atlas&basedundersegmentaon* - people · 2014. 9. 21. · zix ⇠ cat(⇡) ⇡ ⇠ gem(↵)...
TRANSCRIPT
WM GM HC CA PU PA AM AC VE0
20
40
60
80
***
***
***
**
***
***
******
***
Foreground−Background
Ove
r− a
nd u
nder−
segm
enta
tion (
%)
Atlas-‐Based Under-‐Segmenta1on Chris&an Wachinger1,2, Polina Golland1
1Computer Science and Ar&ficial Intelligence Lab (CSAIL), MIT 2MassachuseDs General Hospital, Harvard Medical School
Acknowledgement: We thank Jason Chang and Greg Sharp. This work was supported in part by the Na&onal Alliance for Medical Image Compu&ng (U54-‐EB005149) and the NeuroImaging Analysis Center (P41-‐EB015902).
Introduc1on • Atlas-‐based segmenta&on is commonly used in medical image analysis • Tendency of atlas-‐based segmenta&on to under-‐segment structures • Dice overlap and Hausdorff distance do not measure under-‐segmenta&on • Wang, Yushkevich (2012) proposed deconvolu&on of label maps • Contribu&ons:
• Volume overlap measures to quan&fy under-‐segmenta&on • Hypothesis: asymmetry in single organ segmenta1on as cause • Genera&ve model of background to correct under-‐segmenta&on
Conclusions • Significant bias in atlas-‐based segmenta&on to under-‐segmenta&on • Asymmetry in foreground-‐background segmenta&on as new hypothesis • Genera&ve model to par&&on background reduces under-‐segmenta&on
Under-‐Segmenta1on Measures to quan&fy over-‐ and under-‐segmenta&on:
Results • Datasets: • 16 heart MRA images (0.6x0.6x1.5mm) • 18 head and neck CT scans (0.9x0.9x2.5mm) • 39 brain MRI with 1mm isotropic resolu&on
• Comparison to deconvolu&on (Wang, 2012) • Details: • DP-‐GMM, GMM, DP-‐means, k-‐means • Global and local approach • Patch size: 3 x 3 x 3
WM GM HC CA PU PA AM AC VE0
20
40
60
80
*****
Multi−Organ
Ove
r− a
nd u
nder−
segm
enta
tion (
%)
O(S, S) =|S \ S||S|
U(S, S) =|S \ S||S|
↵ ⇡
�1
N
Pix
zix
µk,⌃k
WM GM HC CA PU PA AM AC VE0
20
40
60
80
***
***
**
***
***
Comparison of over− and under−segmentation, Latent Labels
Ove
r− a
nd
un
de
r−se
gm
en
tatio
n (
%)
WM GM HC CA PU PA AM AC VE
−0.02
0
0.02
0.04
****
**
***
** *
** **
*
Dic
e d
iffe
ren
ce
Improvement over foregr−backgr for multi−organ and latent label
IW Deconv g−kmeans g−DPmeans g−GMM g−DPGMM l−kmeans l−DPmeans
0.75
0.8
0.85
0.9
Dic
e
Left Parotid Gland
IW Deconv g−kmeans g−DPmeans g−GMM g−DPGMM l−kmeans l−DPmeans
0.6
0.65
0.7
0.75
0.8
0.85
Dic
e
Right Parotid Gland
IW Deconv g−kmeans g−DPmeans g−GMM g−DPGMM l−kmeans l−DPmeans
0.88
0.9
0.92
0.94
0.96
Dic
e
Left atrium
• Quan&ta&ve segmenta&on results for three applica&ons • Under-‐segmenta&on errors significantly higher than
over-‐segmenta&on errors White maDer (WM), Gray maDer (GM), Hippocampus (HC), Caudate (CA), Putamen (PU), Pallidum (PA), Amygdala (AM), Accumbens (AC), Ventricles (VE)
First column: over-‐segmenta&on Second column: under-‐segmenta&on
Hypothesis for under-‐segmenta1on • Under-‐segmenta&on is caused by asymmetry
in foreground-‐background segmenta&on • Merging all surrounding labels into background creates
a new meta-‐label that dominates the vo&ng process • Mul&-‐organ segmenta&on of brain supports hypothesis
Latent mul1-‐label model of the background • Genera&ve model for the unsupervised separa&on of the
background in K components and simultaneous es&ma&on of K • Dirichlet process Gaussian mixture model (DP-‐GMM) on patches:
• Replace background label with mixture component: zix
2 {1, . . . ,K}S
i
(x) = z
ix
P = {Pix
: x 2 �, Si
(x) = b}
Si(x) = b
Red: manual, Yellow: automa&c, White: voxel of interest
Mul1-‐atlas segmenta1on
Test Image
Registra&on
Training Data
Images
Segmenta&ons
Propaga&on
Test Segmenta&on
?
I = {I1, . . . , In}
S = {S1, . . . , Sn}
l 2 {1, . . . , ⌘}L
l(x) =nX
i=1
p(S(x) = l|Si) · p(I(x)|Ii)
p(S(x) = l|Si) =
⇢1 if Si(�i(x)) = l,
0 otherwise.
p(I(x)|Ii) / exp
��(I(x)� Ii(�i(x)))
2/2�
2�
ˆ
S(x) = argmax
lL
l(x)
L
f (x) > L
b(x)
S
Label maps specify likelihood of each label:
Label likelihood term:
Intensity likelihood term:
Most likely label yields segmenta&on:
Foreground-‐Background segmenta&on:
Red: manual, Yellow: automa&c
Segmenta&on of paro&d glands in CT images
Organ
Background
Merging several structures to one background label
Left Parotid Right Parotid0
10
20
30
40
50
*
**
Ove
r− a
nd u
nder−
segm
enta
tion (
%)
Left Atrium0
10
20
30
40
***
Ove
r− a
nd u
nder−
segm
enta
tion (
%)
Paro&d glands in head & neck CT Lef atrium in heart MRA Nine brain structures in MRI
Foreground-‐background segmenta&on
Mul&-‐organ brain segmenta&on Illustra&on of mul&-‐organ and foreground-‐background segmenta&on
0 1 2 3 4 50.82
0.83
0.84
0.85
Dic
e
Merged Structures
Segmentation of Amygdala
Backgrd AM0
5
10
15
20
25
Vote
s
Frgrd−Bckgrd Voting
Lef Atrium
Left Atrium
0
5
10
15
20
25
30
Ove
r− a
nd u
nder−
segm
enta
tion (
%)
Left Parotid Right Parotid
0
10
20
30
40
50
Ove
r− a
nd u
nder−
segm
enta
tion (
%)
Brain
Paro&d Glands
U
O
Pix
|zix
⇠ N (µz
ix
⌃z
ix
)
zix
⇠ Cat(⇡)
⇡ ⇠ GEM(↵)
(µk
,⌃k
) ⇠ NW(�)
S: Manual ˆS: Automatic segmentation
S
S
WM GM HC AM Other0
5
10
15
20
25
Vo
tes
Multi−Organ Voting