limited res nce it brea classifying rocessing o
ure 3.6 A Po
at one perce used for reg which is 99 mean value
d is then up t in the le
cking
olution, dis aks, the reg g those vox one. It is so
ossible brok
entage of co gion growin
9%.
of classifie pdated. It m
akage, and
stal bronchi gion growin xels with low
called leak
ken airway
onfidence in ng is the me
ed region i means the th d the small
wall are so ng procedu w intensity age or expl
wall.
nterval is th ean value m
is recalculat hreshold hav
er one wil
ometimes br ure previou
in the alve osion into l
he most su multiplied by
ated as show ve 1% toler ll result in
roken after usly describ eoli or bron lung. As a r
itable y one
wn in rance.
poor
(3-5)
some bed is
nchus result,
it is show
entire fill u reach there airwa
occu gettin
the le the a lung.
regio and q regio grow
difficult to wed in figure
There are t e image. It up the entir h lower ord e will be fe
ay.
Therefore, urred, and pr
ng the best The leakag eakage occu airway is dis
.
In this wo on with a sp quantities h on is detect wing process
o identify th e 3.6.
two ways to makes the a re airway lu er bronchi.
ewer chanc
the region revents the one before ge can be de
urred, the v stributed in
ork, leakag pecific thres how dramati ted as equa s.
hose lower
o deal with t airway wall umen, mak Another wa ces to leak
growing pr threshold b leakage.
etected by m olume incre
a small ran
e detection shold volum ically it gro ation 3-6, a
order bron
this issue. O l linked, bu king the reg
ay is to low into lung,
rocedure ha becoming to
monitoring eased drama nge, and ther
n is implem me. It tries to wn. If an ex a leakage is
nchi. A poss
One is to sm ut loses prec
gion growin wer the thres
but still n
s to make s oo high. Fal
the volume atically. The refore it is e
mented by o catch the g
xponential v s identified,
sible broken
mooth or to cession. It is ng procedur shold. With
o chance to
sure that the lling back th
e of classifi e intensity o easy to expl
monitoring growth of c volume gro , and then
n airway w
downsamp s also possib
re impossib lower thres to identify
ere is no lea the threshol
fied region.
of voxels ou lode to the e
g the segm classified vo owth of clas
stops the r wall is
le the ble to ble to shold,
distal
akage d and
Once utside entire
mented olume sified region
(3-6)
3.4
boun segm fragm
morp visua
Postpr
Any tissue ndaries and mented airw
ments.
In order t phology clo al effect and
rocessin
e in the nat d surfaces.
way that p
to reconstru osing operat
d reasonable
ng
ture world Airway is produced by
uct the bou or is applie e segmentat
is rarely to s an organ
y previous
undary of ed. After app
tion.
o be seen w with smo s procedure
airway wa plying this
with sharp o ooth surface es may ha
all and fulf operator, it
or zigzag e e likewise.
ave corners
fill the cav would get b
edges, . The s and
vities, better
Ch
ne-based tra ch points ar Though the this measur the mappin The airway the shape larity of the ndmarks.In order to ts. Forming ch points is ance within
tion which i In this wor s by Thin-P aust search,
ementation It is an imp tion is prop k, the best
duced. It tr
er 4 L
to overcom ansformation
re used to m e similarity
rement. It i ng of gray v y branch poi
of lung w e airway bifu
take the ad g the landm s the way to n landmark is going to b rk, it adopts
Plate-Spline , and stops of image re portant issu posed for th
weighting ries to find t
Lung
me the la n is used to matching the measure is gnores the p value, and ca
ints are ado will be. The furcates. The
dvantage of marks as a
o take bran sets could be minimize s the method e transform s at a pre egistration, e of the com he sake of k
of similar the best sui
Regis
ack of inte o solve the i e structure o used by de physical me ausing the r opted as land e airway is erefore, airw
the local str measureme nch points a d be consid
ed.
d Li et al. p mation [6, 2 defined am but loses ef mbination o keeping bot rity measur it weighting
strati
er-subjects inconsistent of lungs.
fault, it is n eaning of im registration
dmarks, bec the inner way branch
ructure, it h ent in beha as a cost fu
dered as a
proposed w 28]. It mini mount of it
fficient.
of both mea th advantag re and land g by trial an
ion
human lu t shape info
not sufficien mage presen
become me cause the ai structure o points have
has to take u alf of the q unction. Thu measurem
hich tries to imizes all m terations. I
asurements.
ges of each dmark dista nd error, and
ung registr ormation; ai
nt to register nted. It cons eaningless.
irway determ of lungs. It
e to be iden
use of the b quantificatio us, the Eucl ment of the
o register h measureme It is a com
A weighted measure. In ance measu d trying to f
ation, irway
r with siders
mines exist ntified
ranch on of lidian e cost
robu
st weighting
Branch
To take use airway branc In the prev y for a dire cting a voxe The proced etail.
.1 Airwa
The airway tunnel. The le point repr In order to rithm is in n
The thinnin el wide struc
ch point.
The reason essing time ning proced verable. If a
g with good
h Point
e of the bra ching must vious work ect usage. B el represent dures of col
ay Thinn
y segmented e cross secti resent the b o retrieve need [30]. It ng procedu cture, to eas
n not to app e. A postpr dure may b a branch is m
d performan
Sets
anch point s be identifie in chapter Because the as the branc llecting bran
ning
d, is a tree-l ion of the ai ranch point the medial t helps the o ure is to pro
se the diffic
ply pruning rocessing lik better help
miss-pruned nce.
set which re ed.
3, the airw e airway is ch point is a nch points a
like structur irway is a f t [31].
l line of th operator get
oduce a sk culty of retr
algorithms ke morpho than prunin d, a branch
epresent the
ay is succe a tree-like an issue nee are describe
re. There ar flat lumen, s
he airway, tting the ind eleton of th rieving a sin
after thinni logy closin ng, because
point will b
inner struc
ssfully segm air tunnels, eds to overc ed in the fo
re airway w so that it is d
a thinning dex of the br
he airway.
ngle point th
ing procedu ng or smoo e the pruni be lost.
cture of the
mented. It i , the decisi come.
ollowing sec
wall and lum difficult to f
g / skeleton ranch point
It will be a that represen
ure is to sav oth filters b ing result i
lung,
is not on of
ctions
men in find a
nizing . a one nt the
ve the before is not
contr
In this wor ributed this -review
on-.2 Branc
Branch po ge now is a erse the air S).
Mori et al.
ne thing ha e and unwan e [25].
In this wor ility.
.3 Branc
There are l the major b Twenty-thr y are locate chus, right chus 4 + 5, chus 10. On
rk, the meth implement line publica
ch Points
ints are ea a skeleton o rway tree. I
proposed a as to consid nted false b
rk, the branc
ch Point
ots of bifur branch poin ree branch p ed at trache
bronchus 1 right bronc n the other
od Homann ation to the ation in med
s Identifi
asy to ident of the airw It can be B
an automatic der, the airw branching [6
ch points ar
Sets Ma
cations in th ts are consi points with
a or carina , right bron chus 7 + 8, side, they
n et al. prop e Insight Jou
dical image
fication
tify manual way. All the
Breadth-firs
c branch po way might 60], so that
re identified
atching
he airway. I idered.
low branch , right main nchus 3, righ right bronc are located
posed is ado urnal which processing
lly. After t e identificat st search (B
oint retrievin be miss-seg Tschirren e
d manually
In order to f
hing order a n bronchus ht intermed chus 9 + 10 d at left mai
opted. Homa h is an open and visuali
he thinning ion procedu BFS) or De
ng algorithm gmented. It et al. tried t
on purpose
find commo
are selected , 1st divisio diate, right b , right bron in bronchus
ann et al. al access and ization [16]
g procedure ure has to epth-first s
m [29]. But t causes loo to overcom
e of accurac
on branch p
for recogn on of right bronchus 6, nchus 9 and s, 1st divisi
ready d open
.
e, the do is search
there
left m 6, lef
may there
4.2
regis align comp
and t comp
equa
and r
main bronch ft basal, left
In this wor be encoun e are still roo
Prereg
In order t stration is a ned, preregi putational e In this wor translations putational r There are ation 4-1.
The rigid-b rotation tran
hus, left upp t bronchus 8 rk, the airwa ntered in pr oms for refi
gistratio
to speed up a time consu istration re expense.
rk, a rigid-b . It is one o esources.
six param
body transfo nsform matr
per division 8, left bronc ay branch p revious pro ine as ment
on
p the entir uming imag educes time
body transfo of the simpl
meters to de
orm matrix i rix, where
T
n, left lingu chus 9 + 10, oint sets are ocedures. Th
tioned in pre
re process, ge processi e spent by
ormation is lest transfor
escribe the
M TR
is equal to c
1 0 0
0 1 0
0 0 1
0 0 0
ular division , left bronch e matched m hough auto evious secti
a preregis ing techniqu
the later
employed.
rmations tho
rigid-body
combining t
1 2
3
1 q q q
n, left inferi hus 9 and le manually, co omatic algo
ion.
stration is ue. If imag
registration
It consists o ose do not r
y transform
translation t
ior, left bron eft bronchus
onsidering e orithms do
adopted. I ges have rou n which is
of only rota require too
mation matr
transform m nchus s 10.
errors exist,
Image ughly high
ations much
rix as
(4-1)
matrix
(4-2)
and
R
4.3
propo rema facto adde
41 0
0 cos 0 sin
0 0
q q
Registr
The propos osed [42], ain the sam ors those ar
d because it
4 4
4 4
0 sin cos 0
q
q q
ration A
sed method the compo me as const re different
t is required
Figure 4.1 Ima
4
0 cos 0
0 sin 1
Architec
is designed nents, mult traints and from the c d by the cor
1 The Multi age is retrie
5
5
s 0
0 1
n 0
0 0
q
q
cture
d in order to ti-resolution
fixed facto competing m rresponding
i-resolution eved from th
5
5
sin 0
0 0
cos 0
0 1
q
q
o compete w n pyramid, ors. The co
method. On cost functi
registration he ITK man
660 cos 0 sin
0 0
0 q
q
with the met optimizer ost function ne of the tra
on.
n approach.
nual.
66sin cos 0 0 q q
thod Staring and interpo ns are contr
ansformatio 0 0 0 0 1 0 0 1
(4-3)
g et al.
olator rolled ons is
4.3.
that r entire
imag user 4.2.
4.3.
way
Non-landm param
.1
Multi-In this wor registering f e registratio A Gaussian ge is blurred specific py
.2 Transf
Betke et al [37]. Now -rigid transf
According mark-based meters. Onl
Figure 4.2 Ima
-Resolut
rk, a multi-r from low re on as shown n blur filter d then subs yramid level
form
l. proposed wadays, rig
formation is to the me d registratio ly source an
The multi-r age is retrie
tion
resolution a esolution to n in figure 4 r is used to
ampled to a l reached. A
a lung rigi gid registrat
s in need.
ethod Rex on [40]. T nd target lan
resolution r eved from th
approach is original res 4.1.
build the m a lower lev A multi-reso
id registrati tion is not
Cheung et PS transfo ndmarks ar
registration he ITK man
adopted. It solution for
multi-resolu el pyramid.
olution fram
ion using It t good eno
t al. propos ormation do re required.
framework nual.
t is a coarse r the sake of
ution pyram . Repeating mework is s
terative Clo ough for lu
sed, TPS i oes not req
Therefore, k.
e to fine m f speeding u
mid. The ori g this step u showed in f
osest Point ung registr
is just fitte quire too there is no
ethod up the
iginal until a figure
(ICP) ation.
ed for many need
to fin
and A
trans regis it is n in eq
4.3.
It is adap AGS
4.3.
local
ne tune any In this wor Arbel propo
A postregi sformation.
stration puts necessary to quation 4-5.
.3 Optim
In this wor a stochast tive step si SD has an ad
.4 Interp
Trilinear in l axial data o
It is a fa
parameters rk, the TPS osed [39, 47
istration is It is becau s both image
o adopt B-S
mizer
k, Adaptive tic gradient ize predicti daptive step
polator
nterpolation on lattice po ast algorith
s.
implement 7]. The trans
required se that the es on finite Spline regist
e Stochastic t descent o ion. In con p size which
n is used in oints.
hm with lo
tation is bas sformation i
∑
to fitting TPS only t grid which tration to m
∑ ∈
c Gradient D optimization ntrast to Re h helps the c
n this work.
ow comput
sed on the m is illustrated
the shape tries to mat
is denser th match the bo
Descent (AS n method f egular Step
convergent
. It approxi
tational co
method Dav d in equatio
informatio ch the land han landmar oundary. B-S
SGD) optim for image r Gradient D speed faster
imates the v
mplexity a
vis et al., B on 4-4.
on by B-S dmarks. B-S rks. As the Spline is sh
mizer is used registration Decent (RS r than ever.
value withi
against B-s rooks
(4-4)
Spline Spline result, howed
(4-5)
d [51].
with SGD),
in the
spline
interp interp but p
4.3.
contr and l
withi meas
wher
dista struc landm
weig
polation an polated vox produces bet
.5 Cost F
The cost fu ribution is t landmark di The simila in images.
sure in this w
re Ω is the The landm ance of two cture matche
marks.
The goal o ghting. A we
nd window xels. It is a tter result.
Function
unction is th to provide a istance mea arity measur
Mean of S work. The M
MSD T
lung region mark distanc o point set ed. It compu
LDM
of this thes eighted cost
ed sinc int little bit sl
n
he most im a practical asure.
re is shown Squared Di MSD is defi
; ,
n and μ is t ce is showe s with kno utes the dist
M T ; ,
sis is to me t function is
terpolation, lower than
mportant and weighting f
n in equatio ifferences ( fined as,
Ω ∑ ∈
the instance ed in equat own corresp
tance from
∑
erge two m s shown in e
and result Nearest Ne
d the core p for combini
on 4-6. It i MSD) is e
e of transfor tion 4-7. It pondence.
target landm
x T x
measuremen equation 4-8
ts enough eighbor (NN
part of this t ing both sim
s the gray mployed fo
rmation.
t is aimed t It makes t marks to tra
nts as a cos 8,
good quali N) Interpol
thesis. The milarity me
value diffe or the simi
to minimiz the local ai ansformed s
st function ity of
ation,
main easure
erence ilarity
(4-6)
ze the irway ource
(4-7)
by a
wher respe
disco avail imag be di
coeff
re C is the ectively. Ea The cost fu
C
While norm overed. The lable in DIC ges are matc istributed ne In this case ficient is im
C T
e weighted ch is arr unction can
T ; I , I
malizing th e maximum COM, and th
ched. There ear zero. Th e, cost func mplicit and ta
T ; I , I
cost functi ranged with be expande
ω Ω
ω ∑
he measure m of MSD he minimum efore, the sim he LDM is s
tions are w aken place b
∑ ∑
ion, and h a weightin ed as equatio
1
Ω ∈
x T
ements, the is equal to m is zero. A milarity is h similar to M weighted wit
by weightin
ω c T ; I
and are ng for to
on 4-9.
T x
maximum o square of After perform
high. If the MSD.
thout norma ng directly.
, I
e target and otal N equal
m and minim f twice max ming image
MSD is no
alization. Th
d source im al to two.
mum shou ximum inte e registratio ormalized, i
he normaliz (4-8)
mages
(4-9)
ld be ensity
n, the it will
zation
Ch
In this chap e time.
All of the 3 GHz proc ght Toolkit ( essing [55].
on Microsof The registr ce software to deal w ix. A modif lastix. The c 5. A bug was Images are ges, and des tal Imaging versity Hosp
loyed.
Volumes ar [9, 53, 54]
master thes