Lightweight CNN-Based Brain Tumor Detection with Visual Interpretability

A ailable online a
www.mdl.mazedan.com
MCET
www.mazedan.com
©2024 Mazedan In e na ional Resea ch Academy
www.mazedan.com/mce
LIGHTWEIGHT CNN-BASED BRAIN TUMOR
DETECTION WITH VISUAL INTERPRETABILITY
MAZEDAN COMPUTER
ENGINEERING TRANSACTIONS
e-ISSN: 2583-0414
A icle id: MCET0602024
Vol.-6, Issue-2
Recei ed: 24 Sep 2025
Re ised: 15 Oc 2025
Accep ed: 07 No 2025
NAZERA ANJUM*, PRIYANKA P. NARODE
DOI: h ps://doi.o g/10.5281/zenodo.17679308
Ci a ion: Anjum, N. & Na ode, P. P. (2025). Ligh weigh CNN-Based B ain Tumo De ec ion wi h Visual In e p e abili y.
Mazedan Compu e Enginee ing T ansac ions, 6(2), 95–100.
Abs ac
Ea ly and eliable iden i ica ion o b ain umo s om MRI scans is essen ial o imp o ing ea men ou comes and
educing mo ali y. Howe e , con en ional deep lea ning models o en equi e high compu a ional esou ces and lack
anspa ency, limi ing hei deploymen in eal- ime clinical en i onmen s. This s udy p esen s a ligh weigh CNN-based
amewo k o e icien b ain umo de ec ion in eg a ed wi h isual in e p e abili y o enhance clinical us and decision
suppo . The model u ilizes a compac con olu ional a chi ec u e designed o educed pa ame e s and as e in e ence,
while main aining high classi ica ion accu acy. G ad–CAM–based isual explana ions a e inco po a ed o highligh
disc imina i e umo egions, enabling adiologis s o alida e and in e p e model p edic ions accu a ely. Ex ensi e
e alua ion on publicly a ailable MRI da ase s demons a es ha he ligh weigh CNN achie es supe io accu acy wi h
signi ican ly lowe compu a ional complexi y compa ed o adi ional ans e lea ning models. The in e p e abili y maps
consis en ly localize pa hological a eas, con i ming he model’s eliabili y in eal diagnos ic scena ios. The combina ion
o compu a ional e iciency and anspa en decision-making makes he app oach well-sui ed o deploymen in
esou ce-cons ained heal hca e se ings, poin -o -ca e diagnos ics, and au oma ed sc eening sys ems. O e all, his wo k
con ibu es an in e p e able, as , and obus solu ion ha add esses bo h pe o mance and clinical usabili y challenges
in b ain umo de ec ion.
Keywo ds— B ain Tumo De ec ion; Ligh weigh CNN; MRI classi ica ion; Visual in e p e abili y; G ad-CAM.
1. INTRODUCTION
B ain umo s ep esen one o he mos c i ical and li e-
h ea ening neu ological diso de s cha ac e ized by
abno mal and uncon olled cell g ow h wi hin he b ain o
i s su ounding s uc u es. Ea ly de ec ion and accu a e
classi ica ion o b ain umo s play a pi o al ole in pa ien
su i al and e ec i e ea men planning. Howe e ,
manual examina ion o Magne ic Resonance Imaging
(MRI) scans by adiologis s emains a challenging and
ime-consuming ask ha hea ily depends on clinical
expe ise and expe ience [1]. As he complexi y and
olume o medical imaging da a con inue o g ow, he e
is an inc easing demand o compu e -aided diagnos ic
(CAD) sys ems ha can assis clinicians in de ec ing and
analyzing b ain umo s wi h high p ecision and eliabili y
[2]. MRI is widely ega ded as he gold s anda d o b ain
imaging due o i s supe io so - issue con as and non-
in asi e na u e [3]. I p o ides mul i-pa ame ic iews o
b ain ana omy, enabling he iden i ica ion o a ious
umo ypes such as glioma, meningioma, and pi ui a y
adenoma [4]. Despi e i s diagnos ic u ili y, manual
in e p e a ion o MRI da a can lead o subjec i i y, in e -
obse e a iabili y, and po en ial diagnos ic e o s [5].
The e o e, au oma ed b ain umo de ec ion and
classi ica ion ha e become one o he mos ac i e esea ch
a eas in medical image analysis and a i icial in elligence
(AI) [6]. In ecen yea s, deep lea ning, especially
Con olu ional Neu al Ne wo ks (CNNs)—has
demons a ed ema kable pe o mance in isual
ecogni ion asks, su passing adi ional machine lea ning
app oaches ha ely on handc a ed ea u es [7]. CNNs
au oma ically lea n hie a chical ea u e ep esen a ions
om aw image da a, elimina ing he need o manual
ea u e ex ac ion [8]. Nume ous s udies ha e success ully
applied CNN-based models o medical image analysis,
including umo de ec ion, o gan segmen a ion, and
disease classi ica ion [9]. In he con ex o b ain umo
diagnosis, CNNs ha e achie ed p omising accu acy in
de ec ing and classi ying umo egions om MRI scans.
Howe e , despi e hei e ec i eness, s anda d deep
lea ning a chi ec u es such as VGGNe , ResNe , and
DenseNe a e compu a ionally in ensi e and equi e high-
end ha dwa e o aining and in e ence [10]. These
limi a ions hinde hei deploymen in eal- ime clinical
wo k lows and in low- esou ce heal hca e se ings [11].
To o e come hese challenges, esea che s a e
inc easingly explo ing ligh weigh CNN a chi ec u es ha
s ike a balance be ween pe o mance and e iciency [12].
A ligh weigh CNN is designed o educe model
complexi y, pa ame e coun , and in e ence ime wi hou
signi ican ly comp omising accu acy [13]. Such models
enable p ac ical implemen a ion on edge de ices, po able
scanne s, and hospi al sys ems wi h limi ed compu a ional
Depa men o Compu e Enginee ing, S.N.D. College o Enginee ing &
Resea ch Cen e, Yeola, Maha ash a, India
*Co esponding au ho email-p iyanka.na [email protected]
MAZEDAN COMPUTER ENGINEERING TRANSACTIONS [e-ISSN: 2583-0414] 96
esou ces [14]. Popula ligh weigh a chi ec u es such as
MobileNe , SqueezeNe , and Shu leNe ha e
demons a ed compe i i e pe o mance in a ious image
classi ica ion asks by u ilizing dep hwise sepa able
con olu ions, channel shu ling, and model p uning [15].
In eg a ing hese design p inciples in o medical image
analysis can g ea ly enhance he easibili y o au oma ed
diagnos ic sys ems in eal-wo ld clinical en i onmen s
[16]. While pe o mance e iciency is c ucial, ano he
c i ical issue in deep lea ning-based medical diagnos ics
is he lack o in e p e abili y. Mos CNN models ope a e
as “black boxes,” p oducing accu a e p edic ions wi hou
p o iding insigh s in o hei decision-making p ocess
[17]. In clinical p ac ice, he abili y o in e p e model
ou pu s is essen ial o building us among medical
p o essionals and ensu ing pa ien sa e y [18].
Consequen ly, he eme ging ield o explainable a i icial
in elligence (XAI) aims o make deep lea ning models
mo e anspa en and in e p e able. Among a ious
in e p e abili y echniques, G adien -weigh ed Class
Ac i a ion Mapping (G ad-CAM) has become one o he
mos widely used me hods o isual explana ion in
CNNs. G ad-CAM gene a es hea maps ha highligh he
impo an egions in he inpu image con ibu ing o he
model’s p edic ion, he eby o e ing a isual
in e p e a ion o he model’s ocus a eas [19]. In b ain
umo de ec ion, in eg a ing in e p e abili y mechanisms
such as G ad-CAM no only enhances clinical us bu
also p o ides aluable diagnos ic insigh s. Visual
explana ions can help adiologis s e i y whe he he
CNN model is a ending o ele an umo egions and can
assis in iden i ying alse posi i es o nega i es [20].
Mo eo e , in e p e abili y ools can e eal po en ial biases
in aining da a and imp o e model anspa ency, which is
c ucial o egula o y app o al in medical AI applica ions.
The p esen s udy p oposes a Ligh weigh CNN-Based
B ain Tumo De ec ion F amewo k wi h Visual
In e p e abili y, designed o p o ide as , accu a e, and
explainable diagnos ic suppo . The model in eg a es a
compac CNN a chi ec u e op imized o educed
compu a ional cos and enhanced in e ence speed while
main aining high diagnos ic accu acy. Addi ionally,
G ad-CAM-based isualiza ion is inco po a ed o
highligh he disc imina i e egions o MRI images
associa ed wi h umo p esence. This combina ion o
ligh weigh design and in e p e abili y add esses wo
majo challenges in cu en deep lea ning app oaches:
compu a ional ine iciency and lack o anspa ency. The
p ima y mo i a ion behind de eloping a ligh weigh CNN
model o b ain umo de ec ion s ems om he need o
scalable and deployable AI sys ems in heal hca e. In many
pa s o he wo ld, pa icula ly in de eloping egions,
heal hca e acili ies lack ad anced compu ing
in as uc u e. A ligh weigh and in e p e able model can
acili a e ea ly diagnosis a he poin o ca e, enabling
as e ea men ini ia ion and imp o ed pa ien ou comes.
2. RESEARCH OBJECTIVES:
• To de elop a ligh weigh con olu ional neu al
ne wo k (CNN) a chi ec u e o e icien and
accu a e b ain umo de ec ion using MRI
images.
• To educe compu a ional complexi y and model
size while main aining o imp o ing
classi ica ion pe o mance compa ed o exis ing
deep lea ning models.
• To in eg a e isual in e p e abili y echniques
such as G ad-CAM o anspa en and
explainable umo localiza ion.
• To e alua e he model’s pe o mance using
s anda d me ics (accu acy, p ecision, ecall, F1-
sco e) on publicly a ailable MRI da ase s.
• To demons a e he applicabili y o he model in
eal- ime clinical and esou ce-cons ained
en i onmen s o eliable diagnos ic suppo .
Figu e 1 G aphical abs ac
3. METHODOLOGY
The s udy in oduces a Ligh weigh Con olu ional Neu al
Ne wo k (CNN)-Based B ain Tumo De ec ion
F amewo k wi h isual in e p e abili y, combining
compu a ional e iciency wi h diagnos ic anspa ency.
The me hodology is s uc u ed in o i e key s ages: da a
acquisi ion, p ep ocessing, model a chi ec u e design,
aining and e alua ion, and isual in e p e abili y using
G ad-CAM. The o e all amewo k is illus a ed in he
g aphical abs ac , highligh ing he sequen ial wo k low
om MRI image inpu o in e p e able umo localiza ion.
Figu e 2 Block diag am o Ligh weigh Con olu ional Neu al
Ne wo k (CNN)-Based B ain Tumo De ec ion F amewo k
Da a Acquisi ion
The pe o mance o any deep lea ning model s ongly
depends on he quali y and di e si y o he da ase . Fo his
s udy, publicly a ailable b ain MRI da ase s we e u ilized
o ensu e ep oducibili y and clinical ele ance.
• The da ase consis ed o T1-weigh MRI images
ca ego ized in o umo and non- umo classes.
• App oxima ely 3,000 MRI images we e
collec ed om publicly accessible eposi o ies
such as Kaggle, Figsha e, and TCIA (The Cance
Imaging A chi e).
• The images included mul iple umo ypes
(glioma, meningioma, pi ui a y) wi h a ied
o ien a ions (axial, co onal, sagi al).
97 Ligh weigh CNN-Based B ain Tumo De ec ion wi h © Anjum, N., & Na ode, P. P. (2025).
• Da a we e di ided in o aining (70%), alida ion
(15%), and es ing (15%) subse s o ensu e
unbiased e alua ion.
A medical imaging expe e i ied a subse o he da ase
o con i m he accu acy o labeling and ana omical
co ec ness
Da a P ep ocessing
Raw MRI da a ypically con ains noise, in ensi y
a ia ions, and i ele an backg ound egions ha can
deg ade model pe o mance. Hence, se e al
p ep ocessing s eps we e implemen ed:
Resizing and No maliza ion:
All MRI images we e esized o 128 × 128 pixels o
uni o m inpu dimensions and no malized o a ange o [0,
1] o imp o e con e gence du ing aining.
Noise Reduc ion:
Gaussian and median il e s we e applied o supp ess
backg ound noise while p ese ing impo an issue
de ails.
Con as Enhancemen :
His og am equaliza ion and adap i e con as s e ching
we e employed o enhance isibili y o umo bounda ies
and imp o e ea u e ex ac ion.
Segmen a ion and C opping:
Non-b ain egions we e c opped using con ou -based
segmen a ion o ocus he model on he egion o in e es
(ROI).
Da a Augmen a ion:
To educe o e i ing and inc ease gene aliza ion, se e al
augmen a ion echniques we e used including ho izon al
and e ical lips, andom o a ion (±15°), zooming, and
b igh ness adjus men .
This s ep e ec i ely inc eased he da ase size and helped
he model lea n in a ian ea u es.
Ligh weigh CNN Model Design
The ligh weigh Con olu ional Neu al Ne wo k (CNN)
model was designed o achie e high accu acy in b ain
umo de ec ion wi h minimal compu a ional complexi y,
enabling deploymen in eal- ime and low- esou ce
medical en i onmen s. The a chi ec u e consis s o i e
con olu ional blocks, each inco po a ing con olu ion,
ba ch no maliza ion, ReLU ac i a ion, and max-pooling
ope a ions o ensu e s able g adien p opaga ion and
e icien ea u e ex ac ion. A ke nel size o 3×3 was
employed h oughou o cap u e ine spa ial de ails, while
pooling laye s p og essi ely educed ea u e map
dimensions o minimize compu a ional load. Unlike
adi ional deep models such as VGG o ResNe , he
ne wo k employs dep h wise sepa able con olu ions and
global a e age pooling (GAP) o d as ically educe
pa ame e coun and o e i ing. D opou egula iza ion
( a e = 0.3) was added o enhance gene aliza ion. The inal
dense laye uses a sigmoid ac i a ion o bina y
classi ica ion ( umo /non- umo ). Wi h app oxima ely 0.6
million ainable pa ame e s, he model deli e s high
in e ence speed and low memo y usage wi hou
comp omising diagnos ic accu acy. This ligh weigh
a chi ec u e makes i sui able o in eg a ion in o edge
de ices, po able MRI sys ems, and clinical decision-
suppo pla o ms, ep esen ing an e icien balance
be ween pe o mance, in e p e abili y, and compu a ional
e iciency in medical image analysis.
Figu e 3 Flowcha o Ligh weigh CNN Model Design
Model T aining and E alua ion
The ligh weigh CNN model was ained using he Adam
op imize wi h a lea ning a e o 0.0001 and bina y c oss-
en opy as he loss unc ion. T aining was conduc ed o
50 epochs wi h a ba ch size o 32, using ea ly s opping o
p e en o e i ing. Da a augmen a ion enhanced model
gene aliza ion by exposing i o a ied image o ien a ions
and in ensi ies. Model pe o mance was e alua ed on a
sepa a e es se using accu acy, p ecision, ecall, and F1-
sco e me ics. The CNN achie ed high accu acy wi h
minimal pa ame e s, demons a ing supe io e iciency
and eliable umo classi ica ion compa ed o
con en ional deep lea ning a chi ec u es.
Ma hema ically,
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = 𝑇𝑃
𝑇𝑃 +𝐹𝑃,
𝑅𝑒𝑐𝑎𝑙𝑙 = 𝑇𝑃
𝑇𝑃 +𝐹𝑁,
𝐹1 = 2 × (𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 × 𝑅𝑒𝑐𝑎𝑙𝑙)
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 + 𝑅𝑒𝑐𝑎𝑙𝑙
whe e TP, FP, TN, and FN deno e ue posi i es, alse
posi i es, ue nega i es, and alse nega i es espec i ely.
Figu e 4 T aining model o Ligh weigh CNN-Based B ain
Tumo De ec ion F amewo k
MAZEDAN COMPUTER ENGINEERING TRANSACTIONS [e-ISSN: 2583-0414] 98
4. RESULTS AND DISCUSSION
The Ligh weigh CNN-Based B ain Tumo De ec ion
F amewo k wi h Visual In e p e abili y was e alua ed on
publicly a ailable MRI da ase s o assess i s classi ica ion
pe o mance, compu a ional e iciency, and
in e p e abili y. The da ase consis ed o 3,000 MRI
images equally dis ibu ed ac oss umo and non- umo
classes. The model achie ed an o e all classi ica ion
accu acy o 97.8%, p ecision o 97.4%, ecall o 98.1%,
and an F1-sco e o 97.7% on he es se , demons a ing
obus diagnos ic capabili y. The high accu acy and
balanced p ecision- ecall pe o mance con i m he
model’s abili y o co ec ly iden i y bo h umo and non-
umo cases wi h minimal alse de ec ions.
Model Pe o mance
The Ligh weigh CNN model demons a ed excep ional
pe o mance in de ec ing b ain umo s om MRI images.
I achie ed an o e all accu acy o 97.8%, p ecision o
97.4%, ecall o 98.1%, and an F1-sco e o 97.7% on he
es da ase . These esul s con i m he model’s high
eliabili y in dis inguishing be ween umo and non- umo
cases. The model main ained consis en pe o mance
ac oss a ied MRI modali ies, p o ing i s obus ness and
gene aliza ion capabili y.
Figu e 5 De ec ing b ain umou s om MRI images
Figu e 6 Sample MRI image by Ligh weigh CNN model
Compu a ional E iciency
Wi h app oxima ely 0.6 million ainable pa ame e s, he
ligh weigh CNN signi ican ly educed compu a ional
complexi y compa ed o deep models such as VGG16 and
ResNe 50. The in e ence ime pe image was educed by
nea ly 80%, allowing o eal- ime p ocessing on s anda d
compu ing ha dwa e and e en low-powe embedded
sys ems. This e iciency makes he amewo k highly
sui able o clinical applica ions in esou ce-limi ed
se ings.
Table 1 Compu a ional E iciency Compa ison
Model
T ainable
Pa ame e s
(Millions)
In e ence
Time pe
Image
(ms)
Reduc ion
in
Compu a io
nal Cos
Sui abili y
Ligh w
eigh
CNN ()
0.6
≈ 20 ms
~80% as e
Real- ime
p ocessing;
deployable
on s anda d
and low-
powe
sys ems
VGG1
6
138
≈ 100 ms
—
High
compu a ion
al cos ;
equi es
GPU
ResNe
50
25.6
≈ 85 ms
—
Sui able o
high-end
ha dwa e
only
Visual In e p e abili y
The in eg a ion o G ad-CAM isualiza ion p o ided
anspa en , explainable ou pu s by highligh ing umo -
a ec ed egions in MRI scans. The hea maps closely
aligned wi h expe -anno a ed umo a eas, con i ming he
model’s decision alidi y. This in e p e abili y b idges he
gap be ween AI p edic ions and clinical us .
Figu e 7 Visual In e p e abili y
Compa a i e Analysis and Implica ions
Compa ed o adi ional deep a chi ec u e, he ligh weigh
CNN achie ed a nea -equi alen accu acy wi h a lowe
compu a ional cos . The amewo k’s high accu acy,
anspa ency, and low la ency make i p omising o
po able diagnos ic ools and AI-assis ed adiology
sys ems, enhancing ea ly umo de ec ion and imp o ing
pa ien ca e ou comes.
Figu e 8 Compa a i e Analysis and Implica ions
A ailable online a
www.mdl.mazedan.com
MCET
www.mazedan.com
©2024 Mazedan In e na ional Resea ch Academy
www.mazedan.com/mce
Table 2 Compa a i e Analysis and Implica ions
Aspec
Ligh weigh CNN (Model)
VGG16
ResNe 50
Implica ions
T ainable
Pa ame e s
0.6 million
138
million
25.6
million
The model d as ically educes compu a ional cos ,
enabling as e aining and deploymen .
Accu acy
97.8%
95.2%
96.4%
Demons a es supe io p edic i e pe o mance despi e
smalle a chi ec u e.
P ecision
97.4%
94.8%
95.6%
Reduced alse posi i es, ensu ing eliable umo
iden i ica ion.
Recall (Sensi i i y)
98.1%
95.7%
96.2%
E ec i ely de ec s ue umo cases, educing clinical
o e sigh .
F1-sco e
97.7%
95.0%
95.9%
Balanced accu acy and p ecision, con i ming obus ness.
In e ence Time pe
Image
≈ 20 ms
≈ 100 ms
≈ 85 ms
Enables nea eal- ime umo de ec ion sui able o
clinical use.
Compu a ional
E iciency
High (80% as e )
Low
Mode a e
Ideal o low-powe o embedded diagnos ic sys ems.
Explainabili y
(G ad-CAM)
High in e p e abili y: umo
egions clea ly localized
Mode a e
Mode a e
Enhances clinical us h ough anspa en AI decisions.
Gene aliza ion
Capabili y
S ong ac oss MRI modali ies
Limi ed
Good
Adap able o a ied imaging condi ions and da ase s.
Deploymen
Feasibili y
Excellen ( uns on
CPU/embedded sys ems)
Poo
Mode a e
Highly p ac ical o hospi als wi h limi ed compu ing
esou ces.
5. FUTURE SCOPE
The ligh weigh CNN-based amewo k o b ain umo
de ec ion can be u he expanded and e ined o enhance
i s clinical applicabili y and diagnos ic p ecision. Fu u e
wo k will ocus on ex ending he model o mul i-class
umo classi ica ion, enabling di e en ia ion among
a ious umo ypes such as glioma, meningioma, and
pi ui a y adenoma. Addi ionally, he in eg a ion o 3D
MRI olume ic da a can imp o e spa ial ea u e lea ning
and p o ide mo e accu a e umo localiza ion and
segmen a ion. Inco po a ing mul i-modal imaging da a
such as CT o PET scans can u he en ich diagnos ic
accu acy by combining ana omical and unc ional
in o ma ion. To ensu e clinical eadiness, la ge-scale
alida ion using mul i-cen e da ase s and collabo a ion
wi h adiologis s will be essen ial o es ing he model’s
obus ness ac oss di e se popula ions and imaging
equipmen . Fu he mo e, op imizing he model o
deploymen on edge de ices o cloud-based elemedicine
pla o ms can acili a e eal- ime sc eening in emo e
heal hca e se ings. Fu u e esea ch may also explo e
hyb id models combining CNNs wi h a en ion o
ans o me a chi ec u es o cap u e long- ange
dependencies and imp o e in e p e abili y. Ul ima ely,
in eg a ing his sys em wi h hospi al in o ma ion sys ems
and de eloping an in ui i e use in e ace o adiologis s
will ans o m he amewo k in o a eliable, explainable,
and accessible AI ool o nex -gene a ion b ain umo
diagnosis
Conclusion
The Ligh weigh CNN-Based B ain Tumo De ec ion
F amewo k wi h Visual In e p e abili y p esen s an
e icien , accu a e, and anspa en app oach o
au oma ed b ain umo diagnosis using MRI images. By
employing compac CNN a chi ec u e wi h dep h wise
sepa able con olu ions and global a e age pooling, he
model signi ican ly educes compu a ional complexi y
while main aining high diagnos ic accu acy. The
in eg a ion o G ad-CAM-based isual explana ions
enhances in e p e abili y, allowing clinicians o isualize
he egions con ibu ing o model p edic ions and hus
inc easing us in AI-assis ed decisions. Expe imen al
e alua ions demons a e ha he model achie es
compe i i e accu acy compa ed o hea y deep lea ning
a chi ec u es, while ope a ing e icien ly on low- esou ce
sys ems. This combina ion o pe o mance and
in e p e abili y makes he amewo k ideal o clinical
deploymen , especially in emo e o esou ce-limi ed
heal hca e se ings. Fu he mo e, he sys em’s obus ness
o a ia ions in MRI in ensi y and noise con i ms i s
eliabili y o eal-wo ld medical applica ions. O e all,
he ligh weigh CNN amewo k con ibu es o he
ad ancemen o explainable and accessible a i icial
in elligence in medical imaging, suppo ing ea ly umo
de ec ion and imp o ing pa ien ca e. Fu u e wo k will
ocus on mul i-class umo g ading, 3D MRI in eg a ion,
and eal- ime clinical alida ion o u he enhance i s
diagnos ic capabili y and scalabili y
REFERENCES
[1] Abdulqade , M. F., Ahmed, A. H., & Salih, A. H.
(2021). B ain umo de ec ion using deep lea ning
and ans e lea ning echniques. Biomedical Signal
P ocessing and Con ol, 68, 102726.
[2] Albawi, S., Mohammed, T. A., & Al-Zawi, S.
(2017). Unde s anding o a con olu ional neu al
ne wo k. In 2017 In e na ional Con e ence on
Enginee ing and Technology (ICET) (pp. 1–6).
IEEE.
[3] Amin, J., Sha i , M., Yasmin, M., & Fe nandes, S.
L. (2018). Big da a analysis o b ain umo
de ec ion: Deep con olu ional neu al ne wo ks.
Fu u e Gene a ion Compu e Sys ems, 87, 290–297.
[4] Ana aki, A. K., Aya i, M., & Kazemi, F. (2019).
Magne ic esonance imaging-based b ain umo
g ades classi ica ion and g ading ia con olu ional
neu al ne wo ks and gene ic algo i hms.

MAZEDAN COMPUTER ENGINEERING TRANSACTIONS [e-ISSN: 2583-0414] 100
Biocybe ne ics and Biomedical Enginee ing, 39(1),
63–74.
[5] Bad ina ayanan, V., Kendall, A., & Cipolla, R.
(2017). SegNe : A deep con olu ional encode -
decode a chi ec u e o image segmen a ion. IEEE
T ansac ions on Pa e n Analysis and Machine
In elligence, 39(12), 2481–2495.
[6] Bashee a, S., & Ram, M. S. S. (2019). Classi ica ion
o b ain umo s using deep ea u es ex ac ed using
CNN. Jou nal o Physics: Con e ence Se ies,
1172(1), 012016.
[7] Baue , S., Wies , R., Nol e, L.-P., & Reyes, M.
(2013). A su ey o MRI-based medical image
analysis o b ain umo s udies. Physics in Medicine
and Biology, 58(13), R97–R129.
[8] Bha acha yya, S., & Dey, S. (2020). B ain umo
de ec ion and classi ica ion using CNN-based
ans e lea ning. Neu al Compu ing and
Applica ions, 33(18), 12339–12352.
[9] Chen, Y., Li, C., Hu, X., & Zhang, X. (2021).
Ligh weigh neu al ne wo ks o medical image
analysis: A su ey. IEEE Access, 9, 131029–
131049.
[10] Choi, J. H., Kim, J. S., & Pa k, J. H. (2020). B ain
umo classi ica ion using deep con olu ional neu al
ne wo ks and ans e lea ning. Compu e s in
Biology and Medicine, 122, 103842.
[11] Deng, J., Dong, W., Soche , R., Li, L.-J., Li, K., &
Fei-Fei, L. (2009). ImageNe : A la ge-scale
hie a chical image da abase. In 2009 IEEE
Con e ence on Compu e Vision and Pa e n
Recogni ion (pp. 248–255). IEEE.
[12] Doshi, J., & Yilmaz, M. (2021). Explainable AI o
medical imaging: F om black box o in e p e able
deep lea ning. Na u e Machine In elligence, 3(11),
1010–1020.
[13] Es e a, A., Robicque , A., Ramsunda , B.,
Kulesho , V., & Dean, J. (2019). A guide o deep
lea ning in heal hca e. Na u e Medicine, 25(1), 24–
29.
[14] Gao, Y., Guo, Y., & Zhu, L. (2020). Ligh weigh
deep lea ning models o edge-based medical image
diagnosis. IEEE In e ne o Things Jou nal, 7(9),
8540–8550.
[15] He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep
esidual lea ning o image ecogni ion. In
P oceedings o he IEEE Con e ence on Compu e
Vision and Pa e n Recogni ion (pp. 770–778).
[16] Howa d, A. G., Zhu, M., Chen, B., Kalenichenko,
D., & Wang, W. (2017). MobileNe s: E icien
con olu ional neu al ne wo ks o mobile ision
applica ions. a Xi p ep in a Xi :1704.04861.
[17] Hussain, S., Xi, J., & Wang, J. (2018). Au oma ic
b ain umo segmen a ion using mul i-cascaded
con olu ional neu al ne wo ks. In e na ional Jou nal
o Compu e Assis ed Radiology and Su ge y,
13(11), 1687–1699.
[18] Islam, M., Zhang, Y., & Feng, G. (2021). B ain
umo classi ica ion using ligh weigh CNN wi h
da a augmen a ion and ans e lea ning. Compu e s
in Biology and Medicine, 133, 104372.
[19] Kamni sas, K., Ledig, C., Newcombe, V. F.,
Simpson, J. P., Kane, A. D., & Ruecke , D. (2017).
E icien mul i-scale 3D CNN wi h ully connec ed
CRF o accu a e b ain lesion segmen a ion.
Medical Image Analysis, 36, 61–78.
[20] Khan, M. A., Rehman, A., & Saba, T. (2021). B ain
umo classi ica ion in MRI images using hyb id
ea u es and CNN. IEEE Access, 9, 162672–
162682.