=Paper=
{{Paper
|id=Vol-3338/ICCS_CVMLH_02
|storemode=property
|title=Performance Measurement of Prostate Gland Segmentation Approaches in Transrectal Ultrasound and Magnetic Resonance Images
|pdfUrl=https://ceur-ws.org/Vol-3338/ICCS_CVMLH_02.pdf
|volume=Vol-3338
|authors=Kiran Ingale,Pratibha P. Shingare,Mangal M. Mahajan,Pushkar S. Joglekar
}}
==Performance Measurement of Prostate Gland Segmentation Approaches in Transrectal Ultrasound and Magnetic Resonance Images==
https://ceur-ws.org/Vol-3338/ICCS_CVMLH_02.pdf
Performance Measurement of Prostate Gland Segmentation
Approaches in Transrectal Ultrasound and Magnetic Resonance
Images
Kiran Ingale a, Pratibha P. Shingare a, Mangal M. Mahajan b and Pushkar S. Joglekar c
a
College of Engineering, Shivajinagar, Pune, 411005, India.
b
Bharati Vidyapeeth Medical College, Pune, 411043, India.
c
Vishwakarma Institute of Technology, Pune, 411037, India.
Abstract
In the Twenty-first century cancer is a serve threat to human lives. As per cancer facts and
figures of the American Cancer Society, 26% estimated new cases of prostate cancer in males
will be diagnosed in 2021, whereas 11% will die because of this disease. To diagnose such
disease segmentation plays a vital role in detection and treatment. To determine problems in
the prostate, transrectal ultrasound and magnetic resonance images are employed. Finding the
exact shape and size of the prostate gland is a big challenge in the field of medical image
segmentation. This research presented various techniques through which one can segment the
shape and size of the prostate gland for further diagnosis and treatment. The proposed work
first developed the region-based segmentation method. Secondly used the level set function
and optimized it by using a genetic algorithm, and at last, used the k-mean machine learning
approach to optimize results. This paper evaluates the performance of segmentation
algorithms on transrectal ultrasound and magnetic resonance images by using the
performance matrix-like accuracy, sensitivity specificity, mean square error, and Dice
similarity coefficient. The k- mean clustering machine learning approach has given the best-
optimized performance with the extraction of accurate shape and size of the prostate with the
accuracy of 96.3%.
Keywords 1
Prostate cancer, segmentation, level set, region growing, genetic algorithm, clustering.
1. Introduction
In the twenty-first century, cancer is a serve threat to human lives. As per cancer facts and figures
of the American Cancer Society, 248,530 estimated new cases of prostate cancer in males will be
diagnosed in 2021, whereas 34,130 will die because of the disease in the US [1]. As per records of
prostate cancer incidences in India, the rate of prostate cancer has increased [2].
The prostate gland is a part of men’s reproductive system. The prostate gland acts as a muscle and
it generates seminal fluid to protect the sperms [3][4]. Prostate cancer is found in the prostate gland
when an anomalous cell increases more rapidly than a normal prostate and may be converted into a
malignant tumor. For finding the abnormality, initially, the doctor uses a prostate antigen test, digital
rectal examination, and, biopsy. For detecting prostate cancer, there are some common symptoms like
feeling uneasy in the pelvic area with pain, prostatitis, frequent urination, pain at the time of urination,
blood in urine, men’s infertility, and, hematospermia. Figure 1 shows the structure of the prostate
gland located beneath of bladder [5].
CVMLH-2022: Workshop on Computer Vision and Machine Learning for Healthcare, April 22 – 24, 2022, Chennai, India.
EMAIL: kiraningalecoep9@gmail.com (Kiran Ingale)
ORCID: 0000-0001-7137-0576 (Kiran Ingale)
©️ 2022 Copyright for this paper by its authors.
Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR Workshop Proceedings (CEUR-WS.org)
15
Prostate segmentation performs a vital role in various phases of medical treatment. The size,
shape, and volume of the prostate can be directly determined by prostate segmentation. Such shape,
size, and volume of the prostate help in the diagnosis of prostate diseases [4]. The finding of the
perimeter, volume, and counter of the prostate is very helpful in the various treatments and
monitoring of prostate diseases like BPH, prostatic brachytherapy [4]. Parallel to this, segmentation of
the prostatic gland accelerated the procedure of tumor localization in a biopsy and radiotherapy. Even
so, manual prostate segmentation is not a simple job, which is always sensitive to internal and
external observer dispersion. Recently, automatic computerized techniques are being explored to
carry out the segmentation prostate and its performance evaluation.
Figure 1: Prostate gland below the bladder
2. Transrectal Ultrasound and Imaging Method
The transrectal imaging method is also recognized as prostate ultrasound or endorectal
sonography. The key role of the endorectal sonography method is to bring out men’s prostate gland to
detect and diagnose symptoms such as prostatitis, prostate cancer [6], pelvic pain, urinary tract
infections, hematospermia, and problems related to men’s infertility [3]. Because of its low cost,
handy, systematic procedure transrectal ultrasonography is the most popular image modality.
The transrectal ultrasound imaging modality employed the concept of the ultrasound system.
Ultrasound images of the prostate were captured with the help of transducer probes covered with a
safely replaceable cover with gel. Doctors used to examine the prostate-specific antigen level to
forecast the symptoms brought out due to deformities in the prostate gland if the PSA level is high
[3].
Figure 2: (a) Transrectal ultrasound prostate image. (b) Abnormalities in Prostate gland shown in
Trans-rectal image marked by an expert. (c) Normal magnetic resonance prostate image. (d)
Delineated prostate magnetic resonance image.
16
3. Magnetic Resonance Imaging
The magnetic resonance imaging techniques employ radio frequency pulses, magnetic fields, and
computers to construct a completed image of the human body. Radiologists and oncologists utilize
prostate MRI to evaluate the area of prostate cancer and detect whether it has spread or not. For the
diagnosis of conditions, you have like infection and enlargement of the prostate, doctors may employ
such images [9]. The use of standard T2 weighted MRI in the diagnosis, staging, and treatment of
prostate cancer is maturing and extends from center to center. Normally, MR images have poor
specificity and slightly good sensitivity [10].
4. Theoretical Consideration
The purpose of processing image segmentation and extraction of a particular prostate region is
referred to as a region of interest. This can help the radiologist to identify the regions which require
examination. In prostate TRUS and MR imaging, the objective of extracting a region of interest is to
find the region present in the prostate which is likely to have prostate cancer. The region obtained
should contain the region marked by experts [8]. Prostate image segmentation can be accomplished
with the help of a counter segmentation process, region-based segmentation, machine learning-based
supervised, and unsupervised image segmentation [7][4].
5. Seeded Region Growing Technique
It is a basic image segmentation technique. The seeding is used to start the process of
segmentation. The seeding point may be automatic or manual. The main working principle of seed
region growing is that the process starts using a single image’s pixel and it continuously grows a
region [11].
Region growing based image segmentation is used to extract the region from transrectal ultrasound
prostate and MR prostate images. The region is the segmented area extracted from an image. This
process includes all nearby pixels of images that show homogeneity with starting a single image’s
pixel [11]. Vinicius R, P. Borges, Maria Cristina F. de Oliveira proposed region growing based image
segmentation for 2D microscopy digital images [12].
Erwin, Saparudin, in this paper proposed performance analysis of seed point-based region growing
method [13]. The comparison is with image thresholding. Segmentation is done for the Berkeley
segmentation database (BSDS). The author concluded that the region growing segmentation
procedure results in clearer and more accurate boundary detection than the thresholding technique
implemented for the same objective.
The region growing consists of the splitting and merging process. This process divides an image
into uniform regions The splitting process divides an input image into several small regions generated
using manual seeding. The process of spitting continues until no further split occurs. The reverse
process of image split to generate an extracted region homogeneity defines uniformity in the splitting
process [11][13].
The Euclidean distance between the two similar pixels is given by
𝑚𝑖𝑛(∑ 𝑞𝜀𝑊(𝑃)𝑑(𝑞𝑐𝑖 )) (1)
𝜎(𝑃, 𝑐) =
|𝑊(𝑃)|
When the standard deviation of intensity is less than a threshold value, a statistically homogeneous
region is defined for intensity level images. The standard deviation is given as
17
𝑛
1 (2)
𝜎=[ ∫ (𝑥𝑖 − 𝑥)2 ]
𝑛 − 1 𝑗=1
6. Related Work: Level Set Technique
In this technique, an interesting contour is a zero-level set of a level set equation. This status is
fulfilled by the signed distance equation │dѱ│=1. The entire domain of an energy function gives a
signed distance property. When the function is set as │dѱ│=0. Then the function is found far from
the zero-level set [15]. The motion equation of the LS is
𝑑𝜓 (3)
= 𝜈. 𝛻𝛹 = 0
𝑑𝑡
Here 𝑎 = (u, v, w), shows the fitness function, and u, v, and w are velocity or fitness fields
concerning i, j, and k coordinates. The LS modified equation is
ψ(x + 1) − ψ(x) (4)
+ ν⃗(x). ν∇ψ
Δt
According to a special derivative, the equation becomes
𝜓(𝑥 + 1) − 𝜓(𝑥) (5)
+ 𝑢(𝑥)𝜓𝑖 (𝑥) + 𝑣(𝑥)𝜓𝑗 (𝑥) + 𝑤(𝑥)𝜓𝑘 (𝑥) = 0
𝛥𝑡
The level-set technique is employed to a substantial extent in the area of medical image analysis.
As compared to segmented prostates by manual methods, a technique like optimization provides good
results concerning shape training images and mean shape images [16].
7. Genetic Algorithm
A genetic algorithm is applied in two steps. The first step involves image training which obtains
different textures, shapes, and regions. The training data performance of training data is obtained in
the second stage where a genetic algorithm produces contour selection, crossover, and mutation. This
is the stage that produces the next generation. To generate a new generation, the section of an
individual is an important step. According to the fitness function, the individual gets selected
[4][15][16].
𝐴[ 𝑆𝑒𝑙𝑒𝑐𝑡𝑖𝑛𝑔 𝑡ℎ𝑒 𝑖𝑡ℎ 𝑖𝑛𝑑𝑖𝑣𝑖𝑑𝑢𝑎𝑙 ] = 𝑏 ′ (1 − 𝑏)(𝑟 − 1) (6)
Where: b = probability of best individual selection, A = population size, r = individual rank
𝑏 (7)
𝑏′ =
1 − (1 − 𝑏)𝑝
Where p = probability of an individual
Stopping criteria are achieved until the genetic algorithm generates a new generation According to
the study, the most accepted stopping criteria are maximum population generation. [4][17].
8. Unsupervised Cluster-based Image Segmentation
The cluster-based segmentation method is the same as the classification method except that the
clustering technique is an unsupervised technique. To perform the segmentation of the TRUS image,
18
the cluster-based technique utilizes the iteration mechanism between the properties of each class. The
process of TRUS image segmentation can be carried out using a well-known k-means unsupervised
algorithm [18][19][20][21][22]. A cluster-based algorithm is used for the classification of a class of
data set with predefined clusters. It is basically to cluster N data points with M-dimensional feature
space into clusters. The classification space F can be defined as
𝐹𝑎 = [𝑓𝑎1 , 𝑓𝑎2 , … … . . 𝑓𝑎𝑁 ]𝑇 , 𝑤ℎ𝑒𝑟𝑒 𝑎 = 1,2,3 … . . 𝑀 (8)
The goal of this technique is to calculate the lower value of the addition of the square distance
within the cluster between the pixel and the cluster center by moving pixels from one cluster to
another cluster. The objective function for a cluster is given as
∈= ∑𝑥𝑏=1 𝐷 2 (𝑏, 𝑐) (9)
where D (j, k) is the Euclidean distance of the pixel from the mean of the cluster. The cluster is
defined as
𝐷 2 (𝑏, 𝑐) = ∑𝑀 2
𝑎=1[𝑓𝑎1 − 𝑈(𝑏, 𝑎) ] ∗ 𝑊(𝑏, 𝑎) (10)
Here, u (b, c) = features mean with the cluster and W (b, a) = weighing factor [8].
9. Experimental consideration/materials and methods
This research has been conducted with images captured with the approval of the Institutional
Ethics Committee and with the taking permission of patients. The proposed experiments have been
performed on a Windows machine with a configuration of Intel i5 CPU (3.4 GHz). The graphical user
interface has been developed in MATLAB
T2W axial magnetic resonance images were generated using a human body. 1.5T GE Medical
Systems (SIGNA HDxt) at an image size of 512 x 512. The 2-D TRUS images were acquired using
Philips Medical Systems with a CURVED LINEAR transducer probe. The 2-D TRUS image size is
600 x 800. The TRUS and MR image databases contained 400 images each.
10. TRUS and MR images prostate database with ground truth
The required ground truths were generated by experienced radiologists and oncologists to check
the performance of prostate segmentation algorithms.
19
Figure 3: TRUS and MR prostate images database with and without ground truth
The implementation of prostate gland extraction consists of four separate techniques. The process
of region extraction and detection is performed by a region growing level set, level set genetic
algorithm, and k-means clustering. These techniques are used for the segmentation of TRUS prostate
and MR prostate images.
11. Seed guided region growing process
The seeding region growing has a region split and region merge process. The split rapidly divides
an input image into small homogeneous regions.
Split Phase: - A spilled phase provides the generation of a quadtree structure where nodes at each
level result in the split of images into homogeneous sub-images with regular interval sizes. The set of
small regions in the equal and regular structure created by the split process will be the set of input
regions for the merge phase.
Merge Phase: - In the region merge process, nearby small regions are generated by image split.
The iteration process helps to merge the small regions generated by the split to build a large image.
The process of merging will continue until no more merges are possible.
20
Figure 4: Experimentation output for Sees guided region growing process
12. Level Set Genetic Algorithm
LSM was used for segmenting prostatic images. Segmented filtered images were utilized as input
images. The classification of the objects is achieved by feature extraction. To deal with the contours
of the prostate, the features are extracted from the training set. A genetic algorithm is applied to
optimize which decreases the need for an energy equation by the fitness function. It is a value of the
individual population. A new generation is obtained using this value. After performing numerous
operations, a contour is detected on the prostate gland. We exercise the following steps to obtain
desired level-set genetic algorithm.
“Training Images: Manually Segmented image
• Extract the feature
• Mean Shape
• Shape variability
• Texture
• Mean position
Describe shape with help of the LSM
(11)
𝜓[𝑤, 𝑝](𝑥, 𝑦) = 𝜓⃗ (𝑥, 𝑦) + ∑𝑘𝑗=1 𝑤𝑗 𝜓𝑗 ((𝑥, 𝑦).
Set a threshold of fitness or number of generations > 100
for
Is fitness > number of generations > 100
→ Describe ‘n’ individuals of the genetic population as pose and weight parameters of the
above equation. The curve is segmented by each individual.
→ Derive fitness score (0-1000) on test images
→Fitness score == 0: contour found outside the prostate.
else
→ Fitness score == 1000: the prostate encircles by segmenting counter
→ Execute genetic selection, crossover, and mutation to generate a new generation of
segmenting contours
→ Number of generations = number of generations + 1
End.
21
Figure 5: Experimentation output for level set and level Set genetic algorithm
13. Unsupervised k-means clustering process
To segment and extract the prostate area from the TRUS image, a multi-stage segmentation
algorithm is proposed. This consists of five different stages.
1. Build knowledge-based rules before performing extraction and segmentation of the prostate.
Some rules are applied to the proposed technique.
a. TRUS image consists of different regions namely: The prostate, background, and
tissues around the prostate.
b. Background of the image should be black.
c. Prostate Gray level should be low in comparison to tissues around the prostate.
d. The prostate has a smooth curvature shape.
2. Perform image enhancement to achieve better segmentation output. Noise has to be
minimized and images for further processing should be enhanced.
3. Select the cluster by defining its centroid. The k-means cluster process works on the
computation of Euclidean distances between the data sets of distinct classes created by
defining clusters and each cluster must have a centroid value.
4. Detection of prostate boundaries. The process of extraction of the prostate has a clustering
technique is used in the process of extraction of the prostate. Initially, centroid and cluster are
initiated. Calculate the Euclidean distance between the data sets of each class.
5. Extraction of the prostate image after computation of the Pythagorean distance between
points from data sets. The algorithm works along boundaries where two distinct classes are
produced. The final stage has two classes where one class represents the prostate and the
other class represents the background.
Figure 6: Experimentation output for k-mean clustering for TRUS and MR images
22
14. The Result, Analysis and Discussion
To find the performance of the experiment, ground truth images were developed as per the
MICCAI prostate the challenges. Manual segmentations are performed by competent radiotherapists
and confirmed by a professional urologist consultant. In the field of radiology and urology, these
professionals have 12 years of experience. The proposed methods were evaluated for the patient’s
TRUS and MR images collected from the hospital.
Table 1: Overall performance parameters evaluation of prostate TRUS images
Performance Segmentation algorithm
parameters Region growing Level set Level set genetic k-mean
algorithm clustering
MSE (mm) 0.129 --- --- 0.179
Accuracy 0.188 --- --- 0.737
Sensitivity 0.046 --- --- 0.977
Specificity 0.989 --- --- 0.268
DSC (mm) 0.083 --- --- 0.983
Table 2: Overall performance parameters evaluation of prostate MR images
Performance Segmentation algorithm
parameters Region growing Level set Level set genetic k-mean
algorithm clustering
MSE (mm) 0.059 1.46±0.3 1.48±0.3 0.200
Accuracy 0.045 0.429 0.705 0.963
Sensitivity 0.210 0.325 0.390 0.977
Specificity 0.970 0.956 0.913 0.195
DSC (mm) 0.956 0.485 0.496 0.984
The four proposed methods region growing, level set, level set genetic algorithm and clustering
methods have been compared and evaluated performance parameters for the patient’s TRUS prostate
and MR prostate images that were collected from Bharti Vidyapeeth Hospital Pune India. The study
protocol was reviewed. A quantitative study has been done for the proposed methods using
performance parameters MSE, DSC, Specificity, Sensitivity, and Accuracy. Values of performance
parameters for all prostate images are shown in Tables 1 and 2.
The first level set method is used on test images. After the experimentation, only 42.9163%
accuracy was achieved along with, MSE 1.4696±0.3mm, and DSC 0.4853mm. Finally, it is observed
that when a genetic algorithm is applied to the level set function, then achieved optimized results in
terms of performance parameters as shown in table 2. Accuracy increased to 70.5180%. In the end,
the LSM and a genetic algorithm are integrated and compared, showing optimized results.
Here we employed an unsupervised machine learning approach to get a more accurate result. After
the experimentation, the k means clustering method segmented and detected the prostate shape and
size accurately. K means the clustering approach gives a good result on TRUS images. On the
transrectal ultrasound prostate images, it achieved approximately 73.7 % accuracy with MSE and
DSE values of 0.179 mm, 0.983, respectively.
Finally, we achieved approximately 96% segmentation accuracy of prostate magnetic resonance
images along with overall MSE and DSE values of 0.200mm, 0.984mm respectively, using k-mean
clustering.
23
The scope of this research is limited to image processing and Machine learning Technique. Here
we faced difficulties in implementing a level set genetic algorithm on TRUS images. In the next step
in development, we can use deep learning approaches to improve performance.
15. Conclusion
Prostate cancer region extraction and segmentation using manual seeded region growing
technique, level sets, level set genetic algorithm, and unsupervised cluster-based technique have been
implemented. The performance parameters for each of these techniques presented and compared. The
size of an extracted and segmented region from the TRUS and MR prostate images depends upon the
starting seeding points. The region grows along the seed point, computing nearby similar pixel data
elements present in an image. In this experiment, it has been observed that regardless of image
similarities and noise, techniques were used here to segment the prostate gland surrounding tissues
properly. Here we tried a genetic algorithm on a level-set function, that involves complex features
because of that applying segmentation of prostate MR images becomes tiresome. In the level set, the
energy fitness function is used, but optimization procedures like genetic algorithms keep away from
its uses. In the comparison of level set and level set genetic algorithm segmentation techniques, the
level-set genetic algorithm technique provides better and optimized results. The cluster technique
generates the new small clusters and their centroids and produces small regions of similar image pixel
data sets. It is concluded that the region extraction and segmentation using the automatic unsupervised
k-means cluster technique is efficient and more accurate than the manual seed point region growing,
level set, and levels set genetic algorithm method.
16. Acknowledgment
The authors would like to convey their appreciation to the Bharati Vidyapeeth Medical College,
Pune, India, for providing accurate delineations of transrectal ultrasound prostate images and MR
images. We would also like to thank our colleagues for reviewing my work and providing very useful
comments and suggestions.
17. References
[1] The American cancer society, Atlanta, Georgia, Cancer facts and figures 2020. URL:
www.cancer.org. Last accessed 2021/09/05.
[2] B. Yeole, ”Trends in the Prostate Cancer Incidence in India.” Asian Pacific Journal of Cancer
Prevention 9, 141-144 (2008).
[3] J. Sedellar, J. Rosette, H, Beerlage, H. Wiikstra, F. Debruyne, R. Aarnink. ”Transrectal
ultrasound imaging of the prostate: review and perspectives of recent developments.” Prostate
Cancer and Prostatic Diseases 2, 241-252 (1999).
[4] S. Ghosh, A. Olivera, R. Mart, L. Xavier, C. Joan, J. Vilanovac, J. Mitraa, D. Siduded, F.
Mariaudeaub. ”A Survey of Prostate Segmentation Methodologies in Ultrasound, Magnetic
Resonance and Computed Tomography Images.” Preprint submitted to Computer Methods and
Programs in Biomedicine, Elsevier April 11 (2012).
[5] https://www.hopkinsmedicine.org/health/conditios-and-diseases/benign-prostatic-hyperplasia-
bhp. Last accessed 2021/09/05.
[6] R. Alonso, B. Fernando. ”Prostate cancer.” Instituto Nacional de Ciencias M´edicas y Nutrici´on
Salvador Zubir´an. Ciudad de M´exico, M´exico 2(58), (2016)
[7] W. Qiu, J. Yuan, E. Ukwatta. ”Rotationally Resliced 3D Prostate TRUS Segmentation Using
Convex Optimization With Shape Priors.” Medical Physics 42, 877 (2015).
[8] K.Ingale, P. Shingare. ”Transrectal Ultrasound Prostate Image Segmentation using Machine
Learning-Based Unsupervised Clustering Technique.” Journal of critical reviews, 07(18), 4026-
4031(2020).
[9] https://www.radiologyinfo.org/en/info-mr_prostate. Last accessed 2021/09/09.
24
[10] H. Hou, D. Swanson, A. Barqawi, ”Modalities for Imaging of Prostate Cancer.” Advanced in
Urology, Hindawi Publishing Corporation 2009, Article ID 818065, 12 (2009).
[11] K. Ingale, P. Shingare, M. Mahajan. ”Medical Image Segmentation: A Comparision Between
Unsupervised Clustering and Region Growing Technique for TRUS and MR Prostate Images.”
International Journal of Computer Science and Network Security 21, 1-8,(2021).
[12] V. Borges, F. Cristina, D. Oliveira, G.Silva, A. Fellow, B. Hamann. ” Region Growing for
Segmenting Green Microalgae Images.” Journal of Latex Class file 9(13), (2014).
[13] S. Erwin, A. Nevriyanto, D. Purnamasari. ” Performance Analysis of Comparison between
Region Growing, Adaptive Threshold and Watershed Methods for Image Segmentation.”
Proceedings of the International Multiconference of Engineers and Computer Scientists 2018, vol
1, IMECS, March 14-16, Hong Kong (2018).
[14] A.Tsai, A. Yezzi, A Wells, C. Tempany, D. Tucker, A. Fan, E. Grimson, A. Willsky. ” A
Shape-Based Approach to the Segmentation of Medical Imagery Using Level Sets.” IEEE
Transactions on Medical Imaging 22, 137-154 (2003).
[15] S. Osher, R. Fedkiw. ”Level set methods and dynamic implicit surfaces.” Springer, New York
(2002).
[16] K. Ingale, P. Shingare. ”An Optimisation for the Level-Set Method for by Genetic Algorithm
using Prostate Gland Images.” Solid State Technology 6(63), 11691-11699 (2020).
[17] S. Chabrier, C. Rosenburger, B. Emile. ” Laurent, H.: Optimization Based Image Segmentation
by Genetic Algorithms.” EURASIP Journal on Video and Image Processing, 1-23 (2008).
[18] A. Bala, A. Sharma. ”Color Image Segmentation Using K-means Clustering and Morphological
Edge Detection Algorithm.” International Journal of Latest Trends in Engineering and
Technology, (IJLTET).
[19] L. Caponetti, G. Castellano, V. Corsini. ” MR Brain Image Segmentation: A Framework to
Compare Different Clustering Techniques.” MDPI, Information 8, 138, (2017).
[20] K. Senthil, K. Venkatalakshmi, K. Karthikeyan. ”Lung Cancer Detection Using Image
Segmentation by means of Various Evolutionary Algorithms.” Hindawi Computational and
Mathematical Methods in Medicine 2019, Article ID 4909846, (2019).
[21] M. Hamada, Y. Kanat, A. Abiche, A. ” Multi-Spectral Image Segmentation Based on the K-
means Clustering.” International Journal of Innovative Technology and Exploring Engineering
(IJITEE), ISSN: 2278-3075, 2(9), (2019).
[22] D. Kesavaraja, R. Balasubramanian, R. Rajesh, D. Sasireka.” Advance Cluster-Based Image
Segmentation. ICTACT.” Journal on Image and Video Processing, 02(02), (2011).
25