Recei ed: 2 June 2023
-
Re ised: 30 July 2023
-
Accep ed: 24 Augus 2023
-
CAAI T ansac ions on In elligence Technology
DOI: 10.1049/ci 2.12276
ORIGINAL RESEARCH
A deep lea ning usion model o accu a e classi ica ion o b ain
umou s in Magne ic Resonance images
Nechi an Asaad Zeba i
1
|Chi a Nadhee Mohammed
2
|Dilo an Asaad Zeba i
3
|
Mazin Abed Mohammed
4,5,6
|Diya Qade Zeeba ee
7
|
Hayda Abdulamee Ma hoon
8,9
|Ka a Hameed Abdulka eem
10
|
Sei edine Kad y
11
|Wa ana Vi iyasi a a
12
|Jan Nedoma
5
|Radek Ma inek
6
1
Depa men o In o ma ion Technology, Lebanese F ench Uni e si y, E bil, I aq
2
Depa men o Compu e Science, Uni e si y o Zakho, Zakho, Ku dis an Region, I aq
3
Depa men o Compu e Science, College o Science, Naw oz Uni e si y, Duhok, Ku dis an Region, I aq
4
Depa men o A i icial In elligence, College o Compu e Science and In o ma ion Technology, Uni e si y o Anba , Ramadi, I aq
5
Depa men o Telecommunica ions, VSB‐Technical Uni e si y o Os a a, Os a a, Czech Republic
6
Depa men o Cybe ne ics and Biomedical Enginee ing, VSB‐Technical Uni e si y o Os a a, Os a a, Czech Republic
7
Depa men o Compu e Ne wo k and In o ma ion Secu i y, Technical College o In o ma ics – Ak e, Duhok Poly echnic Uni e si y, Duhok, I aq
8
In o ma ion and Communica ion Technology Resea ch G oup, Scien i ic Resea ch Cen e , Al‐Ayen Uni e si y, Thi‐Qa , I aq
9
College o Compu e Sciences and In o ma ion Technology, Uni e si y o Ke bala, Ka bala, I aq
10
College o Ag icul u e, Al‐Mu hanna Uni e si y, Samawah, I aq
11
Depa men o Applied Da a Science, No o Uni e si y College, K is iansand, No way
12
Facul y o Comme ce and Accoun ancy, Chulalongko n Business School, Chulalongko n Uni e si y, Bangkok, Thailand
Co espondence
Mazin Abed Mohammed.
Email: mazinalshujea y@uoanba .edu.iq
Funding in o ma ion
Minis y o Educa ion, You h and Spo s o he
Chezk Republic, G an /Awa d Numbe s: SP2023/
039, SP2023/042; he Eu opean Union unde he
REFRESH, G an /Awa d Numbe : CZ.10.03.01/
00/22_003/0000048
[Co ec ion added on 23 Jan 2024, a e i s online
publica ion. The a ilia ion and a ilia ion designa o
is upda ed].
Abs ac
De ec ing b ain umou s is complex due o he na u al a ia ion in hei loca ion, shape,
and in ensi y in images. While ha ing accu a e de ec ion and segmen a ion o b ain u-
mou s would be bene icial, cu en me hods s ill need o sol e his p oblem despi e he
nume ous a ailable app oaches. P ecise analysis o Magne ic Resonance Imaging (MRI) is
c ucial o de ec ing, segmen ing, and classi ying b ain umou s in medical diagnos ics.
Magne ic Resonance Imaging is a i al componen in medical diagnosis, and i equi es
p ecise, e icien , ca e ul, e icien , and eliable image analysis echniques. The au ho s
de eloped a Deep Lea ning (DL) usion model o classi y b ain umou s eliably. Deep
Lea ning models equi e la ge amoun s o aining da a o achie e good esul s, so he
esea che s u ilised da a augmen a ion echniques o inc ease he da ase size o aining
models. VGG16, ResNe 50, and con olu ional deep belie ne wo ks ne wo ks ex ac ed
deep ea u es om MRI images. So max was used as he classi ie , and he aining se
was supplemen ed wi h in en ionally c ea ed MRI images o b ain umou s in addi ion o
he genuine ones. The ea u es o wo DL models we e combined in he p oposed model
o gene a e a usion model, which signi ican ly inc eased classi ica ion accu acy. An
openly accessible da ase om he in e ne was used o es he model's pe o mance, and
he expe imen al esul s showed ha he p oposed usion model achie ed a classi ica ion
accu acy o 98.98%. Finally, he esul s we e compa ed wi h exis ing me hods, and he
p oposed model ou pe o med hem signi ican ly.
This is an open access a icle unde he e ms o he C ea i e Commons A ibu ion License, which pe mi s use, dis ibu ion and ep oduc ion in any medium, p o ided he o iginal wo k is
p ope ly ci ed.
© 2024 The Au ho s. CAAI T ansac ions on In elligence Technology published by John Wiley & Sons L d on behal o The Ins i u ion o Enginee ing and Technology and Chongqing
Uni e si y o Technology.
790
-
CAAI T ans. In ell. Technol. 2024;9:790–804. wileyonlinelib a y.com/jou nal/ci 2
KEYWORDS
b ain umou , deep lea ning, ea u e usion model, MRI images, mul i‐classi ica ion
1
|
INTRODUCTION
The b ain se es as he con ol and communica ion hub o he
body. I does his by u ilising a as ne wo k o connec ions
and neu ons o con ol all o he essen ial p ocesses ha occu
wi hin he body. A b ain umou is a condi ion ha could cause
li e‐ h ea ening consequences. The b ain's mal unc ion is
caused by he mul iplica ion o abno mal cells wi hin he b ain,
which also a ec s he neu ological sys em and he cen al
spine. I expands slowly, has clea ly delinea ed bo de s, and
a ely mo es beyond hem [1]. Because he e a e billions o
ac i a ed cells in he b ain, analysis can be p e y challenging.
B ain umou s a e one o he leading causes o mo ali y
among people. A ound 250,000 people wo ldwide a e diag-
nosed wi h p ima y b ain umou s e e y yea ; howe e , hese
umou s accoun o ewe han 2% o all cance s [2]. Con-
sis ency is p esen in benign umou s, which do no include any
cance cells ac i ely sp eading; in con as , malignan umou s
ha e s uc u al he e ogenei y and ac i e cance cells. Malignan
umou s also con ain ac i e cance cells. E en hough benign
umou s g ow slowly, he e is always a chance ha hey will
p og ess in o dange ous malignan umou s. Low‐g ade u-
mou s a e ano he e m ha may e e o benign umou s.
Low‐g ade umou s, e e ed o as benign umou s, can be
u he di ided in o wo dis inc ypes: gliomas and meningi-
omas. Glioblas oma and as ocy oma a e he wo ypes o high‐
g ade b ain umou s ha all in o he same ca ego y as ma-
lignan umou s, also known as high‐g ade b ain umou s. A
benign umou is no haza dous since i is no p e alen in
o he po ions o he b ain. Howe e , malignan umou s a e
inc edibly ha m ul because o hei quick g ow h. They can be
emo ed, and once hey a e gone, hey a ely come back.
Cance ous umou s can be di ided in o se e al unique sub-
ypes acco ding o a ious c i e ia, including he g ow h si e,
he deg ee o i s malignancy, and he kind o issue om which
i o igina ed [3].
Many ea men s a e a ailable o b ain umou s, and he
ea men chosen will ely on he posi ion, size, and kind o
umou . Due o he lack o unin ended side e ec s associa ed
wi h su ge y, i is now he mos common and p e e ed emedy
o b ain umou s compa ed o o he ea men s [4]. The
Magne ic Resonance Imaging (MRI) echnique is he mos
e ec i e and widesp ead in b ain umou diagnosis. The MRI
echnique p o ides in o ma ion on he s uc u e o human so
issue, which can be enla ged o o e p ecise iews in e e y
di ec ion. In he ield o medical imaging, MRI is applied o
display dis inc ions in a ious so issues [5]. The MRI image
can be in e p e ed o indica e he p esence o b ain umou s.
Magne ic Resonance Imaging images can ha e a ious cha -
ac e is ics, depending on he in e nal ana omical s uc u es
being examined. The MRI echnique is well‐known o i s
supe io pic u e cla i y and p ecise umou appea ance [6].
Diagnosing a b ain umou equi es a signi ican in es men o
ime and depends signi ican ly on he expe ience and expe ise
o he adiologis . Inc easing he pa ien numbe will cause a
signi ican inc ease in he amoun o da a ha has o be p o-
cessed. This has caused con en ional me hods o be ine icien
as well as expensi e. P oblems a ise due o signi ican a ia-
ions in umou size, o m, and se e i y and he p esen a ion o
o he diseases. These a iances can occu e en wi hin he same
pa ien . Con en ional manual me hods o iden i ying b ain
umou s and moni o ing hei p og ession o e ime a e
labo ious and e o ‐p one [7]. As a esul , au oma ed solu ions
a e necessa y o eplace con en ional human p ocedu es.
Recen e olu ions in a i icial in elligence and machine
lea ning (ML) ha e made i possible o diagnose b ain malig-
nancies ea lie . The e has been a ecen up ick in in e es in he
e olu ion o au oma ed sys ems o he analysis o images o
o e come he es ic ions ha come wi h manually diagnosing
pa ien s. Recen yea s ha e seen he de elopmen o se e al
compu e ‐aided diagnosis sys ems in ended o diagnose b ain
umou s au oma ically. Many academics ha e esea ched how
o classi y b ain umou s and implemen ed a ious me hods o
e alua e MRI scans o ex ac po en ial cha ac e is ics om he
da ase using ML and Deep Lea ning (DL) echniques. The
undamen al objec i e o his esea ch and analysis is o
imp o e p ocesses o he ea ly de ec ion o b ain umou s.
When de ec ing b ain cance s, p edic ing he cou se o he
umou , and ea ing b ain umou s, he au oma ic classi ica-
ion o medical images is signi ican [8]. When a b ain umou
is ound a an ea ly s age, i is mo e likely o ha e a quick e-
ac ion o ea men , which assis s in inc easing he su i al a e
o pa ien s. I equi es signi ican ime and labou o sea ch o
and ca ego ise images o b ain umou s manually con ained in
sizeable medical image a chi es [9].
Misdiagnosing a b ain umou can ha e signi ican e-
pe cussions and lowe he indi idual's likelihood o su i al. In
ligh o he g a i y o he si ua ion, he e is an u gen need o a
echnique ha elies en i ely on au oma ed p ocesses o iden-
i y b ain umou s. The manual p ocedu e o assessing se e al
pho og aphs p oduced in a clinic is a labou ‐in ensi e, ime‐
consuming app oach ha mus be e ised o comp ehend
how a ious cance s beha e ully. I is necessa y o de elop
echnology o de ec ing and iden i ying umou s ha a e mo e
accu a e and compu e ‐based o unde s and and sol e his
complica ed issue. Nowadays, a ious a emp s ha e been un-
de aken, all o digi ise his p ocess o in es iga e he many
di e en machine‐lea ning app oaches ha a e now a ailable.
The e has been a enaissance in in e es in he u ilisa ion o
deep ne wo ks o he iden i ica ion o cance cells in a manne
ha is bo h mo e accu a e and dependable [10]. Nume ous
me hods ha e been a emp ed o de elop an accu a e and
eliable compu e ised diagnosis o b ain umou s. Ye , i e-
mains challenging due o signi ican in e ‐and in a‐shape
a iances, ex u e di e ences, and con as di e ences. T adi-
ional ML me hods ely on manually p oduced ea u es,
ZEBARI ET AL.
-
791
24682322, 2024, 4, Downloaded om h ps://ie esea ch.onlinelib a y.wiley.com/doi/10.1049/ci 2.12276 by Technical Uni e si y Os a a, Wiley Online Lib a y on [25/09/2024]. See he Te ms and Condi ions (h ps://onlinelib a y.wiley.com/ e ms-and-condi ions) on Wiley Online Lib a y o ules o use; OA a icles a e go e ned by he applicable C ea i e Commons License
limi ing he solu ion's abili y o wi hs and challenges, while
DL‐based me hods can au oma ically ex ac use ul in o ma-
ion o imp o ed pe o mance. The no el y o his wo k lies
in he inno a i e combina ion o con olu ional deep belie
ne wo ks (CDBN) and ResNe 50 models as pa o a usion
model o b ain umou classi ica ion. The usion p ocess
in elligen ly combines local and global ea u es cap u ed by
CDBN and ResNe 50, espec i ely, esul ing in a mo e
comp ehensi e and disc imina i e ep esen a ion o b ain u-
mou s. This s udy de eloped a model‐based DL me hod o
classi y b ain cance s in o gliomas, meningiomas, pi ui a y u-
mou s, o no‐ umou classes. Ou con ibu ions include he
de elopmen o a DL‐based p ocess ha au oma ically ex ac s
aluable knowledge o imp o ed pe o mance in classi ying
b ain cance s, add essing he challenges posed by adi ional
ML me hods ha ely on manually p oduced ea u es. The key
con ibu ions o his s udy a e:
�The s udy de eloped a ully au oma ic app oach o mul i‐
classi ying b ain umou s using deep con olu ional neu al
ne wo ks (CNN) o ea u e ex ac ion.
�Deep CNN ne wo ks we e u ilised o ex ac he mos
ele an in o ma ion ea u es om he image and o be e
gene alisa ion. A model was buil o inco po a e deep ea-
u es and de elop he usion model o achie e be e pe -
o mance in he mul i‐classi ica ion accu acy o b ain
umou s.
�De elopmen o a DL usion model o eliably classi y b ain
umou s, signi ican ly inc easing classi ica ion accu acy
pe o mance.
�U ilisa ion o da a augmen a ion echniques o inc ease he
da ase size o aining models, enabling he DL model o
achie e good esul s.
�Compa ison o he p oposed model wi h exis ing me hods
demons a ed ha he p oposed model ou pe o med hem
signi ican ly, indica ing he e ec i eness o he p oposed
DL usion model o b ain umou classi ica ion.
The emaining sec ions o he pape a e laid ou as ollows.
A discussion o he mos ecen ela ed wo ks is p esen ed in
Sec ion 2. The p oposed mechanism includes all i s phases in‐
dep h, as discussed in Sec ion 3. The expe imen al esul s o
he p oposed model and a compa ison wi h he mos ecen
exis ing wo ks in he li e a u e a e gi en in Sec ion 4. A dis-
cussion ela ed o he p oposed me hod and p o ided esul s
ha e been p esen ed in Sec ion 5. The las Sec ion o his pape
discusses his esea ch's conclusion and ea u e wo k.
2
|
RELATED WORK
O e he pas decade, ML and DL me hods ha e signi ican ly
con ibu ed o medical image analysis, making medical pe -
sonnel's wo k mo e e o less and e icien . When add essing
issues ela ed o b ain umou s, mos esea ch publica ions
ha e ocused on DL me hods such as CNNs, ans e
lea ning, and classical neu al ne wo ks. These app oaches
ha e p o en e ec i e in p o iding solu ions o b ain umou
classi ica ion.
A model‐based CNN was p esen ed by Khwaldeh e al.
[11] o dis inguish be ween umou and non‐ umou b ain MRI
images in addi ion o high and low g ades o glioma umou s.
They achie ed an accu acy a e o 91% by modi ying he
AlexNe ne wo k and u ilising i as he ounda ion o hei
ne wo k a chi ec u e. Despi e he c i ical wo k being done in
his ield, mo e e o is s ill equi ed o es ablish a obus and
p ac ical s a egy o classi ying MRI images o he b ain. Das
e al. [12] success ully iden i ied many o ms o b ain malig-
nancies using a CNN con aining 3064 MRI images. These
umou s included glioma umou s, meningioma umou s, and
pi ui a y umou s. In aining, he CNN ne wo k was ained o
use se e al con olu ional and pooling p ocesses. They could
do his by scaling he con olu ional ne wo k o co espond
wi h con olu ional il e s and ke nels o a ying sizes. As a
esul , hey achie ed an accu acy a e o 94.39%. In an in e-
g a ed app oach, Hashemzehi e al. [13] examined he abili y o
a CNN and neu al au o eg essi e dis ibu ion es ima ion
model o de ec b ain umou s om MRI da a. They employed
3064 images, which hey assessed o de e mine wi h a 96%
le el o accu acy h ee dis inc ypes o b ain cance .
A mul i‐pa hway CNN design was p esen ed by F ancisco
e al. [14] o segmen se e al kinds o umou s au oma ically.
They es ed hei sugges ed model by applying i o a T1‐
weigh ed, con as ‐enhanced MRI da ase made accessible o
he public and ound ha i had an accu acy o 97.3%. In
con as , he p ocess o model aining was qui e expensi e. A
deep incep ion esidual ne wo k was p oposed in e e ence
[15], which discussed a h ee‐class classi ica ion sys em o
b ain umou s. The laye ha is conside ed o be he ne wo k's
ou pu in ResNe V2 has been modi ied. The p oposed
app oach achie es he highes possible accu acy o classi ying
b ain umou s. The sugges ed model was alida ed by applying
i o a b ain imaging da ase ha can be accessed openly online
and con ains 3064 images. The pe o mance esul o he
p esen ed me hod is 99.69% be e han he mos ad anced
echniques in he li e a u e. Abd El Kade e al. [16] c ea ed a
deep‐CNN model o iden i y b ain umou s as no mal and
abno mal o MRI images. Employing a di e en ial ope a o in
he ained ne wo k allows o he gene a ion o an addi ional
di e en ial ea u e map, which can be ob ained om he CNN
ea u e map ha was ini ially used. The consequence o using
he app oach o de i a ion is an imp o emen in he e ec-
i eness o he me hod being p o ided, which is de e mined by
he ou come o he e alua i e pa ame e .
To classi y b ain umou s, ImageNe ‐based Vision T ans-
o me s (ViT) ha ha e been lea n and op imised we e sug-
ges ed by Tummala e al. [17]. Pe o mance alida ion, as well
as es ing, was conduc ed using h ee di e en classes o
umou da ase ha was ob ained om Figsha e. This da ase
includes 3064 T1w con as ‐enhanced MRI images ha con-
ained h ee ypes o umou s. A a size o 384 �384 pixels,
he L/32 model pe o med he bes , ea ning a sco e o 98.2%
on o e all accu acy. The e iciency o he ensembles o all ou
ViT me hods a he exac esolu ion exceeded he e iciency o
792
-
ZEBARI ET AL.
24682322, 2024, 4, Downloaded om h ps://ie esea ch.onlinelib a y.wiley.com/doi/10.1049/ci 2.12276 by Technical Uni e si y Os a a, Wiley Online Lib a y on [25/09/2024]. See he Te ms and Condi ions (h ps://onlinelib a y.wiley.com/ e ms-and-condi ions) on Wiley Online Lib a y o ules o use; OA a icles a e go e ned by he applicable C ea i e Commons License
each echnique and hei ensemble a esolu ion 224 �224.
The cos umes o he ou di e en ViT algo i hms achie ed a
es ing accu acy o a ound 98.7%. To di ide 3260 images in o
ou dis inc classes, Nayak e al. [18] de eloped an E i-
cien Ne model ha uses no malisa ion. This was done o
classi y he images. The newly cons uc ed model is a modi-
ica ion o E icien Ne ha includes adding wo o he laye s:
dense and d op‐ou . Simila ly, hey u ilised da a augmen a ion
in conjunc ion wi h no malisa ion. One o he ad an ages o
using a dense CNN model is ha i can e ec i ely classi y a
cons ained image da abase. As a consequence o his, he
echnique ha was sugges ed o e s ou s anding o e all pe -
o mance. Acco ding o he indings o he expe imen s, he
p oposed model had an accu acy o 99.97% when i was being
ained, and i had an accu acy o 98.78% when i was being
es ed. The egion‐based con olu ional neu al ne wo k
(RCNN) me hod was u ilised o c ea e a new ne wo k o b ain
umou iden i ica ion, and his a chi ec u e was alida ed by
using wo da ase s eely a ailable on Kaggle [19]. Resea che s
p o ided a echnique o diagnosing b ain umou s ha uses a
less complex ne wo k o minimise he equi ed ime o p o-
cessing by a con en ional RCNN. They i s employed a wo‐
channel CNN, a low‐complexi y model ha enhances accu acy
by 98.21%. This allowed hem o dis inguish be ween heal hy
and unheal hy umou MRI images. A e ha , his model is
used as a ea u e ex ac ion model in an RCNN o de ec
umou egions in a glioma class om he da ase classi ied
om an ea lie phase. Then, he umou egion is enclosed in
boxes. This me hod has also been applied o ea meningi-
omas and malignancies o igina ing in he pi ui a y gland. Thei
app oach, which had an o e all con idence le el o 98.8%, was
able o a ain a sho execu ion ime compa ed o o he
me hods conside ed o be s a e‐o ‐ he‐a .
The e a e nume ous me hods o iden i ying b ain u-
mou s, bu se e al limi a ions ha e ye o be de e mined. The
p ima y objec i e o his esea ch is o p o ide an e ec i e
diagnosis me hod o b ain umou s by u ilising di e en CNN
ne wo ks. Due o a lack o eadily a ailable knowledge and
da a, many adiologis s need help wi h his classi ica ion. P e-
ious esea ch has shown ha DL app oaches p o ide much
highe accu acy o b ain MRI classi ica ion han classical ML
echniques. Howe e , a as amoun o da a is equi ed o ain
DL models, which canno be achie ed wi h adi ional ML
echniques. Recen esea ch has es ablished DL app oaches as
a mains eam componen o specialis and in elligen sys ems
and medical image analysis. When dealing wi h b ain umou
classi ica ion, i is essen ial o conside he many cons ain s
associa ed wi h he discussed me hodologies.
3
|
PROPOSED MODEL
This sec ion desc ibes ou p oposed app oach in gene al.
Then, in he succeeding subsec ions, we desc ibed he speci ics
o ou c i ical componen s in de ail. The p oposed sys em
includes p ep ocessing (such as image smoo hing, c opping,
esizing, and no malisa ion), da a augmen a ion, DL ea u e
ex ac ion, deep ea u e usion, and classi ica ion. These i e
undamen al phases make up he p oposed app oach. Figu e 1
p esen s a diag amma ic ep esen a ion o he gene al diag am
FIGURE 1 Rep esen a ion o all s eps o he p oposed model o b ain umou iden i ica ion.
ZEBARI ET AL.
-
793
24682322, 2024, 4, Downloaded om h ps://ie esea ch.onlinelib a y.wiley.com/doi/10.1049/ci 2.12276 by Technical Uni e si y Os a a, Wiley Online Lib a y on [25/09/2024]. See he Te ms and Condi ions (h ps://onlinelib a y.wiley.com/ e ms-and-condi ions) on Wiley Online Lib a y o ules o use; OA a icles a e go e ned by he applicable C ea i e Commons License
o ou model o iden i ying b ain umou s. The pe o mance
e alua ion o he sugges ed model has been es ed using T1‐
weigh ed da ase s. S anda d pe o mance measu emen s, such
as accu acy, ecall, p ecision, and F1‐sco e, we e employed o
es he pe o mance o he p oposed model.
3.1
|
P ep ocessing
Image p ep ocessing can also add ess o he signi ican in insic
MRI acquisi ion abno mali ies, such as in ensi y non‐
uni o mi y and Gaussian noise [20]. This is possible because
o he na u e o MRI. In ensi y non‐uni o mi y, in ensi y in-
homogenei y, o bias o he magne ic ield will show up as a
low‐ equency signal on an MRI due o a ious causes, such as
luc ua ion o he magne ic ield. Se e al di e en hings can
cause his signal. In o he olumes, di e en p ep ocessing
pipelines we e gi en so ha he ela ionship be ween image
p ep ocessing pa e ns and noise il e ing o MRI could be
s udied. This in es iga ion aimed o imp o e he ep oduc-
ibili y and eliabili y o adionics cha ac e is ics. In he MRI‐
based adiomics analysis, he lack o p e‐speci ic image no -
malisa ion me hods may impac he ea u es' consis ency and
dependabili y.
The Gaussian il e (GF) and he skull s ipping me hod
a e he p ima y s eps du ing he p e‐p ocessing o he p o-
posed model. The GF me hod has se e al ad an ages,
including educed noise, a mo e manageable design p ocess,
au oma ic il e ing, and o a ional symme y [20]. The image
can ha e Gaussian noise, sal and peppe noise e c. The in-
o ma ion con ained in ou da ase is p ese ed in a way ha is
analogous o noise emo al applica ions. To educe noise om
he image, he GF is used. This il e u ilises a 2D Gaussian
dis ibu ion unc ion, and his unc ion can be desc ibed as
ollows:
Gði;jÞ ¼ 1
2πσ2e−i2þj2
2σ2ð1Þ
Whe e he s anda d de ia ion dis ibu ion is ep esen ed by σ,
he alue o a Gaussian ke nel is used when designing a con-
olu ional il e o achie e he desi ed e ec . Magne ic Reso-
nance Imaging images need he applica ion o con olu ional
il e s o e e y single pixel in he image. In MRI images, ma ix
mul iplica ions a e pe o med on each pixel's b igh ness and
ke nel elemen s. This helps de ec b ain umou s ea lie . As a
consequence, he noise p esen in he MRI is elimina ed, and
a e ha , he image is e ined wi h he u ilisa ion o Gaussian
il e ing.
The pe o mance o iden i ying almos all o he images in
MRI da ase s o b ain umou s could be be e due o he
p esence o unwan ed oids and a eas in he images. Thus, i is
essen ial o c op he images o emo e any unnecessa y pa s
and ex ac only he ele an da a om he image. This wo k
employed he c opping me hod desc ibed in e . [21], which
in ol es calcula ing ex eme poin s. I is ad ised ha he MRI
images in ou da ase be esized o ha e he same wid h and
heigh o ob ain he bes possible esul s. This is because he
wid hs, heigh s, and sizes o he MRI images in ou da ase
a y. Since he size o he images ed in o Deep ne wo ks is
224 �224 pixels, we educed he MRI images o 224 �224
pixels o his ask. The excep ion is CDBN, which equi es he
inpu images o ha e a size o 128 �128 pixels.
The MRI image om he pa ien 's da abase needs o
p o ide mo e cla i y. The b ain umou s in MRI images
con ain a ce ain amoun o ambigui y. As a esul , b ain im-
ages equi e no malisa ion be o e any u he p ocessing can
ake place. Usually, images ob ained om an MRI a e g ayscale
in appea ance. As a esul , he images may be eadily no -
malised, which assis s in easing image quali y and educes he
likelihood o making e o s in classi ica ion. Nayak e al. [18]
used he membe ship unc ion wi h he mo phological no ion
o iden i y b ain cance s. The ollowing is an example o he
membe ship unc ion ha was u ilised o he s udy:
¼d−mn
mx −mn ð2Þ
While is conside ed a no malised image, i equals a
double image and is equal o min (min(image)) and max(max
(image)), espec i ely. Wi h a ange om 0 o 1, his mem-
be ship unc ion's p ima y pu pose is o s anda dise he image
in p epa a ion o enhancemen .
3.2
|
Deep ea u e ex ac ion
The deep ea u es we e ex ac ed om he MRI scans o b ain
umou s using h ee di e en deep‐lea ning ne wo ks:
VGG16, ResNe 50, and CDBN. A e ha , he ea u es a e
in eg a ed o p oduce da a ep esen a ions ha a e mo e ac-
cu a e. Because ea u e desc ip o s based on DL ha e he
po en ial o lea n undamen al ea u es au oma ically con ained
wi hin an MRI image, he u ilisa ion o hese desc ip o s can
conside ably lessen he necessi y o handc a ed ea u e
ex ac ion. This wo k aims o de ise an app oach o dis-
inguishing be ween MRI images o a ious ypes o b ain
umou s. A he end o he p ocess, hese ep esen a ions o
he ea u es a e ed o he classi ie laye , which hen ca ego-
ises hem in o one o se e al ca ego ies [22, 23].
3.2.1
|
ResNe 50
One o he mos sophis ica ed deep ne wo ks cu en ly a ail-
able o image classi ica ion is called ResNe 50. ResNe 50 can
handle he di icul y o imp o ing pe o mance accu acy while
concu en ly dec easing he pa ame e s whene e he model
canno be ained. Thus, a ea u e ex ac o ‐based deep con-
olu ional ha we u ilise is ResNe 50 [24]. ResNe 50 inc eases
he ne wo k's dep h, and he classi ica ion pe o mance based
on esidual lea ning is no impac ed. The emaining modules
each ha e wo possible ou pu s: one p oduces he esidual o
he inpu ea u e by pe o ming wo o h ee con olu ions on
794
-
ZEBARI ET AL.
24682322, 2024, 4, Downloaded om h ps://ie esea ch.onlinelib a y.wiley.com/doi/10.1049/ci 2.12276 by Technical Uni e si y Os a a, Wiley Online Lib a y on [25/09/2024]. See he Te ms and Condi ions (h ps://onlinelib a y.wiley.com/ e ms-and-condi ions) on Wiley Online Lib a y o ules o use; OA a icles a e go e ned by he applicable C ea i e Commons License

ha ea u e, and he o he makes he ou pu di ec ly om he
supplied ea u e. These wo pa hs' combined esul s a e
conside ed he esidual module's ou pu . Figu e 2 e eals ha
he ex ac ion o ea u es was accomplished wi h he help o
ResNe 50. When he ne wo k is ini ialised, he weigh s p e i-
ously p e‐ ained on MRI b ain umou da ase s a e employed.
The lea ned weigh s o deepe laye s, such as ResNe 50's
en i ely linked laye , end o be mo e class‐speci ic. We we e
cu ious abou how well he ou pu ec o s o he o me
con olu ional laye s migh be ca ego ised, and hen we an
some es s on hem. When u ilised p ope ly, ne wo ks ha
ha e deep con olu ional laye s a e key ea u es. As ea u e
ec o s, we used he inal esidual uni ou pu s gene a ed by he
con olu ional laye s 3, 4, and 5. The dimensions o he ea u es
ound in he hi d laye a e less han hose ound in he i h
laye . Table 1demons a es ha we conduc ed ou analysis
using ResNe 50.
3.2.2
|
CDBN model
When u ilising a Con olu ional DBN ne wo k, i is possible o
c ea e alid p obabilis ic conclusions e icien ly om he
bo om up and he op down [25]. On op o his s uc u e a e
nume ous laye s o max‐pooling con olu ional es ic ed
bol zmann machines (CRBMs), and aining is ca ied ou
using he g eedy laye ‐wise app oach, jus as i would be in a
ypical DBN. Using a CDBN model, one can lea n high‐le el
p ope ies such as s oke g oups o objec sec ions. Du ing
he es phase, we ained he CDBN u ilising wo laye s o
CRBM and elied on eed‐ o wa d app oxima ion o make ou
in e ences. On op o he CRBM is whe e he CDBN was
cons uc ed. I is easible o each he CDBN me hod by
ca ying ou a sequence o CRBMs, each eeding in o ano he
CRBM in he aining p ocess. The s uc u e o he CDBN is
depic ed in Figu e 3; i s isible and hidden laye s a e linked by
g oups o local and sha ed cha ac e is ics o he CRBM's
a chi ec u al design.
To de i e deep ea u es om an image, he DBN applies
se e al laye s o RBM on op o each o he in a mul ilaye .
Equa ion (3) p o ides his in o ma ion isually by g aphically
ep esen ing he join po en ial alloca ion in he isible laye
be ween he inpu da a and he l‐laye hidden laye hk. The
unsupe ised g eedy me hod is used o calcula e he weigh o
he da a. T aining he RBM's i s laye o calib a e i s i s
laye 's aining pa ame e s is he i s hing ha mus be done.
A e ha , he ou pu o he hidden laye o he i s RBM laye
is u ilised as he inpu o he RBM o he second laye , and he
pa ame e s o he i s laye a e g adually lea n . The inal
hidden laye is linked o he So Max eg ession classi ie , and
he supe ised G adien Descen algo i hm is used o inish he
ine‐ uning [26].
W�x;y1
;y2
;…;yl�¼ X
l−2
n¼0
Wyn��ynþ1�!W�yl−1
;yl�ð3Þ
P(y
l−1
,y
l
) has been desc ibed as he dis ibu ion o he p ob-
abili y be ween he wo le els o he opmos RBM, which a e
known as he isible and he bu ied laye s.
This wo k u ilises h ee con olu ional laye s wi h h ee
max‐pooling laye s. Table 2summa ises he pa ame e s o he
CDBN model in his s udy. We de e mined 2 �2 as he ke nel
window size in his wo k. Fo aining images, he numbe o
il e s in each laye has been inc eased by including a mo e
FIGURE 2 The ep esen a ion o implemen ed a chi ec u e o ResNe ‐50 ne wo k.
TABLE 1A chi ec u e o implemen ed RenNe ‐50 model o b ain
umou classi ica ion.
Laye Ou pu size 50‐Laye
Con 1 112 �112 7 �7�64, s ide 2
3�3 max pooling, s ide 2
Con 2‐�56 �56 2
4
1�1�64
3�3�64
1�1�256 3
5�3
Con 3‐�28 �28 2
4
1�1�128
3�3�128
1�1�512 3
5�4
Con 4‐�14 �14 2
4
1�1�256
3�3�256
1�1�1024 3
5�6
Con 5‐�7�72
4
1�1�512
3�3�512
1�1�2048 3
5�4
FLOPs 1 �1 A e age pooling, 1000 FC, so max
Abb e ia ion: FC, ully connec ed.
ZEBARI ET AL.
-
795
24682322, 2024, 4, Downloaded om h ps://ie esea ch.onlinelib a y.wiley.com/doi/10.1049/ci 2.12276 by Technical Uni e si y Os a a, Wiley Online Lib a y on [25/09/2024]. See he Te ms and Condi ions (h ps://onlinelib a y.wiley.com/ e ms-and-condi ions) on Wiley Online Lib a y o ules o use; OA a icles a e go e ned by he applicable C ea i e Commons License
complex pa e n. In he es ing phase, he CDBN model u i-
lised an image size o 128 �128 and a ba ch size 200.
3.2.3
|
VGG16 model
VGG16 is a deep con olu ional neu al ne wo k wi h 16 laye s,
and i is known o i s simplici y and e ec i eness. The
ne wo k akes an inpu image o a ixed size (224 �224 �3).
VGG16 consis s o mul iple con olu ional blocks, each con-
aining con olu ional laye s ollowed by max‐pooling laye s [8].
The con olu ional laye s ha e small 3 �3 il e s, and he max‐
pooling laye s use 2 �2 il e s wi h a s ide o 2 o down-
sample he spa ial dimensions. A e se e al con olu ional
blocks, he ne wo k has ully connec ed laye s o make he inal
p edic ions. The inal laye o he ne wo k p oduces he clas-
si ica ion p obabili ies o he di e en classes. The ep esen-
a ion o he implemen ed a chi ec u e o he VGG16 Ne wo k
is p esen ed in Figu e 4.
3.3
|
Fusion model
A mo e obus classi ica ion is achie ed h ough he usion
echnique, which in u n helps o inc ease he pe o mance o
he classi ica ion. Recen s udies demons a e ha he usion
p ocess imp o es classi ica ion pe o mance; ne e heless, he
mos signi ican downside is ha i akes much ime o
compu e, al hough ou p ima y objec i e is o acili a e he
pe o mance e alua ion o b ain umou s. Figu e 5p o ides a
isual ep esen a ion o he p oposed echnique o he usion
p ocess.
The classi ica ion sco es a e combined in o a single inal
sco e using his esea ch‐based sco e‐le el usion me hod,
which mo e de ini i ely de e mines he class. Th ough
decision‐le el usion, which may be u ilised o conclude, i is
easible o combine he classi ica ion decisions gene a ed by
se e al ea u e ec o s and di e en classi ie s in o a single
decision. The ou comes o he h ee models' decisions may be
inco po a ed in o a single ec o . The ea u es o wo DL
models (ResNe 50 and CDBN) we e in eg a ed o achie e
usion a he decision le el. Many o he p oblems ha ha e
been plaguing ML ha e ound an answe in usion modelling.
This is because i imp o es o e all pe o mance by combining
he p edic i e skills o se e al models in o one model. Se e al
aining da ase s o app oaches a e u ilised in he usion p o-
cess, which addi ionally inco po a es he p edic ed esul s o
each base model o o e a single an icipa ed pe o mance. As a
esul , he ac s may be mo e ai h ully po ayed. The pu pose
o combining many models is o educe he amoun o in o -
ma ion ha can be gene alised abou he o ecas . When many
models a e used, he inaccu acy in he classi ica ion diminishes,
p o ided ha he base models a e di e se and independen o
one ano he . Deep Lea ning syn hesis is an app oach ha aims
o inc ease p oduc i i y by consolida ing he diagnos ic con-
clusions d awn om a ious models in o a single, uni ied de-
cision. Addi ionally, da a usion is a subse o usion aining,
and he deg ee o in eg a ion be ween he wo da a ypes a -
ec s he classi ie used. Inco po a ion a his le el is an ici-
pa ed o imp o e classi ica ion pe o mance because he
ea u e se o a model con ains mo e in o ma ion abou he
MRI images han all in eg a ed classi ie s. A decision usion, on
he o he hand, comp ises a p ojec ed decision ha is p o ided
o classi ying he ou come.
In he b ain umou classi ica ion wo k desc ibed in he
p o ided in o ma ion, So Max is used as he inal classi ie in
he DL usion model. A e he ea u e ex ac ion and usion
p ocess using CDBN and ResNe 50, he used ea u e ec o is
ed in o he So Max unc ion in he ou pu laye o mul iclass
classi ica ion. The So Max unc ion akes he used ea u e
FIGURE 3 The ep esen a ion o implemen ed a chi ec u e o CDBN ne wo k. CDBN, con olu ional deep belie ne wo ks.
TABLE 2Implemen ed CDBN a chi ec u e.
Laye s Inpu size
Inpu size 128 �128
Numbe o laye s 2
Con ‐1 7 �7 [32 �124 �124]
Max‐pooling laye 2 �2 [32 �62 �62]
Con ‐2 5 �5 [64 �58 �58]
Max‐pooling laye 2 �2 [64 �29 �29]
Con ‐3 6 �6 [128 �24 �24]
Max‐pooling laye 2 �2 [128 �12 �12]
Ba ch size 200
Epoch 50
Lea ning a e 0.001
Abb e ia ion: CDBN, con olu ional deep belie ne wo ks.
796
-
ZEBARI ET AL.
24682322, 2024, 4, Downloaded om h ps://ie esea ch.onlinelib a y.wiley.com/doi/10.1049/ci 2.12276 by Technical Uni e si y Os a a, Wiley Online Lib a y on [25/09/2024]. See he Te ms and Condi ions (h ps://onlinelib a y.wiley.com/ e ms-and-condi ions) on Wiley Online Lib a y o ules o use; OA a icles a e go e ned by he applicable C ea i e Commons License
ec o as inpu . I ans o ms i in o a p obabili y dis ibu ion
o e he ou b ain umou classes: gliomas, meningiomas,
pi ui a y umou s, and no‐ umou class [18]. The So Max
unc ion calcula es he p obabili y ha he inpu image belongs
o each class, ensu ing ha he sum o all p obabili ies equals 1.
The class wi h he highes p obabili y ou pu by he So Max
unc ion is conside ed he p edic ed class o he inpu b ain
umou image. This So Max classi ica ion allows he model o
make con iden and calib a ed p edic ions ac oss mul iple
classes, p o iding he p obabili ies associa ed wi h each class
o a mo e in o ma i e ou pu . By employing So Max as he
classi ie , he model can e ec i ely handle he mul i‐class na-
u e o he b ain umou classi ica ion ask, p o iding accu a e
and in e p e able p edic ions o di e en umou ypes. The
use o So Max as he inal ac i a ion unc ion aligns wi h bes
p ac ices o mul iclass classi ica ion asks in DL. I is likely a
key ac o con ibu ing o he high e alua ion me ics achie ed
in he s udy. The So Max unc ion is a ma hema ical unc ion
used o mul iclass classi ica ion. Gi en a ec o o aw sco es
o each class, he So Max unc ion con e s hese sco es in o
a p obabili y dis ibu ion, ep esen ing he likelihood o he
inpu belonging o each class; in compa ison, he inpu logi s
o each class a e deno ed as a ec o z=[z1, z2, …, zK],
whe e K is he numbe o classes. The So Max unc ion o
he class kis de ined as ollows: Addi ionally, da a usion is a
subse o usion aining, and he deg ee o in eg a ion be-
ween he wo da a ypes a ec s he classi ie used. Inco po-
a ion a his le el is an icipa ed o imp o e classi ica ion
FIGURE 4 The ep esen a ion o implemen ed a chi ec u e o VGG16 ne wo k.
FIGURE 5 Rep esen a ion o he aining s eps o VGG16, ResNe 50, CDBN, and p oposed usion models. CDBN, con olu ional deep belie ne wo ks.
ZEBARI ET AL.
-
797
24682322, 2024, 4, Downloaded om h ps://ie esea ch.onlinelib a y.wiley.com/doi/10.1049/ci 2.12276 by Technical Uni e si y Os a a, Wiley Online Lib a y on [25/09/2024]. See he Te ms and Condi ions (h ps://onlinelib a y.wiley.com/ e ms-and-condi ions) on Wiley Online Lib a y o ules o use; OA a icles a e go e ned by he applicable C ea i e Commons License
pe o mance because he ea u e se o a model con ains mo e
in o ma ion abou he MRI images han all in eg a ed classi-
ie s. A decision usion, on he o he hand, comp ises a p o-
jec ed decision ha is p o ided o classi ying he ou come.
So maxðzkÞ ¼ expðzkÞ
PðexpðziÞÞ;i¼1 o K ð4Þ
He e, exp(zk) is he exponen ial unc ion. The So Max
unc ion applies he exponen ial unc ion o each logi and
hen di ides i by he sum o he exponen ials o all logi s. This
no malisa ion s ep ensu es ha he ou pu alues a e non‐
nega i e and sum up o 1, c ea ing a alid p obabili y dis i-
bu ion. The ou pu o he So Max unc ion is a ec o o
p obabili ies [p1, p2, …,pK], whe e each elemen p_i ep e-
sen s he p obabili y o he inpu belonging o class i. The class
wi h he highes chance is ypically chosen as he p edic ed
class o he in o ma ion. Ma hema ically, So Max mul i‐class
classi ica ion can be gi en as ollows: Inpu : z=[z1, z2, …,
zK] (logi s o each class) Ou pu : P obabili y dis ibu ion [p1,
p2, …, pK]whe e p_i=exp(z_i)/∑(exp(z_i)) o i =1 o K.
4
|
RESULTS AND DISCUSSIONS
4.1
|
Da ase
A da ase downloaded om he Kaggle websi e was used in
his s udy o alida e he pe o mance e alua ion o he sug-
ges ed model [27]. Table 3p o ides an o e iew o he in-
o ma ion associa ed wi h he da ase . Figu e 6p o ides
illus a i e examples o some samples. The collec ion con ains
3264 MRI images, each ep esen ing one o ou di e en
classes o b ain umou s. The alida ion o he expe imen is
ca ied ou in a speci ic manne by pa i ioning he da ase in o
wo hal es acco ding o he p opo ion o he size o he
aining da ase o he es ing da a, which was 2870 aining
da a samples o 394 es ing da a samples. Py hon 3.6.5 was
used as he ool o do he simula ion o he sugges ed model.
Some lib a ies ha e been used o w i e he code, including
Ke as and Tenso Flow. The ha dwa e sys em employed had
32 GB o RAM and an i5 9 h Gene a ion p ocesso . The
ollowing alues ha e been speci ied o each pa ame e :
lea ning a e was 0.01, d opou was 0.5, ba ch size was 5, epoch
coun was 50, and ac i a ion was ReLU.
4.2
|
Pe o mance me ics
This sec ion con ains he ou e iciency me ics applied o
assess how well he p oposed me hod pe o med. The o he
pe o mance measu es ha a e accessible a e all di e en i -
e a ions o he speci ied pe o mance measu emen s. Accu acy
is one o he ou leading pe o mance indica o s because i
p o ides in o ma ion abou he o e all pe o mance o he
sugges ed app oach. By u ilising Equa ions (5)–(8), addi ional
pe o mance me ics, including Recall, P ecision, and F‐Sco e,
a e employed o analyse he sugges ed model's alid posi i e
and ac ual nega i e a es.
The accu acy measu e o e s de ini ions o he glioma,
meningioma, pi ui a y, and non‐ umou classes depic ed in he
labelled images. Equa ion (5) is he ac o ha de e mines how
accu a e he p oposed model is. In Equa ion (5), T ue Posi i e
deno es cases ha ha e been success ully iden i ied as posi i e,
and T ue Nega i e indica es cases ha ha e been success ully
ecognised as nega i e. Samples ha ha e been success ully
iden i ied as being imp ope ly classi ied a e deno ed by he
no a ion False Nega i e. In con as , False Posi i e ep esen s
cases ha ha e been w ongly iden i ied as being co ec ly
iden i ied. Recall ha i is u ilised o quan i y he pe cen age o
genuine posi i e cases we could success ully o esee by
employing ou model. In o he wo ds, i can show us how
accu a e ou p edic ions we e. To compu e he ecall me ic,
Equa ion (6) is u ilised. P ecision is a measu emen used o
de e mine how many o he accu a ely classi ied cases ac ually
u ned ou o be posi i e. Equa ion (7) is employed o calcula e
he a e o p ecision. F‐Sco e is a measu e o ep esen ing he
TABLE 3The dis ibu ion o he da ase .
Class T aining Augmen ed Tes ing
Glioma‐ umou 826 2478 100
Meningioma‐ umou 822 2466 115
Pi ui a y‐ umou 827 2481 74
No‐ umou 395 1185 105
To al 2870 8610 394
FIGURE 6 Da ase samples used o e alua e he p oposed model.
798
-
ZEBARI ET AL.
24682322, 2024, 4, Downloaded om h ps://ie esea ch.onlinelib a y.wiley.com/doi/10.1049/ci 2.12276 by Technical Uni e si y Os a a, Wiley Online Lib a y on [25/09/2024]. See he Te ms and Condi ions (h ps://onlinelib a y.wiley.com/ e ms-and-condi ions) on Wiley Online Lib a y o ules o use; OA a icles a e go e ned by he applicable C ea i e Commons License