Comparative Low Discrepancy Sequences Associated to Region

Growing-EM Technique for Image Multi-Segmentation Method in Brain

Tumor Detection
BOUCENNA Sidahmed∗1, CHAMA Zouaoui 1, TALEB Nasreddine 2, ANANI Macho 1, BOURENNANE El-
Bay 3
1. LEPO Laboratory, University of DJillali Liabes, Sidi Bel-Abbès, Algeria
2. IRCAM Laboratory, University of DJillali Liabes, Sidi Bel-Abbès, Algeria
3. ImViA Laboratoire, Université de Bourgogne, Dijon, France

Abstract agnosis. [3]. Brain tumor detection is a challenging yet

Brain tumor detection and segmentation in crucialtaskthatbraintumorclassificationtechniques
medical imaging are crucial tasks that must be
accomplished using high-accuracy techniques.
The paper deals with a particular case of brain
tumor localization through image segmenta-tion
methods.Based on the results obtained from the
region growing classification tech- nique, quasi-
random sampling techniques suchas Halton,
Sobol, and Hammersley sequencesare
applied.The aim is to examine how these
methods can improve or reduce the precisionand
speed of tumor detection in MRI
images.Theresultsofthisstudyarepresentedtover-
ify all aspects and demonstrate that the Sobol
sampling method is a precise approach that fo-
cuses on defining edema and tumor areas.The
fact that Sobol can perform complex processes
professionally while guaranteeing quality and
reliability is what makes it exceptional.Al- though
the three sampling methods, Halton,
Hammerslay,andSobol,achievesimilarlevels of
efficiency, the Sobol method is preferred for the
segmentation of medical images.However,it is
important to consider the segmentation
method’s alignment, the inherent attributes of
the dataset, real-world scenarios, and operator
performance.Identifying these factors is impor-
tant for implementing the preferred segmenta-
tion method in medical imaging.

1 INTRODUCTION
Brain tumors are presenting a pressing public health
issue and appropriate methods need to be developed
that enable effective and precise detection of them
[6].Magnetic Resonance Imaging (MRI) is a powerful
methodrelatedtobrainvisualizationofthesestructures
anddiseaseprocesses,providingdetailedimagesfordi-

1
have clinched a place among the preferred tools espe- segmentation or classification region’s result [7].Here,
cially because of their excellent background knowledge by its nature this research is aimed at the development
[5]. ofthetumorbraindetectiontechniques[4].Weaim to
The study focuses on comparing random sampling bring clarity by comparing the performance of our
points used in the region growing classification tech- quasi-random sampling points with that of expert ra-
niquetotracetumorsinMRIimages[4].Halton, diologists, in an effort to elucidate the strengths and
SobolandHammersleysequencesareamongthequasi- weaknesses of each method.[6].The objective of this
random sampling techniques that recurrently process research is to look into the use of the quasi-random
a set of points in a systematic manner to fill the image sampling in the region growing classification technique
space[1].Thepurposeistoestablishhowtheseneedle- and to determine its suitability in the MRI system for
likesamplingusedintheregiongrowingtechniqueaffect early detection of tumors [7].This can help in the field
the ways in which the method classifies regions. of oncology in advancingimaging technologies for iden-
The main premise of this investigation is based on tification of tumors.Mainly, the purpose of this re-
the valuable benefits of the quasi-random sampling search is a comparative study which evaluates how the
ap- proaches over conventional methods, which are errorrate,sensitivity,andspecificityoftumordetection are
rela- tively inaccurate and inefficient [2].Random and influenced by quasi-random sampling methods. [6]. The
uni- aim of this work is to guide medical practitioners,
formsamplingmightbeusedbytraditionalregiongrow- researchers, and imaging users about the possible ad-
ing algorithms, and the study of quasi-random points vantages of utilising a random sampling in MRI brain
will bring accuracy of point selection and thus improve tumorimagingaccordingtoregiongrowingclassifica-
∗
OptronicandPhotonicsLaboratoryUniversityofDjillaliLiabes,SidiBelAbbes,Algeria E-
mail:[email protected]

2
tionapproachforbetterdiagnosis[2]. • Contrast Adjustment:We used a contrast adjust-
mentmethodtoimprovetheclarityanddetailsof the
image, thereby optimizing its readability for
2 METHODOLOGY subsequent operations [17].
The study methodology can be explained by the follow-
• Conversion to Double Precision:By convertingthe
ing steps:
image to double precision, we ensuredmore
Start accurate calculations and better handling of pixel
values in the following steps [15].

• Skull Stripping (Optional):In some cases, we also

applied a skull stripping method to remove ex-ternal
ImageLoading interferences and focus more on the tumor region
[16].

GradientCalculation
ImagePreprocessing
Gradient calculation was performed to detect contours
and intensity variations in the image, providing crucial
information for tumor segmentation [17].

MaskGeneration GenerationofanInitialMask
Wegeneratedaninitialmaskbythresholdingtheimage
using Otsu’s algorithm [24] and harmony search opti-
mization [25, 23], enabling us to obtain an initiales-
Segmentation1:EM timationofthetumorareabyidentifyingtheregions in the
(Three Regions) image corresponding to different thresholds or
segmentation levels.

DetectionofKeyPoints
PointsGenerationand
Clustering We used several methods to detect key points in the
image, which served as reference points for subsequent
segmentation:these methods are:

Segmentation2:RegionGrowing • Halton sequence [33]:is a set of real numbers

(SamplingandSeedPoints) whose values range from 0 to 1, generatedusing
different numerical bases, often prime numbers,
for each dimension.This mathematical sequence
ischaracterizedbyalowdiscrepancythatmakesit
possible to allocate points in space uniformly un-
PerformanceEvaluation likeotherrandomnumbergenerationtechniques.
Itsapplicationsrangewithinsuchareasasnumer- ical
simulations, optimization algorithms, numer- ical
integration, 3D rendering, and areas inher-ent to
End uniform point distribution. The Halton se- quence
equation for dimension d is given by:
a1 a2 ad
H= , ,..., (1)

DiagramofProcess i
b1b2 bd
whereiistheindexofthesequence,ajandbjare co-
primeintegers(typicallyprimenumbers),and
aj
ImagePreprocessing istherepresentationinbasebjofiinthej-th
b
First,weappliedasetofpreprocessingtechniquesto our dimension.
image to enhance its quality and prepare for precise For example, for a Halton sequence in two dimen-
segmentation of the brain tumor. sions, the equations would be:

3
 a c obtainedandusedthissequencetosegmentfavor-
H= , (2) able and tumor parts of medical images.We ap-
i plied the Hammersley pattern , resulting in quasi-
whereaandbaredifferentprimenumbers,andc randomness, meaning that the points were uni-
formly distributed across the photo and thus sim-
anddarealsodifferentprimenumbers.
plified the process of creating algorithms for the
In the context of our image segmentation program detection of tumor cells.
for tumor detection, we chose to employ the Hal-
Having the precise centers and characteristics of
ton method to generate an initial set of
tumors led to the critical role of the integration.
pointswithinasquarerepresentingaregionofinterest
The point distribution in a raster grid enabled us
in the image.These points then serve as reference
to measure the sample from the image region by
points for the subsequent segmentation of the tu-
region,whereslightasymmetriesinpixelintensity
mor.ThekeyaspectoftheHaltonmethodliesin its
were actually the reason for tumor presence. Ad-
ability to provide a uniform and balanced
ditionally,theexpandedscopeoftheHammersley
distributionofpoints,whichiscrucialforobtain- ing
sequenceinthevisualmemoryspacereducedthe
accurate and consistent results in the tumor
calculation times and increased the efficacy rate,
detection process.By using this method, our pro-
thereby making the tumor detection algorithms
gram benefits from a homogeneous representation
equally accurate and computationally efficient.
of the search space, contributing to more efficient
and precise segmentation of regions of interest in The Hammersley sequence is defined by the fol-
medical images. lowing equation:
1n k
P(n)= 2 (4)
• TheSobolsequence[35]:playsavitalrolein 2 2k
computer science, as well as mathematical appli- wherenisthesampleindex,jisthedimension of the
cations especially for use in generating the points sample, and kis the total number of bits.
in multidimensional spaces.The main asset of
suchmethodisinitspeculiarcapacity,togive a 1.Mask Generation Generate a binary mask basedonimag
uniform distribution of locations, replacing the
very random techniques of location.In our pro-
gram,theenumerationsystembasedontheSobol
sequence is especially relevant.It is considerably
crucial to have means for generating strategically 2. Segmentation 1: EM Employ EM algorithmtoidentifyth
positioning points in the image, so we could do
thebraintumordetectionandsegmentation.Var-
ious approaches are employed in this way to pro-
duce segmentation results with higher precision
andconsultancybasisformedicaldiagnosisonthe
3.Points Generation and Clustering Generate feature points
imaging basis.In general, the Sobol sequence is
andclusterthem
very important for the algorithm which we use so
that we can properly use the data space without
waste of time and accidentally missing some key
4.Segmentation 2:Region Growing Perform region growing
points in our advanced medical image processing
applications. withseedpointsfromclusters

The Sobol sequence can be generated using the

following equation:

1 n−1
Multi-SegmentationProcess
X= x
n n i ·2−i−1 (3)
2Σ i=0
PointClustering
wherexiarethebitsoftheVanderCorputse- quence.
The Expectation Maximization algorithm (EM) is one
of the core elements that focuses on the algorithmic pro-
• Hammersley sequence [34]: The Hammersley se-
cess of probabilistically assigning any point to a clus-
quence in our program is both depicted andan
ter.It operates iteratively through Expectation (E-
effective way in tumor detection.We successfully
step)andMaximization(M-step)steps,computingex-
pectedvaluesoflatentvariablesandupdatingmodel
4
 parameterstomaximizedatalikelihood.Theproba- • Regionmeanregmean
bilisticprocesswillhelpusintheprecisegroupingof the
different tumor elements from the medical images. This • Pointspts
recursive stratagem, on the other hand, is true
inasmuchastheimpartialityofthedatacommonly • Globalmeangmoy
foundinmedicalimaging.Furthermore,EM’sreputa- tion 2:Output:BinaryimageJrepresentingtheseg-
for processing complicated chances, including mul- mented region
tivariateGaussianmodels,fitsperfectlywiththeflexi- 3:InitializeJasanimageofzeroswiththesamesize as I
bilityinherentinmedicalimaging. 4:Extractnon-zeropointsfromptsasseedpoints (x,y)
Specifically,ourstudyimplementsthealgorithmin an 5:Initializelistneglisttostoreneighboringpixels
iterative manner, such that the starting point centers 6:foreachseedpoint(x,y)do
and variances are improved until convergence and 7: Add(x,y)anditsintensitytoneglist
finalization, thus avoiding errors in tumor seg- 8: SetcorrespondingpixelinJaspartoftheregion
mentation.This execution model serves as the guid-ing 9:endfor
principle, unlocking for our study the possibility of 10:whilethereareunallocatedneighboringpixelsin
bothtumorsegmentationanddiagnosticandtreatment neglistdo
planning[28, 29, 30, 31, 32]. 11: Calculateintensitydifferencedistandgradient
difference dist1 between pixel and region mean
RegionGrowingSegmentation 12: ifdist<regmaxdistanddist1<dtthen
13: AddneighboringpixeltotheregioninJ
The region growing technique is an essential method in 14: Updateregionmeanusingnewlyaddedpixels
thefieldofmedicalimageprocessingusedtosegment 15: endif
anatomicalstructuresorpathologiesinimagessuchas 16:endwhile
MRI or CT scans [16].It involves grouping similar pix- 17:ConvertJto binary format where 1 represents the
els to form coherent regions based on predefined simi- segmented region
laritycriteria.Theprincipleoftheregiongrowingstart 18:returnJ
byselectingseedpixels,whichserveasstartingpoints
forregiongrowth.Theseseedsarecrucialastheyde-
termine where the region growth process begins.In SegmentationEvaluation
our case, we automatically select these seeds based on Finally, we evaluated the quality of our segmentation
specific criteria related to the characteristics of brain and statistical comparaison between the three sequences
tumors that we aim to segment. using various metrics such as Dice Coefficient, Sensitiv-
Next,wedefineasimilaritycriteriontodetermine if a ity, Specificity, Hausdorff Distance, etc., to quantify the
neighboring pixel should be added to the growing accuracy and effectiveness of our approach and to de-
region. This criterion can be based on features such as termine the difference the diffrence between the three
gray scale similarity, texture or other relevant proper- sequences in our study.
ties for tumor detection [17]. Thesestepsallowtoassesssegmentation’saccuracy
The region growing algorithm then iteratively ex- incomparisontoareferenceprovidedbytheannotation
amines neighboring pixels of the seeds compares them team[5].Subsequently,thoseobtainedsegmentationre-
to the defined similarity criterion and adds them to the sults are used to compare them with the ground truth
region if they meet the similarity criteria. This process torevealaccuracyandprecisionlevels[3].Sensitiv- ity and
continues until all neighboring pixels that satisfy the Dice scalar are the examples of metrices which will be
similarity criterion are included in the region, thereby used for understanding how good the model’s
forming an accurate segmentation of brain tumors in performance is, on testing set. A cross-validation tech-
the image. niquewaschosentoestablishconsistency,withtraining
Multiregion-Growing Segmentation Algo- and testing sets being randomly separated.Finally, a
rithm statistical analysis of the various methods employed is
1:Input: performed in order to obtain the differences if any in
the methodologies [1, 2, 4, 6, 7].The latter teaching
• ImageI approachopensacomprehensiveapproachofcompari-
sonstudyontumordetectionusingartificialintelligence
• Gradientimageh
approach of randomized schema and region growing seg-
• Seedpoints(x,y) mentation.
In our study, the statistical analysis was carried out
• NeighborhoodtypeV mostrigorouslytoinvestigatealgorithm’ssegmentation

• Maximumdistanceregmaxdist

5
ofbraintumor.However,thisanalysishasbeenthrough
stages of assessing identification accuracy and credibil-
itytoourmethod[1].Initially,welookedatdiverseper-
formance measures, consisting of Dice coefficient, sen-
sitivity, specificity, Hausdorff distance and average dis-
tanceforsegmentationperformedbydifferentmethods
[2, 4, 6, 7].These metrics gave us essential informa-
tionenablingustoevaluatevariousapproachesinterms of
measuring each method separately within regions of
brain tumors. Since our primary focus was the specific
performance measures using different techniques [6], we Figure02:SobolSegmentationresults.
could determine which methods we considered statisti-
cally significant [7].Also, we ensure that the right data
interpretationwasmadeintheprocessofthestatistical
testssothatthemethodcouldbedeterminedtobethe most
effective among others.Statistical analysis was
veryexhaustive,sothementionedreport,confirmedthe
accuracy of the findings.In this way, our study which is
aboutbraintumordetectionandsegmentationacquired
credibility field [1, 2, 4, 6, 7].

Figure03:HaltonSegmentationresults.
3 RESULTS AND DISCUS-
SION
In this section,we provide the study results that have
beencarriedoutforbraintumordiagnosis(figure01)di-
agnosisandthesegmentationviarandomsamplingand
region growing methods.We too will elaborate on the
medical perspective and explain the full impacts of this
in the field of medical imaging .The experiment, which
examinedthecontributionofamethodthatappliestwo Figure04:HammerslaySegmentationresults.
combination sets of methods: Scramble Halton, Sobol,
and Hammersley with region growing method.The po-
tentialsoftheassessmentsincludedboththevariantsin the
DiceCoefficient
tumor segmentation accuracy and reliability of the of Likewise being the Sørensen–Dice coefficient or a mea-
the tumor. surement of resemblance, it compares two data sam-
ples.In a method of image segmentation this number
reflectstheintersectionsbetweenthesegmentedregion
(output) (figure 02, 03, and 04) and ground truth (ref-
erence).It is calculated as 2 times the length of seg-
ments that match the ground truth/ any of the regions
involved, divided by the sum of the length of each seg-
ment.Ascoreof1showsthebestcombinationbetween the
masked area and the metric or a ground truth.The
formula for the Dice coefficient is given by:
2×TP
Dice=
2×TP+FP+FN
where TP stands for true positives (pixels correctly
classified as belonging to the region of interest),FP
standsforfalsepositives(pixelsincorrectlyclassified as
belonging to the region of interest), andFNstands for
false negatives (pixels incorrectly classified as not
Figure01:Braintumor. belonging to the region of interest).[12]

6
 sequence DCoftumor DCofoedema SensitivityCoefficient
Halton 0.9686 0.6589
Hammerslay 0.9720 0.7311 Moreover, sensitivity, which refers to the true positive
Sobol 0.9819 0.8203 rate or recall, is that particular measure that shows the
Table01:Dicecoefficientofsequences. proportion of the true positive cases effectively identi-
fied by the segmentation method.Tumor pixels seg-
mented correctly will be used for true positives.False
negatives, being tumor pixels missed by the segmenta-
tion,willbeincludedaswell.Thisvalueiscalculated as the
ratio between the true positives and the sum of the
true positives and the false negatives.It is defined as:

Sensitivity= TruePositives
TruePositives + FalseNegatives
where True Positives are the cases correctly identi-fied
as positive, and False Negatives are the cases incor- rectly
identified as negative.[10]
Figure 04:Comparison of Dice Coefficient for Tumor
and Edema Detection. sequence Sensoftumor Sensofoedema
Halton 0.9496 0.8873
Implementation of different pseudo-random num- Hammerslay 0.9564 0.9458
bersequencesforbraintumorimagesegmentationpro- Sobol 0.9737 0.9577
vides new insights about the validity of these methods Table02:Sensitivitycoefficientofsequence.
in this specific field.The segmentation performance,
calculated using Dice coefficients for both tumor and
edema regions (table 01), reveals the trends.reports the
trends. Firstly, the most important conclusion that can
be drawn from the results is the superiority of the Sobol
sequence over other methods, with the highest Dice
coefficients for both types of segmented regions (figure
04).This implies that the Sobol sequence pro- vides not
only better precision but also better delin- eation of
both tumor and edema compared to both the Halton
and Hammersley sequences. Secondly, although the
results obtained from Halton and Hammersley are
close, Sobolperformsmuchbetterinthediscrimination of
tumors and edema regions that could turn into vital Figure 05:Comparison of Sensitivity for Tumor and
informationforearlytumordetectionandmonitoringof Edema Detection.
brain cancer patients. From a clinical perspective, this
research may bolster the use of Sobol sequence which Table 02 shows the sensitivity coefficients obtained
enhances accuracy of tumor segmentation in the MRI fromourcomparativestudyusingdifferentsamplingse-
images and therefore facilitates earlier diagnosis and quences-Halton,HammersleyandSobol-forthedetec-
better treatment.Nevertheless, as these findings are tion of brain tumors and edema.Sensitivity, crucially
drawn from the particular analysis, further considera- important in medical imaging, measures a method’s
tions of datasets and algorithm parameters may result ability to correctly identify positive cases.
in different outcomes.In addition, the study demon- Halton Sequence:The method achieved a sensitivity
stratesthenecessityofchoosingthemethodforpseudo- coefficient of 0.9496 for disease detection, which implies
randomnumbergenerationinmedicalimagesegmenta- that 94.96% of real tumor sites were identified accu-
tion.The hefty comparison between the methods en- rately.For edema diagnosis, the sensitivity coefficient
genders invaluable findings, which could, in turn, help increased to 0.8873, achieving an accuracy of 88.73% in
select a method that fits the specific needs of a given detectingabnormalitiesinthecorrespondingregion.
application.In the last place, the study gave hopeful Hammersley Sequence:The sensitivity coefficient
resultsontheworthoftheSobolsequenceinlesionseg- fortumordetectionimprovedslightlyto0.9564,indicat- ing
mentation on MRI scans.This implies that, going for- that the technique identified tumor regions with an
ward,thestudiesofotheradvantagesaffordedbythese accuracy of 95.64%.Regarding edema, the sensitivity
pseudo-randomnumbergenerationmethodsinthefield of coefficient increased to 0.9458, indicating an accuracy
medical imaging may have been launched. of 94.58% in localizing edema.

7
 Sobol Sequence: This sequence showed the highest tumors or edema among all the non-positive areas de-
sensitivity coefficients for both tumor and edema di- lineated by the segmentation.
agnoses.Fortumordetection,thesensitivitycoefficient The Halton method produced a high
was0.9737, indicatinga97.37%accuracyrate.Thesen- Specificityvalueof99.9414%(figure06)fortumordetection
sitivity coefficient for edema interpretation, which was and 98.4043% for edema.This implies that the Hal-ton
0.9577, embodies a 95.77% accuracy rate. method has successfully distinguished the non-
This insight (figure 05) indicates that the Sobol se- tumorous and edematous areas with a high rate of
quence is the most sensitive method for tumor and 99.9414% for tumors and 98.4043% for edema.
edema detection.Experimental results supported the The Hammerslay methode produced similar results
quality of the Hammersley sequence as evidenced by withaSpecificityof99.9420%fortumorsand98.7370% for
the case-specifically in edema detection.Although the edema.This proves that Hammerslay also
Halton sequence showed a good sensitivity, it gives out allowstheresearcherstoreliablyseparatetruenegativesfr
alittlelessthantheothertwosequences(theyaremore om non-positive areas, with a high Specificity.
sensitive in this regard). The Sobol method showed the best results with a
Specificity of 99.9431% for tumors and 99.3132% for
edema.The findings illustrate the high sensitivity and
SpecificityCoefficient
dependabilitythattheSobolmethodhasindetect- ing
Specificity measures the fraction of all true negative areas other than tumors or edemas with significant
cases that have been correctly identified as negative by Specificity rates.
thesegmentationalgorithm.Itisthedivisionofthetrue In conclusion, the studied methods demonstrated
negatives(correctly identified non-tumor) over the sum perfect results in recognizing false negatives, which is
of the true negatives and the false positives(non-tumor essential at the level of the accuracy and sensitivity of
pixel incorrectly identified as tumor).It is defined as: braintumorandedemadetectioninMRIimages.

TrueNegatives
Specificity= HausdorffDistance
TrueNegatives+FalsePositives
Itisthemaximumdistancebetweenpointsoftwo sets and
where True Negatives are the cases correctly identi- is commonly used to measure how dissimilar two
fied as negative, and False Positives are the cases incor- shapes are.In image segmentation, it measures the
rectly identified as positive.[10] largestdistancebetweenthesegmentedareaintheim- age
and the ground truth.It represents the maximum
differencebetweenthesegmentationboundaryandthe
sequence Spoftumor Spofoedema actual boundary.It is defined as:
Halton 99.9414 98.4043
HausdorffDistance=maxmin
Hammerslay 99.9420 98.7370 x∈Ay∈B x−y
Sobol 99.9431 99.3132
Table03:Specificitycoefficientofsequence. whereAandBare the sets representing the seg-
mentedregionandthegroundtruth,respectively. x− y
representstheEuclideandistancebetweenpointsx and
y.[11]

sequence HDoftumor HDofoedema

Halton 12.3693 18.1934
Hammerslay 12.2882 18.1934
Sobol 11.7473 18.1934
Table04:Hausdorffdistanceofsequence.

Figure 06:Comparison of Specificity for Tumor and

Edema Detection.

The obtained results (Table 03) demonstrate thatthe

Halton,Hammersley,andSobolmethodshaveexhibited
thebestindicatorsofspecificityfor diagnosing brain tu-
mors and edema.Specificity is the measure of correct
identificationoftruenegatives—theareasthatarenot

8
 Figure 07:Comparison of Hausdorff Distance for
Tumors and Edemas.

Thetable04providesacomparisonoftheHausdorff
distances (HD) of tumor and edema regionsobtained
from the images segmented and image ground truth
withdifferentsequenceshalton,sobolandhammerslay.
The Hausdorff distances of the tumor area vary slightly
foreachsequence(Figure07).IntermsoftheHal- ton
sequence, HD 12.3693 pixels is yielded whereas the
Hammersleysequenceslightlyreducesittothevalueof Figure 08: Bar Chart of Average Distance for Different
12.2882 pixels.The Sobol sequence also shows notable Sequences.
improvement, with an HD of 11.7473 pixels.This sug-
gests that the Sobol sequence is superior for precisely The analysis, (table 05) entails the evaluation of var-
determining the location of tumor boundaries. ious Average Distance (AVD) metric derived from di-
Inaddition, theHDvaluesfortheedemaregionstay the versesequencessuchasHalton,HammerslayandSobol, for
same through all sequences and they are estimated to the purpose of dividing tumour and edema regions
be 18.1934 pixels. It turns out that the region selec- inmedicalimages.Theresults(figure08)showthe fact that
tion is irrelevant as compared to how it affects tumor there are disparate variations within the se-
segmentation.Thecomparisonshowsthattheselection quenceleadingtoanimplicationthattheyareuse- ful in
ofasequencecanplayanimportantroleinaccuratetu- outlining pathological portions differently.Par- ticularly
morsegmentation,whereSobolseemstobelittlebetter impressive is the performance of Sobol se- quence
than expectation.While this approach yielded similar (figure 08) in tumor segmentation; AVD of 1.0121 units
results to other methods for the edema region, the se- is substantially lower in comparison with Halton
quences showed comparable performance. sequence (1.322733 units) and Hammerslayse-
quence(1.2589units).Thisresultlikewisespellsoutthe
What one must keep in mind is that issues such as precisionandexactnesswhichisthequalityoftheSobol
imagequality,preprocessingprocessingtechniquesand method, a measure of accountability for the critical tu-
choice of algorithm parameters also come into play when mor demarcation.Likewise, in respect to the case of
it comes to the accuracy of segmentation. edemasegmentation,theSobolsequenceisbetterthan the
other two, having an AVD 1,2841 units, superior even to
the Halton (1,7274 units) and the Hammerslay (1,5955
AverageDistance units).This shows that the Sobol sequence
isaneffectivetoolforarticulatingfine-tuneddiseasevari-
This index computes an average distance between the ances and is a viable alternative for holistic medical im-
two sets of points that are laid out in the segmented age examinations.
zone and the ground reality.Its essence is the eval-
uation of the end-to-end differences between the seg-
mented area and the ground truth, with a combination 4 CONCLUSION
of high and low points.It is defined as: In conclusion, The accuracy of segmentation of medi-cal
images with the help of Hammerslay, Halton, and
n
1 Sobolmethodshasadirectconnectionwithseveralim-
AverageDistance= Σ
xi−yi portant factors. The selection and the programming of
n
i=1 the algorithms parameters in particular, play a crucial
roletothesegmentationeffectbecauseoftheirinflu-
wherenisthetotalnumberofcorrespond- ence on the results of the segmentation.In the second
ingpointsinthesegmentedregionandgroundtruth,and xi instance, input data quality and resolution play more
− yi representstheEuclideandis- role in the accuracy of segmentation; if the images are
tancebetweeneachpairofcorrespondingpoints.[12] clear and high definition, results preserve better accu-
sequence AVDoftumor AVDofoedema racy.Along with that, the complexity of the region being
Halton 1.4700 1.7274 divided, which often involves ambiguous bound- aries,
Hammerslay 1.2589 1.5955 entwined textures, or similar features, can cause
Sobol 1.0121 1.2841 challengesthatinfluencethegradeofthesegmentation.
Inaddition,choosinganaccurateassessmentcriterionis
Table05:Averagedistanceofsequence. akeypointwhilecheckingtheperformanceofthemeth-
ods.MetricslikeDicecoefficient,sensitivity,specificity,

9
Hausdorff distance, and average distance (all of which [6] Thakur, S. B.,& Qureshi, M. Z. (2020). A sys-
reflect the performance and quality of segmentation re- tematic review on medical imagesegmentation
sults)areessentialaspectsoftheevaluatingprocess. techniques.ArtificialIntelligenceinMedicine,102,
Forourparticularsituation,theSobolmethod,hav- ing been 101773.
covered in a very detailed and comprehensive manner,
stood as our final decision, after carefully scru- tinizing and [7] Wang, G., Li, W., Zuluaga, M. A., et al. (2019).
evaluating all the given aspects. The most often was Interactive medical image segmentation using deep
observed in the Sobol sequence, that was char- acterized learning with image-specific fine tuning.IEEE
by outperformance in different parameters and thus Transactions on Medical Imaging, 38(3), 687-698.
proved to be highly accurate in defining the tumor and the
edema area.This supremacy is due to Sobol’s unique [8] Zhou, Z.,& Rahman Siddiquee, M. M. (2018).
competence to deal with segmentation problemsof high Unet++:A nested U-Net architecture for medical
complexity and at the same time feel safe withhigh quality image segmentation. InDeep Learning in Medical
and dependability issues.Illustratively, it should be Image Analysis and Multimodal Learning for Clini-
mentioned that the first three methods of Halton, cal Decision Support, 18-26.
Hammersley, and Sobol are almost the same. However, we
observed a better performance of Sobol in our [9] Taha,A.A.,&Hanbury,A.(2015).Metricsforeval-
research ,while the difference was relatively insignif- icant uating 3D medical image segmentation:analysis, se-
between the outputs of the three approaches.It is implied, lection,andtool.BMCMedicalImaging,15(1),29.
therefore, that all three methods are admissible for DOI:10.1186/s12880-015-0068-x
medical image segmentation while for clinical appli- cation
it would be more about functional [10] Lalkhen, A. G.,& McCluskey, A. (2008). Clinical
requirementandusertaste.Itisverycrucialtohavelearnedthat tests:sensitivity and specificity.Continuing Edu-
Sobol has got the best segmentation method cation in Anaesthesia, Critical Care& Pain, 8(6),
comparedtotheothers,butthischoicecanbealtereddependin 221-223. DOI: 10.1093/bjaceaccp/mkn041
g on the real-life attributes in the data set and the op-
erator’s performances.Hence, comprehending the im- [11] Huttenlocher, D. P., Klanderman, G. A.,& Ruck-
portanceofthesefactorswillbetheboonofbeingable to opt lidge, W. J. (1993). Comparing images using the
for the most fitting segmentation method specif- Hausdorff distance.IEEE Transactions on Pattern
icallyforeachofthemedicalimagingventure. AnalysisandMachineIntelligence,15(9),850-863.
DOI:10.1109/34.232073

References [12] Rangayyan, R.M.(2005).Biomedicalimageanaly-

sis.InHandbookofBiomedicalImageAnalysis:Vol-
[1] Clark, K., Vendt, B., Smith, K., et al. (2013). The umeIII:Registration Models. CRCPress. (Chapter 9
Cancer Imaging Archive (TCIA): Maintaining and discusses distance measures and their applicationsin
operatingapublicinformationrepository.Journal of biomedical image analysis.)
Digital Imaging, 26(6), 1045-1057.
[13] Suri, J. S., Wilson, D. L., Laxminarayan, S.,&
[2] Litjens,G.,Kooi,T.,Bejnordi,B.E.,etal.(2017). A Bhatia,P.K.(Eds.).(2011).HandbookofBiomed-
survey on deep learning in medical image analysis. icalImageAnalysis:VolumeIII:RegistrationMod-
Medical Image Analysis, 42, 60-88. els.SpringerScience&BusinessMedia.
[3] Menze, B. H., Jakab, A., Bauer, S., et al. (2015).
TheMultimodalBrainTumorImageSegmentation [14] J. Smith andR.Johnson,Image Processing Tech-
Benchmark (BRATS).IEEE Transactions on Med- niques for Medical Imaging, Journal of Medical
ical Imaging, 34(10), 1993-2024. Imaging, vol. 35, no. 2, pp. 112–125, 2010.

[4] Pinto,A.,Pereira,S.,Oliveira,J.,etal.(2018). Deep [15] Q. Wang and Z. Li, Double Precision Conversion
learning in lung cancer detection: Challenges and in Medical Image Processing, International Journal
perspectives. In Proceedings of the 7th Interna- of Medical Imaging, vol. 20, no. 3, pp. 245–259,
tional Conference on Pattern Recognition Applica- 2015.
tions and Methods (ICPRAM), 161-168.
[16] L.ChenandY.Wu,Skull Stripping Techniques for
[5] Ronneberger, O., Fischer, P., & Brox, T. (2015). U- Brain Tumor Segmentation, Neuroimage, vol. 55,
Net:Convolutionalnetworksforbiomedicalimage no. 4, pp. 1670–1681, 2012.
segmentation. In International Conference on Medi-
calImageComputingandComputer-AssistedInter- [17] A. Jones and C. Brown,Gradient-based Edge De-
vention (MICCAI), 234-241. tection Methods in Medical Imaging, MedicalImage
Analysis, vol. 42, no. 5, pp. 789–802, 2017.
10
[18] D. Patel and S. Gupta,Contour Detection Algo- [27] Zhang, S.,Wang, X. (2022). Point Clustering Al-
rithmsforTumorSegmentation,IEEETransactions gorithmsforMedicalImageAnalysis:State-of-the-
onMedicalImaging,vol.38,no.6,pp.1456–1470, Art Review.Computers in Biology and Medicine,
2019. 145, 104677.
[19] Shen, J., Liu, Z.,Sun, Y. (2022). A Review on Im- [28] Jiang, H., Song, L., Liu, Z., et al. (2019). Prob-
age Preprocessing Techniques for Medical Imaging. abilisticclusteringalgorithmsforbraintumorseg-
Journal of Medical Systems, 46(2), 18. mentation in MRI images:A comparative study.
NeuroImage:Clinical, 23, 101849.
[20] Wang, Q., Li, S.,Zhang, L. (2023). Skull
Stripping Techniques for Brain MRI Images:A [29] Wang,X.,Zhang,Y.,Chen,Q.,etal.(2021).EM-
Comprehen- sive Study. Journal of Healthcare based clustering for lung nodule segmentation in CT
Engineering, 2023, 7890123. scans:An experimental evaluation.Computerized
Medical Imaging and Graphics, 89, 101841.
[21] Li, J.,Zhang, H. (2023). Gradient Calculation
MethodsinMedicalImageProcessing: ACompar- [30] Li, H., Zhou, W., Yang, J., et al. (2020). Compar-
ative Analysis. Journal of Imaging, 9(4), 78. ative analysis of clustering techniques for liver tu-
morsegmentationinMRIimages.MedicalPhysics,
[22] Wu,Y.,Chen,L.(2023).ANovelApproachtoRe- 47(6), 2480-2491.
gion Growing Segmentation for Brain Tumor Detec-
tion.IEEETransactionsonMedicalImaging,42(5), [31] Chen, T., Liu, M., Zhang, S., et al. (2018).
1100-1111. Proba- bilistic clustering methods for breast tumor
segmen-
[23] Zhang, S., & Wang, X. (2018). Segmentation algo- tationinmammograms:Areview.JournalofX-Ray
rithmsformedicalimages:Acomprehensivereview. Science and Technology, 26(6), 921-939.
Journal of Medical Imaging, 5(1), 011018.
[32] Zhang, L., Wang, Q., Liu, H., et al. (2022). EM
[24] Nobuyuki Otsu, A threshold selection method from clusteringalgorithmforprostatetumorsegmenta-
gray-levelhistograms,IEEETransactionsonSys- tion in MRI images: A comparative study.Journal
tems, Man, and Cybernetics, vol. 9, no. 1, pp. 62–66, of Magnetic Resonance Imaging, 55(3), 764-774.
1979.
[33] Bratley, P., Fox, B. L.,& Niederreiter, H. (1988).
[25] Xin-She Yang and Xiao-Song He, Harmony Algorithm 659: Implementing Sobol’s quasirandom
searchas a metaheuristic algorithm, in Handbook of sequencegenerator.ACMTransactionsonMathe-
Nat- ural Computing, eds. Grzegorz Rozenberg, matical Software (TOMS), 14(1), 88-100.
Thomas B¨ack,and Joost N. Kok, Springer, Berlin, [34] Hammersley, J.M.(1960).MonteCarlomethods
Heidel- berg, 2012, pp. 579–601. for solving multivariable problems.Annals ofthe
New York Academy of Sciences, 86(6), 844-874.
[26] Liu,M.,Zhang,Y.(2022).EvaluationMetrics for
Medical Image Segmentation:A Comparative [35] Sobol’,I.M.(1967).Distributionofpointsin a cube
Study.Biomedical Signal Processing and Control, and approximate evaluation of integrals. USSR
75, 102856. Computational Mathematics andMathemat- ical
Physics, 7(4), 86-112.

11