UNIVERSITAT POLIT`
ECNICA DE CATALUNYA
B ain lesion segmen a ion using
Con olu ional Neu onal Ne wo ks
by
Cla a Bonn´ın Rossell´o
In pa ial ul ilmen o he equi emen s o he deg ee in
Ci`encies i Tecnologies de Telecomunicaci´o enginee ing
in he
Escola T`ecnica d’Enginye ia de Telecomunicaci´o de Ba celona
Image P ocessing G oup
Ad iso s: Ve ´onica Vilaplana and Ad i`a Casami jana
May 2018
UNIVERSITAT POLIT`
ECNICA DE CATALUNYA
Abs ac
Escola T`ecnica d’Enginye ia de Telecomunicaci´o de Ba celona
Image P ocessing G oup
by Cla a Bonn´ın Rossell´o
Con olu ional neu al ne wo ks (CNN) a e powe ul ools o lea ning ep esen a ions om im-
ages. They a e being used in a la ge ange o applica ions, being he s a e-o - he a in many
compu e ision asks. In his wo k, we s udy he b ain umo segmen a ion p oblem using
CNNs and he publicly a ailable B aTS da ase . One o he key ac o s o his ask is which
aining scheme is used since i should deal wi h memo y cons ain s and should alle ia e he
high-imbalance na u e be ween heal hy and lesion issue in he b ain.
Thus, he pu pose o his p ojec is o p opose a compa ison be ween se e al aining schemes
and ex ensi ely analyze and e alua e hem in e ms o he dice sco e. We e alua e dense-
aining agains pa ch-sampling, and pa icula ly, ixed- ule agains adap i e sampling scheme.
Fu he mo e, a ian s and modi ica ions o he exis ing aining schemes ha e been p oposed in
o de o enhance hei pe o mance. Finally, se e al loss unc ions o each aining scheme ha e
been analyzed.
Acknowledgemen s
I would like o exp ess my since e g a i ude o se e al people. Fi s ly, I would like o hank
specially Ad i`a Casami jana and Ve ´onica Vilaplana, he supe iso s o his hesis, o hei
ad ice and pa ience du ing he whole p ojec and o gi ing me he chance o lea n deep lea ning
in his challenging wo k ha combines biomedicine and enginee ing. O cou se, I would like o
hank o he collabo a ion schola ship ecei ed o ealize his hesis. I would like o ake his
oppo uni y o hank he ETSETB eache s and also my colleges o his las wonde ul 4 yea s
in he UPC. Finally, I would like o dedica e his p ojec o my amily and o `
0sca , o hei
suppo and unde s anding.
ii
Con en s
Abs ac i
Acknowledgemen s ii
Lis o Figu es
Lis o Tables ii
Abb e ia ions iii
Symbols ix
1 In oduc ion 1
1.1 S a emen o Pu pose ................................. 1
1.2 Ou line o he wo k ................................... 3
1.3 Technical Rema ks ................................... 3
2 S a e o he a 4
3 Me hodology 6
3.1 Sys em A chi ec u e .................................. 6
3.1.1 A chi ec u e .................................. 6
3.1.1.1 Masked V-NET ........................... 6
3.1.1.2 Deep Medic Ne wo k ........................ 6
3.1.2 Loss unc ions ................................. 8
3.1.2.1 C oss-en opy ............................ 8
3.1.2.2 Dice Simila i y Coe icien ..................... 8
3.1.2.3 Gene alised Dice Sco e ....................... 10
3.1.2.4 Weigh ed loss ............................. 10
3.1.3 Me ics ..................................... 10
3.1.3.1 Dice Sco e .............................. 10
3.1.3.2 Con usion ma ix .......................... 11
3.2 T aining scheme .................................... 11
3.2.1 Dense- aining ................................. 11
3.2.2 Pa ch sampling ................................. 11
3.2.2.1 Baseline: Fo eg ound-backg ound ................. 11
3.2.2.2 Pe -Class sampling scheme ..................... 12
3.2.2.3 Cu iculum Adap i e Sampling ................... 12
3.2.2.4 Baseline Adap i e Sampling Scheme ................ 13
iii
Con en s i
4 Expe imen s and Resul s 14
4.1 Da ase ......................................... 14
4.2 Dense- aining ..................................... 15
4.3 Pa ch sampling ..................................... 16
4.3.1 Fixed- ule sampling schemes ......................... 16
4.3.1.1 Loss unc ion: c oss-en opy .................... 17
4.3.1.2 Loss unc ion: weigh ed c oss-en opy ............... 18
4.3.2 Adap i e sampling schemes .......................... 19
4.3.2.1 CASED ................................ 20
4.3.2.2 BaseASS ............................... 22
4.4 Discussion ........................................ 25
5 Budge 28
6 Conclusions and u u e de elopmen 29
A Code o he p ojec 31
B Dense- aining 32
B.1 Se up .......................................... 32
B.2 T aining Cu es ..................................... 32
C Pa ch sampling 34
C.1 Se up .......................................... 34
C.2 T aining Cu es ..................................... 34
Bibliog aphy 37

Lis o Figu es
1.1 MRI T1 ......................................... 2
1.2 MRI T1c ........................................ 2
1.3 MRI T2 ......................................... 2
1.4 MRI FLAIR ....................................... 2
1.5 Mul imodal MRI images om subjec B a s17 −CBICA −ALX ........ 2
1.6 Glioma sub- egion. .................................. 2
3.1 V-Ne .......................................... 7
3.2 Two-pa h Deep Medic a chi ec u e .......................... 9
3.3 Example o eal s cap u ed dis ibu ion in he aining da a o BRATS 2015 . . 12
3.4 Schema ic diag am o CASED amewo k ...................... 13
4.1 Loss unc ion con e gence o dense- aining ..................... 16
4.2 G ound T u h s P edic ion o subjec B a s17 −T CIA −444 −1....... 16
4.3 Dice Whole o eg ound-backg ound .......................... 18
4.4 Dice Co e o eg ound-backg ound ........................... 18
4.5 Dice Enhance o eg ound-backg ound ........................ 18
4.6 T aining / Valida ion Dice Sco e E olu ion o baseline o eg ound-backg ound . 18
4.7 Dice Whole Pe -Class sampling scheme ........................ 18
4.8 Dice Co e Pe -Class sampling scheme ......................... 18
4.9 Dice Enhance Pe -Class sampling scheme ...................... 18
4.10 T aining / Valida ion Dice Sco e E olu ion o Pe -Class sampling scheme . . . . 18
4.11 Dice Whole pe o eg ound-backg ound ....................... 19
4.12 Dice Co e pe o eg ound-backg ound ........................ 19
4.13 Dice Enhance pe o eg ound-backg ound ...................... 19
4.14 T aining / Valida ion Dice Sco e E olu ion o o eg ound-backg ound sampling
scheme and weigh ed loss ............................... 19
4.15 Dice Whole Pe -Class sampling scheme ........................ 19
4.16 Dice Co e Pe -Class sampling scheme ......................... 19
4.17 Dice Enhance Pe -Class sampling scheme ...................... 19
4.18 T aining / Valida ion Dice Sco e E olu ion o Pe -Class sampling scheme and
weigh ed loss ...................................... 19
4.19 Pa ch dis ibu ion e olu ion du ing aining in baseline CASED model ...... 21
4.20 Pa ch dis ibu ion e olu ion du ing aining in slowed down model ........ 21
4.21 Valida ion Dice Whole CASED ............................ 22
4.22 Valida ion Dice Co e CASED ............................. 22
4.23 Valida ion Dice Enhance CASED ........................... 22
4.24 Tes DSC o all h ee CASED a ian s ....................... 22
4.25 Valida ion Dice Whole BaseASS ........................... 23
4.26 Valida ion Dice Co e BaseASS ............................ 23
4.27 Valida ion Dice Enhance BaseASS .......................... 23
4.28 Tes DSC o all h ee BaseASS a ian s ....................... 23
4.29 Pa ch dis ibu ion e olu ion o he BaseASS model ................ 24
Lis o Figu es i
4.30 Pe cen age o class oxels pe ba ch pe epoch o BaseASS ............ 24
4.31 E o maps om subjec B a s17 −CBICA −AAL wi h BaseASS ........ 25
4.32 O e all Dice Whole Tumo Compa ison. Valida ion is done using he whole sub-
jec as inpu . ...................................... 26
4.33 O e all Dice Co e Tumo Compa ison. Valida ion is done using he whole subjec
as inpu . ......................................... 27
4.34 O e all Dice Enhance Tumo Compa ison. Valida ion is done using he whole
subjec as inpu . .................................... 27
B.1 Dice Whole ....................................... 32
B.2 Dice Co e ........................................ 32
B.3 Dice Enhance ...................................... 32
B.4 Dice Sco e E olu ion o c oss-en opy loss unc ion ................. 32
B.5 Dice Whole ....................................... 33
B.6 Dice Co e ........................................ 33
B.7 Dice Enhance ...................................... 33
B.8 Dice Sco e E olu ion o DSC loss unc ion ..................... 33
C.1 Dice Whole ....................................... 35
C.2 Dice Co e ........................................ 35
C.3 Dice Enhance ...................................... 35
C.4 Dice Sco e E olu ion o he baseline CASED model ................ 35
C.5 Dice Whole ....................................... 35
C.6 Dice Co e ........................................ 35
C.7 Dice Enhance ...................................... 35
C.8 Dice Sco e E olu ion o he al e ed dis ibu ion CASED model .......... 35
C.9 Dice Whole ....................................... 35
C.10 Dice Co e ........................................ 35
C.11 Dice Enhance ...................................... 35
C.12 Dice Sco e E olu ion o he slowed down dis ibu ion CASED model ...... 35
C.13 Dice Whole ....................................... 35
C.14 Dice Co e ........................................ 35
C.15 Dice Enhance ...................................... 35
C.16 Dice Sco e E olu ion o BaseASS model ....................... 35
C.17 Dice Whole ....................................... 36
C.18 Dice Co e ........................................ 36
C.19 Dice Enhance ...................................... 36
C.20 Dice Sco e E olu ion o BaseASS model and median e o ............ 36
C.21 Dice Whole ....................................... 36
C.22 Dice Co e ........................................ 36
C.23 Dice Enhance ...................................... 36
C.24 Dice Sco e E olu ion o BaseASS and gene alised DSC .............. 36
C.25 Tes DSC o BaseASS and Deep Medic ne wo k .................. 36
Lis o Tables
3.1 Con usion Ma ix Example .............................. 11
4.1 Compa ison o mean alida ion DSC me ics o dense- aining .......... 15
4.2 Con usion ma ix o whole subje using DSC cos unc ion ( alida ion) ..... 15
4.3 Compa ison o mean alida ion DSC me ics o ixed- ule sampling schemes . . 17
4.4 Con usion ma ix o baseline o eg ound-backg ound ( alida ion) ........ 17
4.5 Con usion ma ix Pe -Class sampling scheme ( alida ion) ............. 18
4.6 Con usion ma ix o o eg ound-backg ound sampling using weigh ed loss ( ali-
da ion) ......................................... 19
4.7 Con usion ma ix Pe -Class sampling scheme using weigh ed loss ( alida ion) . . 19
4.8 Compa ison o mean alida ion DSC me ics o CASED .............. 20
4.9 Con usion ma ix baseline CASED scheme ( alida ion) .............. 20
4.10 Con usion ma ix slowed down CASED scheme ( alida ion) ............ 21
4.11 Compa ison o mean alida ion DSC me ics o BaseASS ............. 23
4.12 Con usion ma ix BaseASS wi h c oss-en opy loss ( alida ion) .......... 25
4.13 Con usion ma ix BaseASS wi h Gene alised DSC loss ( alida ion) ........ 25
4.14 Con usion ma ix o BaseASS and Deep Medic ne wo k ( alida ion) ....... 25
4.15 Mean Dice sco e me ics om he main expe imen s ca ied on .......... 26
5.1 P ojec Budge ..................................... 28
6.1 Compa ison be ween ou app oach and some o 2017 MICCAI B aTS Challenge 30
ii
Abb e ia ions
BaseASS Baseline Adap i e Sampling Scheme
CNN Con olu ional Neu al Ne wo k
CASED Cu iculum Adap i e Sampling o Ex eme Da a Imbalance
DSC Dice Simila i y Coe icien
FLAIR Fluid A enua ed In e sion Reco e y
HGG High G ade Glioma
ROI Region O In e es
LGG Low G ade Glioma
MRI Magne ic Resonance Imaging
WL Weigh ed Loss
iii
Me hodology 5
he eme gence o di e en sampling schemes (dense- aining, pa chwise aining, e c..) and cos
unc ions.
Pe o mance o CNN is signi ican ly in luenced by he s a egy o ex ac ing aining samples. A
common app oach is selec ing image pa ches equally sampled om each class. Ano he app oach
is o equally sample backg ound and o eg ound segmen s. On he o he hand, by employing
dense- aining o by sampling pa ches uni o mly, i migh su e om se e e class imbalance.
Hence, mul iple cos unc ions ha e been p oposed o alle ia e his issue.
The loss unc ion p oposed is c oss-en opy. Howe e , when he aining da a is se e ely unbal-
anced, his o mula ion can lead o a s ongly biased es ima ion owa ds he majo i y class. A
weigh ed c oss-en opy (B osch e al. [19]) is p oposed o ackle his p oblem, whe e he weigh s
a e in e sely p opo ional o he class equencies. Also, a di e en iable e sion o he Dice Sco e
Coe icien (DSC), p oposed by Mille a i e al. [20], is used as loss unc ion as i measu es he
o e lap o he egion o in e es (ROI). Recen ly, wo no el loss unc ions ha e been p esen ed:
he Gene alised DSC [21] and he Wasse s ein Dice .
This wo k ocuses on he segmen a ion o b ain umo s ollowing he guidelines indica ed by
B aTS Challenge, which began in 2012. The me hods submi ed in hese las yea s can be ound
in [17], [22]. The MICCAI (The Medical Image Compu ing and Compu e Assis ed In e en ion
Socie y) B aTS Challenge has an upda ed leade boa d1o he models wi h he bes pe o mances
acco ding o ollowing me ics: Dice sco e, sensi i i y, speci ici y and Hausdo dis ance. UCL-
TIG is in he i s posi ion on he anking wi h Ensembles o Mul iple Models and A chi ec u es
o Robus B ain Tumou Segmen a ion [18]. I achie es a dice whole umo o 0.90499, a dice
co e umo o 0.83779 and and a dice enhancing umo o 0.78585.
1h ps://www.cbica.upenn.edu/B aTS17/lboa dValida ion.h ml

Chap e 3
Me hodology
3.1 Sys em A chi ec u e
3.1.1 A chi ec u e
Two di e en a chi ec u es ha e been s udied in o de o disce n which one combined wi h o he
con igu a ions has he bes beha io : The Masked V-Ne [12] and he Deep Medic [11].
3.1.1.1 Masked V-NET
The masked V-Ne [12] is a modi ied e sion o he V-Ne a chi ec u e [20], which consis s o a
downsampling o encode pa h in cha ge o compac ing he signal and an upsampling o decode
pa h ha combines coa se ea u es om he encode ou pu wi h ine ea u es om hidden, in-
e media e le els o he encode o p o ide a segmen ed image o he same esolu ion as he inpu
image. The modi ica ions include he use o small ke nels o size 33, ba ch no maliza ion a e he
con olu ion and hen ReLU as non-linea i y. I was also in oduced a modi ied exp ession o he
esidual connec ions ha aim o p ese e he inpu signal h ough all he ne wo k. Max-pooling,
epea ed up-sampling o spa ial co espondence and 1x1x1 con olu ion a e a ian s in oduced
o ensu e dimensions ma ched in he addi ion laye . No e ha he ROI mask in oduced be o e
he inal p edic ions was only used o he dense- aining expe imen s. See igu e 3.1.
3.1.1.2 Deep Medic Ne wo k
The Deep Medic a chi ec u e p oposed in [11] consis s in cap u ing con ex ual and spacial in-
o ma ion h ough wo pa allel pa hs wi h di e en ea u e esolu ion ha sa es compu a ional
cos s and a oids pooling which could a ec on he accu acy o ou sys em. The 11-laye a chi ec-
u e p oposed by Deep Medic [11] is buil as shown in igu e 3.2. A high esolu ion pa h is able
o cap u e he mos complex de ails wi hin a small local neighbou hood, while a pa allel low
esolu ion pa h cap u es image-le el ea u es such as localiza ion o umo size. The di e ence in
6
Me hodology 7
Figu e 3.1: V-Ne
Me hodology 8
esolu ion is achie ed wi h di e en ecep i e ields: bo h pa hs a e buil upon a conca ena ion
o con olu ional laye s bu he la e has a pooling module a he e y beginning.
The ke nels on he con olu ional laye s o bo h high and low esolu ion pa hs a e o size 33. The
esul ing ma ices o he con olu ional laye s a e i s combined in o wo ull classi ica ion laye s
and hen inally classi ied.
3.1.2 Loss unc ions
Di e en loss unc ions ha e been conside ed in o de o s udy which one i s be e o ou
segmen a ion p oblem. The h ee cos unc ions ha will be s udied a e: c oss-en opy, Dice
Simila i y Coe icien (DSC) and Gene alised DSC.
3.1.2.1 C oss-en opy
C oss-en opy loss [23] o a mul i-class se ing can be exp essed as ollows
L(ˆy, y) = −
N
X
n=1
ynlog( ˆyn) (3.1)
No e ha Ns ands o he numbe o classes and yand ˆya e N-dimensional ec o s, whe e
yna e he ue labels o he class nand ˆyna e he so max ou pu s o he ne wo k o hose
ue labels. The mean c oss-en opy o e he whole ba ch is used as he cos unc ion a each
i e a ion and is compu ed as ollows:
¯
L(ˆy, y) = −1
|Y|X
y
X
n
ynlog( ˆyn) (3.2)
whe e Y e e s o aining samples om he ba ch.
3.1.2.2 Dice Simila i y Coe icien
Dice Sco e o mul i-class segmen a ion [24] is a measu e o simila i y be ween wo bina y se s:
he g ound u h G and he p edic ion P. Each se consis s o a egion o in e es (ROI) and
backg ound and he dice sco e, D[0,1], is he a io be ween he in e sec ion o he g ound- u h
and p edic ion ROIs and he sum o he a eas o bo h ROIs. The ollowing exp ession explains
he idea behind his loss unc ion.
D=2|P∩G|
|G|+|P|(3.3)
Then, a con inuous and di e en ial app oxima ion can be done by using so max p edic ions in-
s ead o he p edic ions hemsel es. The esul o a e aging and adap ing he p e ious exp ession
Me hodology 9
Figu e 3.2: Two-pa h Deep Medic a chi ec u e
Me hodology 10
o ou mul i-class p oblem is as ollows
¯
L(ˆy, y) = 1
|N|X
nN
2Piyi
nˆ
yi
n
Pi(yi
n+ˆ
yi
n)(3.4)
3.1.2.3 Gene alised Dice Sco e
The gene alised Dice Sco e is p oposed as loss unc ion in [21]. I s a modi ied e sion o he DSC
and is gi en by:
¯
L(ˆy, y)=1−2PN
n=1 αnPiyi
nˆ
yi
n
PN
n=1 αnPi(yi
n+ˆ
yi
n)(3.5)
whe e αnis a weigh o balance he impac in he loss unc ion o class n. Weigh ing by he
in e se o he class’ olume co ec s he con ibu ion o each label and educes he co ela ion
be ween he egion size and he Dice sco e. I is calcula ed as he in e se o he squa ed sum o
all he oxels o class n:
αn=1
(Piyi
n)2
3.1.2.4 Weigh ed loss
To u he elimina e he nega i e impac o he class imbalance, a weigh ed loss Lwis p oposed
as ollows. Whe e Lnis he speci ic loss o a ce ain class nN and kαnkis he he p obabili y
o appea ance o ha ce ain class. The e o e, by in e ing his p obabili y, we manage o gi e
mo e weigh o classes ha appea much less equen ly han o he s.
Lw=X
nN
1
αn
Ln(3.6)
The class equency αnis calcula ed as he sum o all he oxels o class n di ided by he sum
o all he oxels o he N classes:
αn=Piyi
n
PN
n=1 Piyi
n
(3.7)
3.1.3 Me ics
The me ics used o e alua e he pe o mance o each o he expe imen s ca ied on a e de ailed
below.
3.1.3.1 Dice Sco e
The e alua ion o he me hod using he dice sco e was calcula ed acco ding o 3.4 o di e en
ROI de ini ions: umo co e (classes 1 and 3), whole umo (classes 1,2,3) and enhancing umo
(class 3). This e alua ion amewo k is imposed by he Mul imodal B ain Tumo Segmen a ion
Challenge [3].

Me hodology 11
3.1.3.2 Con usion ma ix
The con usion ma ix is a able which allows o e alua e i ou segmen a ion sys em is mislabelling
one class as ano he . Each ow o he able ep esen s he ue class and each column he
p edic ed class. This able is use ul o gi e an idea o which classes a e well classi ied and which
a e w ongly con used wi h he o he s. I gi es us insigh s o how a model can be imp o ed. The
con usion ma ices in his documen will p esen he ollowing s uc u e:
G P 0 1 2 3
0
1
2
3
Table 3.1: Con usion Ma ix Example
3.2 T aining scheme
3.2.1 Dense- aining
Be o e analyzing how we explo e he pa ch wise aining scheme, which is he main con ibu ion
o ou wo k, we s udy dense- aining. We denomina e dense- aining he s a egy ha uses he
whole MRI image and he ou modali ies as inpu in o he ou Con olu ional Neu al Ne wo k.
The pe o mance o his aining scheme will be analyzed o wo di e en loss unc ions.
3.2.2 Pa ch sampling
Pa ch-sampling aining scheme consis s o eeding he ne wo k wi h small h ee-dimensional
pa ches o each subjec . Each o hese slices is associa ed wi h 4 di e en modali ies: T1,
T2, T1C and FLAIR and i s size is se o 643. The sampling scheme can be c i ique o any
medical applica ion and an exhaus i e analysis o di e en me hods is pe o med h oughou
he manusc ip wi h u he nume ical compa ison. In he case o b ain umo segmen a ion,
he high imbalance be ween backg ound and umo egions and sub egions may equi e lexible
me hod ha balances he inpu aining dis ibu ion o he di e en classes.
3.2.2.1 Baseline: Fo eg ound-backg ound
The aining scheme used in [11] ies o sol e he imbalance p oblem by a sampling scheme ha
samples he cen al- oxel o each pa ch wi h equip obabili y be ween o eg ound ( umo egions)
and backg ound. Hence, each ba ch is build by he same numbe o pa ches whose cen al oxel
is o eg ound and backg ound. No e ha no dis inc ion is made be ween umo sub egions. This
is o main ain he ela i e dis ibu ion o he o eg ound classes and a he same ime accoun
o he imbalance p oblem be ween heal hy issue and umo issue.
Me hodology 12
Figu e 3.3 om [11] shows how he ela i e dis ibu ion o he o eg ound classes is closely
p ese ed and he imbalance in compa ison o he heal hy issue is au oma ically alle ia ed.
Figu e 3.3: Example o eal s cap u ed dis ibu ion in he aining da a o BRATS 2015
This o eg ound-backg ound sampling scheme has been selec ed o be ou baseline scheme on
pa ch sampling.
3.2.2.2 Pe -Class sampling scheme
This aining scheme p oposes sampling pa ches acco ding o a ule ha ensu es equip obabili y
be ween all classes in he cen al- oxel a each epoch. The idea is o accoun o he imbalance
be ween backg ound and each one o he umo sub egions. I al e s he balance be ween each
class and hence, i cap u es a a he di e en dis ibu ion om he o iginal.
3.2.2.3 Cu iculum Adap i e Sampling
Cu iculum Adap i e Sampling (CASED) [25] scheme i s ’s objec i e is o ackle he p oblem
o class imbalance. The basic ideas o his sys em a e:
1. Lea n ea u es ela ed o umo : s a in oducing only pa ches wi h umo in o he ne -
wo k.
2. As he ne wo k is aining, s a adding backg ound pa ches o lea n heal hy issue p op-
e ies.
3. In he end, uni o m sampling is eached o mimic eal da a dis ibu ion.
CASED scheme can be di ided in wo pa s: Cu iculum and Adap i e Sampling.
Cu iculum is he pa ha ackles he class imbalance p oblem con olling he inpu pa ches o
he ne wo k by deciding be ween umo -sampling and uni o m-sampling gene a o s. I aining
was pe o med using only umo pa hces, i could esul in o e i ing because i would no lea n
how o ep esen he main pa o he MRI image, he backg ound class. Then, he cu iculum
pa is esponsible o dec easing (as unc ion o he aining examples seen) he numbe o pa ches
wi h umo un il i eaches he eal dis ibu ion. The h eshold pxis gi en by Equa ion 3.8. Fo
alues g ea e han px umo pa ches a e selec ed, o he wise i is picked any andom pa ch.
px+1 =px∗1
M
(i e ∗epochs∗K)−1
(3.8)
Expe imen s and Resul s 13
wi h
p0= 1
whe e Mis he numbe o segmen s, i e he numbe o i e a ions, epochs he o al numbe o
aining epochs and Kis a cons an o speed up o slow down he h eshold cu e.
Adap i e Sampling is equi ed o e ine he p e ious pa . E en using he cu iculum, solu ions
wi h alse posi i es would s ill happen. Mo eo e , mos ly all oxels on b ain images could be
con iden ly classi ied as backg ound. The e o e, he adap i e sampling encou ages aining inpu s
whose p edic ions a e alse posi i es.
Figu e 3.4: Schema ic diag am o CASED amewo k
3.2.2.4 Baseline Adap i e Sampling Scheme
Adap i e Sampling Scheme o E icien ly T ain Fully Con olu ional Ne wo ks [26], om now
on BaseASS (baseline adap i e sampling scheme), sugges s a pa chwise aining scheme ha
p e ends o adap i ely build aning samples a each epoch by looking a he ne wo k aining
e o . Fo each subjec , e o maps Eia e build concu en ly a he end o each epoch as:
Ei= 1 −CNN(w, In(x))Ln(x)(3.9)
whe e CNN(w, In(x))Ln(x) ep esen s he so max p edic ions calcula ed using he aining
weigh s wo e an image In. Thus, a pa ch is accep ed in o he ba ch acco ding o i s ela-
ion o he h eshold de ined by:
Ei(c)> U(0,1) −(3.10)
whe e c is he cen al oxel o he pa ch, U(0,1) is a andom uni o m a iable and a pa ame e
o calib a e he algo i hm: = 0 means a comple ely adap i e scheme and = 1 he uni o m
sampling scheme.
Chap e 4
Expe imen s and Resul s
This chap e p esen s and compa es he esul s ob ained om applying he me hodologies men-
ioned in Chap e 3.
4.1 Da ase
In his hesis, he da a used o ain he ne wo k has been ob ained om he MICCAI B aTS
2017 Challenge [3]. This da ase is composed by 210 HGG (high g ade glioma) and 75 LGG (low
g ade glioma) subjec s o which 171 a e used o aining (60%) and 116 a e used o alida ion
pu poses (40%). No da a augmen a ion was used in any expe imen . The ou modali ies o each
MRI image (na i e T1, pos -con as T1-weigh ed, T2-weigh ed and T2 FLAIR) a e co- egis e ed
o he same ana omical empla e and in e pola ed o he same esolu ion (1 mm3).
T aining he ne wo k elies on he p ope selec ion o hype pa ame e s, a w ong choice can lead
o o e i ing, unde i ing o simply no aining. Each expe imen should ha e been op imized
indi idually un il ob aining he bes esul s. Besides he ac ha his is oo compu a ionally
demanding, we ha e no done his in o de o be able o compa e all he models unde he same
condi ions. Fi s , he lea ning a e, which ells he op imize how a has o mo e he weigh s in
he di ec ion o he g adien , was se o 0.0005. The momen um was se o 0.99. Regula iza ion
p e en s he coe icien s o o e i and i depends on wo a iables: L2 is a ac o mul iplying
he sum o he squa e o he weigh s, while L1 is he ac o mul iplying he absolu e sum o
he weigh s. The alues chosen we e L1=0.00001, L2=0.005. The numbe o aining epochs is
a iable o each expe imen . Masked V-Ne was used as he base a chi ec u e o all expe imen s
excep one, whe e Deep Medic Ne wo k was used. The ini ializa ion o he weigh s was done
acco ding o [27].
Fo pa ch sampling, we used 600 segmen s/epoch o aining and 400 segmen s/epoch o ali-
da ion wi h a ba ch size o 10.
14
Expe imen s and Resul s 21
G P % 0 1 2 3
0 0,99877 0,00016 0,00095 0,00012
10,10293 0,57165 0,24711 0,07831
20,15830 0,04181 0,78414 0,01576
30,05627 0,04030 0,06982 0,83361
Table 4.10: Con usion ma ix slowed down CASED scheme ( alida ion)
aking in o accoun he cen al oxel o decide which pa ch ype is i . I we conside ed all he
pa ch, calcula ing he class wi h maximum p esence we would always ob ain class 0. The hi d
plo is he cu iculum cu e om equa ion 3.8. O e all, he slowed down e sion was ound o
ha e he bes pe o mance.
Figu e 4.19: Pa ch dis ibu ion e olu ion du ing aining in baseline CASED model
Figu e 4.20: Pa ch dis ibu ion e olu ion du ing aining in slowed down model

Expe imen s and Resul s 22
Finally, he model was used o in e he es da a segmen a ion. Boxplo s om igu e 4.24
show he compa ison in DSC ac oss he h ee models. The ou lie s had been checked and we e
mainly due o MRI inpu images om he da ase wi h poo condi ions. In conclusion, a ian
2 demons a es o be be e han baseline CASED o only including a ian 1 acco ding o
he mean DSC me ics, he con usion ma ix and he alida ion boxplo s. Un il his poin , his
esul ou pe o ms all he o he schemes used p e iously. This is because in he las i e a ions he
aining dis ibu ion app oaches he ue dis ibu ion and he e o e ge s a be e gene aliza ion.
Figu e 4.21: Val-
ida ion Dice Whole
CASED
Figu e 4.22: Valida-
ion Dice Co e CASED
Figu e 4.23: Vali-
da ion Dice Enhance
CASED
Figu e 4.24: Tes DSC o all h ee CASED a ian s. F om le o igh : Baseline CASED,
modi ied sub egion dis ibu ion and slowed down CASED. Valida ion is done using he whole
subjec as inpu .
4.3.2.2 BaseASS
The baseline adap i e sampling scheme was implemen ed as men ioned in Chap e 3. The base-
line me hod uses c oss-en opy and he cen al oxel e o as he c i e ia o selec ion. Mo eo e ,
we explo e his model and p opose h ee al e na i es: 1) pa ch selec ion acco ding he median
e o alue o he whole pa ch (Eq. 4.1) ins ead o only he e o alue o he cen al oxel (Eq.
3.10); 2) use he Gene alised DSC loss unc ion ins ead o c oss-en opy; 3) use ano he a chi ec-
u e: Deep Medic ne wo k ins ead o masked V-Ne . Fo his las expe imen , he cos unc ion
chosen was c oss-en opy as i had shown o pe o m be e in his con igu a ion. This classical
a chi ec u e was chosen because, unlike Masked V-Ne [12], i does no ha e any max-pooling
laye . We wan o a oid max-pooling beacuse i educes he spa ial size o he ep esen a ion and
he eby i educes he numbe o aining pa ame e s, which migh con ain ele an in o ma ion
o ou segmen a ion ask.
median(Eipa ch )> U(0,1) −(4.1)
To be as ai as possible when compa ing, he es o pa ame e s emained ixed. We ained
baseline and he i s wo models wi h Adam Op imiza ion me hod [28] wi h a lea ning a e o
0.0005. The hi d model (Deep Medic Ne wo k) was ained using RMSp op op imize and a
lea ning a e o 0.0001. Fo mo e de ails consul he annex.
The esul s epo ed on Table 4.11 show ha BaseASS wi h classical c oss-en opy loss ou pe -
o ms Gene alised DSC. Howe e , BaseASS using he cen al oxel’s e o o he median e o
Expe imen s and Resul s 23
Scheme Dice Whole Dice Co e Dice Enhance
BaseASS 0,86286 0,74418 0,67466
BaseASS + Gene alised DSC 0,81759 0,69727 0,65100
BaseASS + Median e o as selec ion
c i e ia
0,85087 0,75052 0,657198
BaseASS + Deep Medic ne wo k 0,65379 0,55027 0,48243
Table 4.11: Compa ison o mean alida ion DSC me ics o BaseASS
o he pa ch show simila pe o mance. This sugges s ha using only he cen al oxel is enough
o ge a gene al ep esen a ion o he pa ch e o . This esul s can be clea ly con i med when
looking a he boxplo compa ison in Figu e 4.28. Finally, he masked V-Ne ne wo k ob ains
be e esul s han he Deep Medic Ne wo k. Ne e heless, he la e hype -pa ame e s had no
been op imized and we can’ conclude ha i is beha ing wo se ye . The las expe imen was
no wo h o be included in he boxplo compa ison 4.28, bu can be ound in he Appendix in
Figu e C.25.
Figu e 4.25: Val-
ida ion Dice Whole
BaseASS
Figu e 4.26: Val-
ida ion Dice Co e
BaseASS
Figu e 4.27: Vali-
da ion Dice Enhance
BaseASS
Figu e 4.28: Tes DSC o all h ee BaseASS a ian s. F om le o igh : BaseASS using
c oss en opy loss, BaseASS using median e o and BaseASS using gene alised DSC loss.
Valida ion is done using he whole subjec as inpu
Figu e 4.29 shows he numbe o segmen s o each class pe aining epoch o he BaseASS
model. Fo cla i ica ion, i is conside ed ha he pa ch class is de ined as he class o he cen al
oxel ( he one whose e o alue is aken as selec ion c i e ia). In compa ison o CASED scheme,
sec ion 4.3.2.1 (Figu e 4.19), he e he numbe o lesion pa ches is no educed in each epoch bu
emains cons an o all he aining. In con as , he numbe o backg ound pa ches dec eases
in each i e a ion. I can be seen ha he po ion o class 1 pa ches is sligh ly highe han ha
o class 2 and 3. This leads us o belie e ha i is a wise decision o gi e a li le mo e weigh o
class 1 han 2 o 3.
While Figu e 4.30 shows he aining dis ibu ion o he ba ch in each i e a ion. I can be
obse ed ha despi e o choosing di e en numbe o segmen s o each class, he aining dis-
ibu ion emains almos cons an o all he epochs. Mo eo e , his dis ibu ion is close o he
eal one.
Quali a i e esul s o he e o maps calcula ed in each i e a ion (Figu e 4.31) demons a e he
impo ance o selec ing he igh lesion pa ches ( hose wi h high e o alues) agains selec ing
any umo pa ch wi hou any c i e ia. I can be seen how he pe o mance o he ne wo ks
Expe imen s and Resul s 24
Figu e 4.29: Pa ch dis ibu ion e olu ion o he BaseASS model
Figu e 4.30: Pe cen age o class oxels pe ba ch pe epoch o BaseASS
imp o es om epoch 1 (Fig. 4.31(b)) o epoch 30 (Fig. 4.31(d)). We can see ha om he 15 h
o he 30 h epoch, he ne wo k wo ks o e ine he segmen a ion. Mainly, he e o lies in he
bounda ies be ween wo di e en classes.
Table 4.12, Table 4.13 and Table 4.14 show he con usion ma ices o he BaseASS wi h he wo
di e en loss unc ions and di e en a chi ec u e.
Expe imen s and Resul s 25
(a) G ound T u h
abo e MRI T1
(b) E o map in
epoch 1
(c) E o map in
epoch 15
(d) E o map in
epoch 30
(e)
Colo
scale
Figu e 4.31: E o maps om subjec B a s17 −CBICA −AAL wi h BaseASS. Ligh blue
is equi alen o ze o e o , yellow means maximum e o alue.
G P % 0 1 2 3
0 0,99914 0,00013 0,00066 0,00007
10,09024 0,61238 0,22317 0,07421
20,15804 0,06343 0,76245 0,01608
30,05271 0,04601 0,06876 0,83252
Table 4.12: Con usion ma ix BaseASS wi h c oss-en opy loss ( alida ion)
G P % 0 1 2 3
0 0,99825 0,00020 0,00140 0,00015
10,11032 0,62370 0,18580 0,08018
20,13554 0,062621 0,78282 0,01903
30,05122 0,04287 0,06184 0,84407
Table 4.13: Con usion ma ix BaseASS wi h Gene alised DSC loss ( alida ion)
G P % 0 1 2 3
0 0,99361 0,00074 0,00469 0,00096
10,12832 0,52653 0,29216 0,05300
20,11627 0,09170 0,76764 0,02439
30,04074 0,13621 0,07489 0,74816
Table 4.14: Con usion ma ix o BaseASS and Deep Medic ne wo k ( alida ion)
4.4 Discussion
In his wo k, we ha e analyzed and e alua ed 13 sampling schemes. An o e all compa ison is
done in Table 4.15. Fi s , we concluded ha o whole subjec (dense- aining) DSC pe o med
be e han c oss-en opy since he second ea s all aining oxels equally and his is no help ul
when he ne wo k has di icul ies in lea ning ep esen a ions o he mino i y classes. DSC does
an implici e-weigh ing o he oxels alle ia ing his issue.
In he con ex o pa ch sampling, ixed- ule sampling schemes ha e appea ed o be isky models
as hey o e -modi iy he aining dis ibu ion om he o iginal one, in such manne ha hey
change he model beha io making i di icul o adjus o he alida ion dis ibu ion (uni o m).
Budge 26
The bes esul was o o eg ound-backg ound sampling wi hou weigh ed loss unc ion, he
scheme among he ou which less modi ies he dis ibu ion. This esul s ha e been checked o
c oss-en opy, bu we canno ensu e wha would happen using o he loss unc ions.
Rega ding adap i e sampling, bo h CASED [25] and BaseASS [26] achie e e y good esul s.
I ’s isky o claim which one is be e han he o he because, al hough he BaseASS p esen s
he highe dice sco e and e en a li le mo e be e esul s on he con usion ma ix, he beha io
is e y simila and we do no know i we could achie e be e esul s adjus ing he models wi h
new modi ica ions. The key o he success o bo h is ha he inal dis ibu ion a he aining
end is e y close o he eal one. Hence, his gi es hem s eng h o alle ia e class imbalance
and gene alize co ec ly.
Scheme Dice Whole Dice Co e Dice Enhance
Dense- aining + C oss-en oy 0,67032 0,41302 0,50329
Dense- aining + DSC 0,80534 0,70092 0,65296
Fo eg ound-backg ound sampling scheme 0,80156 0,59277 0,59949
Pe -Class sampling scheme 0,74112 0,53220 0,567712
Fo eg ound-backg ound sampling scheme
+ WL
0,72799 0,49146 0,58795
Pe -Class sampling scheme+ WL 0,30150 0,19794 0,47847
CASED 0,84130 0,73418 0,65403
BaseASS 0,86286 0,74418 0,67466
Table 4.15: Mean Dice sco e me ics om he main expe imen s ca ied on
Boxplo s 4.32,4.33 and 4.34 show qua ile anges o he DSC sco es (Whole Tumo , Co e Tumo
and Enhancing Tumo espec i ely) on he es da ase s, do s indica e ou lie s and he ed
line indica es he median alue. They gi e us an o e iew o he beha iou o each o he
expe imen s. I could be a gued ha whole subjec wi h DSC has a compa able ou come o
BaseASS o CASED, howe e , i we look a qua ile 25 o he i s scheme, i is much lowe han
he qua ile 25 o he las wo.
Figu e 4.32: O e all Dice Whole Tumo Compa ison. Valida ion is done using he whole
subjec as inpu .

Budge 27
Figu e 4.33: O e all Dice Co e Tumo Compa ison. Valida ion is done using he whole
subjec as inpu .
Figu e 4.34: O e all Dice Enhance Tumo Compa ison. Valida ion is done using he whole
subjec as inpu .
Chap e 5
Budge
This p ojec has been ca ied in he Image P ocessing G oup, ETSETB, UPC. Deep Lea ning is
highly compu a ionally demanding, consequen ly a GPU was needed: The GPU GeFo ce GTX
Ti an Black has an app oxima e cos o 920 e, howe e UPC p o ided i o us wi hou any cos .
Thus, he main cos o his p ojec comes om he sala y o he esea che s and he ime spen
in i . The eam o he de elopmen o his hesis is o med by wo p o esso s who we e ad ising
me as senio enginee s and mysel as junio enginee . The o al du a ion o he p ojec was o
33 weeks. The budge o he p ojec can be calcula ed:
Amoun
Wage/hou
Dedica-
ion
To al
Junio
Enginee
1 8,00e/h 25h/week 6,600 e
Senio
Enginee
2 20,00e/h 4h/week 5,280 e
Table 5.1: P ojec Budge
28
Chap e 6
Conclusions and u u e
de elopmen
The main goal o his p ojec was o apply di e en s a e-o - he a me hodologies o b ain
umo segmen a ion o make a compa a i e s udy o all hem. We s udied how dense- aining
and pa ch sampling pe o med in he b ain umo segmen a ion. Fo his eason, we p oposed
a ian s o ixed- ule and adap i e sampling schemes.
In sec ion 4.2 we p o ed ha he DSC ou pe o ms c oss-en opy abili y o classi y umo /non-
umo oxels in dense- aining. Howe e , in pa ch sampling, c oss-en opy demons a ed o be
be e han weigh ed loss unc ions as weigh ed c oss-en opy and gene alised DSC.
Then, in sec ion 4.3, we show ha p ope ly designed pa ch sampling ou pe o ms dense- aining
schemes. Mo eo e , we conclude ha i we al e signi ican ly he aining dis ibu ion om
he eal, such ha using pe -class sampling scheme, we inc ease gene aliza ion e o e en i
aining imp o es due o he misma ch be ween aining and es ing dis ibu ions. Mo eo e ,
no el adap i e aining schemes a e shown o u he imp o e he pe o mance compa ed o
he ixed- ule schemes o he b ain umo segmen a ion ask. We obse ed how only adap i e
sampling ob aines good esul s al e ing he aining dis ibu ion in such a way ha achie es
lea ning he ea u es o hose segmen s ha a e mo e di icul o classi y. Ou bes esul s a e:
dice sco e o whole umo o 0,862 , dice sco e o co e umo o 0,744 and a dice sco e o
enhancing umo o 0,67.
We compa e ou esul s wi h he leade s o he MICCAI challenge anking and he esul s
p esen ed by he UPC his las edi ion (Table 6.1). Fo dice sco e o whole umo , ou me hod
is e y close o he esul s ob ained by he op pe o ming me hods while, ou me hod achie es
low dice sco e o enhancing umo and co e umo . BaseASS has simila pe o mance o he
app oach p esen ed by he UPC, in which hey ackle he p oblem wi h a pipeline o wo masked
V-Ne s and dense- aining. Howe e , ou esul s a e s ill well below he ensemble model p oposed
by UCL-TIG.
29
Appendix 30
Scheme Dice Whole Dice Co e Dice Enhance
UCL-TIG 0,9 0,83 0,78
UPC 0,87 0,63 0,71
BaseASS 0,86286 0,74418 0,67466
Table 6.1: Compa ison be ween ou app oach and some o 2017 MICCAI B aTS Challenge
Finally, he choice o di e en hype -pa ame e s o he op imiza ion and egula iza ion can
hea ily a ec he pe o mance o a model. I is o en obse ed ha he choice o op imize and
i s con igu a ion, o ins ance egula i za ion o he lea ning a e, o a la ge ex en , de e mine
whe he a good o a bad segmen a ion is ob ained. The sensi i i y o all hese hype -pa ame e s
is magni ied by he ac ha e-using he same se ing does no gua an ee o beha e well among
di e en ne wo k a chi ec u es, o e en on di e en asks and da a. Hence, i is o en di icul o
d aw gene ic and con iden conclusions wi hou spending a huge amoun o ime in op imizing
he expe imen al se ings.
In he u u e, we a e in e es ed in ying aining wi h a di e en a chi ec u es as we belie ha
his one migh be a bo leneck. We a e in e es ed in ying dila ed con olu ions as hey a e
able o in oduce sys ema ically mul i-scale con ex ual in o ma ion wi hou loosing esolu ion.
HighResNe [16] is a ne wo k which uses dila ed con olu ion and esidual connec ions. The
a chi ec u e is based on he ac ha dila ed con olu ions suppo exponen ial expansion o he
ecep i e ield wi hou loss o esolu ion and a oid max-pooling.
Now ha we al eady know ha BaseASS p o ides sucess ul esul s i would be in e es ing o do
mo e ese ach on i . The BaseASS and also he CASED s ill ha e o he lea ning pa ame e s ha
could be explo ed, o ins ance gi e p io i y o pa ches o he modali y wi h he g ea es impac .
I would also be ele an o conside a ion make adap i e o he pa ame e s as he lea ning a e
o he pa ch-size. I also has been le ying o do an ensemble o he bes me hods in his hesis.
Finally, adapa i e sampling schemes om his wo k migh be powe ul in o he segmen a ion
asks whe e imbalance is also a p oblem, like whi e ma e hype in ensi ies (WMH) segmen a ion.
Bibliog aphy
[1] E.C. Holland. P ogeni o cells and glioma o ma ion.
[2] Bjoe n Menze, And as Jakab, S e an Baue , Jayash ee Kalpa hy-C ame , Key an Fa a-
hani, Jus in Ki by, Yuliya Bu en, Nicole Po z, Johannes Slo boom, Roland Wies , Le en e
Lanczi, Elisabe h Ge s ne , Ma c-And e Webe , Tal A bel, B ian A an s, Nicholas Ay-
ache, Pa icia Buendia, Louis Collins, Nicolas Co die , Jason Co so, An onio C iminisi,
Tilak Das, He ´e Delinge e, Caga ay Demi alp, Ch is ophe Du s , Michel Doja , Senan
Doyle, Joana Fes a, Flo ence Fo bes, Ezequiel Ge emia, Ben Glocke , Polina Golland, Xi-
ao ao Guo, Andac Hamamci, Khan I ekha uddin, Raj Jena, Nigel John, Ende Konukoglu,
Danial Lashka i, Jose An onio Ma iz, Raphael Meie , Se gio Pe ei a, Doina P ecup, S. J.
P ice, Tammy Riklin-Ra i , Syed Reza, Michael Ryan, Law ence Schwa z, Hoo-Chang
Shin, Jamie Sho on, Ca los Sil a, Nuno Sousa, Nagesh Subbanna, Gabo Szekely, Thomas
Taylo , Owen Thomas, Nicholas Tus ison, Gozde Unal, Flo Vasseu , Max Win e ma k,
Dong Hye Ye, Liang Zhao, Binsheng Zhao, Da ko Zikic, Ma cel P as awa, Mau icio Reyes,
and Koen Van Leempu . The Mul imodal B ain Tumo Image Segmen a ion Benchma k
(BRATS). IEEE T ansac ions on Medical Imaging, 34(10):1993–2024, Oc obe 2014. doi:
10.1109/TMI.2014.2377694. URL h ps://hal.in ia. /hal-00935640.
[3] Mul imodal b ain umo segmen a ion challenge 2017. URL h p://www.med.upenn.edu/
sbia/b a s2017/da a.h ml.
[4] B. H. Menze, A. Jakab, S. Baue , J. Kalpa hy-C ame , K. Fa ahani, J. Ki by, Y. Bu en,
N. Po z, J. Slo boom, R. Wies , L. Lanczi, E. Ge s ne , M. A. Webe , T. A bel, B. B.
A an s, N. Ayache, P. Buendia, D. L. Collins, N. Co die , J. J. Co so, A. C iminisi, T. Das,
H. Delinge e, . Demi alp, C. R. Du s , M. Doja , S. Doyle, J. Fes a, F. Fo bes, E. Ge emia,
B. Glocke , P. Golland, X. Guo, A. Hamamci, K. M. I ekha uddin, R. Jena, N. M. John,
E. Konukoglu, D. Lashka i, J. A. Ma iz, R. Meie , S. Pe ei a, D. P ecup, S. J. P ice, T. R.
Ra i , S. M. S. Reza, M. Ryan, D. Sa ikaya, L. Schwa z, H. C. Shin, J. Sho on, C. A. Sil a,
N. Sousa, N. K. Subbanna, G. Szekely, T. J. Taylo , O. M. Thomas, N. J. Tus ison, G. Unal,
F. Vasseu , M. Win e ma k, D. H. Ye, L. Zhao, B. Zhao, D. Zikic, M. P as awa, M. Reyes,
and K. Van Leempu . The mul imodal b ain umo image segmen a ion benchma k (b a s).
IEEE T ansac ions on Medical Imaging, 34(10):1993–2024, Oc 2015. ISSN 0278-0062. doi:
10.1109/TMI.2014.2377694.
[5] Fei-Fei Li, Jus in Johnson, and Se ena Yeung. Cs231n: Con olu ional neu al ne wo ks o
isual ecogni ion., 2017. URL h p://cs231n.s an o d.edu/.
37

Bibliog aphy 38
[6] Mohammad Ha aei, Axel Da y, Da id Wa de-Fa ley, An oine Bia d, Aa on C. Cou ille,
Yoshua Bengio, Ch is Pal, Pie e-Ma c Jodoin, and Hugo La ochelle. B ain umo segmen-
a ion wi h deep neu al ne wo ks. CoRR, abs/1505.03540, 2015. URL h p://a xi .o g/
abs/1505.03540.
[7] e al Ad i`a Casami jana. 3D Con olu ional Neu al Ne wo ks o B ain Tumo Segmen a ion:
a compa ison o mul i- esolu ion a chi ec u es. In e na ional Wo kshop on B ainlesion:
Glioma, Mul iple Scle osis, S oke and T auma ic B ain Inju ies. Sp inge , Cham, 2016.
[8] F an¸cois Cholle e al. Ke as. h ps://ke as.io, 2015.
[9] Ian Good ellow, Yoshua Bengio, and Aa on Cou ille. Deep Lea ning. MIT P ess, 2016.
h p://www.deeplea ningbook.o g.
[10] S´e gio Pe ei a, Ad iano Pin o, Vic o Al es, and Ca los A. Sil a. Deep con olu ional neu al
ne wo ks o he segmen a ion o gliomas in mul i-sequence m i. In Alessand o C imi,
Bjoe n Menze, Oska Maie , Mau icio Reyes, and Heinz Handels, edi o s, B ainlesion:
Glioma, Mul iple Scle osis, S oke and T auma ic B ain Inju ies, pages 131–143, Cham,
2016. Sp inge In e na ional Publishing. ISBN 978-3-319-30858-6.
[11] Kons an inos Kamni sas, Ch is ian Ledig, Vi ginia F. J. Newcombe, Joanna P. Simpson,
And ew D. Kane, Da id K. Menon, Daniel Ruecke , and Ben Glocke . E icien mul i-
scale 3d CNN wi h ully connec ed CRF o accu a e b ain lesion segmen a ion. CoRR,
abs/1603.05959, 2016. URL h p://a xi .o g/abs/1603.05959.
[12] Ma cel Ca `a, Ad i`a Casami jana, I ina S´anchez, Ma c Combalia, and Ve ´onica Vilaplana.
Masked -ne : an app oach o b ain umo segmen a ion.
[13] ¨
Ozg¨un C¸i¸cek, Ahmed Abdulkadi , Soe en S. Lienkamp, Thomas B ox, and Ola Ron-
nebe ge . 3d u-ne : Lea ning dense olume ic segmen a ion om spa se anno a ion. CoRR,
abs/1606.06650, 2016. URL h p://a xi .o g/abs/1606.06650.
[14] Ola Ronnebe ge , Philipp Fische , and Thomas B ox. U-ne : Con olu ional ne wo ks o
biomedical image segmen a ion. CoRR, abs/1505.04597, 2015. URL h p://a xi .o g/
abs/1505.04597.
[15] Rupesh Kuma S i as a a, Klaus G e , and J¨u gen Schmidhube . Highway ne wo ks.
CoRR, abs/1505.00387, 2015. URL h p://a xi .o g/abs/1505.00387.
[16] Wenqi Li, Guo ai Wang, Lucas Fidon, S´ebas ien Ou selin, M. Jo ge Ca doso, and Tom
Ve cau e en. On he compac ness, e iciency, and ep esen a ion o 3d con olu ional
ne wo ks: B ain pa cella ion as a p e ex ask. CoRR, abs/1707.01992, 2017. URL
h p://a xi .o g/abs/1707.01992.
[17] P ep oceedings 2017 In e na ional MICCAI B aTS Challenge, 2017. MICCAI,CBICA.
URL h ps://www.cbica.upenn.edu/sbia/Spy idon.Bakas/MICCAI_B aTS/MICCAI_
B aTS_2017_p oceedings_sho Pape s.pd .
[18] Kons an inos Kamni sas, Wenjia Bai, Enzo Fe an e, S e en G. McDonagh, Ma hew Sin-
clai , Nick Pawlowski, Ma in Rajchl, Ma hew C. H. Lee, Be nha d Kainz, Daniel Ruecke ,
Bibliog aphy 39
and Ben Glocke . Ensembles o mul iple models and a chi ec u es o obus b ain umou
segmen a ion. CoRR, abs/1711.01468, 2017. URL h p://a xi .o g/abs/1711.01468.
[19] Tom B osch, Youngjin Yoo, Lisa Y. W. Tang, Da id K. B. Li, An hony T aboulsee, and
Roge Tam. Deep con olu ional encode ne wo ks o mul iple scle osis lesion segmen a ion.
In Nassi Na ab, Joachim Ho negge , William M. Wells, and Alejand o F. F angi, edi o s,
Medical Image Compu ing and Compu e -Assis ed In e en ion – MICCAI 2015, pages 3–
11, Cham, 2015. Sp inge In e na ional Publishing. ISBN 978-3-319-24574-4.
[20] Faus o Mille a i, Nassi Na ab, and Seyed-Ahmad Ahmadi. V-ne : Fully con olu ional
neu al ne wo ks o olume ic medical image segmen a ion. CoRR, abs/1606.04797, 2016.
URL h p://a xi .o g/abs/1606.04797.
[21] Ca ole H. Sud e, Wenqi Li, Tom Ve cau e en, S´ebas ien Ou selin, and M. Jo ge Ca doso.
Gene alised dice o e lap as a deep lea ning loss unc ion o highly unbalanced segmen a-
ions. CoRR, abs/1707.03237, 2017. URL h p://a xi .o g/abs/1707.03237.
[22] Mul imodal B ain Tumo Image Segmen a ion Benchma k: Change De ec ion, 2016. MIC-
CAI,CBICA.
[23] Cs229: Addi ional no es on backp opaga ion, 2017. URL h p://cs229.s an o d.edu/
no es/cs229-no es-backp op.pd .
[24] Lucas Fidon, Wenqi Li, Luis C. Ga c´ıa-Pe aza-He e a, Jinend a Ekanayake, Neil Ki chen,
S´ebas ien Ou selin, and Tom Ve cau e en. Gene alised wasse s ein dice sco e o imbalanced
mul i-class segmen a ion using holis ic con olu ional ne wo ks. CoRR, abs/1707.00478,
2017. URL h p://a xi .o g/abs/1707.00478.
[25] Jesson A., Guiza d N., Ghalehjegh S.H., Goblo D., Soudan F., and Chapados N. Cased:
Cu iculum adap i e sampling o ex eme da a imbalance. 10435, 2017.
[26] Lo enz Be ge , Eoin Hyde, M. Jo ge Ca doso, and S´ebas ien Ou selin. An adap i e sampling
scheme o e icien ly ain ully con olu ional ne wo ks o seman ic segmen a ion. CoRR,
abs/1709.02764, 2017. URL h p://a xi .o g/abs/1709.02764.
[27] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Del ing deep in o ec i ie s:
Su passing human-le el pe o mance on imagene classi ica ion. CoRR, abs/1502.01852,
2015. URL h p://a xi .o g/abs/1502.01852.
[28] Diede ik P. Kingma and Jimmy Ba. Adam: A me hod o s ochas ic op imiza ion. CoRR,
abs/1412.6980, 2014. URL h p://a xi .o g/abs/1412.6980.
[29] Thesis. URL h ps://gi hub.com/cla abonnin/segmen a ion_DLMI_cla a.