Ci a ion: S obodo a, P.; Se hia, K.;
S akos, P.; Va yso a, A. Au oma ic
Hepa ic Vessels Segmen a ion Using
RORPO Vessel Enhancemen Fil e
and 3D V-Ne wi h Va ian Dice Loss
Func ion. Appl. Sci. 2023,13, 548.
h ps://doi.o g/10.3390/
app13010548
Academic Edi o s: A sushi Te amo o
and Tomoko Ta eyama
Recei ed: 18 Oc obe 2022
Re ised: 20 Decembe 2022
Accep ed: 21 Decembe 2022
Published: 30 Decembe 2022
Copy igh : © 2022 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/).
applied
sciences
A icle
Au oma ic Hepa ic Vessels Segmen a ion Using RORPO Vessel
Enhancemen Fil e and 3D V-Ne wi h Va ian Dice
Loss Func ion
Pe a S obodo a †, Khya i Se hia †, Pe S akos *,† and Alice Va yso a
IT4Inno a ions, VSB—Technical Uni e si y o Os a a, 70800 Os a a-Po uba, Czech Republic
*Co espondence: pe [email p o ec ed]
† These au ho s con ibu ed equally o his wo k.
Abs ac :
The segmen a ion o hepa ic essels is c ucial o li e su gical planning. I is also a
challenging ask because o i s small diame e . Hepa ic essels a e o en cap u ed in images o low
con as and esolu ion. Ou esea ch uses il e enhancemen o imp o e hei con as , which helps
wi h hei de ec ion and inal segmen a ion. We ha e designed a speci ic usion o he Ranking
O ien a ion Responses o Pa h Ope a o s (RORPO) enhancemen il e wi h a aw image, and we
ha e compa ed i wi h he usion o di e en enhancemen il e s based on Hessian eigen ec o s.
Addi ionally, we ha e e alua ed he 3D U-Ne and 3D V-Ne neu al ne wo ks as segmen a ion
a chi ec u es, and ha e selec ed 3D V-Ne as a be e segmen a ion a chi ec u e in combina ion wi h
he essel enhancemen echnique. Fu he mo e, o ackle he pixel imbalance be ween he li e
(backg ound) and essels ( o eg ound), we ha e examined se e al a ian s o he Dice Loss unc ions,
and ha e selec ed he Weigh ed Dice Loss o i s pe o mance. We ha e used public 3D Image
Recons uc ion o Compa ison o Algo i hm Da abase (3D-IRCADb) da ase , in which we ha e
manually imp o ed upon he anno a ions o essels, since he da ase has poo -quali y anno a ions
o ce ain pa ien s. The expe imen s demons a e ha ou me hod achie es a mean dice sco e o
76.2%, which ou pe o ms o he s a e-o - he-a echniques.
Keywo ds:
li e essel segmen a ion; hepa ic essel segmen a ion; 3D U-Ne ; 3D V-Ne ; Hessian;
F angi; Sa o; RORPO; imp o ed anno a ions; Weigh ed Dice Loss unc ion
1. In oduc ion
Li e essel segmen a ion om Compu ed Tomog aphy (CT) images is impo an
be o e li e su ge y esec ion. Clinicians mus know he li e ’s mo phology and i s enous
sys em loca ion and diame e o plan he pa h o he su gical cu ing. Hence, he segmen a-
ion o he small s uc u es such as hepa ic eins is a c ucial ask. These s uc u es ha e a
small diame e , and low image con as and esolu ion. Today, accu a e hepa ic essel labeling
s ill elies p ima ily on doc o s’ manual segmen a ion, which is ime-consuming and depends
on he specialis s’ expe ise and skills. As a esul , an au onomous, obus , and accu a e
hepa ic ascula segmen a ion algo i hm is c i ical and highly desi ed.
In his wo k, we add ess he issues o low-con as essels and labou -in ensi e manual
essel segmen a ion. We de elop a ully au oma ic segmen a ion me hod ha p o ides
high accu acy and p ecision o hepa ic eins segmen a ion. The solu ion does no equi e
a doc o ’s assis ance, such as in semi-au oma ic me hods.
Al hough many essel segmen a ion app oaches ha e been in es iga ed by o he au-
ho s, such as he h eshold me hod, egion g owing, and mo phology-based me hods, he
s a e-o - he-a me hods use Con olu ional Neu al Ne wo ks (CNNs) and Deep Lea ning
(DL) s a egies. We use he DL app oach in ou me hod as well. Typically, enhancemen
il e s such as hose based on he Hessian ma ix a e used o imp o e he con as and
isibili y o he ascula s uc u es. As opposed o ha , ou me hod is based on he usion
Appl. Sci. 2023,13, 548. h ps://doi.o g/10.3390/app13010548 h ps://www.mdpi.com/jou nal/applsci
Appl. Sci. 2023,13, 548 2 o 22
o he aw image wi h he RORPO (The Ranking O ien a ions Responses o Pa h Ope a o s)
[
1
] enhanced image. This is hen u ilized by he segmen a ion a chi ec u e o 3D U-Ne
o 3D V-Ne . The inal segmen a ion pipeline consis s o RORPO and 3D V-Ne , which
p o ide he bes segmen a ion esul s.
In he pape , se e al expe imen s a e pe o med, which es he main pa s o he
p oposed me hod and make compa isons wi h al e na i e s a e-o - he-a app oaches.
The i s expe imen compa es i e classical essel enhancemen il e s sepa a ely: F angi,
Hessian, Meije ing, RORPO, and Sa o. The ollowing wo expe imen s e alua e a usion o
di e en essel enhancemen il e s wi h wo a ian s o segmen a ion models (3D U-Ne
and 3D V-Ne ). The usion o he RORPO enhancemen il e wi h he CT image o de ec
he li e essel s uc u es suppo ed by he 3D V-Ne is p o en o be he mos e icien . We
also sugges he bes alue o he blending coe icien be ween he RORPO enhanced image
and he aw CT image. In such a se up, we u he expe imen wi h di e en a ian s o
dice loss unc ions o sol e he p oblem o pixel imbalance be ween he o eg ound and
backg ound ca ego ies. The ollowing expe imen hen pe o ms he abla ion s udy o he
selec ed a chi ec u e. In he las expe imen , a compa ison be ween ou p oposed algo i hm
and o he algo i hms is p o ided.
The p oposed algo i hm is p o en o be e ec i e, obus , and accu a e o li e essel
segmen a ion, e en o images wi h low con as and high noise. Expe imen s a e pe -
o med on he public 3D-IRCADb [
2
] da ase ; he a e age dice sco e and p ecision a e
76.2% and 77.7%, espec i ely.
Du ing he wo k on his opic, we ha e used 3D Slice [
3
] so wa e, o which we ha e
c ea ed a so wa e ex ension ha can p o ide doc o s and o he heal hca e p o essionals
wi h ools o au oma ic essels segmen a ion, and ools o alida ing he segmen ed
da ase s. Employing deep lea ning and 3D Slice , we can segmen ascula s uc u es
au oma ically, imp o e he anno a ions, au oma ically collec he anno a ed da ase s, and
hen ine- une he neu al ne wo k models so hey can be quickly exposed o u he use in
au oma ic segmen a ion.
2. Rela ed Wo k
The end o he p oposed segmen a ion me hods leads o he au oma ion o li e
essel segmen a ion [
4
,
5
]. The me hods o li e essel segmen a ion om CT abdominal
images can be di ided in o se e al g oups: acking-based algo i hms, ac i e con ou s,
and machine lea ning [
6
]. The popula app oach o segmen ing li e essels is based on
machine lea ning me hods and Deep Lea ning (DL) speci ically.
The applica ion o 3D U-Ne [
7
–
10
] is o en used o he segmen a ion o li e essels
om CT images. The 3D U-Ne is composed o analy ical and syn he ical pa s [
11
], simila
o he s anda d U-Ne a chi ec u e, bu i uses 3D ope a o s ins ead. Yu e al. [
7
] desc ibed
he 3D Residual U-Ne echnique, which is buil on 3D U-Ne wi h a 3D mo phological
closu e ope a ion in he pos p ocessing phase. The algo i hm was es ed on hei p i a e CT
abdominal da ase . Huang e al. [
8
] used 3D U-Ne wi h a combina ion o he a ian dice
loss unc ion. The segmen ed eins we e mo e con inuous and comple e when aining was
ca ied ou using he enhanced manual expe anno a ions a he han he o iginal da ase .
The es s we e ca ied ou using he 3D-IRCADb, SLi e 007 [
12
], and p i a e da ase s. A -
ane e al. [
9
] expe imen ed wi h h ee dis inc 3D U-Ne app oaches: basic U-Ne , Mul iRes
U-Ne , and Dense U-Ne . On he 3D-IRCADb da ase , hey de e mined ha Mul iRes U-Ne
a chi ec u e was supe io o segmen ing hepa ic blood eins.
Golla e al. [10]
employed
an ensemble echnique. They combined he p edic ions o ne wo ks in o ensemble E and
a e aged he esul ing p obabili y o each class. P obabili y maps o all ne wo ks (2D
U-Ne , 3D U-Ne , 2D V-Ne , and 3D V-Ne ) and ensembles we e esampled back o he
o iginal da a, esolu ion and he maximum p obabili y was applied o he da a o ex ac
he p edic ed segmen a ion. The app oach was es ed on he publicly a ailable da ase s
3D-IRCADb and MICCAI [12], wi h unsui able da a being excluded.
Appl. Sci. 2023,13, 548 3 o 22
V-Ne [
13
] is he nex commonly used neu al ne wo k a chi ec u e ha has been s ud-
ied in ecen yea s o segmen ing li e blood essels. Su e al. [
14
] p oposed he DV-Ne
algo i hm. DV-Ne ep esen s a V-Ne wi h a dense block s uc u e. They u ilize a combina-
ion o DCDS (dense connec ion downsampling app oach) and D-BCE loss unc ion o
cap u ing blood essel s uc u e. The me hod was es ed on 3D-IRCADb.
Yang e al. [15]
used an imp o ed V-Ne o segmen li e blood essels on he 3D-IRCADb da ase . Im-
p o ed V-Ne is based on in e -scale dense connec ions in he decode .
Al ini e al. [16]
used 2.5D V-Ne o he segmen a ion o li e blood essels. They ained he me hod on
he 3D-IRCADb da ase and es ed i on hei da ase . They used T e sky index-based loss
unc ion in combina ion wi h 2.5D V-Ne . The T e sky index ensu es a sho con e gence
ime due o he unbalanced oxels p oblem.
A di e en app oach han he common u iliza ion o ei he he U-Ne o V-Ne a -
chi ec u e is applied in [
17
,
18
]. In [
17
], he au ho s p opose a usion ne wo k called
T ans usionNe . This me hod is based on a ans o me o seman ic segmen a ion. The
au ho s use he 3D-IRCADb and LiTS [
19
] da ase s o p e- aining, and hei da a o
ine- uning he segmen a ion ne wo k. Xu e al. [
18
] applied a deep neu al ne wo k based
on he boo s apping echnique o he 3D-IRCADb da ase . A con ex combina ion o he
model p edic ions and p elimina y p edic ions wi h he subsequen applica ion o a noise
il e is used o segmen he li e blood essels.
The e is a ull body o wo k ha ocuses on image enhancemen ha can be applied
be o e aining he segmen a ion model. This is especially ue i complex s uc u es such
as essels a e segmen ed. Shahid and Taj [
20
] apply a se o p ep ocessing ope a ions,
including il e ing ia he F angi il e , o enhance e inal essels be o e hei segmen a ion.
Soom o e al. [21]
use s eps o supp ess i egula illumina ion and o imp o e low and
a ying con as ia con as -limi ed adap i e his og am equaliza ion o he ask o e inal
essel segmen a ion. Blaiech e al. [
22
] s udied he e ec o enhancing he 2D images
o co ona y a e y segmen a ion. Di e en enhancemen me hods (F angi, CLAHE, and
RORPO) we e used in no mal condi ions and in he p esence o noise. Lamy e al. [
23
]
compa ed se en di e en essel enhancemen il e s on eal (3D-IRCADb) and syn he ical
(VascuSyn h [
24
]) da ase s. RORPO has been e alua ed as being he bes -pe o ming
me hod on he whole li e a ea when applied o he eal 3D-IRCADb da ase .
Wi h a ocus on he use o essel-enhancing il e s as a pa o he pipeline o hepa ic
essel segmen a ion, Su a achakan e al. [
25
] applied ou enhancemen il e s (Hessian,
F angi, Sa o, and Meije ing), and p oposed o use hei ou comes in wo di e en segmen-
a ion designs. The p oposed me hods a e e alua ed on he clinical OSLO-COMET da ase .
In ou esea ch, we p opose an algo i hm ha e ec i ely uses he essel enhancemen
echnique ep esen ed by RORPO [
1
] wi h he aw CT image, and handles he p oblem
o pixel imbalance by employing a a ian o he dice loss unc ion o op imize he seg-
men a ion e en u he . E alua ion is pe o med on he publicly a ailable 3D-IRCADb
da ase . Compa ed o he abo e-men ioned ela ed wo ks [
16
,
23
,
25
], ou p oposed me hod
can p o ide be e segmen a ion esul s, and i adop s a much simple app oach o com-
bining p ope ly he o iginal aw image wi h he enhanced image o ackle he p oblem o
essel segmen a ion. Nume ous ela ed wo ks likewise ail o accoun o he o eg ound–
backg ound pixel misma ch. We use Weigh ed Dice Loss o p o ide penal ies o he
numbe o inco ec ly ca ego ized oxels o ackle he p oblem. Addi ionally, some o he
ela ed esea ch igno es he da ase ’s in e io quali y anno a ions o some pa ien s. To
ensu e ha no essels a e missed and ha he segmen a ion accu acy p o ides an accu a e
po ayal o a ious algo i hms, we esol e his by enlis ing he aid o medical p o essionals
o anno a e any missed essels.
3. Segmen a ion Pla o m
Toge he wi h he p oposed algo i hm, we ha e de eloped a plugin ex ension o
3D Slice [
26
] ha (i) p o ides a emo e AI-Assis ed Anno a ion se ice (AIAA) o med-
ical doc o s om a High-Pe o mance Compu ing (HPC) clus e , allowing hem o use
Appl. Sci. 2023,13, 548 4 o 22
s a e-o - he-a me hods o pe o m he au oma ic segmen a ion o desi ed issue om
medical images, and (ii) p o ides a mechanism o collec he segmen ed and alida ed da a
gene a ed in s ep (i). We can use he s o ed da a o ine- une he exis ing neu al ne wo k
models. These enhanced models can hen be employed o au oma ic issue segmen a ion
in s ep (i).
Since he cu en s a e-o - he-a medical image p ocessing me hods a e based on Deep
Lea ning (DL), and ypically, DL algo i hms a e ained wi h a la ge amoun o da a, we
allow o he use o mul iple GPUs du ing he aining phase o p oduce models o he
equi ed quali y in a easonable amoun o ime. We use he mul i-GPU nodes o an HPC
clus e o ain models om sc a ch, as well as single GPU nodes o p o ide model in e ence
using AIAA. All o he connec ions and da a a e enc yp ed using Secu e Shell (SSH).
The en i e concep is depic ed in Figu e 1. Two main sec ions can be dis inguished:
one uns a a medical doc o ’s si e in a local hospi al ( on end), and he o he ope a es a an
HPC clus e acili y (backend). The on end allows he doc o o load, iew, and pe o m
au oma ic segmen a ion, and imp o e anno a ion on he medical da a using HPC esou ces.
The backend pa p o ides he compu a ional powe and o he equi ed ea u es.
Figu e 1. The main concep o he ool o medical image p ocessing and analysis.
The p o ided concep adop s he Cla a T ain SDK [
27
] om NVIDIA. I con ains se -
e al APIs, such as hose o AIAA, and a aining amewo k o DL-based model aining.
We use Cla a T ain SDK, e sion 4.0, which is solely PyTo ch-based. The Cla a T ain-
ing F amewo k builds on he open-sou ce MONAI amewo k [
28
], which is speci ically
de o ed o deep lea ning in heal hca e imaging.
Cla a’s AIAA is a clien -se e -based a chi ec u e ha deli e s a C++ o Py hon clien
API. Many medical image iewe s, such as 3D Slice (see Figu e 2) can be in e aced as
clien s [
29
] o ob ain he AIAA se ices o e ed by he se e unning on he ne wo k. Based
on he p e-loaded models on he se e , he AIAA can pe o m he au oma ic segmen a ion
o speci ic issues and display he esul s. Doc o s can edi and pos p ocess he esul ing
asks. A modi ied e sion o his ex ension, accommoda ed o HPC clus e usage, has
been used by medical doc o s o p o ide imp o ed anno a ions o he 3D-IRCADb da ase .
Appl. Sci. 2023,13, 548 5 o 22
Figu e 2. Slice ’s Segmen Edi o wi h he AIAA ex ension.
4. The Ha dwa e Used
Fo aining he models, we u ilized clus e nodes wi h 8
×
NVIDIA A100 (40 GB
HBM2) pe node. Each node is a powe ul x86–64 compu e , equipped wi h 2
×
64-co e
AMD Zen 3 EPYC™ 7763, 2.45 GHz p ocesso and 1024 GB o memo y.
5. Da ase s and Imp o ed Anno a ions
Publicly a ailable da ase s ha con ain labeled hepa ic essels a e lis ed in Table 1.
3D-IRCADb is he mos commonly used da ase . We use i in ou expe imen s as well. This
pa icula da ase is desc ibed in mo e de ail in he ollowing subsec ion.
Table 1. O e iew o publicly a ailable da ase s.
Name o da ase Published Numbe o pa ien s
3D-IRCADb [2] 5 May 2019 22 pa ien s
Medical Segmen a ion Deca hlon (MSD) [30] 20 Decembe 2020 430 pa ien s
Vascula Syn hesize (VascuSyn h) [24] Ma ch 2013 120 syn he ic samples
The MSD [
30
] is ano he publicly a ailable da ase . Howe e , i s g ound u h is o
poo quali y, a in e io o he 3D-IRCADb da ase . The e o e, we ha e no used his
da ase a all in his wo k.
We a e also awa e o syn he ically gene a ed ascula da ase s ha can be p o ided by
he VascuSyn h [
24
] so wa e. Howe e , since hese da ase s canno eplace clinical da a
whe e pa ien s ha e cu ilinea essels, we ha e no used hem ei he .
5.1. Da ase 3D-IRCADb
The p oposed me hod has been e alua ed on he publicly a ailable 3D-IRCADb [
2
]
da ase . We ha e speci ically used i s subse 3D-IRCADb-01, which is used in mos o he
simila esea ch, and hus allows o he compa ison. The da ase con ains 20 CT olumes

Appl. Sci. 2023,13, 548 6 o 22
(10 male and 10 emale). Fi een cases (75%) o his da ase ha e hepa ic umo s. The da ase
con ains a ious classes. The classes used in ou s udy a e po al ein and enoussys em.
We combined hem in o one class o eins. The pixel spacing a ies om 0.56 o 0.84 mm,
and he slice hickness a ies om 1 o 4 mm.
The e a e se e al challenges wi hin he da ase . In some cases, he in e io ena ca a
is no pa o he li e mask. Besides ha , one pa ien had many li e umo s ha co e ed
he majo i y o he li e essels. Ano he pa ien had inne me allic objec s, such as s en s,
ha signi ican ly a ec ed he b igh ness o he olume. Despi e hese bo lenecks, we
kep all o he olumes om he da ase in he expe imen o ha e mo e aining da a.
Fu he mo e, sligh abno mali ies make he ne wo k mo e obus , e en i he ou pu quali y
su e s. Al hough low-quali y da a we e used, ou echnique was s ill able o achie e a
high p edic ion quali y. The aining and es ing se s should equally include cases o li e
essels o simila appea ance (bo h con aining di icul and easy cases). To ensu e his, we
spli he da ase as ollows. T aining se con ains pa ien s—2, 5, 7, 12, 10, 11, 13, 16, 18, 19.
Valida ion includes pa ien s—4, 9, 20, 6, 17. Tes ing se is ep esen ed by pa ien s—1, 3, 8,
14, 15.
Imp o ing Anno a ions o 3D-IRCADb
Since he e a e some limi a ions in he anno a ions o hepa ic essels in he 3D-IRCADb
da ase , we decided o imp o e hem. The limi a ions a e mainly because o inadequa ely
anno a ed hepa ic eins, al hough hey a e clea ly isible and could be ma ked co ec ly.
Simila p oblems ha e been ound and ackled by [
8
,
31
]. Some o he CT olumes in
3D-IRCADb a e unde -segmen ed, while o he s a e o e -segmen ed. These issues migh
lead o he misin e p e a ion o esul s when segmen ed essels appea as alse posi i es
o alse nega i es while es ing he neu al ne wo ks [
18
]. Ano he majo limi a ion is he
inconsis ency o he anno a ed da a o he ena ca a. The ena ca a is isible in some
images bu no in o he s. I also makes essel enhancemen di icul [8].
Because he anno a ed da ase s used in [
8
,
31
] a e no publicly a ailable, he addi ional
labeling o he missing eins in 3D-IRCADb has been pe o med using he Slice buil -in
ools and he ex ension explained in Sec ion 3. I inc eased he quali y o anno a ions; see
Figu e 3. The compa ison o he dice sco es wi h he o iginal and imp o ed da ase is in he
expe imen al pa o he pape . I i is no s a ed di e en ly, he expe imen s used e ined
anno a ions since he essel p edic ion is be e and mo e con inuous wi h hem.
Figu e 3.
Anno a ions on Pa ien 1 om 3D-IRCADb da ase . (
a
) CT image; (
b
) O iginal Anno a ion;
(c) Imp o ed Anno a ions; (d) 3D o O iginal Anno a ion; (e) 3D o Imp o ed Anno a ions.
6. Me hods
The co e o ou solu ion lies in he use o enhancemen il e s and deep lea ning
me hods o essel segmen a ion. We pe o m expe imen s on he da ase , as explained in
Sec ion 5.1. We apply di e en il e s o imp o e he con as be ween he li e and hepa ic
essels, and use he enhanced da a as inpu o a segmen a ion ne wo k.
6.1. Vessel Enhancemen Fil e s
We ha e used i e il e s o enhance he ubula s uc u es and o he s uc u al in o ma-
ion o he eins. Simila o Su a achakan e al. [
25
], we ha e applied ou Hessian-based
il e s. Besides ha , a speci ic mo phological il e has been used as well. Namely, we ha e
used Hessian, F angi, Meije ing, and Sa o as he Hessian-based il e s, and RORPO as he
mo phological il e . Figu e 4shows he enhancing e ec o hese il e s on essels.
Appl. Sci. 2023,13, 548 7 o 22
Figu e 4. Compa ison o di e en essel enhancemen il e s.
6.1.1. Hessian Ma ix Compu a ion
The Hessian ma ix se es as a undamen al compu a ional me hod used in all Hessian-
based il e s. I calcula es he local g ada ion change by pe o ming he second-o de pa ial
de i a i es o he inpu image oxel X = (x, y, z) in nine di ec ions. I is de ined as
H(X) = 

xx xy xz
yx yy yz
zx zy zz
=





∂2
∂x2
∂2
∂x∂y
∂2
∂x∂z
∂2
∂y∂x
∂2
∂y2
∂2
∂y∂z
∂2
∂z∂x
∂2
∂z∂y
∂2
∂z2






(1)
Eigenanalysis is used o a oid compu ing de i a i es in many di ec ions, and o ex ac
only he p incipal di ec ions. Le
e1
be he eigen ec o ep esen ing he axial di ec ion,
le
e2
,
e3
be he c oss-sec ional di ec ion o H(X), and he associa ed
λ1
,
λ2
,
λ3
a e he
eigen alues. The eigen alues should mee he ollowing condi ion:
|λ1|≤|λ2|≤|λ3|
. The
essel cen e line oxel should sa is y he ollowing: λ2≈λ30, λ1≈0.
6.1.2. Hessian Vesselness Fil e
By in eg a ing he di ec ional g adien and he Hessian ma ix, Ng e al. [
32
] sugges ed
a modi ied mul i-scale Hessian il e . I is assumed ha he b igh ness o a hepa ic essel is
high in he middle and p og essi ely declines owa d he end, allowing i o be desc ibed
as a Gaussian s uc u e ans e se o i s axis. The con olu ion be ween he image g adien
ield and he Gaussian ke nel is he basis o each app oxima ion in he Hessian ma ix
H
a
a speci ied scale
σ
. The eigen alues
λ1
,
λ2
a e calcula ed and u ilized o compu e he cu e
de i a ion Rand simila i y measu e S. The cu ilinea likeliness Eis gi en by
E(x,y,σ) = 


0i λ2<0
e
−R
2β2
1[1−e
−s
2β2
2]o he wise
(2)
The esponse o he il e L is se as a maximum o di e en scales ha app oxima e
he size o he idges. I is exp essed as
L(x,y) = max
σmin⩽σ⩽σmax
[E(x,y,σ)] (3)
Appl. Sci. 2023,13, 548 8 o 22
The pa ame e s
β1
,
β2
con ol he sensi i i y o he il e o he measu e
R
and
S
,
espec i ely.
6.1.3. F angi Vesselness Fil e
The F angi [
33
] app oach uses all h ee eigen alues o disc imina e he local o ien a ion
pa e n. These h ee eigen alues se e in he di e en ia ion o blobs
Rb
, and pla e-like
and line-like s uc u es
Ra
. To dec ease he in luence o noise
S
, a Hessian no m measu e
was de eloped.
Rb=|λ1|/q|λ2λ3|(4)
Ra=|λ2|/|λ3|(5)
S=qλ2
1+λ2
2+λ2
3(6)
They p opose he ollowing combina ion o he componen s o de ine a esselness
unc ion,
F=


0i λ2>0o λ3>0,
(1−exp(−R2
a
2α2)) exp(−R2
b
2β2)(1−exp(−S2
2c2)) (7)
The pa ame e s
α
,
β
,
γ
a e he h esholds ha con ol he sensi i i y o he il e o he
measu es Ra,Rb, and S.
6.1.4. Meije ing Vesselness Fil e
To ecognize e y elonga ed s uc u es, Meije ing e al. [
34
] sugges ed a pa ame e - ee
esselness unc ion. I is based on he ollowing modi ied Hessian ma ix H0( ):
H0( ) = 

h11 +α
2(h22 +h33) (1−α
2)h12 (1−α
2)h13
(1−α
2)h21 h22 +α
2(h11 +h33) (1−α
2)h23
(1−α
2)h31 (1−α
2)h32 h33 +α
2(h11 +h22)
(8)
In gene al, α= 1/3. The eigen alues o H0( )wi h espec o H( )is exp essed as
λ0
i=λi+αλj+αλk(9)
o i6=j6=k.
The esselness is de ined by,
F=(λmax /λmin λmax <0
0λmax ≥0(10)
whe e,
λmax =max{λ0
1
,
λ0
2
,
λ0
3}
, which is compu ed a each oxel, and
λmin
is he minimum
o all λmax o he image.
6.1.5. Sa o Vesselness Fil e
Sa o e al. [
35
] sugges ed a line enhancemen il e unc ion ha is esponsi e o a ied
diame e anges. Sa o e al. so ed he eigen alues as
λi
as
λ1≥λ2≥λ3
. The eigen ec o
e1
co esponds o he di ec ion o he pu a i e essel.
|λ1|
< 0 and
|λ1|
< 0 ep esen he
sizes o he c oss-sec ion. The Sa o esselness in oduces a a io o he eigen alues o ob ain
a high esponse in ubula s uc u es. I is p o ided by,
F=








λcexp(−λ2
1
2(α1λc)2))λ1⩽0, λc6=0
λcexp(−λ2
1
2(α2λc)2))λ1>0, λc6=0
0λc=0
(11)
Appl. Sci. 2023,13, 548 9 o 22
whe e
α1<α2
,
λc=min{−λ2
,
−λ3}
. The pa ame e s
α1
and
α2
con ol he asymme ical
s eng h.
6.1.6. RORPO Vesselness Fil e
Di e en ial in o ma ion is no used by he RORPO [
1
,
36
] il e . RORPO, on he o he
hand, is based on he ma hema ical mo phology o he pa h ope a o s, and is hus de ined
using adjacency ela ions. RORPO is semi-global and non-linea . This il e compu es he
pa h openings in he se en p ima y di ec ions, anks hei esponses poin -by-poin , and
ex ac s s uc u es based on low and high esponses a each oxel. This noise- esis an il e
main ains he in ensi y o cu ilinea shapes while educing he in ensi y o o he s uc u es.
Pa h ope a o s wi h hin oxels o ien ed along hei leng h a e exp essed as:
LnR=Lmin × nR−1, (12)
whe e
LnR
is pa h leng h o de ec ,
Lmin
is he minimal pa h leng h, he geome ic sequence
o scales is eR, and he numbe o scales is nR.
6.2. Segmen a ion Pipelines
The h ee main s eps o he segmen a ion pipelines we ha e implemen ed a e image
da a p ep ocessing, essel enhancemen , and image segmen a ion. These a e explained in
he ollowing subsec ions. Fi s , he image p ep ocessing is pe o med, and hen speci ic
enhancemen il e s a e applied and ei he used be o e he segmen a ion o immedia ely
a e he segmen a ion. Figu es 5–7show h ee di e en app oaches.
We ha e implemen ed h ee segmen a ion pipelines. In he i s wo pipelines, he
usion o il e s a e inspi ed by he app oach o Su a achakan e al. [
25
], and hey se e as
a compa ison. The hi d pipeline ha we p opose implemen s a simple bu e y e ec i e
usion o aw and il e ed images be o e pe o ming he segmen a ion.
6.2.1. Da a P ep ocessing
In ou me hod, he li e a ea is ex ac ed and c opped acco ding o he li e mask o
ocus p ima ily on he li e ascula sys em. The li e egion is ex ac ed in he i s s ep
as opposed o he las s ep in Su a achakan e al. [
25
]. This helps o emo e unnecessa y
esul ing bounda ies, and leads o less in e e ence wi h he backg ound. All scans a e
con e ed in o 1
×
1
×
1 mm iso opic esolu ion. The image in ensi ies a e windowed
o lie wi hin he
h
80,220
i
Houns ield Uni (HU) ange, and a e hen mapped o he ange
h
0,1
i
. In addi ion, he 3D aniso opic di usion is applied o lowe he image noise and o
s ill p ese e he signi ican pa s such as edges and lines.
6.2.2. Vessel Enhancemen
The enhancemen o essels is he p ep ocessing s ep used be o e he segmen a ion.
The p ep ocessing s eps a e implemen ed using he sciki -image lib a y [
37
]. This s ep is a
key ac o in ob aining signi ican ly be e segmen a ion esul s. Enhanced images acqui ed
by he applica ion o di e en il e s (see Sec ion 6.1) a e u ilized ei he be o e o ollowing
he segmen a ion.
A me hod ha combines he esul s o indi idual essel enhancemen il e s simul ane-
ously be o e segmen a ion, and which is inspi ed by he app oach o
Su a achakan e al. [25]
,
is depic ed in Figu e 5. The inal enhanced image is ob ained by a e aging he espec i e
pixel alues om all il e s. The esul ing image is hen used as an inpu o he segmen a ion
ne wo k. Two a ian s o his app oach ha e been conside ed. They ei he use Hessian, simila
o [
25
], o RORPO as one o he ou il e s. As p o en by expe imen s, he Hessian il e
has he weakes enhancing e ec on he segmen a ion esul s in e ms o he dice sco e. The
RORPO il e , on he o he hand, has he s onges enhancing e ec . The e o e, he ini ial idea
o his me hod was o use he s onges combina ion o ou il e s ha would ou pe o m he
solu ion p esen ed in [
25
]. We indica e his me hod as Fil e Added, and dis inguish whe he
Hessian o RORPO has been used. The usion o il e s is de ined as:
Appl. Sci. 2023,13, 548 16 o 22
(
β=
0.4) o he o iginal image in bo h 3D U-Ne - and 3D V-Ne -based a chi ec u es; see
Figu e 11. I also shows ha segmen a ion wi h 3D V-Ne p o ides be e dice sco es and
mo e consis en esul s o e a wide spec um o
α
alues han 3D U-Ne . As an accep able
egion o blending alue, he egion be ween 40 and 60% is se .
Figu e 11.
The a e age dice sco e o di e en blending a ios be ween he RORPO and he o iginal
image while using 3D U-Ne and 3D V-Ne .
7.3.2. Compa ison o Di e en Loss Func ions
We ha e e alua ed di e en ypes o loss unc ions and hei e ec s on he segmen-
a ion quali y. The e alua ion was pe o med on he bes -pe o ming se up wi h he 3D
V-Ne as a segmen e . The esul s o all me ics a e shown in Table 6.
Table 6.
Compa ison o di e en loss unc ions es ed on he imp o ed 3D-IRCADb da ase and he
3D V-Ne a chi ec u e.
Loss Func ions 3D V-Ne
DSC ACC SEN PRC
Dice Loss 75.4 ±4.3 99.3 ±0.1 76.9 ±4.1 74.4 ±7.8
Dice-BCE Loss 75.9 ±3.5 99.3 ±0.1 75.5 ±3.1 76.8 ±7.1
Log Cosh Dice Loss 74.6 ±3.3 99.3 ±0.1 73.5 ±4.2 76.3 ±7.4
Focal Loss 74.4 ±3.7 99.3 ±0.1 70.2 ±4.0 79.6 ±8.0
Dice Focal Loss 75.2 ±4.3 99.3 ±0.1 75.2 ±4.3 75.6 ±7.0
T e sky Loss 76.0 ±4.2 99.3 ±0.1 75.7 ±3.9 76.8 ±7.1
Weigh ed Dice Loss βl= 6 76.2 ±3.7 99.3 ±0.1 75.0 ±3.4 77.7 ±6.1
The bes -pe o ming loss unc ion in ou model is he Weigh ed Dice Loss. I ope a es
wi h he pa ame e
βl
(see Table 2) o ackle he p oblem o class imbalance. The highe he
alue o
βl
, he g ea e he penal y o misclassi ied oxels. Un o una ely, his also slows
down he g adien low and inc eases he isk o s icking in he local minima. When
βl
is in
he ange o 5–7, op imal ou comes a e ob ained. We ha e ound
βl=
6 o be he bes in
ou s udy.
7.4. Expe imen 4: Abla ion S udy on he Final Segmen a ion Pipeline
We ha e pe o med he abla ion o a a ying numbe o con olu ional il e s a all
s ages o he 3D V-Ne a chi ec u e a e aining. We ha e ound ha ce ain s ages a e
mo e signi ican han o he s o he quali y o eins segmen a ion.
The V-Ne is di ided in o se e al s ages ( he s ages a e shown in Figu e 9) ha ope a e
a di e en esolu ions du ing he down- as well as upsampling phases. Each s age consis s
o one o h ee con olu ional laye s, wi h each laye ha ing a speci ic numbe o channels
( il e s). We abla ed g oups o simila il e s a e e y s age by inc easing he pe cen age

Appl. Sci. 2023,13, 548 17 o 22
(25%, 50%, and 75%) o simila il e s. Since he numbe o il e s a ies om s age o
s age, i means ha he same pe cen age may co espond o a di e en numbe o abla ed
il e s. The simila i y be ween il e s wi hin a g oup was calcula ed based on he absolu e
Euclidean dis ance o he no malized il e weigh s. Abla ions we e pe o med by manually
se ing he weigh s and biases o all incoming connec ions o a il e o 0, hus elimina ing
any ac i a ion o ha il e .
Figu e 12 shows he dice sco es o abla ions o 25%, 50%, and 75% a e e y s age o he
V-Ne . Each da a poin shows he e ec o a speci ic s age abla ion. The dice sco e is a ec ed
no iceably mo e in s ages ’in 16’, ’down 64’, and ’up 64’, han in he o he s ages. Fu he mo e,
he e ec o he la ges abla ion has a s onge in luence on some s ages han on o he s. Fo
ins ance, s age ’up 32’ shows a no iceably la ge d op in dice sco e o 75% o he abla ed
il e s, compa ed o 50%, while he s ages ’down 256’ and ’up 256’ a e la gely una ec ed.
Figu e 12. E ec s o di e en abla ions amoun s in all 3D V-Ne s ages.
7.5. Segmen a ion Resul s
Figu e 13 shows he segmen a ion esul s o he p oposed algo i hm on he imp o ed
3D-IRCADb da ase . The p o ided me hod can p edic he ascula s uc u e well, and he
esul s a e close o he g ound u h.
Figu e 13.
In each case, he i s ow shows he g ound u h, and he second ow shows he
p edic ions in he axial, sagi al, co onal, and 3D iews.
Appl. Sci. 2023,13, 548 18 o 22
7.6. Expe imen 5: Compa ison o Pe o mance be ween he P oposed Algo i hm and
O he Algo i hms
We ha e compa ed ou segmen a ion app oach wi h he wo k o o he s. Speci ically,
wi h wo ks [
4
,
8
,
14
–
16
,
18
,
23
], which combine di e en app oaches o p o ide li e essel
segmen a ion.
In he case o [
4
,
8
,
14
,
15
,
18
], i was impossible o make a ai compa ison. The sou ce
code o hese me hods was no a ailable, so i was no possible o ain and es hese
me hods on he same enhanced 3D-IRCADb da ase ha we used in ou wo k. We p o ide
only an indi ec compa ison wi h he nume ical esul s as p esen ed in hose pape s. A
compa ison is p o ided in he i s pa o Table 7.
We we e able o objec i ely e alua e he pe o mance o me hods om [
16
,
23
] since
hei code is a ailable. The au ho s o he pape [
23
] published he code on Gi Hub and
ha e also c ea ed an online demons a ion ool o using he il e s. We es ed he RORPO
il e wi h de aul pa ame e se ings. Then, he esul was h esholded. The pa ame e
nbTh esholds =
200 means ha each pa ien is h esholded wi h a 200
x
di e en alue.
This alue is inc eased by a e y small s ep o 0.005 and lies in he ange o
<
0,1
>
. As
opposed o ha , he au ho s o he pape [
16
] implemen ed 2.5D V-Ne wi h T e sky loss
unc ion. They combined pa ien s om he 3D-IRCADb da ase , and CT scans om he
Polyclinic o Ba i. The 2.5D V-Ne app oach p ocesses i e slices as i e channels, since
he ne wo k uses only 2D con olu ional laye s. The andom pa ches o i e consecu i e
slices a e sen o he ne wo k o aining. To es he ne wo k, he sliding window is used.
Howe e , we in e sub- olumes o dimensions 512
×
512
×
5, and only he middle slice is
used o p edic ion on he middle slice. The esul s o hese me hods and ou p oposed
me hod a e p o ided in he second pa o Table 7.
Table 7. Compa ison o he p oposed me hod, wi h ela ed wo ks.
Fil e s Published Dice sco e Accu acy Sensi i i y Speci ici y
Indi ec compa ison
Zhang e al. [4] 2018 67.3 96.4 73.7 97.4
Huang e al. [8] 2018 75.3 97.6 76.7 98.8
Su e al. [14] 2021 75.46 - 76.9 -
Yang e al. [15] 2021 71.6 98.5 75.4 99.5
Xu e al. [18] 2020 68.7 99.8 78.6 99.2
Di ec compa ison ( e-implemen ed ela ed wo ks)
Al ini e al. [16] 2020 65.0 ±9.7 99.8 ±0.4 52.3 ±10.8 88.9 ±6.2
Lamy e al. [23] 2020 59.6 ±5.8 98.9 ±0.3 55.9 ±4.2 64.8 ±12.0
Ou me hod 2022 76.2 ±3.7 99.3 ±0.1 75.0 ±3.4 99.7 ±0.1
8. Discussion
In he i s expe imen , di e en ascula enhancemen il e s we e compa ed. The ask
was o ind ou which o he i e il e s was he bes , in e ms o he segmen a ion quali y
measu es. F om he isualiza ions o he il e s’ enhancing e ec s in Figu e 4, i can be seen
ha he RORPO il e pe o ms e y well. The e is a minimum amoun o image noise. In
addi ion, he e a e no isible bounda ies be ween he li e and he backg ound ha could
a ec he esul . Al hough he image enhanced by he F angi il e seems p omising as
well, he essels a e ela i ely hin. Visible bounda ies and some noise a e ypical o he
Meije ing il e and he Sa o il e . In e ms o he segmen a ion quali y, he dice sco es o
he o iginal and he imp o ed anno a ions p o e ha he RORPO is he mos e ec i e il e ,
and has pe mi ed us o imp o e he dice sco e om 62.7% o 69.7% and o educe he mean
s anda d de ia ion om 13.8% o 4.8%. The imp o emen in dice sco e and educ ion in
s anda d de ia ion shows ha he model can segmen essels mo e e icien ly in di e en
es pa ien s.
Appl. Sci. 2023,13, 548 19 o 22
The second expe imen ocused on le e aging he essel enhancemen e ec s o
di e en il e s in a ious segmen a ion pipelines. In File Added, ou enhancemen il e s
we e combined, and he model was ained on his combined da a. In SegAdded, ou
models we e ained wi h dis inc enhancemen il e s, and he esul s we e la e combined
when a leas wo models we e able o p edic he essel pixel. In linea blending, he
o iginal image was combined wi h he RORPO enhanced image. By compa ing he di e en
pipelines, he linea blending p o ides he bes esul s o all obse ed me ics excep
sensi i i y. He e, he SegAdded app oach pe o ms be e . Resul s we e also ob ained om
di e en segmen a ion models (3D U-Ne and 3D V-Ne ).
In he hi d expe imen , a de ailed analysis o he p oposed me hod was p o ided.
We ha e explo ed he e ec o RORPO on he o iginal image. We ha e se he mixing
a io be ween RORPO and he o iginal image,
α
, om 0% o 100%, wi h a s ep o 10%.
The RORPO il e has a signi ican posi i e e ec on he esul . The bes segmen a ion
is achie ed by using 40% o he o iginal image and 60% o RORPO. The accep able
α
ange is 40–60%. This p o es ha he ole o backg ound in o ma ion is also help ul in
segmen ing he egion o in e es . The esul s a e also pe o med on 3D U-Ne and 3D
V-Ne , and 3D V-Ne pe o ms be e han 3D U-Ne . This shows ha o olume ic inpu ,
3D V-Ne pe o ms be e . Hence, we used V-Ne in all u he es s. Howe e , 3D V-Ne
is mo e expensi e in e ms o compu ing esou ces, and he ime o aining is wice as
long as ha o 3D U-Ne . Ano he e alua ion o he p oposed me hod ocused on he use
o di e en dice loss unc ions. The Weigh ed Dice Loss
βl=
6 ou pe o med o he loss
unc ions, achie ing a dice sco e o 76.2% and a mean s anda d de ia ion o 3.7%. This
loss unc ion can deal wi h unbalanced classes and enhance segmen a ion accu acy and
sensi i i y. Pa ame e βl=6 was used o in oduce a penal y o misclassi ied pixels, and
also o a oid apping he loss unc ion in local minima. The
βl
pa ame e can be in he
ange (0, 9). I he
βl
pa ame e is se o a lowe o highe alue han 6, he esul s a e
wo se o e all, since he coe icien penalizes oo ew o oo many alse nega i es. This
app oach se he inal e sion o he me hod and led o he highes dice sco e, as well as
inc eased p ecision.
The ou h expe imen examined he e ec s o a ying pe cen ages o abla ions on
he segmen a ion pe o mance o he 3D V-Ne a chi ec u e ained on he 3D-IRCADb
da ase . As an icipa ed, he pe o mance ypically declined as he numbe o abla ed il e s
inc eased. The esul s o he expe imen showed ha some il e s con ibu ed mo e o
he pe o mance han o he s. The e ec o abla ion is much s onge o s ages (’down
64’, ’up 64’) o 50% abla ion, and s ages ’up 32’ o 75% abla ion. The s ages (’down
256’, ’up 256’) showed negligible pe o mance loss, as hese s ages may ha e edundan
ea u e ep esen a ion. This expe imen de e mines he in luence o di e en laye s on he
segmen a ion esul s. By iden i ying he ne wo k’s sensi i e a eas, we wan o modi y he
model’s a chi ec u e in u u e esea ch.
In he las expe imen , we compa ed he pe o mance o he p oposed me hod wi h some
o he algo i hms desc ibed in he ela ed wo k. We compa ed indi ec ly wi h
[4,8,14,15,18]
because i was no possible o e-implemen o e alua e hese me hods wi h ou da ase . A
di ec compa ison was made wi h [
16
,
23
]. These me hods we e e-implemen ed based on
he a ailable sou ce code and e alua ed wi h ou da ase . The au ho s in [
23
] jus used a
h esholding app oach ha did no p o ide sa is ac o y essel segmen a ion. The au ho s
o he pape [
16
] used 2.5D V-Ne wi h a T e sky loss unc ion, yielding esul s wi h high
a iabili y and a high mean s anda d de ia ion. We can s a e ha ou me hod p e ails in
dice sco e and speci ici y alue o e all o he o he me hods. In compa ison wi h he wo
eimplemen ed me hods, we domina e in e ms o sensi i i y, meaning ha he e a e ewe
missing and agmen ed essels. In e ms o accu acy, ou me hod lags jus sligh ly behind.
O e all, he me hod ou lined in his pape can au oma ically segmen enough li e
essels o co ec ly ex ac he essels, and i can be u he manually co ec ed using Slice
ools. The au oma ic segmen a ion esul s a e isually p esen ed in Sec ion 7.5. The me hod
wo ks on CT ob ained om di e en machines wi h di e en acquisi ion pa ame e s. Ou
Appl. Sci. 2023,13, 548 20 o 22
me hod is also able o segmen images wi h di e en li e shapes, li e s wi h hinne essels,
and li e s wi h essels a ophied. The speci ic combina ion o he o iginal image wi h he
essel-enhanced image is he o iginali y o ou wo k, and i ou pe o ms o he me hods
p oposed in he li e a u e. The me hod ou lined in his pape ha e some limi a ions.
The esul s in some cases ha e misclassi ied essel endings, a iabili y in essel edges,
and some egions a e no connec ed o he main essels. Finally, we need mo e da a o
make he model mo e obus , es he essel segmen a ion, and lowe he a iabili y in
ou expe imen s.
9. Conclusions and Fu u e Wo ks
The p oposed algo i hm e ec i ely segmen s li e essels wi h a dice sco e o 76.2%
and a p ecision o 77.7%. The ne wo k can au oma ically segmen labeled o e en o he
unma ked li e essels ( ha should be labeled) om he CT images. The RORPO il e
used o essel enhancemen in conjunc ion wi h blending i s ou pu wi h he o iginal CT
image has success ully been p o en o gi e be e esul s compa ed o o he s a e-o - he-a
echniques. Thus, adding his in o ma ion o he enhanced essels has helped o imp o e
he obus ness o he segmen a ion model. The Weigh ed Dice Loss me ic based on he
dice coe icien is used o he loss unc ion compu a ion o imp o e segmen a ion accu acy,
and o deal wi h unbalanced o eg ound and backg ound class oxels. The Weigh ed Dice
Loss unc ion wi h
βl=
6 has been p o en o pe o m he bes i i is combined wi h he
3D V-Ne .
This bes pe o ming model has been made a ailable o he local hospi al’s doc o s,
who can use i o he ini ial anno a ion o new da a using he ool desc ibed in Sec ion 3.
In he u u e, we plan o coope a e wi h medical doc o s mo e closely, ex end he
aining se wi h new alida ed da a, and con inue ou esea ch in he a ea o hepa ic essel
segmen a ion. Addi ionally, in ou u u e s udies, we will ocus on expe imen p uning
and compu a ional cos educ ion by unde s anding he signi icance o abla ed il e s.
Au ho Con ibu ions:
Concep ualisa ion: P.S. (Pe S akos), P.S. (Pe a S obodo a) and K.S.;
me hodology: P.S. (Pe a S obodo a), K.S. and P.S. (Pe S akos); o mal analysis: P.S. (Pe a S o-
bodo a), K.S., P.S. (Pe S akos) and A.V.; in es iga ion: P.S. (Pe a S obodo a), K.S. and P.S. (Pe
S akos); da a cu a ion: P.S. (Pe a S obodo a), K.S., P.S. (Pe S akos) and A.V.; w i ing—o iginal
d a p epa a ion: P.S. (Pe a S obodo a), K.S. and P.S. (Pe S akos); w i ing— e iew and edi ing:
P.S. (Pe S akos), P.S. (Pe a S obodo a) and K.S.; isualiza ion: P.S. (Pe a S obodo a) and K.S.;
supe ision: P.S. (Pe S akos). All o he au ho s ha e ead and app o ed he inal manusc ip
and i s submission o he jou nal. All au ho s ha e ead and ag eed o he published e sion o
he manusc ip .
Funding:
This wo k was suppo ed by he doc o al g an compe i ion VSB—Technical Uni e si y
o Os a a, eg. no. CZ.02.2.69/0.0/0.0/19_073/0016945 wi hin he Ope a ional P og amme Re-
sea ch, De elopmen and Educa ion, unde p ojec DGS/TEAM/2020-008 ‘De elopmen o a ool o
scien i ic da a p ocessing and isualiza ion in VR wi h mul i-use suppo ’.
Ins i u ional Re iew Boa d S a emen : No applicable.
In o med Consen S a emen : No applicable.
Da a A ailabili y S a emen :
No new da a we e c ea ed o analyzed in his s udy. Da a sha ing is
no applicable o his a icle.
Con lic s o In e es : The au ho s decla e no con lic s o in e es .
Appl. Sci. 2023,13, 548 21 o 22
Abb e ia ions
The ollowing abb e ia ions a e used in his manusc ip :
CT Compu ed Tomog aphy
HU Houns ield Uni
SSH Secu e Shell
RORPO Ranking O ien a ion Responses o Pa h Ope a o s
HPC High-Pe o mance Compu ing
GPU G aphics P ocessing Uni
API Applica ion P og amming In e ace
Fil e Added Fil e ed images added
SegAdded Segmen a ion maps added
BN Ba ch No maliza ion
PReLU Pa ame ic Rec i ied Linea Uni
SDK So wa e De elopmen Ki
CNN Con olu ional Neu al Ne wo ks
DL Deep Lea ning
AIAA AI-Assis ed Anno a ion
MONAI Medical Open Ne wo k o AI
Re e ences
1.
Me eille, O.; Talbo , H.; Najman, L.; Passa , N. Cu ilinea s uc u e analysis by anking he o ien a ion esponses o pa h
ope a o s. IEEE T ans. Pa e n Anal. Mach. In ell. 2018,40, 304–317. [C ossRe ] [PubMed]
2.
3D-IRCADb. 2022. A ailable online: h ps://www.i cad. / esea ch/da a-se s/li e -segmen a ion-3d-i cadb-01/ (accessed on
17 Oc obe 2022).
3.
Fedo o , A.; Beichel, R.; Kalpa hy-C ame , J.; Fine , J.; Fillion-Robin, J.; Pujol, S.; Kikinis, R. 3D Slice as an image compu ing
pla o m o he Quan i a i e Imaging Ne wo k. Magn. Reson. Imaging 2012,30, 1323–1341. [C ossRe ]
4.
Zhang, R.; Zhou, Z.; Wu, W.; Lin, C.C.; Tsui, P.H.; Wu, S. An Imp o ed Fuzzy Connec edness Me hod o Au oma ic Th ee-
Dimensional Li e Vessel Segmen a ion in CT Images. J. Heal hc. Eng. 2018,2018, 2376317. [C ossRe ] [PubMed]
5.
Guo, X.; Xiao, R.; Zhang, T.; Chen, C.; Wang, J.; Wang, Z. A no el me hod o model hepa ic ascula ne wo k using essel
segmen a ion, hinning, and comple ion. Med. Biol. Eng. Compu . 2020,58, 709–724. [C ossRe ] [PubMed]
6.
Ciecholewski, M.; Kassja´nski, M. Compu a ional me hods o li e essel segmen a ion in medical imaging: A Re iew. Senso s
2021,21, 2027. [C ossRe ] [PubMed]
7.
Yu, W.; Fang, B.; Liu, Y.; Gao, M.; Zheng, S.; Wang, Y. Li e essels segmen a ion based on 3D esidual U-NET. In P oceedings
o he 2019 IEEE In e na ional Con e ence on Image P ocessing (ICIP), Taipei, Taiwan, 22–25 Sep embe 2019; pp. 250–254.
[C ossRe ]
8.
Huang, Q.; Sun, J.; Ding, H.; Wang, X.; Wang, G. Robus li e essel ex ac ion using 3D U-ne wi h a ian dice loss unc ion.
Compu . Biol. Med. 2018,101, 153–162. [C ossRe ]
9.
A ane, A.; Kucha ski, A.; Chapuis, P.; F eydie , S.; Leb e, M.; Vaca an , A.; Fabija´nska, A. Segmen a ion o li e ana omy by
combining 3D U-ne app oaches. Appl. Sci. 2021,11, 4895. [C ossRe ]
10.
Golla, A.; Baue , D.F.; Schmid , R.; Russ, T.; No enbe g, D.; Chung, K.; Zollne , F.G. Con olu ional Neu al Ne wo k Ensemble
segmen a ion wi h a io-based sampling o he a e ies and eins in abdominal CT scans. IEEE T ans. Biomed. Eng.
2021
,68,
1518–1526. [C ossRe ]
11.
Çiçek, Ö.; Abdulkadi , A.; Lienkamp, S.S.; B ox, T.; Ronnebe ge , O. 3D U-Ne : Lea ning dense olume ic segmen a ion om
spa se anno a ion. In Medical Image Compu ing and Compu e -Assis ed In e en ion, P oceedings o he MICCAI 2016, A hens, G eece,
17–21 Oc obe 2016; Sp inge : Cham, Sw ize land, 2016, pp. 424–432. [C ossRe ]
12.
Segmen a ion o he Li e Compe i ion 2007 (SLIVER07). 2019. A ailable online: h ps://sli e 07.g and-challenge.o g/ (accessed
on 17 Oc obe 2022).
13.
Mille a i, F.; Na ab, N.; Ahmadi, S. V-ne : Fully con olu ional neu al ne wo ks o olume ic medical image segmen a ion. In
P oceedings o he 2016 Fou h In e na ional Con e ence on 3D Vision (3DV), S an o d, CA, USA, 25–28 Oc obe 2016; pp. 565–571.
[C ossRe ]
14.
Su, J.; Liu, Z.; Zhang, J.; Sheng, V.S.; Song, Y.; Zhu, Y.; Liu, Y. DV-ne : Accu a e li e essel segmen a ion ia dense connec ion
model wi h D-BCE loss unc ion. Knowl.-Based Sys . 2021,232, 107471. [C ossRe ]
15. Yang, J.; Fu, M.; Hu, Y. Li e essel segmen a ion based on in e -scale V-ne . Ma h. Biosci. Eng. 2021,18, 4327–4340. [C ossRe ]
16.
Al ini, N.; P encipe, B.; B une i, A.; B une i, G.; T iggiani, V.; Ca nimeo, L.; Casca ano, G.D. A e sky loss-based con olu ional
neu al ne wo k o li e essels segmen a ion. In In elligen Compu ing Theo ies and Applica ion; Sp inge : Cham, Sw ize land,
2020; pp. 342–354. [C ossRe ]

Appl. Sci. 2023,13, 548 22 o 22
17.
Meng, X.; Wang, G.; Zhang, Y.; Shi, X.; Dai, H.; Wang, Z.; Wang, X. Exploi ing ull Resolu ion Fea u e Con ex o Li e Tumo
and Vessel Segmen a ion ia In eg a e F amewo k: Applica ion o Li e Tumo and Vessel 3D Recons uc ion unde embedded
mic op ocesso . a Xi 2022. [C ossRe ]
18.
Xu, M.; Wang, Y.; Chi, Y.; Hua, X. T aining li e essel segmen a ion deep neu al ne wo ks on noisy labels om con as CT
Imaging. In P oceedings o he 2020 IEEE 17 h In e na ional Symposium on Biomedical Imaging (ISBI), Iowa Ci y, IA, USA, 3–7
Ap il 2020; pp. 1552–1555. [C ossRe ]
19.
Li e Tumo Segmen a ion Challenge (LiTS) Da ase . A ailable online: h ps://compe i ions.codalab.o g/compe i ions/17094
(accessed on 17 Oc obe 2022).
20.
Shahid, M.; Taj, I.A. Robus e inal essel segmen a ion using essel’s loca ion map and angi enhancemen il e . IET Image
P ocess. 2018,12, 494–501. [C ossRe ]
21.
Soom o, T.A.; A i i, A.J.; Ali Shah, A.; Soom o, S.; Baloch, G.A.; Zheng, L.; Yin, M.; Gao, J. Impac o image enhancemen echnique
on CNN model o e inal blood essels segmen a ion. IEEE Access 2019,7, 158183–158197. [C ossRe ]
22.
Blaiech, A.G.; Mansou , A.; Ke keni, A.; Bedoui, M.H.; Ben Abdallah, A. Impac o enhancemen o co ona y a e y segmen a ion
based on deep lea ning neu al ne wo k. In Ibe ian Con e ence on Pa e n Recogni ion and Image Analysis; Sp inge : Cham,
Sw ize land, 2019. [C ossRe ]
23.
Lamy, J.; Me eille, O.; Ke au e , B.; Passa , N.; Vaca an , A. Vesselness il e s: A su ey wi h benchma ks applied o li e
imaging. In P oceedings o he 2020 25 h In e na ional Con e ence on Pa e n Recogni ion (ICPR), Milan, I aly, 10–15 Janua y
2021; pp. 3528–3535. [C ossRe ]
24.
Hama neh, G.; Jassi, P. VascuSyn h: Simula ing ascula ees o gene a ing olume ic image da a wi h g ound- u h segmen a-
ion and ee analysis. Compu . Med. Imaging G aph. 2010,34, 605–661. [C ossRe ] [PubMed]
25.
Su a achakan, S.; Pelanis, E.; Khan, Z.A.; Kuma , R.P.; Edwin, B.; Lindse h, F. E ec s o enhancemen on Deep Lea ning based
hepa ic essel segmen a ion. Elec onics 2021,10, 1165. [C ossRe ]
26. 3D Slice . 2022. A ailable online: h ps://www.slice .o g/ (accessed on 28 Decembe 2021).
27.
NVIDIA Cla a Imaging. 2022. A ailableonline: h ps://de elope .n idia.com/cla a-medical-imaging (accessed on
28 Decembe 2021
).
28. P ojec MONAI. 2021. A ailable online: h ps://monai.io/ (accessed on 28 Decembe 2021).
29.
Ai-Assis ed-Anno a ion-Clien . 2022. A ailable online: h ps://gi hub.com/NVIDIA/ai-assis ed-anno a ion-clien / ee/mas e /
slice -plugin (accessed on 28 Decembe 2021).
30.
Simpson, A.L.; An onelli, M.; Bakas, S.; Bilello, M.; Fa ahani, K.; an Ginneken, B.; Kopp-Schneide , A.; Landman, B.A.;
Li jens, G
.;
Menze, B.; e al. A la ge anno a ed medical image da ase o he de elopmen and e alua ion o segmen a ion algo i hms. a Xi
2019, a Xi :1902.09063.
31.
Yan, Q.; Wang, B.; Zhang, W.; Luo, C.; Xu, W.; Xu, Z.; You, Z. A en ion-guided deep neu al ne wo k wi h mul i-scale ea u e
usion o li e essel segmen a ion. IEEE J. Biomed. Heal h In o m. 2021,25, 2629–2642. [C ossRe ]
32.
Ng, C.; Yap, M.H.; Cos en, N.; Li, B. Au oma ic w inkle de ec ion using Hyb id Hessian il e . In Compu e Vision— P oceedings o
he ACCV 2014, Singapo e, 1–5 No embe 2014; Sp inge : Cham, Sw ize land, 2015; pp. 609–622. [C ossRe ]
33.
F angi, A.F.; Niessen, W.J.; Vincken, K.L.; Vie ge e , M.A. Mul iscale essel enhancemen il e ing. In Medical Image Com-
pu ing and Compu e -Assis ed In e en ion—P oceedings o he MICCAI’98, Camb idge, MA, USA, 11–13 Oc obe 1998; Sp inge :
Be lin/Heidelbe g, Ge many, 1998; pp. 130–137.
34.
Meije ing, E.; Jacob, M.; Sa ia, J.; S eine , P.; Hi ling, H.; Unse , M. Design and alida ion o a ool o neu i e acing and analysis
in luo escence mic oscopy images. Cy ome y 2004,58A, 167–176. [C ossRe ]
35.
Sa o, Y.; Nakajima, S.; Shi aga, N.; A sumi, H.; Yoshida, S.; Kolle , T.; Kikinis, R. Th ee-dimensional mul i-scale line il e o
segmen a ion and isualiza ion o cu ilinea s uc u es in medical images. Med. Image Anal. 1998,2, 143–168. [C ossRe ]
36. RORPO GIT. A ailable online: h ps://gi hub.com/pa h-openings/RORPO (accessed on 28 Decembe 2021).
37.
Van de Wal , S.; Schönbe ge , J.L.; Nunez-Iglesias, J.; Boulogne, F.; Wa ne , J.D.; Yage , N.; Gouilla , E.; Yu, T. Sciki -image: Image
p ocessing in py hon. Pee J 2014,2, e453. [C ossRe ]
38.
Wu, M.; Qian, Y.; Liao, X.; Wang, Q.; Heng, P.A. Hepa ic essel segmen a ion based on 3Dswin- ans o me wi h induc i e biased
mul i-head sel -a en ion. a Xi 2021, a Xi :2111.03368.
39.
Sud e, C.H.; Li, W.; Ve cau e en, T.; Ou selin, S.; Jo ge Ca doso, M. Gene alised dice o e lap as a deep lea ning loss unc ion o
highly unbalanced segmen a ions. In Deep Lea ning in Medical Image Analysis and Mul imodal Lea ning o Clinical Decision Suppo ;
Sp inge : Cham, Sw ize land, 2017; pp. 240–248. [C ossRe ]
40.
Jadon, S. A su ey o loss unc ions o seman ic segmen a ion. In P oceedings o he 2020 IEEE Con e ence on Compu a ional
In elligence in Bioin o ma ics and Compu a ional Biology (CIBCB), Viña del Ma , Chile, 27–29 Oc obe 2020. [C ossRe ]
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 .