Building Knowledge Models In RT
Disclosure
Learning Objectives









Research Grant: NIH/NCI
Master Research Grant: Varian Medical Systems
License Agreement: Varian Medical Systems
Speaker Agreement: Varian Medical Systems
Extract Human Planning Knowledge Into Features
Machine Learning And Modeling Techniques
Understand Model Parameters and Physics Parameters
Knowledge Base Refinement and Evolution
Knowledge Model Guided Treatment Planning
Q. Jackie Wu
Department of Radiation Oncology
Duke University Medical Center
1
Building Planning Knowledge Models
Building Planning Knowledge Models
Building Planning Knowledge Models
 Treatment Planning Knowledge
 Experience and Intuition
 Modeling Experience and Intuition
 “Cheat sheet”
 “Tricks” and “Tips”
 “Templates”
 Volume, location, shape, overlap, etc
 Qualitative vs. quantitative vs. discreptive
 Summarized collection vs. individual case memory
Rectum Volume
Rectum DVH
 Correlate patient anatomy features (input) and PTV/OAR
features with dose distribution features (output)
 Correlate plan design features (input) with dose
distribution features (output)
 Understand and translate the knowledge model
parameters to the physics/dose parameters
 Knowledge base evolution and progressive modeling
2
Data
base
ofofHigh
Data
base
Building Knowledge Models
Data
base
ofHigh
High
Database
of
Expert
quality
treatment
quality
treatment
 Knowledge
Platform Design
Base
Patient & Plan
Training
Design Features
quality
treatment
Knowledge
&
plan
plan
plan Cases
Treatment
Building Knowledge Models
Building Knowledge Models
 Platform Design
 Feature Identification and Selection
Dose Parameter
Features
Output Features
Yi
Input Features
Xi
DistanceBased
Machine Learning
Knowledge Base
Y1= F1(X)
Y2=F2(X)
High Order
VolumeBased
Knowledge Base Application
New Patient
Feature
Characterization
Xnew
Dose Parameter
Features
Ynew
New Patient
Feature
Characterization
Xnew
Feature
Extraction
Institutional
Knowledge Base Application
Y=f(X)
Database of
Expert
Knowledge &
Treatment
Cases
Y=f(X)
Dose Parameter
Features
Ynew
DoseBased
3
Building Knowledge Models
Building Knowledge Models
Building Planning Knowledge Models
 Data Size
 Data Size
 Feature Extraction and Dimension Reduction
#2
#3
#4
Parotid
> 38 cc
22 cc to 38 cc
<22 cc
 Data Quality
 Reliable feature extraction
 Large amount data -> features vs. noises
 Principal Component Analysis (PCA)
Parotid
> 38 cc
22 cc to 38 cc
<22 cc
1st Principal Component
Mean Population DVH
Individual Patient’s DVH
Percent Dose (%)
Percent Volume (%)
#1
 Adequate number of cases to represent the clinical case
spectrum
Percent Volume (%)
 Adequate number of cases to represent the clinical case
spectrum
2nd Principal Component
Individual Patient’s DVH
Deviation from Mean
Sum of First Two PCs
Percent Dose (%)
4
Building Planning Knowledge Models
Building Knowledge Models
 Systematic Modeling of Knowledge
 Plan Standardization
Understand The Knowledge Models
Model Parameter
Bladder DVH PCS1 (Median Dose)
 Machine learning, Descriptive statistics, Pattern
classification
Database of Expert
Knowledge &
Treatment Cases
Multiregression
Support
Vector
Regression
Feature
Extraction
Neural
Network
Model
Training
Descriptive
statistics
Significant Factors
R2
Bladder DTH PCS1
0.81
(Median Distance)
2nd Order of Bladder DTH
0.22
PCS1
Combined
0.88
Bladder DVH PCS2 (DVH Slope)
Physics Parameter
Rectum DVH PCS1 (Median Dose)
Significant Factors
Rectum DTH PCS1
(Median Distance)
Volume of Rectum
Overlap Volume
Combined
Rectum DVH PCS2 (DVH Slope)
R2
0.59
0.12
0.08
0.68
Significant Factors
R2
Significant Factors
R2
Out-of-field Volume
0.50
Rectum DTH PCS2 (gradient)
0.32
Overlap Volume
Bladder DTH PCS2
(gradient)
0.33
Out-of-field Volume
0.32
0.30
Overlap Volume
0.12
Combined
0.85
Rectum DTH PCS3
0.12
Combined
0.69
5
Understand The Knowledge Models
Cross-Institution Knowledge Models
Cross-Institution Knowledge Models
Circle: Training Plan
 Cross-institution Knowledge
 If you believe best planning knowledge is shared among
all planners
Prescriptions (Gy)
 LUNG IMRT Pilot Study By RTOG/NRG
Volume (cm3)
Location (side)
• 71 Cases
• 3 Institutions
Mean
67
Median
64
Min
40
Max
74
mean
median
min, max
Institution 1
421
343
62, 1132
Institution 2
595
519
76, 1132
Institution 3
512
379
175, 1161
Total
Left/Left-Medial
Right/Right-Medial
Medial
45
18
21
6
10
4
6
0
16
5
10
1
Cross: Validation Plan
Not contributing to
model
6
Cross-Institution Knowledge Models
Institution 2
Not contributing to model
Cross-Modality Knowledge Models
Parotid DVH
 Cross-Modality Knowledge
Contralateral
Lung DVH
_
_
 If you believe best planning knowledge is independent of
treatment modality
Model
Institution A
Institution B
Plan
•
•
•
•
7-8 min delivery time
Delivery system: Varian IMRT
Planning system: Eclipse
Sequential Boost
•
•
•
•
•
~60 head-and-neck cases
Institution 3
Contributing to model
Cross-Institution Knowledge Models
–
–
7-8 min delivery time
Delivery system: Tomotherapy
Planning system: Tomotherapy
SIB
–
Multiple plans (one plan for 1 PTV)
40-50 Gy and 60-70 Gy
–
•
Tomotherapy Model
vs. IMRT Model
vs. Actual Plan DVH
1 plan (one plan cover all PTVs with
diff. daily doses)
54.25 Gy and 70 Gy
~60 head-and-neck cases
7
Knowledge Model Guided TomoTherapy
Planning
Knowledge Model Guided TomoTherapy
Planning
Original Clinical Plan
Knowledge-Model Guided Plan
Knowledge Model Refinement
 Trade-off Modeling
 Clinical tradeoff among different OARs or between OAR
and PTV is common
 Pareto-front planning (T. Borfeld@MGH, etc)
 Different models may be needed to represent trade-off
options
8
Knowledge Model Refinement
HN Patient and dose
prescription
Y
Meet Tradeoff
Criteria?
Predict parotid DVHs
by “standard model”
R
 Parotid Tradeoff Model
Data
base ofofHigh
 Spine
Shape
Pattern As New Features
Data
Database
base ofHigh
High
SBRT
Tradeoff Trigger threshold
Which side has
lower D50?
Predict R parotid DVH
by “combined model”
Knowledge Model Refinement
Tradeoff Criteria
Predict parotid D50
by “standard model”
N
Knowledge Model Refinement
L
ROC Analysis
TPR = 0.82,
FPR = 0.19
Tradeoff
Model
quality
treatment
quality
treatment
Spine
SBRT
quality
treatment
plan
plan
Treatment
plan Cases
PTV Contours
Dose
Distribution
Active shape
model Xi
Active optical
flow model Yi
Machine
Learning
Y= F(X)
Predict L parotid DVH
by “combined model”
9
Knowledge Model Refinement
Knowledge Model Refinement and Evolution
Knowledge Model Refinement and Evolution
 Active Shape Model: iterative closest point (ICP) algorithm to align
 Optical Flow Model: measures dose variance between a reference
 Machine Learning
PTV contours
image and any other images within the training dataset.
Reference
Image
28
PTV contour space
cord dose space
Optical Flow
Registration
29
10
Knowledge Model Evolution
Basic
Anatomy
Cases
Expanded
Model 1
Add.
Case 2
Expanded
Model 2
Type 2: Anal rectal
Additional Anatomy Case
Add.
Case 4
Expanded
Model 3
Expanded
Model 4
Add.
Case 5
Expanded
Model 5
Add.
Case 6
Expanded
Model 6
DVH Predicting
Type 2: High risk prostate
Bladder Model: gEUD prediction error
Basic Model
Add.
Case 1
Add.
Case 3
Basic Anatomy Case
Prediction Error
For Additional Anatomy Case
Model Training
 Build single knowledge model for multiple cancer types in the pelvic
region
Type1: Low-intermediate risk prostate
Knowledge Model Evolution
New Case
 Progressive Learning: Knowledge Update and Adaptation
11
Summary: Knowledge-Model Guided Planning
Physician
Prescription
Planned vs. Modeled
References
Acknowledgement
 AAPM2014: SU-F-BRD-9: Yuan et al, Lung IMRT Planning Using Standardized
•
Beam Bouquet Templates
 AAPM2014: SU-E-T-49:
Yuan et al, Automatic Beam Angle Determination for
Lung IMRT Planning Using a Beam Configuration Atlas
 AAPM2014: SU-E-T-229:
IMRT/VMAT
Planning
Hu et al, Machine Learning Methods for Knowledge
Based Treatment Planning of Prostate Cancer
 AAPM2014: MO-C-17A-7:
Sheng et al, Building Atlas for Automatic Prostate IMRT
Planning: Anatomical Feature Parameterization and Classification
 AAPM2014: TU-C-17A-11: Lu et al, Progressive Knowledge Modeling for Pelvic
Plan
Evaluation
IMRT/VMAT Treatment Planning
 AAPM2014: TH-A-9A-1:
Liu et al, Active Optical Flow Model: Predicting VoxelLevel Dose Prediction in Spine SBRT
 AAPM 2014: SU-E-T-527:
•
•
•
•
•
•
•
•
•
•
•
•
John Kirkpatrick, MD/PHD (Duke
University)
Brian Czito, MD (Duke University)
W. Robert Lee, MD (Duke University)
Bridget Koontz, MD (Duke University)
David Yoo, MD, PHD (Duke University)
David Brizel, MD (Duke University)
Chris Kelsey, MD (Duke University)
Mark Dewhirst, DVM(Duke University)
Radhe Mohan PHD (MD Anderson)
Zhongxing Liao MD (MD Anderson)
Jaques B. Bluett (MD Anderson)
Xiaodong Zhang PHD (MD Anderson)
Michael Gillin PHD (MD Anderson)
•
•
•
•
•
•
•
•
•
•
•
•
•
Jun Lian, PHD (UNC)
Sha Chang (UNC)
Bhishamjit S. Chera, MD (UNC)
Larry Marks, MD (UNC)
Wilko Verbakel, PHD (Vu University)
Jim Toll, PHD (Vu University)
James Deye PHD (NCI)
James Galvin PHD (RTOG)
Ying Xiao PHD (RTOG)
Kevin Moore PHD (UCSD)
Charles Simone MD (U Penn)
Liyong Lin PHD (U Penn)
Jeffery D. Bradley MD (Wash U)
Lian et al, Prior Knowledge Guided TomoTherapy
Treatment Planning
12
Thank you
13