrast levels
that are cropped at different places are shown in
Figure 1. Images like this with different amounts of
space in the background or images that are cropped
at different levels will yield more biased
comparisons.
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.24
Kishore Kumar Akula, Alexander Gegov,
Farzad Arabikhan
E-ISSN: 2224-2872
209
Volume 22, 2023
Fig. 1: Human lung CT scans with various
diagnoses:
1.1 - A normal lung CT scan acquired at normal
contrast and cropped at the lungs,
1.2 - a CT scan of COVID-19-positive lungs,
cropped at the lungs,
1.3 - Normal CT scan with high contrast and not
cropped at the lungs,
1.4 - Low contrast, normal CT scan for lungs –
uncropped at the lungs.
3.4.4 Sample for Application
To study an application of the MFA method, the
method was applied to a small data set of CT scans
of adult male and female lungs. The data considered
were of a mixed sample with confirmed COVID-19
cases and normal CT scans without COVID-19.
3.5 Usage of Computer Vision
3.5.1 Grayscale of the Image
To get a good estimation of the SSI, all the images
were converted to greyscale. This allows for the
focus of the image to be on the intensity of light
rather than the color of the image. To accomplish
this, all natural colors of the image were converted
to a shade between black to bright white.
3.5.2 Similarity between Two Images
When two similar images are compared, like CT
scans of the lungs of two humans, the images will
be similar by a certain percentage, which is denoted
by k%.
Fig. 2: Cropped images of CT scans so that only the
left or right lung is included.
2.1 - A normal left lung,
2.2 - an infected left lung with COVID-19,
2.3 - a normal right lung,
2.4 - an infected right lung.
3.5.3 Noise
In the context of MFA, the noise is considered
something that affects the similarity between two
images when compared. That is, some regions of an
image may contain too much contrast or other parts
of the objects that we do not want to consider. For
instance, if the goal is to only study the CT scans of
lungs from two individuals, then the presence of
bone in the image will influence the similarity score.
In addition, when two images are compared there
might be similarity due to grey pixels and
dissimilarity due to the presence of unwanted
objects in the background. This noise cannot be
completely avoided, but it can be minimized if the
object of focus in the image is cropped and
extracted before comparison. For instance, in the
application part of MFA, instead of taking an image
consisting of both the left and right lung of a human
and comparing it with another image of both lungs
from other humans, the left lung and right lungs will
be cropped and compared separately with the
corresponding left or right lung images of other
individuals. Panels a and c in Figure 2 show
examples of the normal right and left lungs,
respectively. Images b and d are the right and left
lungs infected by COVID-19.
3.5.4 Structural Similarity Index (SSI)
The main technique used in the study is SSI from
computer vision. This technique is used to compare
two similar structures or objects present in an image
and find the percentage similarity of the first image
to the other. The SSI can be found using the
following formula, [7]:
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.24
Kishore Kumar Akula, Alexander Gegov,
Farzad Arabikhan
E-ISSN: 2224-2872
210
Volume 22, 2023
Here, x and y are two image patches of the same
spatial location of two images being compared and
whose SSI is to be calculated. µx is the mean of x, µy
= mean of y, σx2 and σy2 are mean and variances of
y, and σxy2 is the covariance of x and y. Each part of
the first image will be compared to the
corresponding part of the other image, provided that
the images are of the same dimensions and the
location of the object in the two images is
approximately the same.
The SSI between two images works in such a
way that all parts of one image (x) will be compared
with the corresponding parts of a similar image (y),
pixel by pixel. When two images are compared, the
minimum SSI is 0% if the images are dissimilar, and
it is 100% if the image is compared to itself.
However, under no circumstances will the SSI be
equal to 0. The reason for this is that shapes, as well
as white, grey, and black pixels, are present in any
image. Consequently, most images will have an SSI
that is a real number ranging from 0 and 1.
3.5.5 Acquiring the SSI between a Single
Standard, Normal Image Versus Multiple Images
with Abnormalities
The method formed to get SSI is shown in Figure 3.
The standard image will be compared to images
with abnormalities but are the same type of image
and the same spatial location of the same image.
With this method, the similarity between two
images will be obtained as a percentage. One
characteristic of the SSI to be noted is that no two
images of the same kind are 100% alike, because of
the difference in the objects in the images. For
example, even if two healthy lungs are compared,
the SSI will not be 100%. The reason could be the
size of the healthy lungs, the unmatched color of
pixels of the object or background, or the difference
in the ages of the persons, among other factors.
Fig. 3: A schema of how normal or standard images
are compared one by one with images containing
abnormalities, such as a tumor or infection.
The general instance of this relationship is as
follows:
If SSI = 0%, then the abnormality among the images
=100% approximately
If SSI = k1%, then the abnormality among the
images = (100% - k1%) approximately
If SSI = k2%, then the abnormality among the
images = (100% - k2%) approximately
If SSI = k3%, then the abnormality among the
images = (100% - k3%) approximately
If SSI = kn%, then the abnormality among the
images = (100% - kn%) approximately, where k is
an unknown arbitrary constant. To find the value of
k, the rules based on intelligent systems were used.
3.6 The Multilayer Fuzzy Input
As mentioned previously, the first type of fuzziness
in the image is due to the conversion of a three-
dimensional real object to a two-dimensional image.
In addition, the second type of fuzziness results
from the formation of pixels in the image. Together,
the multilayer fuzzy input (Figure 4) consists of the
images to be studied, such as the CT scans shown in
Figure 1 and Figure 2.
Fig. 4: Flow of the algorithm used in MFA.
3.6.1 The Multilayer Fuzzy Set (DFS)
The multilayer fuzzy set of abnormalities in the
images is given by ‘DFS = {ID, µ(ID)},’ where ID
is the identity of the image or patient, and µ = SSI,
is the membership input. For instance, in the
application part of MFA, the multilayer fuzzy set is
‘{patients ID, {2.3,0.01, 1, …..}}.’
3.7 Defuzzification, Applying Fuzzy Logic,
Intelligence Rules, and Classification of
Images based on the Intensity of
Abnormality in the Images with Respect to
SSI
The abnormalities in the objects of the images are
represented by the SSI score given by the
membership function of DFS, which is a crucial step
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.24
Kishore Kumar Akula, Alexander Gegov,
Farzad Arabikhan
E-ISSN: 2224-2872
211
Volume 22, 2023
in the current MFA project. After testing the code
several times with the data, the numerical values of
k will be decided by checking the images and the
SSI score. As a result, the images will be classified
based on the intensity of similarity using fuzzy logic
in this MFA study as follows.
If SSI ≤ or ≥ k1%, if ‘Yes’ then the abnormality
within the image is ‘severe’, or if ‘No’:
If SSI ≤ or ≥ k2%, if ‘Yes’ then the abnormality
within images is ‘moderate’, or if ‘No’:
If SSI ≤ or ≥ kn%, if ‘Yes’ then the abnormality
within images is ‘none’.
This logic scale can be chosen after testing the
software code with the data. After applying the
above steps, the data will be automatically classified
according to the severity of the abnormality.
3.7.1 Numerical Value of k
MFA will be applied to the data set of CT scans of
lungs mentioned above, to investigate infection with
COVID-19. As part of that, when the normal right
lung (Figure 2(2.1)) was compared with an
abnormal image of the right lung (for example,
Figure 2(2.2)), then a numerical value for k in the
above schema will be obtained.
3.7.2 Rules based on Intelligent Systems and
Computer Vision that Help to Find the k Values
The rules required to help find k values are the
most crucial stage of the MFA study and are given
below.
a. Manual software testing (MST). MST is a
process in which the software code or package
written will be tested to determine whether it is
correctly fulfilling its tasks, and specifications and
acquiring accurate results. To do so, a clear testing
strategy will be created. Subsequently, the code will
be executed several times, and the accuracy of the
results will be tested. If the accuracy is not
satisfactory, then the code will be corrected and run
repeatedly until the correct results are obtained.
In MFA, MST helps find the bounds of the SSI for
classification of the severity of the disorder in the
objects of the images. The main rule to find k values
involves testing the code and adjusting the k values.
Here the value of k is equal to the SSI. For instance,
if SSI is 10%, then the abnormality in the objects of
the image is severe. In addition, if the SSI is greater
than 10% and less than 45%, then the objects in the
image will be normal. Each time the crucified and
classified images are examined, the k value will be
adjusted accordingly.
b. Outliers. Sometimes the software shows greater
similarity or dissimilarity between images than is
the case. This is due to the presence of grey/black
pixels in both images or more contrast or less
brightness in the images. The outliers cause false
similarity or false dissimilarity when two images are
compared.
c. The characteristics of the image. Knowing
about the characteristics of the images being studied
will increase the accuracy of the k values. For
example, since the images considered are CT scans
of the lungs, information about how the scans were
taken, like the angle, contrast, and brightness, is
important. Furthermore, using the same example of
CT scans, the heart is anatomically located
approximately between the lungs. If an observer
who is not a medical professional were to examine
the images, the heart might seem like a tumor or an
infection due to COVID-19.
d. Precautions to be taken to find k values.
i. The size of all images must be the same.
ii. To improve the accuracy of the SSI score, the
object under study in the images has to be cropped
at the same point as shown in Figure 1(1.2) and
Figure 1(1.4), so that only the lung tissue will be
compared to lung tissue and not blank space within
the image.
iii. The accuracy of the results can be improved by
using images with the same contrast and brightness.
iv. It is better to crop the unwanted part of the image
and keep only the required object.
3.7.3 Defuzzification and Multilayer Fuzzy
Output
The images were gathered in terms of SSI based on
the SSI fuzzy output. The images were classified
using their SSI score, and the classified images were
in the form of the fuzzy score, SSI. Each class is a
subset of the original fuzzy set, ‘{ID, µ(ID)},’ and
again, each of these subsets are multilayer fuzzy set.
3.7.4 Multilayer Fuzzy Output
Figure 6 is a sample of the actual images being
considered for the study, which were categorized
with respect to the abnormality.
3.8 Prediction with Random Data
A total of 9 images from a different database, [10],
were considered to test if the MFA model was
working correctly or not, and whether it was able to
make accurate predictions. To check the model, the
data for the prediction was linked to the software
code, and the code was run.
3.9 Software
The software used was Python 3.6.1 of Anaconda3
4.4.0 with Spider 3.1.4 as the graphical user
interface.
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.24
Kishore Kumar Akula, Alexander Gegov,
Farzad Arabikhan
E-ISSN: 2224-2872
212
Volume 22, 2023
3.10 Efficacy of the MFA Method
The efficacy of the MFA method was examined by
considering the CT scans of the same patient taken
at different times. Since the patients developed
COVID-19, when the CT scans were observed, the
spread of the virus can be seen gradually
(Supplementary Figure 1). For the same scans, the
SSI was calculated just to check whether the SSI
score was inversely proportional to the spread of the
virus and whether it gradually decreased with the
spread of the disease.
3.11 Results and Discussion
In the current study, the multilayer fuzzy method,
MFA, was developed using fuzzy systems, along
with the concept of comparison derived from
cognitive science and SSI from computer vision.
The application of the MFA method will be to
detect the presence or absence of COVID-19 in CT
scans of the lungs and classify them according to the
spread of COVID-19 as seen on the CT scans.
3.12 Cognitive Comparison using Computer
Vision to Detect COVID-19 at a Microscopic
Level
The way comparisons were carried out using
computer vision is shown in Figure 5 using a normal
CT scan versus the CT scan of a lung affected by
COVID-19. Every spatial region of the normal
image was compared to the corresponding spatial
region of COVID-19-positive CT scans. The
difference is the spread of COVID-19 and noise if
any.
3.13 SSI Scores and a Multilayer Fuzzy Set
The SSI scores when the normal right lung CT scan
was compared with the abnormal right lung affected
with COVID-19 are given in Supplementary Table
1.
Fig. 5: The method used to find the SSI score. SSI
was calculated for a normal CT scan of a lung
versus a CT scan of a COVID-19-positive lung
using the same spatial location and utilizing the
cognitive science concept of comparing.
3.14 The Multilayer Fuzzy Set for the
Application of the Current Method, MFA, to
Detect COVID-19 in CT Scans
An example of the multilayer fuzzy set for the left
lung CT scans is {ID, SSI(ID)} = {(1, 0.5602), (2,
0.282), … (n,1.00)}, where ID is the patient’s ID,
and SSI is the membership value. Another example
of a multilayer fuzzy set for COVID-19-positive
lungs categorized as having a severe spread of the
virus is given by {ID, k2}, where k2 represents the
SSI scores, a cut-off for the severe abnormality.
3.15 The Detection, Classification, and
Prediction of COVID-19 using MFA
3.15.1 SSI or k Values Found when the Normal
Right Lung is Compared with the Data
After testing the code several times with the data of
CT scans and adjusting the k value according to the
abnormality, the following categorization was
established:
SSI or k ≤ 0.2000, for a very severe abnormality,
0.2000 < SSI or k ≤ 0.2800, for a severe
abnormality,
0.2280 < SSI or k ≤ 0.3030, for a moderate
abnormality,
SSI or k ≥ 0.3030, for a mild or no abnormality.
The classification by considering the above
thresholds after testing the code several times is
shown in Table 1 and some of the classified images
are shown in Figure 6. Although the images in
Figure 6(6.3) and Figure 6(6.4) look alike, the
classification for both is different using MFA, since
the image in Figure 6(6.3) has higher contrast than
the image in Figure 6(6.4).
Table 1. In a random data set of CTs, COVID-19
was successfully detected and classified according
to the severity of the spread of the virus by MFA.
Right lung CTs detected & classified by the degree
of COVID-19 infection by MFA
Total
CTs
Very
severe
Severe
Moderate
Mild or
Normal
21
7
5
2
7
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.24
Kishore Kumar Akula, Alexander Gegov,
Farzad Arabikhan
E-ISSN: 2224-2872
213
Volume 22, 2023
Fig. 6: Detection and categorization of COVID-19
with the help of multilayer fuzzy systems. The left
lung has been cropped and compared with the
normal left lung and categorized as
6.1 - a very severe abnormality,
6.2 - severe abnormality,
6.3 - moderate-severe abnormality (however, this is
due to high contrast) and,
6.4 - mild or normal CT scans.
3.16 Efficacy of the MFA Method
The results from testing the efficacy of the MFA
method (Supplementary Figure 1 with CT scans and
the SSI scores) were derived from a single patient
with COVID-19. The CT scans were taken with a
time gap between each image. The propagation of
the virus in the lungs was gradual as seen in the CT
scans. The scans with the SSI are shown in
Supplementary Figure 1. The CT scan in the first
figure panel showed greater spread of COVID-19,
with an SSI of 0.078, whereas the CT scan with less
spread of the virus had a greater SSI of 0.2659.
Using the multilayer fuzzy set, every spatial
location of an abnormal object within an image can
be compared to the corresponding spatial location of
a normal object (Figure 4). This process generates
the SSI, which represents the abnormalities of the
objects in the image. This is a feature of the MFA
study when contrasted with the current comparator,
a CNN, the latter of which cannot find the
abnormality in the images directly but needs to be
trained and tested using an already categorized data
set with abnormalities. Moreover, using the MFA
method, once the thresholds for different
abnormalities are found, the image can be
categorized automatically (Section 4).
In addition, the same analysis carried out by a
physician to recognize abnormalities in a CT scan
can be performed using MFA. Moreover, the
physician would need time to decide that the CT
shows an abnormality, whereas the present method
can diagnose numerous CTs for abnormalities in a
shorter time frame.
3.17 The MFA Method Versus the CNN
3.17.1 Training, Testing and Validation by a
CNN using the Considered Data Set, as well as
Prediction
To further validate the MFA method, it was
contrasted with the most recent comparator, a CNN.
The same data used in the previous section to test
the MFA method was not large enough to run a
CNN. The training showed 100%, but the validation
accuracy was 0%. In addition, in the further epoch,
the validation fluctuated between 0% and 100%
percent, showing the insufficiency of the data.
However, with the same data, the detection and
classification of the abnormality of COVID-19 was
not only possible but also accurate using the MFA
method (Table 1). The reason the MFA method
works with less data and the CNN does not is that
the MFA method is based on the comparison of two
images, requiring only two similar images. On the
other hand, a CNN has to be trained first, which
needs a greater number of images to learn the
patterns within the images. The predicted images
from a random data set to examine the process of
detection and classification of COVID-19 in the
right lung using the inspection method are presented
in Figure 7.
Fig. 7: Predicted images from a random data set to
examine the process of detection and classification
of COVID-19 in the right lung using the inspection
method.
7.1 - a very severe abnormality (however, this is
likely due to high contrast),
7.2 - a severe abnormality, and
7.3 - a moderate abnormality, and
7.4 - a normal lung.
3.17.2 Prediction Accuracy of MFA Versus the
Current Comparator, CNN, using a COVID-19
Data Set
As the CNN needs a massive, already classified data
set for training, the data considered was insufficient
to run a CNN (Section 3.17.1). Subsequently, 21
random CTs were fed into the CNN, as well as into
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.24
Kishore Kumar Akula, Alexander Gegov,
Farzad Arabikhan
E-ISSN: 2224-2872
214
Volume 22, 2023
the MFA method. It can be observed in Table 3 that
the CNN failed to predict the presence of COVID-
19, whereas the classification and prediction by
MFA were successful with the same small data set
of CTs. Except in the case of a ‘very severe’
abnormality, all the CTs were correctly identified
and classified according to the severity of the spread
of COVID-19. The images categorized as having a
‘very severe’ abnormality contained high contrast,
so these were outliers, as described in previous
sections.
As seen in Table 2, CNN was not successful in
classifying the images with regard to the
abnormalities in the small data set used, as opposed
to MFA which was quite successful. In addition,
CNN usually has a long training time whereas MFA
has a short training time.
Table 2. Comparison of predicted results for
COVID-19 data.
Categories of abnormality in COVID
Very
severe
Severe
Moderate
Mild
No. Of
CTs
9
10
1
1
MFAa
9
(100%)
8
(80%)
1
(100%)
1
(100%)
MFAb
0 (0%)
2
(20%)
0 (0%)
0 (0%)
CNNa
2 (22%)
0 (0%)
0 (0%)
0 (0%)
CNNb
7
(78%)
10
(100%)
1
(100%)
1
(100%)
a. Correctly classified, b.Misclassified
Under the very severe category, the correct
classification by the current method MFA was 78%
greater than the classification by the CNN.
Moreover, under the severe, moderate, and normal
categories, the CNN was not able to classify any
data, whereas the accuracy was 80%, 100%, and
100% greater than a CNN in these categories,
respectively, when classified by MFA. Similarly,
the misclassification by CNN was several times
greater than the MFA method, as seen in Table 3.
This comparison demonstrates that MFA works
better than the current comparator, CNN, for
classification, in part, because the CNN cannot
detect the abnormality in images directly like MFA
and needs already classified data.
Based on these preliminary tests, the MFA
method was found to be accurate for even minor
differences in the abnormality in the image, because
the method is based on the comparison of a pixel in
the normal image with a pixel in the corresponding
spatial location in the image with a potential
abnormality (Section 3.4.4).
Table 3. The difference between the MFA method
and the current comparator, CNN, is in detecting
abnormalities and classifying the abnormality for
any kind of similar images.
MFA
CNN
1. Data set size
needed
small or
massive
Massive
2. To find
abnormalities in
rare images
Successfully
works on rare
disease images
Cannot be used
with rare images,
as the sample
size will not be
large
3. Detection of
abnormality in
the objects of
similar images
Detects
abnormalities
Cannot detect the
abnormalities
4. Minimum
number of
images needed
to form the
method
One (because
it is to
compare)
More - possibly
thousands
5. Required,
already
classified
images
Not required
Mostly Required
The MFA method showed a gradual change in
the SSI score corresponding to the spread of the
abnormality. For the application part of the present
method, the longitudinal scans of a patient with
changes in the spread of COVID-19 over time were
considered. The gradual change in the spread of the
disease (Supplementary Figure 1), as seen on the CT
scan, resulted in gradual changes in the SSI score.
For instance, the SSI score of the first image is
0.078 and the spread of the COVID-19 virus almost
fully occupies the left lung, whereas in the last
image, the infection is spread over a smaller area of
the same lung of the same patient, with an SSI score
of 0.2659. That is, the similarity between the normal
image (healthy lung) and the first image was very
low, compared to the similarity between the same
normal lung and the last image. This shows that the
MFA method can identify the abnormality in the
image, like the current comparator, which is a CNN.
MFA can also identify the rate of increase of the
spread of the virus if the exact time between
acquisitions of the two images is known, which the
current comparator, CNN, cannot.
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.24
Kishore Kumar Akula, Alexander Gegov,
Farzad Arabikhan
E-ISSN: 2224-2872
215
Volume 22, 2023
4 Conclusion
The proposed MFA is an interdisciplinary cross-
application of AI with multilayer fuzzy systems,
intelligent systems, and computer vision. This
approach can detect abnormalities in similar images
from any application domain and classify them by
comparing a normal image with images containing a
similar object with abnormalities. MFA does not
require training and testing as compared with the
comparator, CNN. The prediction accuracy of MFA
for a small COVID data set was significantly higher
than the one of CNN. That shows that MFA can
work effectively even for small data sets and
therefore can be used with data for rare or new
diseases. Overall, the current study presents a
method to convert the abnormality present in the
objects of the image into a quantity or number.
Some of the future applications of this study are
further use of the MFA both in the medical sciences,
as well as outside of medicine. For instance, in the
case of medical studies involving CT scans, an
application of the current study would be the
quantification of the abnormality in the lung
resulting from cancer using disease images. As a
future study, the risk due to the quantified
abnormality that is present in a group of patients or
a single patient can be estimated. Beyond medical
sciences, this method can also be applied, for
instance, to automatically analyze satellite images of
regions of Earth containing water, or analyze
astronomy images of space.
References:
[1] Bloom, B. S.; Engelhart, M. D.; Furst, E. J.;
Hill, W. H.; Krathwohl, D. R.., Taxonomy of
educational objectives: The classification of
educational goals. Vol. Handbook I:
Cognitive domain. New York: David McKay
Company, 1956.
[2] Ciresan, Dan; Ueli Meier; Jonathan Masci;
Luca M. Gambardella; Jurgen Schmidhuber,
Flexible, High-Performance Convolutional
Neural Networks for Image Classification.
Proceedings of the Twenty-Second
International Joint Conference on Artificial
Intelligence, Vol. 2, pp.1237–1242.
[3] Liu S., Deng W., “Very Deep Convolutional
Neural Network Based Image Classification
Using Small Training Sample Size”, 3rd IAPR
Asian Conference on Pattern Recognition,
2015.
https://doi.org/10.1109/ACPR.2015.7486599
[4] Zhao P., Li C., Rahaman M. M., Xu H., Yang
H., Sun H., Jiang T., Grzegorzek M., A
Comparative Study of Deep Learning
Classification Methods on a Small
Environmental Microorganism Image Dataset
(EMDS-6): From Convolutional Neural
Networks to Visual Transformers, Systems
Microbiology, Vol. 13, 2022.
https://doi.org/10.3389/fmicb.2022.792166.
[5] Phung V. H., Rhee E. J.. A High-Accuracy
Model Average Ensemble of Convolutional
Neural Networks for Classification of Cloud
Image Patches on Small Datasets, Applied
Sciences, 9(21), 4500, 2019.
[6] Tamalika Chaira, Medical Image Processing:
Advanced Fuzzy Set Theoretic Techniques,
CRC Press, 2015.
[7] Zhou Wang. Multi-scale structural similarity
for image quality assessment. Proceedings of
the 37th IEEE Asilomar Conference on
Signals, Systems and Computers, Pacific
Grove, CA, 2003.
[8] Alexander Gegov. “Fuzzy Networks for
Complex Systems”, Springer, 2011.
[9] Alexander Gegov. “Complexity Management
in Fuzzy Systems”, Springer, 2007.
[10] National Cancer Institute, 2021. [Online],
https://www.cancerimagingarchive.net/
(Accessed Date: 23/10/2023).
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.24
Kishore Kumar Akula, Alexander Gegov,
Farzad Arabikhan
E-ISSN: 2224-2872
216
Volume 22, 2023
Appendix
Supplementary Table 1.
Abnormalities detected in the right lung
k1 <= 0.2
0.2 < k2
<= 0.28
0.28 <=
k3 <=
0.303
k4 >=
0.303
0.1717, 0.1994,
0.1987, 0.1066,
0.0916, 0.1654,
0.192
0.2742
0.2136
0.227
0.2064
0.2081
0.201
0.221
0.2372
0.2873
0.3015
1
0.3316
0.441
0.421
SSI = 0.078. 0.1957, 0.1924, 0.1893, 0.195, 0.1811,
0.1898
SSI = 0.1974, 0.212, 0.1935, 0.1892, 0.1884, 0.2023, 0.2115
SSI = 0.2132, 0.2199, 0.2145, 0.2079, 0.201, 0.2071, 0.2156
SSI = 0.2157, 0.224, 0.2335, 0.2419, 0.2365, 0.2326, 0.2354
SSI = 0.2423, 0.2523, 0.267 0.2711 0.2659
Supplementary Fig. 1: CT scans of the lungs of a
single patient at different stages to detect COVID-
19. This figure shows the efficacy of the MFA
method at detecting abnormalities in the objects
within an image using a multilayer fuzzy system,
artificial intelligence (AI), cognitive science, and
computer vision.
Contribution of Individual Authors to the
Creation of a Scientific Article (Ghostwriting
Policy)
- Kishore Kumar Akula conceived the study, wrote
the software code, and prepared and
reviewed the manuscript.
- Alexander Gegov supervised the idea and the
research and reviewed the manuscript.
- Farzad Arabikhan supervised the research and
reviewed the manuscript.
Sources of Funding for Research Presented in a
Scientific Article or Scientific Article Itself
No funding was received for conducting this study.
Conflict of Interest
The authors do not have any conflicts of interest to
disclose.
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the
Creative Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en
_US
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.24
Kishore Kumar Akula, Alexander Gegov,
Farzad Arabikhan
E-ISSN: 2224-2872
217
Volume 22, 2023
