Acknowledgements
Auto-segmentation using
Biomechanical Models
Kristy K Brock, Ph.D.
Physicist, Radiation Medicine Program, Princess Margaret Hospital
Assistant Professor, Depts of Radiation Oncology & Medical Biophysics,
University of Toronto
Scientist, Ontario Cancer Institute
Biomechanical Models for
Deformable Registration
Clinicians
Collaboration
MORFEUS Team
• Laura Dawson
• David Jaffray
• Adil Al-Mayah
• Charles Catton
• Michael Sharpe
• Deidre McGrath
• Cynthia Ménard
• Doug Moseley
• Michael Milosevic
• Joanne Moseley
• Anthony Fyles
• Stephen Breen
• Andrea McNiven
• John Waldron
• Jeff Siewerdsen
• Carolyn Niu
• Karen Lim
• James Stewart
• Thao-Nguyen Nguyen
• Young-Bin Cho
• Michael Velec
• Shannon Hunter
Funding
• Roxana Vlad
• NIH R01, NCI Canada – Terry Fox Foundation
• Lin Xu
• Ontario Institute for Cancer Research, Cancer Care
• Jenny Lee
Ontario Research Chair
• Lily Chau
• Elekta Oncology Systems, Philips Medical
Systems, RaySearch Laboratories
MORFEUS for Auto-Segmentation
Morfeus
Exhale Image
Inhale Image
Surface
correlation of
organ subset
• Include biomechanical
properties of tissues
• Enable complex
surface interfaces
• Little dependence on
image intensity
• Proven accuracy
Surface Mesh
Volumetric Mesh
Material Properties
Finite
Element
Analysis
Boundary Conditions
• Enables propagation
of contours using a
smaller subset of
contours
• Accuracy of
propagation of
structures
independent of
image contrast
1
Warning! What are you propagating?
Prior to Deformable Registration
coronal
Use well defined contours to
auto-segment uncertain
structures
sagittal
Before
After
Deformable Registration
GTV Volume
CT = 13.9 cc
MR = 6.7 cc
∆Vol = 7.2 cc
(52%)
Prostate Therapy
Accurate Target Definition
Combining MR-ERC and CT in Prostate Cancer
Eligibility
and
consent
TRUS
+
Fiducial
marker
insertion
ERC-MRI
100cc
0cc
No ERC
Volume
delineation
in Pinnacle
Target
Plan
Morfeus
Treat
Lesion Boost
100cc ERC
Primary Image
Deformed
Image
kV CBCT
Actual No
coil Image
2
Cervix Therapy
Target
Cervix Therapy
Dose:
20
28
38
0 Gy
8
Gy
Treat 48
Plan
H&N Applications
• Combining intensity
based autosegmentation with
biomechanical
models
• Using single point
identification – autosegmentation of
VBs
Target
Plan
Treat
H&N Applications
• Combining intensity
based autosegmentation with
biomechanical
models
• Using single point
identification – autosegmentation of
VBs
D Létourneau, et. al. Med Phys 35(1) 2008
3
H&N Applications
Precision Image-Guidance
Resolve Geometric discrepancies
Surface Accuracy: abs mean
< 2mm (parotids)
< 1mm (tumor)
• Combining intensity
based autosegmentation with
biomechanical
models
• Using single point
identification – autosegmentation of
VBs
New Tumor
Position!
Planning CT [w contrast]
CBCT [w/o contrast]
Accurate Tumor Guidance
12 Liver Patients: 6 Fx Each
Rigid Reg → Deformable Reg
∆ Tumor
AVG
SD
Max
Min
Median
dLR
-0.04
0.10
0.27
-0.34
-0.03
dAP
-0.01
0.15
0.43
-0.65
0.01
dSI
0.01
0.20
0.97
-0.70
0.00
abs(dLR) abs(dAP) abs(dSI)
0.08
0.10
0.10
0.07
0.11
0.17
0.34
0.65
0.97
0.00
0.00
0.00
0.05
0.06
0.04
Use models and limited image
information to auto-segment
structures
• 33% (4/12) Patients had at least 1 Fx with a
∆COM of > 3 mm in one direction
• 15% of Fx had a ∆COM of > 3 mm in 1 dir.
4
Generate Population Liver Model
Generate Population Liver Model
(a)
Patient Exhale
Population
Model
TN Nguyen, Brock Med Phys. 2009 36(4)
Generate Population Liver Model
Generate Population Liver Model
(a)
Patient Exhale
(c)
Patient Exhale
Population
Model
Patient Inhale
(b)
5
Generate Population Liver Model
(c)
Patient Exhale
Patient Inhale
(d)
Generate Population Liver Model
Displacement [cm]
1.90
1.81
1.73
1.65
1.56
1.48
1.40
1.31
1.23
1.15
Y
X
Z
Absolute average ± SD of the population respiratory motion:
Superior / Inferior (z) = 1.2 ± 0.2 cm
Anterior / Posterior (y) = 0.8 ± 0.1 cm
Left / Right (x) = 0.1 ± 0.1 cm
TN Nguyen, Brock Med Phys. 2009 36(4)
1000
500
0
0
2
4
z [cm]
Exhale
6
1500
Point Motion Detection Method
Inhale Image
Intensity
Exhale Image
1500
Intensity
Intensity
Point Motion Detection Method
1000
500
00
2
4
z [cm]
6
1500
∂NC1 = 1.20 cm
∂NC1
Exhale
Inhale
1000
500
00
2
4
z [cm]
6
= motion at
Navigator
Channel 1
Inhale
6
Results: Motion Detection Accuracy
∂NC =
1
1.20 cm
1.27 [cm]
Adaptation
Node number: 457
Point Motion
Detection
Displacement [cm]
1.86
1.72
1.58
1.43
1.29
1.15
0. 86
0.58
0.30
0.15
Y
Population
Motion Model
?
X
Displacement [cm]
1.9
1.8
1.7
1.6
1.5
1.4
1.4
1.3
1.2
1.1
Patient Motion
Y
Displacement [cm]
1.8
1.7
1.5
1.4
1.3
1.1
0.8
0.5
0.3
0.0
X
Y
Z
Z
Mean [cm]
SD [cm]
Max [cm]
Min [cm]
Superior Dome
0.10
0.09
0.31
0.01
Inferior Tip
0.11
0.04
0.20
0.04
Anterior
0.11
0.07
0.21
0.01
Posterior
0.10
0.07
0.26
0.00
k=0
k=1
Relative Nodal
Position
k=1
k=
(xo , y o , z o )edge − ( xo , y o , z o )
(x o , y o , z o )edge − (xo , y o , z o )edge
1
1
k=1
N = # of Patients = 13
Relative Nodal
Position
k=0
Population
Motion Model
k=1
k=1
k=1
∂ ( x o , y o , z o ) = ∆∆((xxoo,, yyoo,, zzoo ))
2
TN Nguyen, Brock Med Phys. 2009 36(4)
Accuracy Metrics
Adaptation
Point Motion
Detection
X
Z
1) MORFEUS
=
Patient Motion
Displacement [cm]
Displacement [cm]
1.9
1.8
1.7
1.6
1.5
1.4
1.4
1.3
1.2
1.1
1.8
1.7
1.5
1.4
1.3
1.1
0.8
0.5
0.3
0.0
2) Tumour COM
Y
Z
∂ NC 1
∆ NC 1
(1 − k ) +
∂ NC 2
∆ NC 2
(k )
X
Exhale
Image
Inhale
Image
3) Blood vessel Bifurcation
Exhale
Image
Inhale
Image
7
Results: Adapted Motion Accuracy
Results: Motion Detection Accuracy
N = # of Patients, n = # Samples
Displacement [cm]
Displacement [cm]
1.86
1.72
1.58
1.43
1.29
1.15
0.86
0.58
0.30
0.15
Y
1.86
1.72
1.58
1.43
1.29
1.15
0.86
0.58
0.30
0.15
X
=
Y
Z
Z
Error
Bifurcations (N=3, n=16)
Patient Full 4D - Navigator-refined
Deformable
Patient Motion
Motion Model
Model
Patient Planning Model → Tx Model
Displacement [cm]
1.9
1.8
1.7
1.6
1.5
1.4
1.4
1.3
1.2
1.1
Y
Z
Displacement
[cm]
2.1
1.9
1.6
1.4
1.2
0.9
0.7
0.5
0.2
0.0
X
• Population Model
• Patient specific
planning model
• Propagate contours
onto Tx (4D CBCT)
images
X
SI
AP
LR
Vector
Mean [cm]
Mean [cm]
Mean [cm]
Mean [cm]
0.26
0.13
0.15
0.33
Tumour COM (N=7, n=14)
MORFEUS tumour COM
0.19
0.15
0.20
0.33
Image tumour COM
0.14
0.13
0.21
0.29
0.25
0.24
0.23
0.43
MORFEUS (N=13)
Liver & Tumor Segmentation: Tx
Planning CT exhale
4D CBCT exhale
2.1
1.9
1.6
1.4
1.2
0.9
0.7
0.5
8
Liver & Tumor Segmentation: Tx
1000
1000
800
800
600
400
200
4D CBCT exhale
600
4D CBCT inhale
400
200
0
0
1
2
3
4
0
0
5
1
2
SI distance across NC [cm]
3
4
5
SI distance across NC [cm]
Navigator Channel Shift = 0.12 [cm]
Planning
Tx
1200
primary
secondary
1000
Intensity [HU]
Intra-Fx Motion
Fx2 CBCT Exhale Intensity Plot
1200
Intensity [HU]
Intensity [HU]
CT Exhale Intensity Plot
1200
800
600
400
200
0
0
1
2
3
4
5
SI distance across NC [cm]
Displacement [cm]
2.1
1.9
1.6
1.4
1.2
0.9
0.7
0.5
0.2
0.0
Intra-Fx Motion
Fx2 CBCT Exhale Intensity Plot
Fx2 CBCT Inhale Intensity Plot
1000
900
800
800
In ten sity [H U ]
900
700
600
500
300
0
Intra-Fx Tumor
Segmentation for
Motion Assessment
700
600
500
400
400
0.5
1
1.5
2
2.5
3
3.5
4
4.5
300
0
5
0.5
1
1.5
SI distance across NC [cm]
2
2.5
3
3.5
4
4.5
2.1[cm]
Respiration motion
5
SI distance across NC [cm]
Respiration motion [cm]
1.9
2.1
2
1.9
Navigator Channel Point Motion = 0.66 [cm]
1.6
1000
1.6
primary
secondary
900
1.4
1.5
1.4
800
Inten sity [H U ]
In ten sity [H U ]
1000
1.2
1.2
700
600
0.9
0.9
500
0.7
0.7
0.5
0.5
0.5
400
300
0
1
0.5
1
1.5
2
2.5
3
SI distance across NC [cm]
3.5
4
4.5
5
Liver
Tumour
9
Accuracy of Adaptation
Summary
Absolute accuracy of Inter-Fx motion [cm]
Mean
SD
LR
0.15
0.11
AP
0.16
0.10
SI
0.15
0.09
Vector
0.14
0.10
Absolute accuracy of Intra-Fx motion [cm]
LR
AP
SI
Mean
0.14
0.15
0.18
Vector
0.17
SD
0.10
0.08
0.09
0.10
• Must be careful what contours we propagate
• Biomechanical models can assist in accurate
propagation of structures lacking significant
intensity variation
• Combination of intensity-based segmentation
of more easily discernable structures and
biomechanical models can improve
segmentation
• Use of limited intensity information and
population (or patient specific models) can
enable automated biomechanical-model
based segmentation
10