Computers, Materials & Continua
DOI:10.32604/cmc.2022.020571
Article
Tech Science Press
Automatic Detection and Classification of Human Knee Osteoarthritis Using
Convolutional Neural Networks
Mohamed Yacin Sikkandar1, * , S. Sabarunisha Begum2 , Abdulaziz A. Alkathiry3 ,
Mashhor Shlwan N. Alotaibi1 and Md Dilsad Manzar4
1
Department of Medical Equipment Technology, College of Applied Medical Sciences, Majmaah University, Al Majmaah,
11952, Saudi Arabia
2
Department of Chemical Engineering, Sethu Institute of Technology, Kariapatti, 626115, Tamilnadu, India
3
Department of Physical Therapy, College of Applied Medical Sciences, Majmaah University, Al Majmaah, 11952,
Saudi Arabia
4
Department of Nursing, College of Applied Medical Sciences, Majmaah University, Al Majmaah, 11952, Saudi Arabia
*
Corresponding Author: Mohamed Yacin Sikkandar. Email: m.sikkandar@mu.edu.sa
Received: 29 May 2021; Accepted: 15 July 2021
Abstract: Knee Osteoarthritis (KOA) is a degenerative knee joint disease
caused by ‘wear and tear’ of ligaments between the femur and tibial bones.
Clinically, KOA is classified into four grades ranging from 1 to 4 based on the
degradation of the ligament in between these two bones and causes suffering
from impaired movement. Identifying this space between bones through the
anterior view of a knee X-ray image is solely subjective and challenging.
Automatic classification of this process helps in the selection of suitable
treatment processes and customized knee implants. In this research, a new
automatic classification of KOA images based on unsupervised local center of
mass (LCM) segmentation method and deep Siamese Convolutional Neural
Network (CNN) is presented. First-order statistics and the GLCM matrix
are used to extract KOA anatomical Features from segmented images. The
network is trained on our clinical data with 75 iterations with automatic weight
updates to improve its validation accuracy. The assessment performed on the
LCM segmented KOA images shows that our network can efficiently detect
knee osteoarthritis, achieving about 93.2% accuracy along with multi-class
classification accuracy of 72.01% and quadratic weighted Kappa of 0.86.
Keywords: Osteoarthritis; segmentation; intensity value; unsupervised;
neural network
1 Introduction
The human knee is an important part of the body that helps to carry out mobility. Also,
it is one of the complex and articulated joints of the body. Knee Osteoarthritis (KOA) is a
common chronic condition known as degenerative knee joint disease caused by ‘wear and tear’ of
ligaments between the femur and tibial bones [1]. KOA is one of the common causes of disability
in older adults [2]. The World Health Organization (WHO) in its 2016 report on osteoarthritis
This work is licensed under a Creative Commons Attribution 4.0 International License,
which permits unrestricted use, distribution, and reproduction in any medium, provided
the original work is properly cited.
4280
CMC, 2022, vol.70, no.3
says that 9.6% of men and 18.0% of women aged over 60 years have characteristic osteoarthritis.
Among those, 80% have problems in mobility, and 25% of them were facing difficulty in carrying
out their day-to-day activities [3]. According to a recent study, by 2032, the proportion of the
population aged ≥45 with doctor-diagnosed OA is estimated to increase from 26.6% to 29.5%
(any location), from 13.8% to 15.7% for the knee and 5.8%–6.9% for the hip [4]. In 2003, a
study conducted on KOA in Al-Qaseem, Saudi Arabia region says that 13% adult population got
affected by this disease and its prevalence steeply increases with increase in age reaching 30.8% in
those aged 46–55 years and 60.6% in the age group 66–75 years [5]. KOA ranks among the top
five causes of disability and represent an increasing economic burden to society, primarily through
lost working hours and healthcare expenses [6]. Fig. 1 pictorially represents the normal knee joint
and osteoarthritis knee joint. It is clinically very important to diagnose this joint accurately and
identify the regions that are affected. There are modalities such as X-ray, CT, and MRI are used to
scan these regions to measure the wear and tear and further clinical procedures such as implanting
and total knee replacement.
Figure 1: Pictorial representation of the X-ray image of human normal knee joint and
osteoarthritis knee joint
The diagnosis of OA is usually based on a clinical (physical) examination, the patient’s
medical history, and changes seen on plain X-rays. The Kellgren–Lawrence (KL) grading is the
main radiography based grading criteria followed over 50 years for assessing the severity and
progression of KOA and making treatment decisions [7]. Radiography (X-ray) imaging is the most
traditional tool for the assessment of OA due to its low cost, availability, high contrast, and spatial
resolution for bone and tissue [8]. Automatic detection of radiographic osteoarthritis in X-ray
knee images based on the Kellgren-Lawrence (KL) classification grades was described by Shamir
et al., in which simple weights were assigned to images and to predict the four grades of KL [9].
Many researchers have developed feature extraction algorithms and artificial neural networks for
the automatic classification of KOA [10–12].
Machine-Learning based approaches have been surveyed by Kokkotis et al. [13] for the
detection of knee osteoarthritis from X-rays. A computer-based system [14] was created that
uses body kinetics as input and produces as the result not just an assessment of the knee
osteoarthritis, also highlights the most discriminating parameters alongside a bunch of rules on
CMC, 2022, vol.70, no.3
4281
how this choice was reached as the output. The chosen input parameters for the system were of
the vertical, anterio-posterial, and medio-parallel GRFs, and, mean, push-off time, and slopes.
Random Forest Regressors were used to map those boundaries through rule acceptance to the
level of knee osteoarthritis. In another study, [15] straightforward and transparent computer-aided
diagnosis method dependent on the Deep Siamese Convolutional Neural Network was employed
to naturally score KOA seriousness based on the Kellgren-Lawrence scale. Using Deep Siamese
Convolutional Neural Network for X-ray images with symmetry can reduce the number of learnable parameters. This method made the model more reliable and less sensitive to noise. Another
way to identify the degree of knee osteoarthritis from radiographs utilizing deep convolutional
neural networks (CNN) is proposed in [16]. The classification accuracy was essentially improved
utilizing convolutional neural network models pre-trained on Image Net and re-training them
using knee osteoarthritis X-ray images. A PC-vision-based framework was proposed by Saleem
et al. [17] that can help the radiologists by studying the radiological manifestations in knee
X-ray images for osteoarthritis. Various image processing tools were used on knee radiographs to
improve the quality of the X-ray images. The knee region i.e., the tibio femoral joint is extracted
with template matching. Then the knee joint space width is calculated, and the radiographs
are classified using neural networks based on the normal knee joint space width. An efficient
computer-aided image analysis strategy [18] based on weighted neighbor distances utilizing a
compound hierarchy of algorithms speaking to morphology (WND-CHARM) was utilized to
study sets of weight-bearing knee X-beams. This stated that WND-CHARM, a computer-based
analysis tool can be used to find the development of osteoarthritis from normal knee X-ray
images. Another methodology to automatically evaluate the degree of knee osteoarthritis utilizing
X-ray images was proposed in [19], which involves two stages: first, automatically limiting the
knee joints, using a fully convolutional neural network (FCN); next, arranging the confined knee
joint images. Then a convolutional neural network (CNN) was trained to consequently evaluate
the degree of knee osteoarthritis by weight updating of two-loss functions. This joint training of
FCN and CNN further improves the general evaluation of knee osteoarthritis. A multi-modular
AI-based osteoarthritis progression forecast model is presented in [20] that uses radiographic information, clinical assessment results, past clinical history of the patient, anthropometric information,
and, alternatively, a radiologist’s assertion (KL-grade). The significant finding of this investigation
is that it is possible to predict knee osteoarthritis progression from a knee radiograph along
with clinical information in a completely programmed way. A deep neural network was used for
distinguishing the different types of osteoarthritis [21] utilizing patient’s statistical information on
clinical and health care data. Principal component analysis with quintile transformer scaling was
employed to create highlights from the patients’ background clinical records and classify the degree
of osteoarthritis [21]. Moreover, the findings of those techniques, level of accuracy in automatic
classification is not as good as clinically expected. These aforementioned approaches have tried to
deal with finding the degree of KOA like an image classification issue. The features from X-ray
images can be found with the help of a computer-aided analysis that would quantify KOA severity,
and also predict the development of knee OA. Deep learning technique especially CNN have
recently shown extraordinary results in a variety of image recognition and classification tasks.
Instead of direct radiographic features, in this work an adaptive learning feature representations
using CNN is proposed and it can be better utilized for assessing KOA images and finding the
degree of KOA prognosis. In this work, CNN is developed from scratch to automatically segment
and classify the KOA with X-ray images as input. This involves three main steps 1. Automatic
detection 2. Localization of ROI (tibio-femoral space) and 3. Extraction of features from ROI
to classify KOA. Knee joint region width is a key feature in evaluating the degree and prognosis
4282
CMC, 2022, vol.70, no.3
of KOA. In many of the KOA detection techniques, isolation of ROI is performed manually.
Hand drawn contours and cropping of the joint area are typically for ROI isolation which is
cumbersome, time consuming and highly subjective leads to human errors. In order to overcome
this, unsupervised methods can be utilized without training information to segment ROI. The
Local Center of Mass (LCM) methodology is an unsupervised image segmentation method that
depends on the calculation of the local center of mass [22]. In this methodology, the pixels of
a one-dimensional signal are grouped, which is used in an iterative algorithm for two-and threedimensional image segmentation. LCM based strategy created less over-segmented regions when
compared to other segmentation algorithms like a Gaussian-mixture-model-based hidden-Markovrandom-field (GMM-HMRF) model, watershed segmentation, and the simple linear iterative
clustering (SLIC) algorithm.
In this work, X-ray radiographic images of KOA will be segmented to identify ROI based
on LCM method. Then feature extraction algorithms are applied namely the object statistics and
texture features. CNN is trained using the features extracted from the segmented X-ray Image.
This features were considered to accurately classify the KOA type and thereby add to the reliability
of the results. After training, the network is tested for its accuracy. Post-training and testing
phases, the CNN was used to classify the KOA of five classes namely; Grade 0 or normal,
Grade 1, Grade 2, Grade 3, and Grade 4. The performance evaluations of proposed method
is presented and compared with existing techniques. The paper is organized as five sections,
first section explained about the various exiating methodology of detection of KOA followed by
database used in the second section. Methodology carried in this research is given in third section.
Results and conclusion is given in fourth and fifth section.
2 Materials and Methods
2.1 Clinical Data Collection
The digital human KOA X-ray images of 350 subjects consisting all four grades in the age
group of 50 ± 30 years including both male and female were collected from Durma and Tumair
General Hospital, Riyadh were used in this work. The X-ray images used in here composed of
both bilateral images (radiograph of both the knees) and unilateral images (radiograph of only
one knee). The bilateral images are partitioned into two individual images, each consisting of a
single knee joint. All of the obtained radiographs are in antero-posterior view. These images were
clinically and manually classified by clinical expert.
3 Methodology
The investigation begins with preprocessing of Knee X-ray image data of all four grades of
osteoarthritis, followed by segmentation using the Local center of mass algorithm, from which the
knee joint is isolated as presented in Fig. 2. Then feature extraction and analysis are done for the
knee joint. Features are extracted based on First order statistics and the GLCM matrix.
These features are used to train an artificial neural network classifier, namely Convolution
Neural Network. Once training is completed, CNN can be used to classify the shown four
degree of knee osteoarthritis in Fig. 3. The proposed algorithm is implemented in the MATLAB
environment.
CMC, 2022, vol.70, no.3
4283
Figure 2: Flow chart of the proposed methodology
Figure 3: Knee X-ray images—(A) Normal, (B) Grade 1, (C) Grade 2, (D) Grade 3, (E) Grade 4
3.1 Preprocessing
The input knee X-ray image is initially checked for the proper quality and then it is
smoothened with anisotropic filter. All the input images are resized into 256 × 256 for the
standardization of all the four categories of KOA.
3.2 Segmentation Using Local Center of Mass Algorithm and Extraction of Region of Interest
The local center of mass calculation method is used as an iterative algorithm for unsupervised
segmentation of the KOA X-ray images. A label of an estimated local center of mass is assigned
to pixels highlighting that center of mass, making the pixels in a similar locale eventually share the
same label. The pixels corresponding to a common point form a group with a specific label. Here,
the labeling of a pixel is done utilizing the data from the entire image which is depicts in Fig. 4.
The detailed methodology of LCM segmentation technique has already been implemented and
published [23,24]. An automatic cropping algorithm is then applied to isolate the knee joint space
from the segmented image using colour scale maping technique, highlighting knee joint space
4284
CMC, 2022, vol.70, no.3
(Fig. 5). Since the tibio femoral joint space is narrow, all the small colour contour segmented
regions are isolated using related colour maping then they are extracted as one single region. This
is then used in the feature extraction.
Figure 4: Segmentation of knee X-ray image using LCM
Figure 5: Isolation and extraction of ROI
3.3 Feature Extraction from Segmented ROI
3.3.1 First Order Statistics Based Approach—Histogram Based Features
First-order statistics measures are calculated from the original image values and don’t consider
pixel neighborhood relationships.
Histogram based method is based on the intensity value concentrations on all or part of an
image represented as a histogram. Common parameter values include moments such as mean,
variance, dispersion, mean square value or average energy, entropy, skewness, and kurtosis.
3.3.2 Spatial Frequencies Based Texture Analysis—Co-occurrence Matrix Based Features
The major statistical method used in image analysis is the one, dependent on the definition
of the joint likelihood distributions of sets of pixels.
The second-order histogram is defined as the co-occurrence matrix hdθ (i, j). When divided by
the total number of neighboring pixels R(d, θ) in the image, this matrix becomes the estimate
of the joint probability, pdθ (i, j), of two pixels, a distance d apart along a given direction θ
having particular (co-occurring) values i and j. Two forms of co-occurrence matrix exist—one
symmetric where pairs separated by d and −d for a given direction θ are counted, and the other
not symmetric where only pairs separated by distance d are counted.
CMC, 2022, vol.70, no.3
4285
This yields a square matrix of dimension equal to the number of intensity levels in the image,
for each distance d and orientation θ. Due to the intensive nature of computations involved, often
only the distances d = 1 and 2 pixels with angles θ = 0◦ , 45◦ , 90◦ , and 135◦ are considered.
As opposed to using a gray level co-occurrence matrix directly to gauge the textures of
images, the matrices can be converted into simpler scalar measures inertia, correlation, entropy
and homogeneity as given in the Eqs. (1)–(4).
Inertia (contrast) =
N−1
N−1
(i − j)2 p(i, j)
(1)
i=0 j=0
Correlation =
N−1
N−1
ijp(i, j) − µx µy
σx σy
(2)
p(i, j)log2 [p(i, j)]
(3)
i=0 j=0
Entropy = −
N−1
N−1
i=0 j=0
Homogeniety =
N−1
N−1
i=0 j=0
p(i, j)
1 + |i − j|
(4)
3.4 Training the Neural Network with Obtained Features
In machine learning, a convolutional neural network (CNN, or ConvNet) is a class of deep,
feed-forward artificial neural networks that has effectively been applied to analyze images. CNN’s
utilize a variation of multilayer perceptron’s designed to minimize preprocessing.
Here, the structure of CNN includes two layers. The first is the feature extraction layer, where
the input of each neuron is connected to the local responsive fields of the previous layer and
extracts the required parameters. Once the parameters are isolated, the positional relationship
between them and other parameters also will be resolved. The other is the feature map layer; each
computing layer of the network is composed of a majority of the feature map. The structure of
the feature map uses the sigmoid function as an activation function. Then there is the output layer
that corresponds to the number of output classes. In this investigation one input layer, four hidden
layers (feature map layers) composed of 8, 16, 32, 64 neurons in each layer and the output layer
has five neurons, corresponding to Grades 1 through 4 and. The CNN is trained using the 280
X-ray images (80% of the X-ray dataset) from the database for 75 iterations with automatic weight
updates to improve its validation accuracy. Fig. 6 shows the training and validation of CNN of
the features obtained from segmented 280 X-ray images. The validation accuracy is obtained as
93.2% and the learning rate is 0.01.
4286
CMC, 2022, vol.70, no.3
Figure 6: Training results of the CNN for KOA classification
4 Results
Post segmentation is the isolation and extraction of the ROI, i.e., the tibio-femoral joint
space, which is done automatically by the algorithm. Figures below (Fig. 7) represent the ROI for
Normal case, KOA Grade 1 through 4.
From ROI, first order statistical features and GLCM-based features were extracted. The
Tab. 1 below describes the feature values obtained for the different grades of knee osteoarthritis,
using which the CNN performs said classification.
4.1 Testing the CNN for Classification of Knee Osteoarthritis
Testing was done using 70 X-ray images, (20% of the X-ray dataset) to estimate the final
performance of the CNN. The trained and tested CNN was then used for the identification of
degree of knee osteoarthritis using test X-ray images.
The evaluation of the CNN classifier was performed on the knee radiograph database to study
the performance of the proposed algorithm. The model was designed to automatically localize
and isolate the knee joints from radiographs. The database was split into a training (80%) and
test (20%) sets. The CNN was trained for 75 iterations depicts in Fig. 7 with automatic weight
updates with a learning rate of 0.01. Considering the morphological variations of the knee joints,
the automatic detection with this CCN was found to be highly accurate with a validation accuracy
of 93.2%. Taking into account that the proposed method is fully automated, the computation
time of this model is appreciable. The trained and tested CNN was then able to classify a knee
X-ray image and identify the grade of osteoarthritis with the help of said features. Fig. 8 depicts
the classification results and obtained feature values for a test X-ray image.
CMC, 2022, vol.70, no.3
Categories
4287
Isolation of ROI
Extraction of ROI
Normal
Knee Joint
(Grade 0)
Grade 1
Grade 2
Grade 3
Grade 4
Figure 7: Isolated knee joint space and extracted region of interests for all cases
Table 1: Representative feature values for each grade of knee osteoarthritis obtained from the
region of interest
Type
Normal
Grade 1
Grade 2
Grade 3
Grade 4
Mean
Variance
Skewness
Kurtosis
Energy
Contrast
0.001465
0.000223
15.05791
235.1202
0.057392
0.589656
1.224056
0.719996
1.229023
0.001985
0.000548
15.60637
247.4674
0.140738
0.557982
1.033884
0.545032
1.04231
0.001752
0.000399
15.43617
243.6679
0.102485
0.515587
1.038621
0.59254
1.041314
0.001798
0.000421
15.33527
241.3825
0.108292
0.62604
1.24575
0.704747
1.248643
0.001706
0.000364
15.37473
242.3292
0.093506
0.612698
1.201
0.666545
1.192609
(Continued)
4288
CMC, 2022, vol.70, no.3
Table 1: Continued
Type
Normal
Grade 1
Grade 2
Grade 3
Grade 4
Correlation
0.970369
0.938534
0.963734
0.938236
0.577546
0.564033
0.574112
0.563942
0.989242
0.977919
0.986912
0.977834
0.976484
0.956463
0.977073
0.956128
0.505721
0.496365
0.506136
0.496215
0.989793
0.981297
0.990018
0.981152
0.977007
0.953797
0.973607
0.953634
0.528481
0.517577
0.526615
0.517529
0.990556
0.98091
0.988868
0.980862
0.971354
0.94321
0.967942
0.943082
0.539416
0.527058
0.537792
0.527012
0.988518
0.976943
0.986596
0.976895
0.97187
0.944877
0.969402
0.945325
0.53988
0.527778
0.538512
0.527929
0.988829
0.978333
0.98787
0.97848
Entropy
Homogeneity
Figure 8: Classification results of the CNN for KOA and its obtained features
To evaluate the performance of the network, the Kappa coefficient and Multiclass accuracy
were also calculated. This methodology yielded a multi-class classification accuracy of 72.01%,
quadratic weighted Kappa of 0.86. These results highlight that this network has the ability to
make accurate classification of KOA with X-ray images.
CMC, 2022, vol.70, no.3
4289
5 Conclusion
The investigation shows that it is possible to identify the grade of KOA based on the
features extracted from the segmented radiograph images. In comparison to other publications,
this method uses specific features related to the ROI, which can be correlated with the degree
of KOA. Typically the isolation of the tibio-femoral joint (ROI) is a tedious task due to its
irregular boundaries and manually performing the segmentation is quite difficult. Thus, by using
an unsupervised segmentation algorithm, namely the Local Centre of Mass algorithm to segment
the ROI has proven to be highly effective. This makes it easier to extract the structure metrics,
i.e., the features from the knee joint space. Feature extraction from the region of interest has been
incorporated to improve the predictive capabilities of the CNN and also with the aim to increase
its computational efficiency.
The assessment performed on the local center of mass-based segmented knee X-ray images
shows that this neural network can efficiently detect knee osteoarthritis, achieving about 93.20%
accuracy and compared to the previous researches and methods, this algorithm has the highest
multi-class classification accuracy of 72.01%, quadratic weighted Kappa of 0.86. Thus, the major
finding of this investigation is that it is possible to predict KOA with high levels of accuracy.
The limitation of this artificial neural network is that it requires large datasets to be trained and
the decision process is often considered to be paradox, meaning it is not easy to understand
how the decision is made. This model was trained using X-ray images that weren’t categorized
according to the age of the subjects. A future extension of this work can be done including the
age factor of the subjects that would also influence the affliction of knee OA. Despite these limitations, the findings of this work imply that this feature-based analysis method can utilize X-ray
images, to detect features present in the radiograph that are conclusive of KOA or its progression
and thus can be utilized for diagnostic purposes. This paper proposed an automated strategy to
detect knee osteoarthritis from x-ray images. The proposed approach has a few advantages. It can
help subjects experiencing knee pain receive a quicker diagnosis. Also, medical services in general
will profit by lessening the expenses of routine work. Thus, knee radiographs can be analyzed
much faster and the algorithm decreases the danger of human error in diagnosis.
Funding Statement: The authors extend their appreciation to the Deputyship of Research and
Innovation, Ministry of Education in Saudi Arabia for funding this research work through the
Project Number IFP-2020-42.
Conflicts of Interest: The authors declare that they have no known competing financial interests or
personal relationships that could have appeared to influence the work reported in this paper.
References
[1]
[2]
[3]
[4]
I. Haq, E. Murphy and J. Dacre, “Osteoarthritis,” Postgrad Medical Journal, vol. 79, pp. 377–383, 2003.
W. Laupattarakasem, M. Laopaiboon, P. Laupattarakasem and C. Sumananont, “Arthroscopic debridement for knee osteoarthritis,” Cochrane Database of System Review, vol. 1, no. CD005118, pp. 1–18,
2008.
J. C. Mora, R. Przkora and Y. C. Almeida, “Knee osteoarthritis: Pathophysiology and current
treatment modalities,” Journal of Pain Research, vol. 11, pp. 2189–2196, 2018.
A. Turkiewicz, I. F. Petersson, J. Bjork, G. Hawker, L. E. Dahlberg et al., “Current and future impact
of osteoarthritis on health care: A population-based study with projections to year 2032,” Osteoarthritis
and Cartilage, vol. 22, no. 11, pp. 1826–1832, 2014.
4290
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
CMC, 2022, vol.70, no.3
A. S. Al-Arfaj, S. R. Alballa and S. A. Saleh, “Knee osteoarthritis in Al-Qaseem,” Saudi Medical
Journal, vol. 24, no. 3, pp. 291–293, 2003, Saudi Arabia.
A. Maetzel, L. C. Li, J. Pencharz, G. Tomlinson and C. Bombardier, “The economic burden associated with osteoarthritis, rheumatoid arthritis, and hypertension: A comparative study,” Annals of the
Rheumatic Diseases, vol. 63, no. 4, pp. 395–401, 2004.
J. H. Kellgren and J. S. Lawrence, “Radiological assessment of osteoarthrosis,” Annals of the Rheumatic
Disease, vol. 16, no. 4, pp. 494–502, 1957.
H. Mingqian and E. Mark, “The role of radiology in the evolution of the understanding of articular
disease,” Radiology, vol. 273, no. 2, pp. 1–22, 2014.
L. Shamir, S. M. Ling, W. W. Scott, A. Bos, N. Orlov et al., “Knee X-ray image analysis method for
automated detection of osteoarthritis,” IEEE Transactions on Biomedical Engineering, vol. 56, no. 2, pp.
407–415, 2008.
D. Deokar, D. Dipali and G. Patil, “Effective feature extraction based automatic knee osteoarthritis
detection and classification using neural network,” International Journal of Engineering and Techniques,
vol. 1, no. 3, pp. 1–6, 2015.
J. Jayanthi, T. Jayasankar, N. Krishnaraj, N. B. Prakash, A. Sagai Francis Britto et al., “An intelligent
particle swarm optimization with convolutional neural network for diabetic retinopathy classification
model,” Journal of Medical Imaging and Health Informatics, vol. 11, no. 3, pp. 803–809, 2021.
D. Venugopal, T. Jayasankar, M. Y. Sikkandar, M. I. Waly, I. V. Pustokhina et al., “A novel deep neural
network for intracranial haemorrhage detection and classification,” Computers Materials & Continua,
vol. 68, no.3, pp. 2877–2893, 2021.
C. Kokkotis, S. Moustakidis, E. Papageorgiou, G. Giakas and D. E. Tsaopoulos, “Machine learning in
knee osteoarthritis: A review,” Osteoarthritis and Cartilage Open, vol. 2, no. 3, pp. 1–13, 2020.
M. Kotti, L. D. Duffell, A. A. Faisal and A. H. McGregor, “Detecting knee osteoarthritis and its
discriminating parameters using random forests,” Medical Engineering & Physics, vol. 43, pp. 19–29,
2017.
A. Tiulpin, J. Thevenot, E. Rahtu, P. Lehenkari and S. Saarakkala, “Automatic knee osteoarthritis
diagnosis from plain radiographs: A deep learning-based approach,” Scientific Reports, vol. 8, no.1, pp.
1–10, 2018.
A. Joseph, K. McGuinness, E. Noel, O. Connor and K. Moran, “Quantifying radiographic knee
osteoarthritis severity using deep convolutional neural networks,” in Proc. 23rd Int. Conf. on Pattern
Recognition, Cancun, Mexico, pp. 1195–1200, 2016.
M. Saleem, M. S. Farid, S. Saleem and M. H. Khan, “X-ray image analysis for automated knee
osteoarthritis detection,” Signal, Image and Video Processing, vol. 14, pp. 1079–1087, 2020.
M. Shari, M. Ling, W. Scott, M. Hochberg, L. Ferrucci et al., “Early detection of radiographic knee
osteoarthritis using computer-aided analysis,” Osteoarthritis and Cartilage, vol. 17, no. 10, pp. 1307–
1312, 2009.
J. Antony, K. McGuinness, K. Moran and N. E. O’Connor, “Automatic detection of knee joints and
quantification of knee osteoarthritis severity using convolutional neural networks,” in Int. Conf. on
Machine Learning and Data Mining in Pattern Recognition, Newyork, USA, pp. 376–390, 2017.
T. Aleksei, S. Klein, M. A. Bierma-Zeinstra, J. Thevenot, E. Rahtu et al., “Multimodal machine
learning-based knee osteoarthritis progression prediction from plain radiographs and clinical data,”
Scientific Reports, vol. 9, no. 1, pp. 1–11, 2019.
L. Jihye, J. Kim and S. Cheon, “A deep neural network-based method for early detection of osteoarthritis using statistical data,” International Journal of Environmental Research and Public Health, vol. 16, no.
7, pp. 1281, 2019.
CMC, 2022, vol.70, no.3
4291
[22] A. Mukesh, G. Harisinghani, R. Weissleder and B. Fischl, “Unsupervised medical image segmentation
based on the local center of mass,” Scientific Reports, vol. 8, no. 1, pp. 1–8, 2018.
[23] Y. Li, X. Ning and Q. Lyu, “Construction of a knee osteoarthritis diagnostic system based on X-ray
image processing,” Cluster Computing, vol. 22, no. 6, pp. 15533–15540, 2019.
[24] I. Aganj, M. G. Harisinghani, R. Weissleder and B. Fischl, “Unsupervised medical image segmentation
based on the local center of mass,” Scientific Reports, vol. 8, no. 1, pp. 1–8, 2018.