symme y
S
S
A icle
An In elligen I is Based Ch onic Kidney
Iden i ica ion Sys em
Sohail Muzamil 1,2,* , Tassadaq Hussain 1,3,* , Amna Haide 3, Umbe Wa aich 3,4,
Umai Ashiq 2and Edua d Ayguadé 5
1Depa men o Elec ical Enginee ing, Riphah In e na ional Uni e si y, Islamabad 46000, Pakis an
2Depa men o Elec ical and Compu e Enginee ing, Abbo abad Campus,
COMSATS Uni e si y Islamabad, Khybe Pakh unkhwa, Abbo abad 22060, Pakis an;
[email p o ec ed]
3UCERD P L d. Islamabad, Islamabad 44000, Pakis an; [email p o ec ed] (A.H.);
[email p o ec ed] (U.W.)
4Depa men o Biomedical Enginee ing, Na owal Campus, Uni e si y o Enginee ing and Technology
Laho e, Punjab, Na owal 54890, Pakis an
5Ba celona Supe compu ing Cen e (BSC-CNS), E08034 Ba celona, Spain; Edua [email p o ec ed]
*Co espondence: [email p o ec ed] (S.M.); assadaq@uce d.com (T.H.)
Recei ed: 20 No embe 2020; Accep ed: 8 Decembe 2020; Published: 12 Decembe 2020


Abs ac :
In ecen yea s, he demand o al e na i e medical diagnos ics o he human kidney o
enal is g owing, and some o he easons behind his ela e o i s non-in asi e, ea ly, eal- ime,
and pain- ee mechanism. The ch onic kidney p oblem is one o he majo kidney p oblems,
which equi e an ea ly-s age diagnosis. The e o e, in his wo k, we ha e p oposed and de eloped
an In elligen I is-based Ch onic Kidney Iden i ica ion Sys em (ICKIS). The ICKIS akes an image
o human i is as inpu and on he basis o i idology a deep neu al ne wo k model on a GPU-based
supe compu ing machine is applied. The deep neu al ne wo k models a e ained while using
2000 subjec s ha ha e heal hy and ch onic kidney p oblems. While es ing he p oposed ICKIS
on 2000 sepa a e subjec s (1000 heal hy and 1000 ch onic kidney p oblems), he sys em achie es
i is-based ch onic kidney assessmen wi h an accu acy o 96.8%. In he u u e, we will wo k o
imp o e ou AI algo i hm and y da a-se cleaning, so ha accu acy can be inc eased by mo e
e icien ly lea ning he ea u es.
Keywo ds: i idology; heal h-ca e; embedded compu e ision; a i icial in elligence
1. In oduc ion
The p ime unc ion o he human u ina y sys em [
1
] is o elimina e was e ma e ials and
p oduc s om he body. In o de o main ain homeos asis, he kidney plays an impo an ole in he
u ina y sys em. The was e ma e ials; mine al [
2
], i amin [
3
], ni ogenous was e [
4
], ammonia [
5
],
and c ea inine [
6
] a e con e ed in o u ine, which is aken ou o he body h ough u e e s, bladde ,
and u e h a [
1
]. Some o he unc ions o he u ina y sys em include: he egula ion o plasma ionic
composi ion [
7
], plasma olume [
8
], sec e ion o ho mones, plasma hyd ogen ion concen a ion (pH),
and plasma osmola i y [9].
The kidney diso de leads o: (a) diabe es, (b) hype ension, (c) he edi a y (inhe i ance),
(d) u ological, (e) acu e enal ailu e, ( ) in ec ion o he u ina y ac , kidney (g) glome ula disease,
and (h) ascula enal [
10
–
13
]. A o esaid p oblems inc ease he isk o majo a al p oblems, such as:
ca dio ascula issues [14], human immunode iciency i us “HIV” [15], and mala ia [16].
The dea h a e due o kidney ailu e om 2005 o 2013 is inc eased by 32% [
17
]. In 2010, 2.3 o
7.1 million people los hei li es because o non-accessibili y o kidney dialysis [
18
]. App oxima ely
Symme y 2020,12, 2066; doi:10.3390/sym12122066 www.mdpi.com/jou nal/symme y
Symme y 2020,12, 2066 2 o 14
1.7 million people died annually only due o acu e kidney inju ies [
19
]. The Wo ld Heal h O ganiza ion
(WHO) epo published in 2018 e ealed ha o e all 5–10 million people los hei li es due o u ina y
diso de s [20]. These s a is ics show ha kidney ela ed diseases a e inc easing day-by-day.
Ch onic kidney disease is one o he majo p oblems o he u ina y sys em. O he kidney- ela ed
p oblems a e: (1) Au osomal Dominan Polycys ic Kidney Disease (ADPKD) [
21
], which is a gene ic
diso de ha is caused by he g ow h o nume ous cys s in he kidneys; (2) he long e m in lamma ion
and sca ing o he glome uli known as ch onic glome uloneph i is (CGN) [
22
]; (3) ch onic in e s i ial
neph i is (CIN) [
23
], in which swelling occu s in be ween he kidney ubules; and, (4) eno ascula
disease (RVD) [24], which is ela ed o he a e ies o he kidneys.
Figu e 1desc ibes he pe cen age o ch onic kidney p oblems in unde de eloped coun ies
(like Pakis an), whe e mos o he popula ion is a ec ed by Ch onic Glome uloneph i is (CGN).
The main cause o CGN is ei he he gene ic ans e o he p oblem o he immune sys em p oblem.
The body mis akenly a acks i s own cells due o he immune sys em p oblem [
25
]. Diabe ic
neph opa hy is ano he se ious kidney p oblem ha is caused by high blood glucose le els [
26
].
Diabe ic neph opa hy is nex o CGN, a ec ing 23% o he popula ion. Ch onic In e s i ial Neph i is
(CIN) is also one o he majo kidney p oblem wi h 17% a ec ees, his p oblem has many di e en
e iologies, which a e ca ego ized based on p oblem-speci ic indings.
In con en ional medical p ocedu es, pa hological ma ke s, abno mali ies in he blood o u ine,
imaging es s, and Glome ula Fil e Ra e (GFR) may e eal he kidney p oblems. Acco ding o
he Kidney Disease Ou comes Quali y Ini ia i e (K/DOQI) o he Na ional Kidney Founda ion
(NKF), he easons o ch onic kidney disease could be: (a) kidney damage o (b) a dec ease in
GFR o 60 mL/min./1.73 m
2
o less, o h ee o mo e han h ee-mon hs [
27
]. The bes a ailable
me hod o es ima ing GFR is o de e mine he le el o c ea inine in blood and u ine [
28
]. The majo
causes o ch onic kidney ailu e a e he delayed diagnosis, inancial p oblems o he pa ien , and he
una ailabili y o expe doc o s in emo e a eas [
29
,
30
]. The ea lie de ec ion o ch onic kidney disease
can educe he mo bidi y [31] and mo ali y a e [32,33].
Al e na i e medicine [
34
] and ea ly diagnoses [
35
] a e he need and demand o u u e heal hca e
echnology. This is because i p o ides he non-in asi e, and eal- ime diagnoses o human
dys unc ional o gans o imp o e heal h quali y. The usage o an al e na i e p e-diagnosis me hod [
36
]
can sa e complica ions, cos , and delay diagnosis. I idology [
37
] o i is s udy is an al e na i e
diagnosing me hod ha is used o p e-diagnose human dys unc ional o gans. I idology ca ego izes
he human i is acco ding o each o gan o he body [
38
]. The weakness o damage o an o gan is
e lec ed by he ma ks, spo s, o pa e ns in he co esponding i is egion [
39
]. These i is-pa e ns
indica e he p oblems o he speci ic body o gans [
40
]. Fo ins ance, i he e is a p oblem wi h he
s omach, hen i shows changes in he i is a ound he pupil. Simila ly, changes in kidneys appea a
he bo om edge o he i is. The igh i is indica es he o gans on he igh side o he human body,
while he le side o he i is e lec s he o gans on he le side o he body. Figu e 2shows di e en
egions ha a e ela ed o almos all human body o gans o unc ions [41].
In his p oposed esea ch wo k, we de elop an I is-based Ch onic Kidney Iden i ica ion Sys em
(ICKIS) o an ea ly assessmen o kidney p oblems. The ICKIS akes i is images as inpu and pe o ms
gene al p e-diagnosis o ch onic kidney o gan diso de ha is based on an i idology cha and a i icial
in elligen algo i hm. The algo i hm is execu ed on a GPU-based supe compu ing machine. The ICKIS
de e mines he se e i y o damage o he human kidney by iden i ying he ex u e and colo o ma k o
lesion (in he i is kidney egion) while using a CNN based a i icial in elligence algo i hm. The ICKIS
is execu ed on GPU based supe compu ing sys em. While es ing ICKIS on 2000 (heal hy and kidney
p oblems) subjec s, he esul s con i m ha he ICKIS iden i ied he ch onic kidney p oblem wi h an
accu acy o 96.8%. The p oposed ICKIS sys em would help clinicians o pe o m quick and ea ly
p e-diagnoses o he human ch onic kidney ha would no only sa e ime, bu also dec ease he
inancial bu den o e he pa ien s and he na ional heal h sys em. The p oposed echnique will open a
new pa adigm in he ield o medicine by p o iding cos -e ec i e, quick, and painless p e-diagnosis.
Symme y 2020,12, 2066 3 o 14
Figu e 1. Pie cha showing pe cen age ch onic kidney diseases in Pakis an [42].
Figu e 2. I idology Cha : Showing Kidney Region o In e es in Le and Righ I ides.
2. Rela ed Wo k
Va ious esea ch has been conduc ed on he human i is o iden i y human heal h dys unc ionali ies
and i idology e ec i eness.
SE. Hussein e al. [
39
] use i idology o diagnose kidney p oblems and ocus on Ch onic Renal
Failu e (CRF). The mul iplex Handheld came a is used o ge i is images o 340 subjec s, among hem
168 do no ha e CRF while 172 ha ing CRF ( es ed by c ea inine es ). Di e en echniques a e used o
segmen a ion and no maliza ion, such as (a) Ci cula Hough T ans o m (CHT) [
43
], (b) Homogeneous
Rubbe Shee Model (RSM), and (c) Gabo Fil e . Fea u es a e being ex ac ed while using 2D-wa ele
T ans o m and, o classi ica ion, an Adap i e Neu o-Fuzzy In e ence Sys em is used. The CRF is ound
82% and 93% co ec ly o each subjec ha ing kidney p oblems o no , espec i ely. This p oposed
esea ch gi es p omising esul s o he p e-diagnosis o kidney CRF.
To diagnose medical condi ions, Sa a Ame i a e al. [
44
] ha e pe o med esea ch wo k while
using human i ises. The CASIA i is image da abase has been used du ing he esea ch wo k. In o de
o de ec ci cula shapes, CHT is used, while Hessian analysis is used o geome ical enhancemen [
45
].
Mul iscale blobness [
46
] is used o spo ou di e en sizes in he i is egion o he kidney egion and
connec ed componen labeling is done in o de o g oup pixels o he candida e egion. Thei esul s
con i m ha he algo i hm iden i ied a subjec wi h CRF and no-CRF wi h an accu acy o 82% and
93%, espec i ely.
Symme y 2020,12, 2066 4 o 14
Agus P ayi no e al. [
47
] unde ake hei s udies on kidney complica ions ha a e caused by
diabe es. The i is egion o in e es (ROI) is used acco ding o he i idology cha . To analyze he ROI,
colo cons ancy [
48
] and independen componen analysis [
49
] a e used. They use Dino-Li e Ve . 2.0
o cap u ing i is images. The esul s show ha 76% o pa ien s a e diabe ic wi h kidney complica ions.
The esul s o he p oposed sys em a e e i ied by he pa ien medical c ea inine le el.
Adhi e al. [
50
] classi y he las s age o kidney CRF using he i idology cha . They cap u e i is
images o 82 subjec s using he Dino-Li e Digi al i is scope AMH-RUT came a sys em ha ing 1.3 M
pixels esolu ion and 30 ame a e. The wa e shed me hod o ea u e segmen a ion is used in o de o
examine he b oken issues in he i is. The au ho uses he mean o he g adien , mean o he wa e shed,
mean o wa e shed bina y, and size o b oken issue as ea u es, which a e inpu ed o Suppo Vec o
Machine (SVM) o machine lea ning-based classi ica ion. Du ing he esea ch, 61 hemodialyses and 21
heal hy pa ien s a e used o analysis and he esul s show ha he algo i hm iden i ied hemodialysis
pa ien s wi h an accu acy o 87.5%.
Saman and Aga wal [
51
] use di e en machine lea ning echniques, including (a) Bina y T ee
model, (b) SVM, (c) Adap i e Boos ing model, (d) Neu al Ne wo k, and (e) Random Fo es model
o classi ica ion and diagnosis o diabe es. A o esaid echniques use I is-SCAN-2 wi h c oss-ma ch
echnology o ob ain g ay in a ed images o he in es iga ion o (a) s a is ical, (b) ex u e, and (c)
disc e e wa ele -based ea u es. The au ho s pe o m modi ied CHT o ex ac he a o emen ioned
ea u es. Rubbe -shee no maliza ion is used o map i is ea u es on ixed ec angula ep esen a ion.
They di ided 338 pa ien s in o wo g oups, one ha ing a sign o diabe es include 180 subjec s, and he
o he wi h no sign ha e 158. The esul s con i m ha he model iden i ied diabe es wi h an accu acy
o 89.63%.
Rossi E win e al., in [
52
], use i idology o de elop a da abase o colon diso de . Au onomic
Ne ous W ea h (ANW) o he i is is ocused on diges i e p oblems. Hough ans o m and RSM a e
used o ea u e ex ac ion. Six y subjec s a e used, in which 25 a e con olled while 35 a e p o en
colon diso de . The esul s show an 8% e o . This esea ch shows p og essi e esul s and concludes
ha , o ob ain be e and eliable esul s, a la ge da a-se is equi ed.
Jamal e al. [
53
] wo k o de elop a ool o de ec ing diabe es h ough he co esponding egion
o he panc eas o gan, while using colo coding and isual inspec ion. Segmen a ion is done by
Daugman’s app oach and no maliza ion is done while using RSM; adap i e his og am equaliza ion
o he spa ial domain me hod, which di ec ly ope a es on pixels, is used o image enhancemen .
Fea u es o 20
×
20 ROI a e ex ac ed while using P incipal Componen Analysis (PCA). Feed
o wa d-back p opaga ion based, A i icial Neu al Ne wo k is applied o classi ica ion wi h he
Le engu g–Ma qua d back-p opaga ion o weigh and bias o ne wo k as aining unc ion. Ou o
10 subjec s, eigh subjec s a e iden i ied as diabe ic subjec s. La e on, all o hese eigh a e e i ied as
diabe ic ype II using an insulin no mali y es . 100 % accu acy is achie ed o he panc eas o gan.
UCERD P i a e Limi ed [
54
], I idology Resea ch G oup [
55
], has de eloped imaging sys em
a chi ec u e [
56
] and algo i hms [
57
] o i idology based dys unc ional o gans iden i ica ion. The g oup
p oposed in elligen algo i hms o heal h-ca e. I uses high-pe o mance a chi ec u e [
58
,
59
] ha
manages an eno mous olume o pa ien images [
60
,
61
] in o ma ion, and e ec i ely uses hem o
make a decision.
3. Me hodology: An In elligen I is Based Ch onic Kidney Iden i ica ion Sys em
This sec ion explains he me hodology o he I is-based Ch onic Kidney Iden i ica ion Sys em
(ICKIS). The sec ion is subdi ided in o ou main sec ions: he Kidney Pa hology, he I is-Acquisi ion
Sys em, he P ocessing Sys em, and he A i icial In elligence (AI) Algo i hm.
3.1. Kidney Pa hology
The sec ion explains he pa hology o he kidney agains ch onic kidney disease while using
con en ional me hods and i idology echniques.
Symme y 2020,12, 2066 5 o 14
3.1.1. Con en ional Me hod
Con en ional me hods use Glome ula Fil a ion Ra e (GFR), in o de o de e mine he s age o
ch onic kidney disease. In GFR, samples o he pa ien ’s blood and u ine a e compa ed. A es o
c ea inine using he blood sample is conduc ed in o de o measu e GFR (i c ea inine alue inc eases,
he GFR dec eases).
GFR measu es millili e s o was e ha is il e ed by kidneys in a minu e. The kidneys o a
heal hy pe son il e o e 90 mL o was e pe minu e [
27
]. The ollowing a e he s ages o ch onic
kidney disease conce ning he GFR desc ibed by he Kidney Disease Ou come Quali y Ini ia i e
(K/DOQI) [62]:
S age1: (GFR > 90 mL/min./1.73 m
2
) kidney unc ioning is no mal up o 90%, bu wi h mino
damage (p o ein in u ine).
S age 2: (GFR 60–89 mL/min./1.73 m
2
) Acu e Kidney Disease, mild loss in kidney unc ioning,
and i is wo king up o 60 o 89%.
S age 3: (GFR 30–59 mL/min./1.73 m
2
) Ch onic Kidney Disease, mode a e loss o kidney
unc ioning, and i is dec eased o 45 o 59%, bu i may lead o a mode a e o se e e loss o kidney
unc ioning, up o 30 o 44%.
S age 4: (GFR 15–29 mL/min./1.73 m
2
) Se e e Ch onic Disease, se e e loss o kidney unc ioning
om 15 o 29%.
S age 5: (GFR < 15 mL/min./1.73 m
2
) End-S age Renal Disease (ESRD) i.e., kidney ailu e s age.
A his s age, he e is a o al loss o less han 15% unc ioning.
The GFR le el keeps on changes wi h ime o di e en age g oups. In hese ea ly s ages,
GFR alone does no secu e he inding ha ’s why o he kidney inspec ion es s like (a) blood, (b) u ine,
and (c) kidney scan can be used in o de o de e mine he kidney anomalies. In s age 4 and s age 5,
he se e e symp oms a e being el by he pa ien s, which leads him/he o ch onic kidney ailu e.
In his esea ch wo k, pa ien s who a e g ouped unde s ages 4 and 5 a e ocused, as indica ed by hei
GFR, and he emphasis is on he p esence o absence o ch onic kidney disease o s age 4 and s age 5.
3.1.2. I idology
In he ligh o clinical assessmen s om pa ien da a and epo s, a ious pa e ns and ea u es o
ch onic kidney a e illus a ed by u ilizing he i idology cha (shown in Figu e 2). The i idology cha
ha is shown in Figu e 2is di ided in o 12 o (360
◦
) po ions. The a ge ed po ion o ch onic kidney
disease (CKD) agains s age 4 and s age 5 lies unde posi ion 5.35–5.95 (252
◦
–268
◦
) o he igh eye
and in posi ions 6.05–6.60 (272
◦
–288
◦
) o he le eye. The igh eye ep esen s he igh enal o gan
and he le eye ep esen s he le enal o gan. Figu e 3shows he egion o in e es (ROI) and s age 4
and 5, ma ks and pa e ns o bo h he le and igh eyes.
(a) (b)
Figu e 3.
I idology egion o in e es (ROI) Ma ks and pa e ns showing kidney p oblems (
a
) Righ
Eyes (b) Le Eyes.

Symme y 2020,12, 2066 6 o 14
I is images o subjec s ha ing ch onic kidney p oblem con ained he ollowing pa e ns and
ma ks: (a) A long solid black line. (b) One o mo e, la ge o small open lesion. (c) Change in colo
o he i is issue. As kidney p oblems ge wo se, he GFR le el dec eases, and, simila ly, he ma ks,
lesions, and c yp s appea in he i is in he co esponding egion. Fu he mo e, in Figu e 3i can be
clea ly seen he a ia ions in he i is pa e n o colou o igh and le eye i ises. Change in he ma ks
and pa e ns on he i ises e lec he a e -e ec s o ch onic kidney disease.
3.2. I is-Acquisi ion Sys em (IAS)
The I is-Acquisi ion Sys em (shown in Figu e 4) collec s he da a om di e en subjec s while
using a high- esolu ion i is came a along wi h he pa ien ’s de ailed medical his o y and pe o ms
s a is ical s udy based on he ch onic kidney p oblem and s o e i in he da a-se . To acqui e he da a
he IAS is linked wi h hospi als and medical labs. Du ing i is acquisi ion, he subjec s wi h ch onic
kidney disease a e ga he ed and labeled. These labelings a e based on subjec medical epo s and
his o y wi h GFR < 30. Figu e 4shows he IAS wo king en i onmen , which acqui es ue RGB colo ed
images in high esolu ion. The i is came a uses a cha ge-coupled de ice (CCD) based image senso 24
Megapixel Digi al Single Lens Re lex (DSLR) Canon Came a wi h a 100 mm Mac o lens ha ga he s
colo images wi h a esolu ion up o 2560
×
1920 pixels. The mac o lens con igu a ion is con igu ed
specially o i idology ha ing ISO se ings be ween 100 o 800 and a na ow ape u e ha esul s in
good exposu e. The IAS has a USB in e ace ha can in e ace easily wi h any compu e de ice and
allows o da a o a el a an a e age o en imes he speed o he no mal pa allel po .
3.3. P ocessing Sys em
The sec ion desc ibes he UCERD GPU-based supe compu e sys em a chi ec u e (shown in
Figu e 5), which is used o execu e he ICKIS algo i hm. UCERD GPU-based supe compu e is an
In el Xeon p ocesso s ha ing enough p ocessing powe and speed o handle he in ensi e p ocessing
and N idia Pascal GPU accele a o s. The UCERD GPU-based supe compu e sys em uses ou
nodes, and each node is equipped wi h 12 co es and wo h eads/co es, a o al o 24 h eads, and 2
NVidia-Pascal GPUs. The dis ibu ed nodes a e in e connec ed using Gigabi in e connec ions and
hey u ilize he Cen OS Linux Ope a ing Sys em. This sys em has a peak pe o mance o 48.5 Te a
loa ing-poin ope a ions pe second.
Figu e 4. Pho og aph o I is Acquisi ion Sys em.
Symme y 2020,12, 2066 7 o 14
Figu e 5. Pho og aph o UCERD GPU-based Supe compu ing Clus e .
3.4. A i icial In elligence (AI) Algo i hm
Figu e 6shows he block diag am o he AI Algo i hm. Py hon 3.6 p og amming language,
OpenCV compu e ision, and Tenso -Flow-2 deep lea ning amewo k a e used o p og am he
I is-based Ch onic Kidney Iden i ica ion Sys em (ICKIS) algo i hm.
Figu e 6. AI algo i hm used o ain and es he da a-se o ch onic kidney p oblem.
The con en ional neu al ne wo k (CNN) is used in his esea ch wo k o ex ac and classi y
co e ea u es au oma ically while using il e s. These ea u es a e being ex ac ed om he egion o
in e es (ROI), excluding o he i is a eas. The size o he il e and weigh s we e au oma ically adjus
based on ini ial aining esul s. La e egula iza ion echniques a e being applied, and he weigh s a e
adjus ed o ine- uning. I is no ed ha , as he il e s inc ease om 96 o he inpu image, he a ge ed
Symme y 2020,12, 2066 8 o 14
pa e ns and ma ks a e no encoded. Addi ionally, he inc ease in il e size does no a ec he accu acy,
bu inc eases he ime. On he o he hand, he weigh s o he il e s a e being added and adjus ed based
on eal- ime ails esul s. The ea u ed in o ma ion goes h ough he successi e con olu ion il e s
and ge s ou inc easingly pu i ied. A he max-pooling s age [
63
,
64
], he mos app op ia e ea u es a e
ex ac ed, and he ully connec ed laye a he end-s age is used o classi ica ion pu poses. The AI
algo i hm has 195,529,698 ainable pa ame e s. The pa ame e iden i ica ion o he ICKIS algo i hm
is added by eading om he sou ce and by eal- ime expe imen s-based expe iences.
The inpu o he AI algo i hm is aining and es ing da a-se s ha consis o i is images,
and each image is labeled wi h heal hy kidney and ch onic kidney classes. The inpu esolu ion
o ou da a-se is 2560
×
1920
×
3, he e o e he inpu laye esizes he da a-se in o
640 ×640 ×3
.
The AI algo i hm uses h ee con olu ion laye s o ex ac he ea u es o he images while using
con olu ion il e s, wi h 2
×
2 max-pooling. The i s con olu ion laye applies 96 il e s wi h a
window size o 9
×
9. The pooling laye 1 (Pooling 1) akes he esul s o con olu ion laye 1 (Con 1)
(as shown in Figu e 6) and down samples he image o 320
×
320 by using 2
×
2 max-pooling ha ing a
s ide o 2. Max pooling pe o ms a educ ion and o e - i ing in dimensions o he image by aking
he maximum pixel alue o a g id. La e , he Con 2 [
65
] laye applies 64 il e s ha ing a window size
o 5
×
5. The Con 3 laye applies 32 il e s wi h a window size o 3
×
3. The ou pu o he con olu ion
laye is ou -dimensional; he e o e, we add 2 Fully Connec ed (FC) laye s ha con e he inpu laye
in o wo-dimensional. La e , a so -max laye is added wi h each FC laye . Subsequen ly, he ou pu
o FC laye s is connec ed wi h he inal laye , which is he so -max laye . The inal laye is used o
classi y he image in o one ca ego y ou o wo gi en ca ego ies acco ding o he ea u es o he image.
The ou pu o he ICKIS algo i hm has wo classes/ca ego ies ch onic kidney and heal hy kidney.
The pa ame e iden i ica ion o he ICKIS algo i hm is added by eading om he sou ce and by
eal- ime expe imen s based expe iences.
4. Resul s and Discussion
In his sec ion, we pe o m he es ing and alida ion o he p oposed ICKIS sys em and discuss
da a-se s, accu acy, and pe o mance o he algo i hm. A e wa ds, we compa e ou ICKIS algo i hm
wi h al eady exis ing algo i hms. This sec ion is di ided in o ou subsec ions: he da a-se ; he aining
me hodology; he alida ion and accu acy; and, compa ison wi h o he p oposed algo i hms.
4.1. I is Da a-Se
The da a we e ga he ed om he IIMCT-Pakis an Railway Hospi al and NIRM-Na ional Ins i u e
o Rehabili a ion Medicine. Wi h mo e han wo di e en ins i u es collabo a ing in his s udy, all o
he esea ch publica ion e hics and anonymi y o subjec s a e espec ed a all s ages o he p ojec .
All pa s o he da a collec ion a e ca ied ou by skilled p o essionals (Biomedical Depa men , Riphah
In e na ional Uni e si y). The au ho s ha e ca ied ou mee ings wi h all pa ies and each g oup
knows hei ole in he p ojec . Regula mee ings a e held be ween all pa ies o o esee and plan o
any issue ha may a ise.
A o al o ou housand subjec s a e used in his esea ch, in which 2000 subjec s ha e ch onic
kidney disease and he emaining 2000 a e subjec s wi h heal hy kidneys. The ch onic kidney disease
subjec s a e in es iga ed h ough hei medical epo s and only hose subjec s a e conside ed ha
ha e GFR < 30. Ten images om each subjec a e aken wi h i e images o he le eye and i e images
o he igh eyes. Each i is image has a esolu ion o 2560 ×1920 ×3 pixels.
The da a-se (as shown in Table 1) holds 40,000 i is images o bo h le and igh i is, ou o which
20,000 images belong o heal hy subjec s, whe eas he emaining 20,000 images belong o subjec s
ha ing ch onic kidney disease. The da a-se s a e u he ca ego ized in o T aining Da a-Se and
Tes ing Da a-Se . Each Da a-Se holds 20,000 i is images, ou o which 50% images comp ise subjec s
wi h ch onic kidney disease, and he emaining 50% images a e o heal hy subjec s.
Symme y 2020,12, 2066 9 o 14
Table 1. T aining and Tes ing Da a-Se .
Da a-Se Heal hy Ch onic To al
Subjec s/Images Subjec s/Images Subjec s/Images
T aining 1000/10,000 1000/10,000 2000/20,000
Tes ing 1000/10,000 1000/10,000 2000/20,000
To al 2000/20,000 2000/20,000 4000/40,000
4.2. T aining Me hodology
In his sec ion, we explain how he ICKIS AI algo i hm (shown in Figu e 6) is ained on T aining
Da a-Se (Table 1) and UCERD-GPU based supe compu e . The AI algo i hm has 195,529,698 ainable
pa ame e s. In o de o ain he AI algo i hm, we applied da a-le el pa allelism ha di ides he
aining da a in o mul iple subse s, each se is execu ed on a sepa a e node. To pe o m synch oniza ion,
he S ochas ic G adien Descen (SGD) is applied ha inco po a es he esul o each compu a ion wi h
he AI algo i hm. We ha e applied aining o he ICKIS AI algo i hm wi h s ochas ic g adien descen
by u ilizing he Ho owa d [
66
] deep lea ning oolki . The dis ibu ed execu ion on a supe compu e
sys em uses eigh eplicas each pe o ming p ocessing on an NVidia Pascal GPU. The aining
me hodology on he UCERD-GPU based supe compu e akes 94 min. o ain he ICKIS AI algo i hm.
The aining me hodology p o ides a as , scalable, in elligen , and high-pe o mance compu ing
based me hod o he p ima y diagnosing o ch onic kidney disease.
4.3. Valida ion and Accu acy
The es ing da a-se (in Table 1) is used in o de o alida e he ICKIS. Because each subjec has 10
images, he subjec is conside ed o be a ch onic kidney pa ien i he AI algo i hm has ch onic kidney
pa e ns in 5 o mo e o i s i is images. I he ICKIS AI algo i hm de ec s ch onic kidney p oblems
ei he in he igh o le i is o he subjec , i is conside ed o ha e ch onic kidney disease. In o de o
ecognize he heal hy kidneys, images o bo h eyes mus ha e a heal hy kidney unc ion, so he subjec
is conside ed o be heal hy i he AI algo i hm has heal hy kidney pa e ns o all o i s en i is images.
The esul s o he es subjec s (1000 subjec s wi h ch onic kidney p oblems and 1000 subjec s wi h
heal hy kidneys) in Table 2, espec i ely, show he co ec classi ica ion, namely subjec s wi h heal hy
kidney and ch onic kidney is 95.4% and 98.2%, espec i ely. On he o he hand, he alse classi ica ions
o subjec s wi h a heal hy kidney a e 4.6%.
False-nega i e classi ica ion is conside ed o be one o he mos impo an ac o s in he
pe o mance o he designed classi ie , and i is ound o be 1.8%. This pe cen age indica es ha ,
ou o 100 subjec s wi h ch onic kidney p oblems, almos wo subjec s may be misdiagnosed.
Table 2.
A i icial In elligence (AI) Algo i hm ou comes o he es ing da a-se o heal hy and
ch onic kidney.
Heal hy Kidney Ch onic Kidney
Numbe o es ed Subjec s 1000 1000
Co ec Classi ied 954 982
False Classi ied 46 18
The ICKIS AI algo i hm is es ed on a sepa a e Tes ing Se (shown in Table 1). The esul s show
(Figu e 7) he accu acy o he ICKIS AI algo i hm. The esul shows ha he algo i hm iden i ies he
ch onic kidney subjec wi h an accu acy o 96.8%. The ICKIS algo i hm iden i ies he imaging esul s
o he i is scan ha ing a iable and non-speci ic pa e ns le el o kidney p oblem. The algo i hm can
also e eal he e olu ion o kidney p oblems wi h he ime and e ec o medica ion on he pa ien .
The c ea inine es o he subjec s e i ies he esul s.