Ci a ion: Ma ínez-Del-Río-O ega, R.;
Ci i -Maso , J.; Luna-Pe ejón, F.;
Domínguez-Mo ales, M. B ain Tumo
De ec ion Using Magne ic Resonance
Imaging and Con olu ional Neu al
Ne wo ks. Big Da a Cogn. Compu .
2024,8, 123. h ps://doi.o g/
10.3390/bdcc8090123
Academic Edi o : Yoichi Hayashi
Recei ed: 19 Augus 2024
Re ised: 8 Sep embe 2024
Accep ed: 19 Sep embe 2024
Published: 21 Sep embe 2024
Copy igh : © 2024 by he au ho s.
Licensee MDPI, Basel, Swi ze land.
This a icle is an open access a icle
dis ibu ed unde he e ms and
condi ions o he C ea i e Commons
A ibu ion (CC BY) license (h ps://
c ea i ecommons.o g/licenses/by/
4.0/).
A icle
B ain Tumo De ec ion Using Magne ic Resonance Imaging and
Con olu ional Neu al Ne wo ks
Ra ael Ma ínez-Del-Río-O ega 1, Ja ie Ci i -Maso 1,2,3 , F ancisco Luna-Pe ejón 1,2,3,4,*
and Manuel Domínguez-Mo ales 1,2,3,4
1E.T.S. Ingenie ía In o má ica, Uni e sidad de Se illa, A da. Reina Me cedes s/n, 41012 Se ille, Spain;
[email p o ec ed] (R.M.-D.-R.-O.); [email p o ec ed] (J.C.-M.); [email p o ec ed] (M.D.-M.)
2Robo ics and Technology o Compu e s Resea ch G oup (TEP-108), A chi ec u e and Compu e Technology
Depa men , E.T.S. Ingenie ía In o má ica, Uni e sidad de Se illa, A da. Reina Me cedes s/n,
41012 Se ille, Spain
3Escuela Poli écnica Supe io (EPS), Uni e sidad de Se illa, 41011 Se ille, Spain
4Compu e Enginee ing Resea ch Ins i u e (I3US), E.T.S. Ingenie ía In o má ica, Uni e sidad de Se illa,
A da. Reina Me cedes s/n, 41012 Se ille, Spain
*Co espondence: [email p o ec ed]
Abs ac : Ea ly and p ecise de ec ion o b ain umo s is c i ical o imp o ing clinical ou comes and
pa ien quali y o li e. This esea ch ocused on de eloping an image classi ie using con olu ional
neu al ne wo ks (CNN) o de ec b ain umo s in magne ic esonance imaging (MRI). B ain umo s
a e a signi ican cause o mo bidi y and mo ali y wo ldwide, wi h app oxima ely 300,000 new cases
diagnosed annually. Magne ic esonance imaging (MRI) o e s excellen spa ial esolu ion and so
issue con as , making i indispensable o iden i ying b ain abno mali ies. Howe e , accu a e
in e p e a ion o MRI scans emains challenging, due o human subjec i i y and a iabili y in umo
appea ance. This s udy employed CNNs, which ha e demons a ed excep ional pe o mance in
medical image analysis, o add ess hese challenges. Va ious CNN a chi ec u es we e implemen ed
and e alua ed o op imize b ain umo de ec ion. The bes model achie ed an accu acy o 97.5%,
sensi i i y o 99.2%, and bina y accu acy o 98.2%, su passing p e ious s udies. These esul s
unde sco e he po en ial o deep lea ning echniques in clinical applica ions, signi ican ly enhancing
diagnos ic accu acy and eliabili y.
Keywo ds: b ain umo s; MRI; con olu ional neu al ne wo ks; deep lea ning; image classi ica ion;
medical imaging
1. In oduc ion
The ea ly and p ecise de ec ion o b ain umo s is c ucial o imp o ing clinical
ou comes and quali y o li e o pa ien s. B ain umo s ep esen a signi ican cause o
mo bidi y and mo ali y wo ldwide. Acco ding o he Wo ld Heal h O ganiza ion (WHO),
app oxima ely 300,000 new cases o b ain umo s a e diagnosed annually, making hem one
o he leading causes o cance - ela ed dea h in child en and young adul s [
1
]. The su i al
a e o b ain umo pa ien s hea ily depends on ea ly de ec ion and he e ec i eness o
he subsequen ea men .
Magne ic esonance imaging (MRI) is an ad anced imaging echnique ha p o ides ex-
cellen spa ial esolu ion and so issue con as , making i an essen ial ool o iden i ying
b ain abno mali ies. MRI can non-in asi ely de ec a ious ypes o b ain umo s and o he
neu ological diso de s. Howe e , in e p e ing MRI scans is a complex ask ha equi es
he expe ise o highly ained adiologis s. This ask is p one o e o s, due o a iabili y in
umo appea ance and o human limi a ions, such as a igue and subjec i i y [
2
]. Addi ion-
ally, in egions wi h a sho age o specialis s, access o accu a e and imely diagnosis can be
limi ed, delaying ea men ini ia ion and ad e sely a ec ing pa ien ou comes.
Big Da a Cogn. Compu . 2024,8, 123. h ps://doi.o g/10.3390/bdcc8090123 h ps://www.mdpi.com/jou nal/bdcc
Big Da a Cogn. Compu . 2024,8, 123 2 o 17
Con olu ional neu al ne wo ks (CNNs) ha e eme ged as a powe ul echnique in he
ield o deep lea ning o image classi ica ion. CNNs can lea n ele an ea u es di ec ly
om image da a, elimina ing he need o manual ea u e enginee ing and enabling highe
classi ica ion accu acy [
3
]. These ne wo ks ha e demons a ed ou s anding pe o mance in
image ecogni ion asks ac oss a ious domains, including he de ec ion o pa hologies in
medical images [
4
]. In medical imaging, CNNs ha e shown p omise in asks such as umo
de ec ion, segmen a ion, and classi ica ion, p o iding a eliable second opinion, helping
p io i ize u gen cases, and allowing physicians o ocus on mo e complex and c i ical
asks [5].
The aim o his wo k was o de elop an image classi ie based on CNNs o de ec ing
b ain umo s in MRI scans, wi h he goal o suppo ing medical diagnosis and imp o ing
pa ien ou comes. Using a public MRI da ase , he p ojec in ol ed p ep ocessing he
da a o ensu e image quali y and consis ency, designing a speci ic CNN a chi ec u e o
b ain umo de ec ion. A g id sea ch me hod was employed o op imize hype pa ame e s,
ensu ing he bes -possible model pe o mance.
The emainde o he pape is s uc u ed as ollows. In he nex sec ion, Ma e ials
and Me hods, he da ase used, he p ep ocessing s eps pe o med, and he p oposed
models, along wi h he hype pa ame e s conside ed o pe o mance analysis, a e desc ibed.
Following his, in he Resul s sec ions, he me ics ob ained a e p esen ed and he bes
models a e iden i ied. Subsequen ly, in he Discussion sec ion, a compa ison wi h ecen
wo ks add essing he bina y p oblem o b ain umo iden i ica ion is included, and he
limi a ions, challenges, and u u e wo k a e ou lined. Finally, he conclusions a e p esen ed.
2. Ma e ials and Me hods
This sec ion p o ides an o e iew o he publicly a ailable da ase , which unde wen
p ep ocessing o aining and es ing he classi ica ion sys em. I also desc ibes he classi-
ie s de eloped and he e alua ion me ics used. Figu e 1summa izes he me hodology
used o achie e hese asks:
Figu e 1. Wo k low o e iew. The da ase is p ep ocessed ( esizing and no maliza ion), spli o
aining and es ing, and used o ain a con olu ional neu al ne wo k (CNN). Hype pa ame e s like
laye s, ba ch size, and lea ning a e a e op imized du ing aining. The bes model is alida ed and
es ed o p oduce he inal esul s.
Big Da a Cogn. Compu . 2024,8, 123 3 o 17
2.1. Da ase
The e a e a ious con ibu ions in he li e a u e p o iding publicly a ailable da ase s
o b ain umo iden i ica ion, each o e ing di e en ad an ages and limi a ions. One
such da ase is he B 35H da ase (h ps://www.kaggle.com/da ase s/ahmedhamada0
/b ain- umo -de ec ion, accessed on Sep embe 2023), which con ains a well-balanced
collec ion o 1500 umo al and 1500 non- umo al images. This balance makes i sui able o
bina y classi ica ion asks; howe e , i s limi a ions include he high con as o he images
and he ac ha all a e cap u ed exclusi ely on he axial plane, which may es ic he
model’s abili y o gene alize ac oss di e en ana omical pe spec i es. Simila ly, he B aTS
da ase , widely used o b ain umo segmen a ion compe i ions since 2014 [
6
], compiles
o e 1000 umo al images bu lacks non- umo al samples, making i mo e app op ia e
o segmen a ion asks han bina y classi ica ion. Ano he da ase , he IXI da ase (h ps:
//b ain-de elopmen .o g/ixi-da ase /, accessed on Sep embe 2023), hough use ul o
a ious MRI analyses, con ains only 600 images, which poses a challenge o deep lea ning
models ha ypically equi e la ge da ase s o e ec i e aining.
The da ase used in his s udy is he publicly a ailable P ee Vi adiya B ain Tumo
Da ase (h ps://www.kaggle.com/da ase s/p ee i adiya/b ian- umo -da ase , accessed
on Sep embe 2023), consis ing o b ain scans ca ego ized in o heal hy and umo -a ec ed
g oups. This da ase con ains a subs an ial numbe o images and is well-balanced, wi h
55% o he images depic ing umo s and 45% ep esen ing heal hy b ain scans. Addi ion-
ally, i includes a di e se ange o ana omical planes and image shi s, which is highly
ad an ageous o deep lea ning models. This di e si y helps mi iga e he isk o o e i ing
and educes he likelihood o he model becoming o e ly specialized o speci ic ea u es
o pa e ns. The inclusion o a ied imaging pe spec i es ensu es ha he model can
gene alize mo e e ec i ely o di e en scena ios, ul ima ely enhancing i s obus ness and
pe o mance when applied o new, unseen da a.
Figu e 2displays ep esen a i e samples om each class, illus a ing no only he
a iabili y and cha ac e is ics inhe en in bo h he heal hy and umo -a ec ed b ain scans
bu also he impac o his di e si y. The p esence o di e en ana omical planes and shi s
plays a c i ical ole in imp o ing he model’s capaci y o adap and pe o m eliably ac oss
a wide ange o clinical cases.
Heal hy
B ain umo
Figu e 2. Example MRI images om he da ase used in his s udy. The op ow shows MRI scans o
heal hy b ain issue, while he bo om ow depic s scans wi h isible b ain umo s. Each ow includes
di e en iews, such as axial, co onal, and sagi al planes, demons a ing he ana omical di e ences
be ween heal hy and pa hological cases.
Big Da a Cogn. Compu . 2024,8, 123 4 o 17
The da ase was spli in o aining, alida ion, and es ing se s o ensu e obus model
e alua ion. The aining se included 70% o he da a, he alida ion se 20%, and he
es ing se 10%. The p opo ion o samples wi hin each class was main ained consis en ly
ac oss all h ee subse s, he eby ensu ing ha he class dis ibu ion emained balanced and
ep esen a i e in each subse . The speci ic dis ibu ion o samples wi hin hese subse s can
be consul ed in Table 1:
Table 1. Da ase used in his wo k and subse s di ision.
Class T ain (70%) Valida ion (20%) Tes (10%) To al
Heal hy 1461 418 208 2087
B ain umo 1759 502 252 2513
To al 3220 920 460 4600
2.2. Da a P ep ocessing
P ep ocessing s eps a e essen ial o enhancing he quali y and consis ency o MRI
images, as highligh ed by Akkus e al. [
7
]. Se e al key echniques we e implemen ed
o p epa e he da a o e ec i e analysis. Ini ially, all he images we e uni o mly scaled
o a esolu ion o 256
×
256 pixels. Fu he mo e, images exhibi ing a i ac s wi hin he
c anial egion we e excluded om u he analysis, o main ain da a in eg i y. Following
his, a min–max global no maliza ion echnique was applied, o adjus he pixel in ensi y
alues o a common scale anging be ween 0 and 1, he eby ensu ing uni o mi y ac oss
all images. The smalles pixel alue in he da ase was ans o med o 0, he la ges o 1,
and he emaining alues we e p opo ionally scaled in be ween. This s ep was c ucial
o mi iga ing he e ec s o a ying acquisi ion p o ocols and o acili a ing consis en
model pe o mance. The e o e, by applying hese global no maliza ion pa ame e s ac oss
he en i e da ase , he in ensi y le els emained consis en , e en o images acqui ed om
di e en MRI sou ces o se ings. This s ep played a c i ical ole in imp o ing uni o mi y
and minimizing biases due o b igh ness o con as a ia ions ac oss he scans. Mo eo e ,
segmen a ion echniques we e employed o isola e egions o in e es (ROIs), speci ically
ocusing on b ain issues while excluding non- ele an pa s o he image. This app oach
enabled he analysis o concen a e on he mos c i ical a eas, educing he in luence
o ex aneous s uc u es and he eby enhancing he accu acy o subsequen asks, such
as classi ica ion.
Images exhibi ing a i ac s wi hin he c anial egion we e also add essed du ing
p ep ocessing. Ini ially, manual inspec ion was conduc ed, iden i ying a o al o 20 images
wi h a i ac s. Since hese a i ac s we e loca ed ou side he egions o in e es (ROIs), i
was possible o e ain hese images o aining and e alua ion by c opping he a ec ed
a eas. In e ms o segmen a ion, he da ase had al eady unde gone p ep ocessing o
emo e pa s o he backg ound un ela ed o he skull. Howe e , an addi ional s ep o edge
c opping was pe o med o u he ocus on he egions o in e es , pa icula ly cen e ing
he ele an b ain a eas. This in ol ed elimina ing po ions o he neck o jaw in images
whe e hey appea ed, which u he e ined he da ase . This ex a e inemen s ep no only
enhanced he model’s abili y o ocus on he mos pe inen a eas bu also educed i ele an
backg ound noise, he eby imp o ing he accu acy and e iciency o he aining p ocess.
2.3. Model A chi ec u e
A con olu ional neu al ne wo k (CNN) was designed speci ically o he ask o
de ec ing b ain umo s. The a chi ec u e consis ed o he ollowing:
• Inpu Laye : Accep ing p ep ocessed MRI images.
•
Con olu ional Laye s: Ex ac ing spa ial ea u es h ough il e s applied o he in-
pu images.
•
Pooling Laye s: Reducing he dimensionali y o he ea u e maps while e aining
impo an in o ma ion.
Big Da a Cogn. Compu . 2024,8, 123 5 o 17
•
Fully Connec ed Laye s: Pe o ming high-le el easoning based on he ex ac ed
ea u es.
• Ou pu Laye : P o iding bina y classi ica ion ou pu s (heal hy o umo ).
2.4. Implemen a ion
The de elopmen en i onmen used Jupy e No ebook wi h Py hon 3.3. Tenso Flow
and Ke as buil and ained he CNN. The s udy was ca ied ou wi h a g id sea ch o
hype pa ame e uning. Nex , he op imal models iden i ied h ough he g id sea ch we e
subsequen ly alida ed using he es subse . Fo his alida ion, we conside ed he a e age
alues o h ee key me ics: accu acy, p ecision, and sensi i i y ( ecall).
A g id sea ch is a sys ema ic me hod o explo ing hype pa ame e space by e alua ing
all possible combina ions o p ede ined hype pa ame e alues. While his app oach can be
compu a ionally in ensi e, i was chosen due o i s ho oughness in iden i ying he op imal
con igu a ion. Gi en he manageable size o ou hype pa ame e space, we we e able o
conduc his sea ch e icien ly using a ailable compu a ional esou ces. Addi ionally, he
use o he g id sea ch allowed us o ensu e ha e e y po en ial combina ion was es ed,
p o iding comp ehensi e co e age o he hype pa ame e space o achie e he bes esul s.
The hype pa ame e s included we e lea ning a e, ba ch size, numbe o con olu ional
and max pooling laye s, and ba ch size. Va ious combina ions o hese hype pa ame e s
we e es ed o iden i y he bes -pe o ming models con igu a ions. The hype pa ame e s
and hei alues a e de ailed in Table 2. The selec ion o alues o he g id sea ch was
guided bo h by ou expe ise and by insigh s gained om p e ious s udies in diagnos ic
suppo h ough image analysis [
8
–
11
]. These s udies, which ha e ocused on op imizing
simila machine lea ning models in heal hca e applica ions, p o ided a ounda ion o
de e mining app op ia e hype pa ame e anges.
Each model was ained using he de ined hype pa ame e combina ions, wi h he
aining p ocess in ol ing mul iple i e a ions and adjus men s o op imize pe o mance.
The g id sea ch was implemen ed h ough exhaus i e loops o explo e all possible com-
bina ions o hype pa ame e s, wi h he esul s sys ema ically eco ded o compa a i e
analysis. To p e en o e i ing, ea ly s opping was applied using a pa ience pa ame e o
5 i e a ions. Speci ically, i he a ge me ic on he alida ion subse did no imp o e a e
5 consecu i e i e a ions, he aining was hal ed, and he model wi h he bes pe o mance
wi hin hose i e a ions was selec ed. Model checkpoin ing was also employed o ensu e
ha he bes -pe o ming model was sa ed h oughou he aining p ocess. These measu es
helped ensu e ha he models we e no o e ained, acili a ing he iden i ica ion o he
op imal hype pa ame e con igu a ions.
Table 2. Hype pa ame e s and hei alues used o g id sea ch.
Hype pa ame e Values
Lea ning a e 0.1, 0.01, 0.001
Ba ch size 10, 20, 30
Numbe o con olu ional and max pooling laye s 2, 3, 4
Numbe o dense laye s 1, 2, 3
2.5. E alua ion Me ics
The models’ pe o mances we e e alua ed using se e al me ics:
• Accu acy: The pe cen age o co ec ly classi ied images ou o he o al images.
• Sensi i i y (Recall): The abili y o he model o co ec ly iden i y posi i e cases ( umo ).
• Speci ici y: The abili y o he model o co ec ly iden i y nega i e cases (no umo ).
• P ecision: The p opo ion o posi i e iden i ica ions ha we e ac ually co ec .
•
F1 Sco e: The ha monic mean o p ecision and ecall, p o iding a single me ic ha
balanced bo h conce ns.
Big Da a Cogn. Compu . 2024,8, 123 6 o 17
These me ics p o ided a comp ehensi e assessmen o he model’s classi ica ion
capabili ies, ensu ing ha he model no only pe o med well on a e age bu also e ec i ely
dis inguished be ween umo and non- umo cases. Acco dingly, he high-le el me ics a e
p esen ed in he ollowing equa ions:
Accu acy =∑
c
TPc+TNc
TPc+FPc+TNc+FNc,c∈classes (1)
Speci ici y =∑
c
TNc
TNc+FPc,c∈classes (2)
P ecision =∑
c
TPc
TPc+FPc,c∈classes (3)
Sensi i i y =∑
c
TPc
TPc+FNc,c∈classes (4)
F1sco e =2×p ecision ×sensi i i y
p ecision +sensi i i y. (5)
3. Resul s
The esul s will be p esen ed sys ema ically, beginning wi h he g id sea ch esul s
o he alida ion and es ing subse s. Following his, he op i e models om all ain-
ing i e a ions will be discussed, and de ailed pe o mance me ics o hese models will
be p o ided.
3.1. G id Sea ch
Du ing his phase, a o al o 81 aining sessions we e conduc ed, encompassing
all possible combina ions o he hype pa ame e s ou lined in he p e ious sec ion. The
esul s o he alida ion subse , eco ded du ing he aining p ocess, a e summa ized in
Figu e 3. Fu he mo e, he esul s o he es ing subse , e alua ed a e aining, a e shown
in Figu e 4. Supplemen a y Ma e ial Tables S1 and S2 con ains de ailed abula ed esul s
o u he e e ence.
F om a global pe spec i e, i is e iden ha he models wi h lowe lea ning a es (e.g.,
0.0001) and mode a e ba ch sizes (e.g., 20) ended o achie e a mo e a o able balance
be ween accu acy and gene aliza ion. This cha ac e is ic is pa icula ly signi ican in he
con ex o b ain umo de ec ion, whe e a model’s abili y o gene alize e ec i ely o new
and unseen da a is c ucial o i s clinical applicabili y.
Fo ins ance, a model ained wi h a lea ning a e o 0.0001 and a ba ch size o
20 demons a ed a alida ion loss o 0.35 and a alida ion accu acy o 90.5%. This speci ic
combina ion o pa ame e s no only acili a ed imp o ed con e gence du ing he aining
p ocess bu also con ibu ed o he model’s s abili y, enabling mo e g adual and p ecise
weigh upda es. Such beha io sugges s enhanced gene aliza ion o unseen da a, which
is essen ial o accu a e umo de ec ion ac oss a ying imaging condi ions and di e se
pa ien popula ions.
In con as , models u ilizing highe lea ning a es (e.g., 0.001) exhibi ed apid con e -
gence; howe e , hey o en displayed a signi ican dispa i y be ween aining accu acy and
alida ion accu acy, indica i e o o e i ing. O e i ing occu s when a model becomes
o e ly a uned o he aining da a, cap u ing noise and idiosync asies speci ic o ha
da ase a he han lea ning pa e ns ha a e b oadly gene alizable. Consequen ly, hese
models end o pe o m poo ly when applied o new da ase s.
Big Da a Cogn. Compu . 2024,8, 123 7 o 17
Figu e 3. A e age alida ion accu acy esul s o di e en hype pa ame e se ings. These ba plo s
illus a e he impac o hype pa ame e choices on he model’s alida ion accu acy. The op-le
plo shows he e ec o ba ch size. The op- igh plo p esen s he in luence o lea ning a e. The
bo om-le plo shows he e ec o dense laye s, while he bo om- igh plo e lec s he impac o
con olu ional laye s.
The s abili y obse ed in models wi h lowe lea ning a es and mode a e ba ch sizes
can be a ibu ed o he ac ha hese se ings allowed he model o make ine adjus men s
du ing he aining p ocess. A mode a e ba ch size, such as 20, p o ides an op imal balance
be ween he equency o weigh upda es and he s abili y o hose upda es. This is pa icu-
la ly c i ical in deep con olu ional neu al ne wo ks, whe e ab up adjus men s can lead o
des abiliza ion du ing aining and p e en he model om achie ing a global op imum.
The op-pe o ming models achie ed accu acies exceeding 90%, demons a ing ex-
cep ional pe o mance in he ask o image classi ica ion o b ain umo de ec ion. These
esul s a e pa icula ly signi ican in he clinical se ing, whe e p ecision and accu acy
a e c i ical o ensu ing diagnos ic con idence. Fo ins ance, a model con igu ed wi h a
ba ch size o 20, a lea ning a e o 0.0001, h ee con olu ional laye s, and wo dense laye s
achie ed a p ecision o 97% and an accu acy o 98%. These me ics unde sco e he model’s
abili y o consis en ly make co ec p edic ions, accu a ely iden i ying bo h ue posi i es
and ue nega i es when applied o he es da ase .
This model’s con igu a ion enabled i o cap u e complex and de ailed ea u es wi hin
he MRI images, which is essen ial o accu a ely dis inguishing be ween heal hy and
umo issues. A p ecision o 97% indica es ha he model co ec ly iden i ied 97% o he
posi i e cases, minimizing he ma gin o e o and educing he likelihood o alse posi i es.
Fu he mo e, an accu acy o 98% signi ies ha he model e ec i ely main ained a s ong
balance in co ec ly iden i ying bo h heal hy and umo s a es. This is c ucial o a oiding
misdiagnoses, which could o he wise lead o inapp op ia e ea men s o delays in medical
in e en ion. Such high pe o mance me ics a e i al in ensu ing ha he model can be
eliably in eg a ed in o clinical p ac ice, whe e accu a e and imely diagnosis is pa amoun .
Big Da a Cogn. Compu . 2024,8, 123 8 o 17
Figu e 4. A e age me ics esul s o di e en hype pa ame e se ings. These ba plo s illus a e he
impac o hype pa ame e choices on he model’s es accu acy, p ecision, and ecall. The op-le
plo shows he e ec o ba ch size. The op- igh plo p esen s he in luence o lea ning a e. The
bo om-le plo shows he e ec o dense laye s, while he bo om- igh plo e lec s he impac o
con olu ional laye s.
Mo eo e , he p ecision and sensi i i y me ics p o ided a mo e nuanced unde s and-
ing o he models’ pe o mances. Fo ins ance, a model con igu ed wi h a ba ch size o
10, a lea ning a e o 0.001, ou con olu ional laye s, and wo dense laye s achie ed a
p ecision o 96% and a sensi i i y o 98%. The 96% p ecision indica es he model’s capabili y
o accu a ely classi ying images as posi i e, he eby minimizing he occu ence o alse
posi i es. On he o he hand, he sensi i i y o 98% demons a es he model’s e ec i eness
in co ec ly iden i ying images ha ac ually con ained umo s, hus educing he likelihood
o alse nega i es. This balance be ween p ecision and sensi i i y is c i ical, as i ensu es
ha he model is no only accu a e in i s posi i e p edic ions bu also eliable in de ec ing
all umo cases.
This model, wi h i s speci ic con igu a ion, success ully achie ed a balance be ween
iden i ying posi i e cases and minimizing alse posi i es. This balance is pa icula ly
signi ican in medical applica ions, whe e alse posi i es can lead o unnecessa y in asi e
p ocedu es and inc eased pa ien anxie y, while alse nega i es may esul in he ailu e o
p o ide imely ea men o pa ien s wi h b ain umo s. The model’s abili y o main ain
high sensi i i y ensu es ha he as majo i y o umo cases a e de ec ed, which is essen ial
o e ec i e disease managemen and ea men .
3.2. Sea ching o he Bes Models
The models wi h he bes accu acy sco es s ood ou o hei abili y o consis en ly
make co ec p edic ions. Table 3shows he esul s o he models wi h he highes accu acy.
These models demons a ed consis en pe o mance, wi h accu acies exceeding 95%,
unde sco ing hei obus ness and eliabili y in classi ying magne ic esonance images
o b ain umo de ec ion. These high accu acy a es a e pa icula ly signi ican , as hey
indica e he models’ abili y o e ec i ely di e en ia e be ween images o heal hy b ains
and hose wi h umo s, he eby minimizing bo h alse posi i es and alse nega i es.
Big Da a Cogn. Compu . 2024,8, 123 9 o 17
Table 3. Lis o he i e models wi h he highes accu acy. The columns ep esen he ba ch size (BS),
lea ning a e (LR), numbe o con olu ional laye s (nCL), numbe o dense laye s (nDL), p ecision,
ecall, and accu acy.
BS LR nCL nDL P ecision Recall Accu acy
30 0.001 3 1 0.97510 0.99156 0.98214
10 0.001 2 3 0.97983 0.98380 0.97991
30 0.001 2 3 0.98393 0.98000 0.97991
10 0.0001 2 3 0.97983 0.97983 0.97767
20 0.001 3 2 0.97580 0.98373 0.97767
Addi ionally, hese models gene ally employed mode a e ba ch sizes and medium
lea ning a es, which sugges s an op imal balance be ween aining e iciency and model
s abili y. Mode a e ba ch sizes, such as 20 o 30, enabled e icien aining by p ocessing
a su icien numbe o samples in each i e a ion o de elop a obus ep esen a ion o he
da a, while also minimizing he in oduc ion o excessi e noise. This app oach helped
s abilize model weigh upda es, allowing o g adual and con olled adjus men s ha
p e en ab up oscilla ions ha could des abilize he aining p ocess.
Medium lea ning a es, such as 0.0001 o 0.001, also played a c ucial ole in he pe o -
mance o hese models. A medium lea ning a e ensu es ha he model makes meaning ul
p og ess du ing each aining i e a ion wi hou making o e ly la ge adjus men s ha could
lead o o e i ing. This balance is essen ial o main aining bo h he accu acy and gene -
alizabili y o he model. The use o hese lea ning a es allowed he models o e ec i ely
adap o he aining da a, he eby enhancing hei abili y o gene alize o new es da a.
The models wi h he bes p ecision sco es excelled in minimizing alse posi i es.
Table 4p esen s he i e models wi h he highes p ecision.
Table 4. Lis o he i e models ha achie ed he bes p ecision esul s.
BS LR nCL nDL P ecision Recall Accu acy
10 0.001 3 3 0.98804 0.96875 0.97544
30 0.0001 3 1 0.98770 0.96787 0.97544
20 0.001 4 1 0.98412 0.97637 0.97767
30 0.001 2 3 0.98393 0.98000 0.97991
20 0.0001 2 1 0.98031 0.98031 0.97767
The model con igu ed wi h h ee con olu ional laye s and h ee dense laye s achie ed
an imp essi e accu acy o 98.8%, unde sco ing i s ema kable e iciency in co ec ly iden-
i ying b ain umo cases. This con igu a ion enabled he model o ex ac in ica e and
complex ea u es om he MRI images, enhancing i s abili y o dis inguish be ween heal hy
and umo issues. The h ee con olu ional laye s we e ins umen al in cap u ing a ying
le els o image ea u es, anging om basic edges and ex u es in he ini ial laye s o mo e
complex and umo -speci ic pa e ns in he deepe laye s. The subsequen h ee dense
laye s consolida ed his in o ma ion, enabling he model o make p ecise inal classi ica-
ion decisions.
In addi ion o hei high accu acy, hese models employed con igu a ions ha p io i-
ized high speci ici y, a c i ical ac o in clinical applica ions. High speci ici y indica es a
low alse-posi i e a e, ensu ing ha he model accu a ely iden i ies images ha do no
con ain umo s. In a clinical con ex , his is pa icula ly impo an , as alse posi i es can
lead o misdiagnosis, causing unnecessa y anxie y o pa ien s and po en ially esul ing in
unwa an ed in asi e p ocedu es.
The models’ abili y o minimize alse posi i es enhances con idence in hei p edic ions,
ensu ing ha only cases wi h a high p obabili y o umo p esence a e lagged o u he
diagnosis and ea men .
Big Da a Cogn. Compu . 2024,8, 123 16 o 17
anspa ency enables heal hca e p o essionals o e i y he ele ance o he model’s ocus,
inc easing us and aiding in clinical decision making. S udies ha e demons a ed ha xAI
can imp o e model adop ion in heal hca e by enhancing in e p e abili y and iden i ying
po en ial biases o e o s [16].
5. Conclusions
This s udy highligh s he e ec i eness o a well-s uc u ed CNN in accu a ely de ec -
ing b ain umo s om MRI scans, wi h pe o mance ha ma ches o exceeds hose epo ed
in simila s udies, bo h in e ms o accu acy and execu ion imes. By employing ad anced
p ep ocessing me hods, comp ehensi e hype pa ame e op imiza ion, and igo ous e alu-
a ion me ics, ou model o e s a scalable, e icien , and p ecise solu ion o b ain umo
de ec ion, enhancing diagnos ic wo k lows. I s po en ial o acili a e la ge-scale popula ion
sc eening, especially in emo e o unde se ed egions, while main aining high accu acy,
unde sco es i s p ac icali y in con empo a y heal hca e se ings. Fu u e imp o emen s,
such as inco po a ing mo e di e se da ase s, adop ing ad anced a chi ec u es, and in eg a -
ing explainable AI (xAI) echniques o p o ide in e p e able esul s, could u he enhance
he abili ies o CNNs in medical imaging.
Supplemen a y Ma e ials: The ollowing suppo ing in o ma ion can be downloaded a h ps://www.
mdpi.com/a icle/10.3390/bdcc8090123/s1, Table S1: Resul s ob ained in he G id Sea ch o he
aining and alida ion subse s; Table S2. De ailed esul s ob ained a e he G id Sea ch o he
es ing subse .
Au ho Con ibu ions: Concep ualiza ion: J.C.-M. and M.D.-M.; me hodology: F.L.-P. and M.D.-M.;
so wa e: R.M.-D.-R.-O.; alida ion: J.C.-M. and F.L.-P.; o mal analysis: M.D.-M.; in es iga ion and
esou ces: J.C.-M. and M.D.-M.; da a cu a ion: R.M.-D.-R.-O.; w i ing—o iginal d a p epa a ion:
F.L.-P. and M.D.-M.; w i ing— e iew and edi ing: R.M.-D.-R.-O.; isualiza ion: F.L.-P.; supe ision:
J.C.-M. and M.D.-M.; p ojec adminis a ion: M.D.-M.; unding acquisi ion: M.D.-M. All au ho s ha e
ead and ag eed o he published e sion o he manusc ip .
Funding: This esea ch ecei ed no ex e nal unding.
Da a A ailabili y S a emen : The da a and code associa ed wi h his s udy a e a ailable upon eques
by con ac ing he co esponding au ho .
Acknowledgmen s: We wan o hank he esea ch g oup “TEP108—Robo ics and Compu e Tech-
nology” om Uni e si y o Se ille (Spain).
Con lic s o In e es : The au ho s decla e no con lic s o in e es .
Re e ences
1.
Osbo n, A.; Louis, D.; Poussain , T.; Linsco , L.; Salzman, K. The 2021 Wo ld Heal h O ganiza ion classi ica ion o umo s o he
cen al ne ous sys em: Wha neu o adiologis s need o know. Am. J. Neu o adiol. 2022,43, 928–937. [C ossRe ] [PubMed]
2.
Bi, W.L.; Hosny, A.; Schaba h, M.B.; Gige , M.L.; Bi kbak, N.J.; Meh ash, A.; Allison, T.; A naou , O.; Abbosh, C.; Dunn, I.F.; e al.
A i icial in elligence in cance imaging: Clinical challenges and applica ions. CA Cance J. Clin. 2019,69, 127–157. [C ossRe ]
[PubMed]
3. LeCun, Y.; Bengio, Y.; Hin on, G. Deep lea ning. Na u e 2015,521, 436–444. [C ossRe ] [PubMed]
4.
K izhe sky, A.; Su ske e , I.; Hin on, G.E. Imagene classi ica ion wi h deep con olu ional neu al ne wo ks. Ad . Neu al In .
P ocess. Sys . 2012,25, 84–90. [C ossRe ]
5.
Es e a, A.; Kup el, B.; No oa, R.A.; Ko, J.; Swe e , S.M.; Blau, H.M.; Th un, S. De ma ologis -le el classi ica ion o skin cance
wi h deep neu al ne wo ks. Na u e 2017,542, 115–118. [C ossRe ] [PubMed]
6.
Menze, B.H.; Jakab, A.; Baue , S.; Kalpa hy-C ame , J.; Fa ahani, K.; Ki by, J.; Bu en, Y.; Po z, N.; Slo boom, J.; Wies , R.; e al. The
mul imodal b ain umo image segmen a ion benchma k (BRATS). IEEE T ans. Med. Imaging 2014,34, 1993–2024. [C ossRe ]
[PubMed]
7.
Akkus, Z.; Galimziano a, A.; Hoogi, A.; Rubin, D.L.; E ickson, B.J. Deep lea ning o b ain MRI segmen a ion: S a e o he a
and u u e di ec ions. J. Digi . Imaging 2017,30, 449–459. [C ossRe ] [PubMed]
8.
Ro h, H.R.; Lu, L.; Liu, J.; Yao, J.; Se , A.; Che y, K.; Kim, L.; Summe s, R.M. E icien alse posi i e educ ion in compu e -aided
de ec ion using con olu ional neu al ne wo ks and andom iew agg ega ion. In Deep Lea ning and Con olu ional Neu al Ne wo ks
o Medical Image Compu ing: P ecision Medicine, High Pe o mance and La ge-Scale Da ase s; Sp inge : Cham, Swi ze land, 2017;
pp. 35–48.
Big Da a Cogn. Compu . 2024,8, 123 17 o 17
9.
Muñoz-Saa ed a, L.; Ci i -Maso , J.; Luna-Pe ejón, F.; Domínguez-Mo ales, M.; Ci i , A. Does wo-class aining ex ac eal
ea u es? a COVID-19 case s udy. Appl. Sci. 2021,11, 1424. [C ossRe ]
10.
Ka hik, R.; Menaka, R.; Ka hi esan, G.; Ani udh, M.; Nagha jun, M. Gaussian d opou based s acked ensemble CNN o
classi ica ion o b eas umo in ul asound images. IRBM 2022,43, 715–733. [C ossRe ]
11.
Gago-Fabe o, Á.; Muñoz-Saa ed a, L.; Ci i -Maso , J.; Luna-Pe ejón, F.; Rod íguez Co al, J.M.; Domínguez-Mo ales, M.
Diagnosis Aid Sys em o Colo ec al Cance Using Low Compu a ional Cos Deep Lea ning A chi ec u es. Elec onics 2024,
13, 2248. [C ossRe ]
12.
Zahoo , M.M.; Khan, S.H.; Alahmadi, T.J.; Alsah i, T.; Maz oa, A.S.A.; Sak , H.A.; Alqah ani, S.; Albanyan, A.; Alshemaim i, B.K.
B ain umo MRI classi ica ion using a no el deep esidual and egional CNN. Biomedicines 2024,12, 1395. [C ossRe ] [PubMed]
13.
Mandle, A.K.; Sahu, S.P.; Gup a, G.P. CNN-based deep lea ning echnique o he b ain umo iden i ica ion and classi ica ion in
MRI images. In . J. So w. Sci. Compu . In ell. (IJSSCI) 2022,14, 1–20. [C ossRe ]
14.
Zeineldin, R.A.; Ka a , M.E.; Bu ge , O.; Ma his-Ull ich, F. Mul imodal CNN ne wo ks o b ain umo segmen a ion in MRI: A
B aTS 2022 challenge solu ion. In P oceedings o he In e na ional MICCAI B ainlesion Wo kshop, Singapo e, 18 Sep embe 2022;
Sp inge : Be lin, Ge many, 2022; pp. 127–137.
15.
Xue, J.; Yao, Y.; Teng, Y. Mul i-modal Tumo Segmen a ion Me hods Based on Deep Lea ning: A Na a i e Re iew. In Quan i a i e
Imaging in Medicine and Su ge y; AME Publishing Company: Hong Kong, China, 2022.
16.
Pawa , U.; O’Shea, D.; Rea, S.; O’Reilly, R. Inco po a ing Explainable A i icial In elligence (XAI) o aid he Unde s anding o
Machine Lea ning in he Heal hca e Domain. In P oceedings o he AICS, Dublin, I eland, 7–8 Decembe 2020; pp. 169–180.
Disclaime /Publishe ’s No e: The s a emen s, opinions and da a con ained in all publica ions a e solely hose o he indi idual
au ho (s) and con ibu o (s) and no o MDPI and/o he edi o (s). MDPI and/o he edi o (s) disclaim esponsibili y o any inju y o
people o p ope y esul ing om any ideas, me hods, ins uc ions o p oduc s e e ed o in he con en .
