Clinical Background of Knowledgebased Models In IMRT/VMAT Planning
Jackie Wu, PhD, FAAPM
Professor
Duke University Medical Center
Department of Radiation Oncology
DUKE University Radiation Oncology
DUKE University Radiation Oncology
Disclosure
n
Research Grant: NIH/NCI
n
Master Research Grant: Varian Medical System
n
Technology License Agreement with Varian.
DUKE University Radiation Oncology
Motivation
n
Extract human expert’s knowledge
n
Model past planning experience
n
Reduce, or even automate the planning process
n
Hypothesis
n
Past knowledge and experience has always be applied to
new patient case
n
We can mathematically extract and model such application
of knowledge and experience
n
Knowledge modeling can help us improve the planning
efficiency, consistency and quality.
DUKE University Radiation Oncology
Elements of IMRT/VMAT Treatment Planning
how the dose distribution
should look like –
experience & knowledge
DUKE University Radiation Oncology
Knowledge Example: Volume vs. Dose
Bladder DVH
Bladder Volume Distribution
Rectum DVH
Rectum Volume Distribution
DUKE University Radiation Oncology
Knowledge Example: Distance vs. Dose
Bladder DVH
Rectum
PTV
Bladder
Dose vs. Distance
Dose vs. Distance
Euclidean distance
Variable distance
DUKE University Radiation Oncology
Zhu et al, Med Phys 38:719-726, 2011
Yuan et al, Med Phys 39:6868-6878, 2012
Knowledge Example: Distance-to-target Histogram (DTH)
Signed Distance-to-target
histogram (DTH)
(overlap: negative distance)
Relative geometrical
relationships between OAR
and PTV
n
Percent Volume (%)
n
Distance to PTV Surface (cm)
=
⋃{ |
∈
and ( ,
∈
}
⋃{ |
DUKE University Radiation Oncology
)≤
}
Zhu et al, Med Phys 38:719-726, 2011
Yuan et al, Med Phys 39:6868-6878, 2012
Knowledge Example: Shape vs. Dose
Spinal
Cord
Case 11
Case
Spinal Cord and PTV DVHs
Spinal
Cord
Case 2
Percent Volume (%)
PTV
Case 2
Case 1
PTV
Percent Dose (%)
DUKE University Radiation Oncology
Anatomy and Dose Features Overview
Site
Prostate
HN
OAR
Rectum
Bladder
Parotids
Oral cavity
Larynx
Pharynx
Spinal cord
Brainstem
Mandible
DUKE University Radiation Oncology
Anatomical And Dosimetric Features
Distance to target histogram (DTH): PCS
Distance to OAR (DOH): PCS
OAR volumes
PTV volume
Fraction of OAR volume overlapping
with PTV (overlap volume)
Fraction of OAR volume outside the
treatment fields (out-of-field volume)
Tightness of the geometric enclosure of
PTV surrounding OAR
Curvature of specific OAR
----------------------------PTV dose homogeneity
PTV hotspot
OAR DVHs
AAPM 2012
Bladder
> 230 cc
100 cc to 230 cc
<100 cc
Parotid
> 38 cc
22 cc to 38 cc
<22 cc
DUKE University Radiation Oncology
Rectum
> 120 cc
60 cc to 120 cc
<60 cc
Brainstem
> 28 cc
20 cc to 28 cc
<20 cc
Modeling Planning Knowledge
n
DVH/DTH Feature Extraction and Dimension
Reduction
n
Principal Component Analysis (PCA)
Mean Population DVH
Individual Patient’s DVH
DUKE University
Radiation
Percent
DoseOncology
(%)
Percent Volume (%)
Percent Volume (%)
1st Principal Component
2nd Principal Component
Individual Patient’s DVH
Deviation from Mean
Sum of First Two PCs
Percent Dose (%)
Predict Dose/DVH Based on Anatomy Features?
Multiobjective Approach To Morphology-based Radiation Treatment Planning
Boonyanit Mathayomachan , PHD Thesis 2005
Case Western Reserve University
DUKE University Radiation Oncology
Modeling Planning Knowledge
n
Correlate patient anatomy features (input) with dose distribution
features (output)
n
Correlate plan design features (input) with dose distribution
features (output)
n
Understand and translate the knowledge model parameters to
the physics/dose parameters
n
Extend and refine the Knowledge base with progressive
modeling and rapid learning technologies
DUKE University Radiation Oncology
Modeling Planning Knowledge
DistanceBased
Database of
Tx Cases
High Order
VolumeBased
Feature
Extraction
Institutional
DoseBased
Prospective
New Pt Case
Planning Guidance
A planning quality evaluation tool for
prostate adaptive IMRT based on
machine learning
DUKE University Radiation Oncology
Medical Physics 38, 719,2011
Model
Training
Predictive
Model
Retrospective
Plan Databases
Quality Analysis
Quantitative analysis of the factors which
affect the interpatient organ-at-risk dose
sparing variation in IMRT plans.
Medical Physics 39, 6868,2012
Modeling Planning Knowledge
n
Machine Learning of Treatment Planning Knowledge
n
Knowing X1, X2, ,,,,Xm, and Y1, Y2,,,,,, Yn,
Solve: Y(1,2,,,,n) = F(X1,2,,,,m)
n
Multi-regression learning
n
Support vector learning
n
Neural Network learning
n
Many other methods
DUKE University Radiation Oncology
Parotids modeling by multiple regression (Green) and SVR (Red)
Plan 1
Plan 2
Plan 3
Plan 13
Plan 14
Plan 15
100
100
100
100
100
100
50
50
50
50
50
50
0
0
50
Plan 4
100
0
0
50
Plan 5
100
0
0
50
Plan 6
100
0
0
50
Plan 16
100
0
0
50
Plan 17
100
0
0
100
100
100
100
100
100
50
50
50
50
50
50
0
0
50
Plan 7
100
0
0
50
Plan 8
100
0
0
50
Plan 9
100
0
0
50
Plan 19
100
0
0
50
Plan 20
100
0
0
100
100
100
100
100
100
50
50
50
50
50
50
0
0
50
Plan 10
100
0
0
50
Plan 11
100
0
0
50
Plan 12
100
0
0
50
Plan 22
100
0
0
50
Plan 23
100
0
0
100
100
100
100
100
100
50
50
50
50
50
50
0
0
50
100
0
0
Plan 25
50
100
0
0
Plan 26
50
100
0
0
50
100
0
0
Plan 37
Plan 27
50
100
0
0
Plan 38
100
100
100
100
100
50
50
50
50
50
50
50
Plan 28
100
0
0
50
Plan 29
100
0
0
50
Plan 30
100
0
0
50
Plan 40
100
0
0
50
Plan 41
100
0
0
100
100
100
100
100
100
50
50
50
50
50
50
0
0
50
Plan 31
100
0
0
50
Plan 32
100
0
0
50
Plan 33
100
0
0
50
Plan 43
100
0
0
50
Plan 44
100
0
0
100
100
100
100
100
100
50
50
50
50
50
50
0
0
50
Plan 34
100
0
0
50
Plan 35
100
0
0
50
Plan 36
100
0
0
50
Plan 46
100
0
0
50
Plan 47
100
0
0
100
100
100
100
100
100
50
50
50
50
50
50
0
0
0
0 Oncology
DUKE
University
50
100
0
50 Radiation
100
0
50
100
0
0
50
100
0
0
50
100
50
Plan 21
100
50
Plan 24
100
50
100
Plan 39
100
0
0
50
Plan 18
100
0
0
50
Plan 42
100
50
Plan 45
100
50
Plan 48
100
50
100
Modeling Planning Knowledge
n
Plan Standardization
DUKE University Radiation Oncology
Complex OAR Sparing Knowledge Modeling
Actual Parotid DVH
Ability to model DVH
Dose (%)
crossover due to
complex OAR/PTV
shape variation
among different
patients Dose (%)
Dose (%)
Dose (%)
DUKE University Radiation Oncology
Modeled Parotid DVH
Example of Bladder DVH Modeling
Prostate
Cases
DUKE University Radiation Oncology
Example of Bladder DVH Modeling
Anal-rectal
Cases
DUKE University Radiation Oncology
Knowledge Model
Training
Data
base
of
Data
base
of
Database
of
Data
base
of
High
quality
High
Highquality
Quality
High
quality
treatment
plan
treatment
plan
Treatment
treatment
plan
Cases
Anatomy
Features
Dose/Plan
Features
PCA
PCA
Descriptor
variables
input : Xi
Response
variables
output : Yi
Model training:
Machine
Learning
Y= F(X)
Knowledge Model Application
New
Patient’s
Anatomy
PCA
Characteriz
ed by Xnew
DUKE University Radiation Oncology
Y=f(X)
PCA
-1
Ynew
DVH of
OARs
Zhu et al, Med Phys 38:719-726, 2011
Yuan et al, Med Phys 39:6868-6878, 2012
From Models To Planning
DUKE University Radiation Oncology
From Models To Planning
DUKE University Radiation Oncology
Cross-Institution Knowledge Base
n
Cross-institution Knowledge
n
If we believe best planning knowledge is shared among all planners
n
LUNG IMRT Pilot Study By RTOG/NRG
n
71 Cases
n
3 Institutions
DUKE University Radiation Oncology
Cross-Institution Knowledge Base
Prescriptions (Gy)
Mean
67
Median
64
Min
40
Max
74
Institution 1 Institution 2
Institution 3
Volume (cm3)
mean
median
min, max
421
343
62, 1132
595
519
76, 1132
512
379
175, 1161
Location (side)
Total
Left/Left-Medial
Right/Right-Medial
Medial
45
18
21
6
10
4
6
0
16
5
10
1
DUKE University Radiation Oncology
Cross-Institution Knowledge Base
Circle: Training Plan
Plan
Cross: Validation
Not contributing to
model
DUKE University Radiation Oncology
Cross-Institution Knowledge Base
Institution 2 Not contributing to model
Contralateral
Lung DVH
_
_
Model
Plan
Institution 3 Contributing to model
DUKE University Radiation Oncology
Cross-Modality Knowledge Base
n
Cross-Modality Knowledge
n
If you believe best planning knowledge is independent of treatment
modality
Institution A
Institution B
•
•
•
•
•
•
•
•
7-8 min delivery time
Delivery system: Varian IMRT
Planning system: Eclipse
Sequential Boost
–
–
•
–
Multiple plans (one plan for 1 PTV)
40-50 Gy and 60-70 Gy
~60 head-and-neck cases
DUKE University Radiation Oncology
7-8 min delivery time
Delivery system: Tomotherapy
Planning system: Tomotherapy
SIB
–
•
1 plan (one plan cover all PTVs with
diff. daily doses)
54.25 Gy and 70 Gy
~60 head-and-neck cases
Lian et al, Med Phys 40:121704, 2013
Cross-Institution Knowledge Base
Parotid DVH
Tomotherapy Model
vs. IMRT Model
vs. Actual Plan DVH
DUKE University Radiation Oncology
Knowledge Model Improves Plan Quality
DUKE University Radiation Oncology
Knowledge Model Improves Plan Quality
Original Clinical Plan
DUKE University Radiation Oncology
Knowledge-Model Guided Plan
Summary: What Knowledge Modeling
May Help
Physician
Prescription
Planned vs. Modeled
IMRT/VMAT
Planning
Plan
Evaluation
DUKE University Radiation Oncology
Other Knowledge Models: Dose Models
n
In Spine SBRT, dose
distributions in cord
are highly correlated
with tumor contour
shapes
Contours
Dose Dist.
DUKE University Radiation Oncology
Liu et al, PMB 60:N83-N92, 2015
Other Knowledge Models: Dose Models
n
Compute correlation between tumor contour shapes and cord
dose distributions
n
Use learned correlations to predict voxel-level dose distributions
Correlation
Contour space
DUKE University Radiation Oncology
Dose space
Liu et al, PMB 60:N83-N92, 2015
Other Knowledge Models: Dose Models
n
Active Shape Model
n
Align the reference tumor contours and all other contours
using the iterative closest point (ICP) algorithm
DUKE University Radiation Oncology
Liu et al, PMB 60:N83-N92, 2015
Other Knowledge Models: Dose Models
n
Active shape models
n
PCA analysis of a set of aligned tumor contours
-3λ1
Mean Shape
3λ1
DUKE University Radiation Oncology
Liu et al, PMB 60:N83-N92, 2015
Other Knowledge Models: Dose Models
n
Optical Flow Dose Distribution Model
n
measures dose variance between a reference image and any other
images within the training dataset
Reference
Image
DUKE University Radiation Oncology
Optical Flow
Registration
Other Knowledge Models: Dose Models
n
Active optical flow dose distribution model
n
PCA analysis of a sequence of optical flow fields
-3λ1
Mean Dose
3λ1
DUKE University Radiation Oncology
Liu et al, PMB 60:N83-N92, 2015
Other Knowledge Models: Dose Models
n
Machine Learning
n
PTV contour space
cord dose space
DUKE University Radiation Oncology
Liu et al, PMB 60:N83-N92, 2015
Other Knowledge Models: Dose Models
DVH: overall cord dose prediction accuracy
Portion of Spinal
Cord volume (%)
n
Relative Dose (%)
DUKE University Radiation Oncology
Relative Dose (%)
Summary
n
Modeling clinic treatment planning knowledge is feasible
n
Various sources of knowledge can be combined
n
Multi-center, multi-modality knowledge modeling will help clinical
practice in large and small centers and clinical trials
n
Knowledge models can assist physicians, physicists and
planners
n
Knowledge modeling can help to improve plan quality,
consistency, as well as efficiency
DUKE University Radiation Oncology
Acknowledgement
n
John Kirkpatrick, MD/PHD (Duke
University)
n
Brian Czito, MD (Duke University)
n
W. Robert Lee, MD (Duke University)
n
Bridget Koontz, MD (Duke University)
n
David Yoo, MD, PHD (Duke University)
n
David Brizel, MD (Duke University)
n
Chris Kelsey, MD (Duke University)
n
Mark Dewhirst, DVM(Duke University)
n
Radhe Mohan PHD (MD Anderson)
n
Zhongxing Liao MD (MD Anderson)
n
Jaques B. Bluett (MD Anderson)
n
Xiaodong Zhang PHD (MD Anderson)
n
Michael Gillin PHD (MD Anderson)
DUKE University Radiation Oncology
n
Jun Lian, PHD (UNC)
n
Sha Chang (UNC)
n
Bhishamjit S. Chera, MD (UNC)
n
Larry Marks, MD (UNC)
n
Wilko Verbakel, PHD (Vu University)
n
Jim Toll, PHD (Vu University)
n
James Deye PHD (NCI)
n
James Galvin PHD (RTOG)
n
Ying Xiao PHD (RTOG)
n
Kevin Moore PHD (UCSD)
n
Charles Simone MD (U Penn)
n
Liyong Lin PHD (U Penn)
n
Jeffery D. Bradley MD (Wash U)
Thank You
DUKE University Radiation Oncology
Treatment planning knowledge models are
20%
1.
Confined to a single institution
20%
2.
Applicable to multiple modalities
20%
3.
Useful for only IMRT
20%
4.
Physician Specific
20%
5.
Useable only with Monte Carlo-based dose
calculation algorithms26
DUKE University Radiation Oncology
10
Treatment planning knowledge models are
0%
1.
Confined to a single institution
0%
2.
Applicable to multiple modalities
0%
3.
Useful for only IMRT
0%
4.
Physician Specific
0%
5.
Useable only with Monte Carlo-based dose
calculation algorithms
DUKE University Radiation Oncology
Treatment planning knowledge models are
n
Answer:
n
2
n
Reference:
Lian et al, Modeling the dosimetry of organ-at-risk in head
and neck IMRT planning: An inter-technique and inter-institutional
study, Medical Physics 2013, 40(12)
DUKE University Radiation Oncology
Machine learning of the knowledge models
is useful
20%
1.
In quantifying the influence of anatomy features to the
dose sparing in the OARs
2.
In defining linac performance
3.
In predicting dose prescription
4.
In collecting past cases as database
5.
In detecting planning errors
20%
20%
20%
20%
DUKE University Radiation Oncology
10
Machine learning of the knowledge models
is useful
0%
1.
In quantifying the influence of anatomy features
to the dose sparing in the OARs
0%
2.
In defining linac performance
0%
3.
In predicting dose prescription
0%
4.
In collecting past cases as database
0%
5.
In detecting planning errors
DUKE University Radiation Oncology
Machine learning of the knowledge models
is useful
n
Answer:
n
1
n
Reference:
Yuan et al, Quantitative analysis of the factors which affect
the inter-patient organ-at-risk dose sparing variation in IMRT plans,
Medical Physics 2012, 39(11)
DUKE University Radiation Oncology
The organ sparing capability predicted by
the knowledge model is
20%
1.
The average value of the sparing in the database
20%
2.
Interpolated among a few similar cases
20%
3.
Independent of prescription dose
20%
4.
Only valid for maximum dose
20%
5.
Patient specific, based his/her anatomy and
physician’s prescription
DUKE University Radiation Oncology
10
The organ sparing capability predicted by
the knowledge model is
0%
1.
The average value of the sparing in the database
0%
2.
Interpolated among a few similar cases
0%
3.
Independent of prescription dose
0%
4.
Only valid for maximum dose
0%
5.
Patient specific, based his/her anatomy and
physician’s prescription
DUKE University Radiation Oncology
The organ sparing capability predicted by
the knowledge model is
n
Answer:
n
5
n
Reference:
Yuan et al, Quantitative analysis of the factors which affect
the inter-patient organ-at-risk dose sparing variation in IMRT plans,
Medical Physics 2012, 39(11)
DUKE University Radiation Oncology
Clinical Background of Knowledgebased Models In IMRT/VMAT Planning
Jackie Wu, PhD, FAAPM
Professor
Duke University Medical Center
Department of Radiation Oncology
DUKE University Radiation Oncology