https://doi.org/10.31449/inf.v47i9.4566 Informatica 47 (2023) 11–16 11 
Optimization of Brain Cancer Images with Some Noise Models 
 
Ali Abdulmunim Ibrahim Al-Kharaz 
Technical College of Management-Baghdad, Middle Technical University, Baghdad, Iraq 
E-mail: ali.al-kharaz@mtu.edu.iq 
Keywords: harris corner detectors, gaussian noise, salt and pepper noise, mann-whitney test, magnetic resonance 
imaging  
Received: December 13, 2022 
 
The diagnosis of brain cancer based on magnetic resonance imaging (MRI) is affected by many 
conditions such as the movement of the patient during capture which leads to the occurrence of noise 
associated with these images. In order to improve corner detection methods, many corner detection 
methods were adopted according to the many noise models. Experimental results showed the effect of 
magnetic resonance imaging capabilities on both corner detection methods and the noise model. This 
study shows that it is possible to rely on other medical images such as Doppler images and neural 
network algorithms to use image features to diagnose cancer. 
Povzetek: Opisana je nova metoda za diagnozo možganskega raka, ki učinkoviteje odpravi šum v MRI 
slikah možganov.  
 
1 Introduction  
Magnetic Resonance Imaging (MRI) has become one of 
the most important tools for mapping brain function [1]. 
Diagnosing brain cancer based on MRI continues to be of 
interest; therefore many researches have used it to 
diagnose patients. In 2009, research by (Rajeev Ratan) 
and others, included the detection of brain cancer by MRI 
by relying on the sensitivity of image edges, but based 
their work on a threshold that represents the limit of the 
possible variation in image units. The experimental results 
showed the ability of the method to sensitize the areas of 
tumors [2]. Research introduced by (Aaswad Sawant) and 
others in 2018 used a set of (1800) magnetic resonance 
images with (900) cancerous and (900) non-cancerous, 
basing their analysis on a convolutional neural network 
(CNN) for the purpose of diagnosing the patient, and the 
results of the diagnoses appeared by (99) percent [3]. In 
2022, research by (Mahsa Arabahmadi) and others 
compared three methods of sensing brain cancer 
(supervised, unsupervised and semi-supervised) and 
applied a number of neural network and CNN methods. 
These methods were applied to the studied group of MRI 
images, and the images compared [4]. Deep learning 
based on a systematic CNN approach was performed on 
1258 MRI images of 60 patients. Accuracy was 47.02% 
using one epoch, increasing to 96% when the number of 
epochs rose to 15 [5]. Although previous studies removed 
noise and classified the MRI images into benign tumors 
and malignant tumors in the brain, the current study added 
two types of noise to the images and then used three 
methods (Features from Accelerated Segment Test, Harris 
Corner Detectors, and Binary Robust Invariant Scalable 
Key) to detect the brain cancer to establish the best 
method for detecting and removing noise. Table 1 
provides a summary of the related work. 
 
 
 
 
Different noise models can affect MRI images of the 
brain, so the diagnosis of brain cancer can be affected by 
noise models and create problems that accompany the 
diagnosis stage. This research aimed to find the best 
corner detection method for detecting brain cancer with 
magnetic resonance images which are less affected by the 
noise model. 
2 Background for digital images  
A digital image can be a (2D) representation of (3D) 
sensing images with each (2D) image consisting of sub-
units, each of which is called a pixel. Digital images can 
be (color, gray or binary) images. 
The most common representation of a digital image is 
[6][7] 
 
𝑓 (𝑥 ,𝑦 )=
[
 
 
 
 
𝑓 (𝑥 0
,𝑦 0
) 𝑓 (𝑥 0
,𝑦 1
) . . 𝑓 (𝑥 0
,𝑦 𝑁 )
𝑓 (𝑥 1
,𝑦 0
) 𝑓 (𝑥 1
,𝑦 1
) . . 𝑓 (𝑥 1
,𝑦 𝑁 )
.
.
𝑓 (𝑥 𝑀 ,𝑦 0
)
.
.
𝑓 (𝑥 𝑀 ,𝑦 1
)
.
.
.
.
.
.
.
.
𝑓 (𝑥 𝑀 ,𝑦 𝑁 )]
 
 
 
 
   ,(1) 
 
Where 𝑓 (𝑥 𝑖 ,𝑦 𝑗 )   represents the pixel element and the size 
of the image is  (𝑀𝑋𝑁 ).  
 
 
 
 
 
 
 
 
 
 
 
 
 

12 Informatica 47 (2023) 11–16 A.A.I Al-Kharaz 
 
Table1: Summary of related work. 
3 Some corner detection methods  
The discovery of the angle within the image is one of the 
methods that depends on computer vision in order to 
search for the image features, and the accurate discovery 
of these angles helps in various applications, especially in 
the field of distinguishing patterns and identifying objects 
within images and in the field of discovering important 
points within the graphic content. Some of these are 
discussed below. 
3.1 Features from accelerated segment test 
(FAST) 
In (2006), the FAST algorithm was proposed by (Edward 
Rosten and Tom) and the algorithm working as a corner 
detection method with the following steps [8]: 
 
 1-Select each pixel to be  point of interest interest 
point or not with proposed intensity to be (𝑓 (𝑥 ,𝑦 ) ) 
2-Select a threshold value to be (𝜏 ) and proposed (𝑘 ) 
should be a positive integer  
3-Select (𝜑 ) which represents the nearest circle pixels 
with (16) pixels, see Figure 1 
4-Now the selected pixel (𝑓 (𝑥 ,𝑦 ) ) will be the corner 
point if there exists (𝑢 ) contiguous pixels in (𝜑 ) which are 
all greater than (𝐿 1
) or less than (𝐿 2
) such that (𝐿 1
=
𝑓 (𝑥 ,𝑦 )+ 𝜏 ) and (𝐿 2
= 𝑓 (𝑥 ,𝑦 )− 𝜏 ) 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Figure 1: The nearest circle pixels with (16) pixels 
 
3.2 Harris corner detectors (HCD) 
   In (1988), the HCD algorithm was proposed by (Chris 
Harris & Mike Stephens) and the algorithm worked by 
finding (𝐺 (𝑢 ,𝑣 ) ) such that [9] 
𝐺 (𝑢 1
,𝑢 2
)= ∑ ∑ 𝑤 (𝑥 1
,𝑥 2
)[𝑓 (𝑢 1
+ 𝑥 1
,𝑢 2
1
𝑥 2
=−1
1
𝑥 1
=−1
+ 𝑥 2
)− 𝑓 (𝑥 1
,𝑥 2
)]
2
  ,(2) 
 
Where  
𝑤 (𝑥 1
,𝑥 2
) represent the window function 
𝑓 (𝑢 1
+ 𝑥 1
,+𝑥 2
) represents the shifted intensity   
𝑓 (𝑥 1
,𝑥 2
) represents the image intensity  
3.3 Binary robust invariant scalable key 
(BRISK) 
This method gives better rotation stability in photo 
recording for image applications with more blur and 
quickly detected steps. This method also has modularity 
vision to combine with any other key and improve 
performance.  
Region (𝑔 (𝑢 𝑖 ,𝑢 𝑗 ) ) can be defined by the formula [6] 
𝑔 (𝑢 𝑖 ,𝑢 𝑗 )= (𝑢 𝑗 − 𝑢 𝑖 ).
𝐼 (𝑢 𝑗 ,𝜎 𝑗 )− 𝐼 (𝑢 𝑖 ,𝜎 𝑖 )
‖𝑢 𝑗 − 𝑢 𝑖 ‖
2
  ,(3)   
 We note that this formula depends on 
(𝐼 (𝑢 𝑗 ,𝜎 𝑗 )&   𝐼 (𝑢 𝑖 ,𝜎 𝑖 ) ), which can be used to indicate the 
local gradient. 
𝐴 = { (𝑢 𝑖 ,𝑢 𝑗 )∈ ℝ
2
× ℝ
2
| 𝑖 < 𝑁 ∧ 𝑗 < 𝑖 ∧ 𝑖 ,𝑗 ∈ 𝑁 }  ,(4) 
 
A subset of minimum-distance for each pair is defined 
within (𝑀𝐼 ) and another subset includes maximum-
distance for each pair defined within (𝑀𝐴 )  
 
They are defined by the formulas 
𝑀𝐼 = { (𝑢 𝑖 ,𝑢 𝑗 ) ∈ 𝐴 |‖𝑢 𝑗 ,𝑢 𝑖 ‖ <∈
𝑚𝑎𝑥
} ⊆ 𝐴  ,(5  ) 
𝑀𝐴 = { (𝑢 𝑖 ,𝑢 𝑗 ) ∈ 𝐴 |‖𝑢 𝑗 ,𝑢 𝑖 ‖ >∈
𝑚𝑖𝑛 } ⊆ 𝐴  ,( 6  ) 
With   (∈
𝑚𝑎𝑥
= 9.75𝑡  ,∈
𝑚𝑖𝑛 = 13.67𝑡 ) such that (t) 
represents the measure of training through the points 
within (𝐿 ) that includes the distinctive abilities within the 
specified direction and for each of the points. 
By the following equation 
 
𝑔 = (
𝑔 𝑥 𝑔 𝑦 ) =
1
𝐿 . ∑ 𝑔 (𝑝 𝑖 ,𝑝 𝑗 )   ,( 7 )
(𝑝 𝑖 ,𝑝 𝑗 )∈ℒ
 
  
Ref. 
Noise 
detection and 
removal 
Methods 
[2] Gradient 
magnitude of an 
intensity 
Sobel edge 
detection for ROI 
and watershed 
segmentation (2D 
and3D) 
[3] Adam 
optimizer 
algorithm 
Machine 
learning 
[4] CNN-based 
deep feature 
extraction, 
classification 
and localization 
Compare 
among 
supervised, 
unsupervised and 
semi-supervised 
research 
[5] Median 
filter 
Deep CNN 
method 
Current 
study 
Accelerated 
Segment Test, 
Harris Corner 
Detector and 
Binary Robust 
Invariant 
Scalable Key 
Mann-
Whitney Test to 
find the best 
corner detection 

Optimization of Brain Cancer Images with Some Noise Models… Informatica 47 (2023) 11–16 13 
4 Noise models  
Noise always appears in digital images during image 
acquisition, encoding, transmission and processing. It is 
very difficult to remove noise from digital photos without 
prior knowledge of the type of noise, therefore it is 
necessary to review the models in the study of image noise 
reduction techniques. Noise is defined as a random signal 
that is used to destroy most or part of the information. 
There are many types of noise including those discussed 
below. 
4.1 Gaussian noise (GN) 
This is one of the most common noises in images. 
Gaussian noise can be defined by using the most common 
statistical distribution which it is normal distribution.  
It can be defined by the following formula [10] 
 
𝑓 (𝑥 )=
1
√2𝜋 𝜎 2
  𝑒 −
(𝑥 −𝜇 )
2
2𝜎 2
   ,(8) 
 
Where 𝑥  represents the random variable,
μ,σ  represent the parameters. 
4.2 Salt and pepper noise (SPN) 
This noise model is caused in image transmission. The 
following formula represents the noise equation [10] 
 
𝑓 (𝑥 )= {
𝑝𝑎      𝑓𝑜𝑟   𝑧 = 𝑎 𝑝𝑑     𝑓𝑜𝑟      𝑧 = 𝑏 0                    0.𝑤 .
    ,(9) 
 
5 Mann-Whitney test (MWT) 
This is a nonparametric test based on the null hypothesis 
test that assumes that the two random variables (𝑋 ,𝑌 ) 
came from the same statistical distribution (that is, the 
statistical distribution of the first variable is similar to the 
statistical distribution of the second variable) [11]. The 
test is calculated by [12] 
Let (𝑥 1
,𝑥 2
,𝑥 3
,… …,𝑥 𝑛 )     be the first identically 
independent distribution and  (𝑦 1
,𝑦 2
,𝑦 3
,… ,𝑦 𝑛 ) be the 
second identically independent distribution with both 
samples independent of each other. Then, the Mann-
Whitney Test is defined as: 
 
𝜏 = ∑ ∑ 𝜑 (𝑥 𝑖 ,𝑦 𝑗 )
𝑛 𝑗 =1
𝑛 𝑖 =1
   ,(10) 
Where  
𝜑 (𝑥 𝑖 ,𝑦 𝑗 )= {
1     𝑖𝑓   𝑥 𝑖 > 𝑦 𝑗   
0.5     𝑖𝑓     𝑥 𝑖 = 𝑦 𝑗 0      𝑖𝑓  𝑥 𝑖 < 𝑦 𝑗  ,(11) 
 
6 Methods 
6.1 The suggested optimization model 
(SOM) 
The proposed model includes (normal images and 
noise images), where two types of noise were applied 
(Gaussian Noise, Salt and Pepper Noise) through 
three detection algorithms (Harris, BRISK, FAST). 
The following block diagram shows the structure of 
the suggested optimization model. 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Figure 2: Block diagram of the suggested         
optimization model (SOM). 
 
6.2 Experimental dataset 
The study data included a number of magnetic 
resonance images that represent three patients with 
benign brain tumors and three patients with 
malignant brain tumors. Figure 3 shows the images 
of two patients. The left image is for benign brain 
tumors (BT) and the right image for malignant brain 
tumors (MT). 
GN      
SPN 
 
   Start 
Input 
Image 
     
Adding 
Noise 
Yes No 
FAST/HCD/ 
BRISK 
Corner Detection 
 
MWT 
End 
Image Data 
 

14 Informatica 47 (2023) 11–16 A.A.I Al-Kharaz 
 
Figure 3: Experimental data set including (left: benign         
tumors (BT)) and right: (malignant tumors (MT)). 
 
7 Experimental results and discussion 
Applying the proposed system (SOM) to the research data, 
table 2 contains the testing results. 
Table 2: Average of Mann-Whitney test and testing results 
(R: reject; A: accept) 
Corner 
Detection 
Method 
Noise 
Model 
Tumor 
Kind 
MWT P-value Testi
ng 
Resul
ts 
FAST With 
out  
Benign 1.15234
× 10
8
 
0.00187 R 
Malignant 1.22034
× 10
8
 
0.00025 R 
GN Benign 1.88204
× 10
8
 
0.64223 A 
Malignant 1.98023
× 10
8
 
0.73165 A 
SPN Benign 2.25312
× 10
8
 
0.00012 R 
Malignant 2.26434
× 10
8
 
0.53423 A 
HCD With 
out  
Benign 3.54926
× 10
8
 
0.00026 R 
Malignant 4.34589
× 10
8
 
0.00012 R 
GN Benign 3.76723
× 10
8
 
0.00334 R 
Malignant 4.67726
× 10
8
 
0.00654 R 
SPN Benign 7.00598
× 10
8
 
0.00398 R 
Malignant 3.43312
× 10
8
 
0.54434 A 
BRISK With 
out  
Benign 7.41287
× 10
8
 
0.00767 R 
Malignant 7.33287
× 10
8
 
0.00026 R 
GN Benign 4.86498
× 10
8
 
0.00042 R 
Malignant 4.66532
× 10
8
 
0.36078 A 
SPN Benign 3.78554
× 10
8
 
0.55676 A 
Malignant 7.55812
× 10
8
 
0.00346 R 
 
Testing the results depended on the P-value with  
Reject the (Null Hypothesis) if the (P − value < 0.05) 
Accept the (Null Hypothesis) if the (P − value ≥ 0.05) 
With the null hypothesis (𝐻 0
) states that the two random 
variables (𝑋 ,𝑌 ) follow the same statistical distribution 
and the alternative hypothesis (𝐻 1
) states that the two 
random variables (𝑋 ,𝑌 ) follow a different statistical 
distribution. 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
Figure 4: Average Mann-Whitney test and testing results 
for each group 
 
 
 
 
 
 
 
 
Figure 5: P-value and testing results for each group 
 
By observing Table 2 and Figures 4 and 5, it becomes 
clear that the (Harris) method is less affected by (Gaussian 
Noise) and (Salt and Pepper Noise), while the (FAST) and 
(BRISK) methods are affected more by the noise and its 
different models. 
The results showed that all corner detection methods were 
affected by the noise model as the comparisons, according 
to the results of the Mann-Whitney test, (18) compared, 
of which (6) accepted (12) reject for the hypotheses and 
according to their values, and thus the percentages of 
acceptance and rejection were according to the following 
Table 3. 
 
Table 3: The number of accepted and rejected groups 
 
 
The findings of the present study reveal the Harris Corner 
Detectors (HCD) to be the best noise detection method 
using corner pixels, and can be used as a training feature 
for support vector machine along with other features for 
classification [13]. In addition, the Harris Corner method 
performs better with a recognition rate and can be used as 
an effective tool for localization and detection [14].  
 
 
 
 
 
Group Case Number 
of times  
% 
A Reject 12 0.66 
B Accept 6 0.34 
 
 

Optimization of Brain Cancer Images with Some Noise Models… Informatica 47 (2023) 11–16 15 
8 Conclusions and future work 
The experimental results of (SOM) offer many 
conclusions including that corner detection methods are 
influenced by the type of image or image contents. Corner 
detection methods are influenced by the type of noise, and 
the Harris method gives the best results according to the 
overall average. In future work, other noise models such 
as (Rayleigh & Speckle) can be used to understand the 
effect of noise on images. Furthermore, the covariance 
and rotation of images can be changed to see their effect 
on detection methods. Finally, it is possible to use more 
images to achieve more accurate results. 
References 
 
[1] Xiong, Jinhu, Jia‐Hong Gao, Jack L. Lancaster, and 
Peter T. Fox (1996). Assessment and optimization of 
functional MRI analyses.  Human Brain Mapping, 4, 
3, pp. 153-167. https://doi.org/10.1002/(SICI)1097-
0193. 
[2] Ratan R, Sharma S, Sharma S. (2009).Brain tumor 
detection based on multi-parameter MRI image 
analysis, ICGST-GVIP Journal,9, 3, PP.9-17. 
https://doi.org/10.1007/978-3-642-20998-7_38. 
[3] Sawant A, Bhandari M, Yadav R, Yele R, Bendale 
MS (2018). Brain cancer detection from mri: A 
machine learning approach (tensorflow). Brain,5, 4. 
[Google Scholar] 
[4] Arabahmadi M, Farahbakhsh R, Rezazadeh J 
(2022). Deep Learning for Smart Healthcare—A 
Survey on Brain Tumor Detection from Medical 
Imaging. Sensors, 22, 5, PP.1-27. 
https://doi.org/10.3390/s22051960. 
[5] Mohammed, B.A. and Al-Ani, M.S. (2020). An 
efficient approach to diagnose brain tumors through 
deep CNN. Math. Biosci. Eng, 18, pp.851-867. 
https://doi.org/10.3934/mbe.2021045 
[6] Gonzalez RC, Wintz P (1977). Digital image 
processing.  Addison-Wesley Publishing Co, Inc 
Applied Mathematics and Computation. 
https://doi.org/10.1007/978-3-642-95298-2_1. 
[7] Vlaardingerbroek MT, Boer JA (2013). Magnetic 
resonance imaging: theory and practice. Springer 
Science & Business Media. 
https://doi.org/10.1002/(sici)1097-
458x(199709)35:9%3C656::aid-
omr138%3E3.0.co;2-i 
[8] Viswanathan DG (2009). Editor Features from 
accelerated segment test (fast). Proceedings of the 
10th workshop on image analysis for multimedia 
interactive services, London, UK. 
https://doi.org/10.1109/wiamis15130.2009. 
[9] Orguner U, Gustafsson F, editors (2007). Statistical 
characteristics of harris corner detector.  IEEE/SP 
14th Workshop on Statistical Signal Processing; 
2007: IEEE. Pp. 571-575. 
https://doi.org/10.1109/SSP.2007.4301323. 
[10] Bar L, Sochen N, Kiryati N, editors (2005). 
Image deblurring in the presence of salt-and-
pepper noise. International Conference on 
Scale-Space Theories in Computer Vision; 2005: 
Springer. https://doi.org/10.1007/11408031_10. 
[11] Ghisu, T., Parks, G.T., Jaeggi, D.M., Jarrett, J.P. 
and Clarkson, P.J.(2010). The benefits of 
adaptive parametrization in multi-objective Tabu 
Search optimization. Engineering 
Optimization,42,10, pp.959-981. 
https://doi.org/10.1080/03052150903564882. 
[12] Fagerland MW, Sandvik L (2009). The 
wilcoxon–mann–whitney test under scrutiny. 
Statistics in medicine, 28, 10, pp1487-1497. 
https://doi.org/10.1002/sim.3561. 
[13] Taheri, M.(2017). Enhanced breast cancer 
classification with automatic thresholding using 
support vector machine and Harris corner 
detection. 
https://doi.org/10.1145/2987386.2987420. 
[14] Basha, C.Z., Reddy, M.R.K., Nikhil, K.H.S., 
Venkatesh, P.S.M. and Asish, A.V.(2020). 
March. Enhanced computer aided bone fracture 
detection employing X-ray images by Harris 
Corner technique. In 2020 Fourth International 
Conference on Computing Methodologies and 
Communication (ICCMC) (pp. 991-995). IEEE. 
https://doi.org/10.1109/iccmc48092.2020.iccmc
-000184. 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

16 Informatica 47 (2023) 11–16 A.A.I Al-Kharaz