dr. ying xiao: radiation therapy oncology group bioinformatics
DESCRIPTION
On July 25. Dr. Ying Xiao delivered a presentation titled "RTOG - Radiation Therapy Oncology Group Bioinformatics."TRANSCRIPT
1
RTOG Bioinformatics
Y. Xiao, Ph.D.
RTOG, ACR
Radiation Oncology, Jefferson
Medical College
2
Evidence Based
Radiation Oncology
Radiation Therapy Oncology Group (RTOG)
Improve The Survival Outcome And Quality Of Life
Evaluate New Forms Of Radiotherapy Delivery
Test New Systemic Therapies In Conjunction With Radiotherapy
Employ Translational Research Strategies
3
RTOG Bioinformatics Mission
To facilitate the development and to
develop personalized predictive models
for radiation therapy guidance from
specific characteristics of patients and
treatments with integrated clinical trial
databases, bridging clinical science,
physics, biology, information technology
and mathematics
4
BIOINFORMATICS ELEMENTS AND
PROCEDURES Available to BioWG
Database
RT Dose/Images/Clinical Data
Genomic/Proteomic
Biomarker
Data Analysis
Protocol development/Protocol operation
support/Trial Outcome-Secondary Analysis
Validation/Development/Research
5
DATA/
DATA Integration
6
RTOG DATA for BioWG Investigation Protocol N Endpoints
0022 oropharyngeal cancer 60 Salivary function
0117 lung 73 Pneumonitis and esophagitis
0126 prostate ~1500 Erectile dysfunction; rectal bleeding
Fecal incontinence vs dose
0225 nasopharyngeal 60 Salivary function
0232 prostate brachytherapy
0234 head and neck 230 TCP? Ongoing, not recruiting
0236 lung SBRT 52 Ongoing: TCP, toxicity
0321 prostate HDR brachy 110 Late/Acute GU/GI
0522 head and neck Local control
0529 IMRT anal canal cancer 59 GI/GU acute
9311 lung ~150 NIH R01 (Deasy). Toxicity: esophagitis; pneumonitis
9406 EBRT prostate 800 NIH R01 (Tucker) toxicity
9803 3D CRT GBM 40 Brain toxicity
7
Rapid Learning
CAT (Computer Assisted Theragnostics)
MAASTRO/RTOG Collaboration
Andre Dekker, PhD
MAASTRO Knowledge Engineering
8
Why Rapid Learning/CAT?
[..] rapid learning [..] where we can learn from each patient to guide
practice, is [..] crucial to guide rational health policy and to contain costs
[..].
Lancet Oncol 2011;12:933
Personalized medicine
• Explosion of data
• Explosion of decisions
• Decision support
• Evidence base
Personalized medicine improves survival and quality of life.
9
Prediction by MDs?
• Non Small Cell Lung
Cancer
• 2 year survival
• 30 patients
• 8 MDs
• AUC: 0.57
• Retrospective
10
How to get
data for rapid learning
11
Challenges to share data
[..] the problem is not really technical […]. Rather, the problems are
ethical, political, and administrative. Lancet Oncol 2011;12:933
1.Administrative (time)
2.Political (value, authorship)
3.Ethical (privacy)
4.Technical
12
CAT approach
CAT is a research project in which
we develop an IT infrastructure -> technical
to make radiotherapy centers
semantic interoperable (SIOp*) -> administrative
while the data stays inside your hospital -> ethical
under your full control -> political
* SIOp level 3 = Machine Readable ->Data in common syntax and with common meaning
13
Components
ETL
Deident. & Filter
Ontology
XML DICOM
Application Sharing
Distributed Learning
CTMS PACS
Export
Query
14
Key features
• No sharing of data, truly federated
• Machine learning (retro.) & clinical trials (prosp.)
• NCI Thesaurus with formal additions
• 5 languages, 5 countries & 5 legal systems
• Focus on radiotherapy
• Inclusion of non-academic centers
• Industry involvement
15
Network 11/2011
Active or funded CAT partners (10)
Prospective centers (4)
2
5
Map from cgadvertising.com
16
Laryngeal carcinoma model
• 994 MAASTRO patients
• 1990-2005
• www.predictcancer.org
• Input parameters
– Age
– Hemoglobin
– T-stage
– EDQ2T (Gy)
– Gender
– N+
– Tumor location
• Output parameters
– Overall survival
www.predictcancer.org, Egelmeer et al., Radiother Oncol. 2011 Jul;100(1):108
17
Larynx Query
18
Validation - Result
19
Distributed
Learning
Architecture
Update
Model
Learn Model
from Local Data
Central Server
Model Server
RTOG
Send Model Parameters
Final Model Created
Learn Model
from Local Data
Learn Model
from Local Data
Model Server
MAASTRO
Model Server
Roma
Send Model Parameters
Send Model Parameters
Send Average Consensus Model
Send Average Consensus Model
Send Average Consensus Model
Only aggregate data is exchanged between the Central Server and the local Servers
20
Distributed Learning Results
21
Web-based Documentation
System
with Exchange of DICOM RT for
Multicenter Clinical Studies
in Particle Therapy
Priv.-Doz. Dr. med. Stephanie E. Combs, DEGRO 2012
22
PARTICLE THERAPY
• particle therapy is a new and promising technique in radiation
oncology for cancer treatment
• compared to standard radiation therapy (RT) with photons, it is with
either
– ions (e.g. Carbon C12, Helium He2, Oxygen O16) or
– protons (H1)
• advantages:
− more precise dose delivery
to the target
sparing of normal tissue and
organs at risk
− enhanced radiobiological
effectiveness (carbon ions) increase in local tumor control
23
• First results of clinical phase I and II trials performed at NIRS support the assumption that carbon ions provide an enhanced biological effectiveness in adenoid cystic carcinomas, H/N melanomas, lung and liver tumors, large soft tissue sarcomas, chordomas / chondrosarcomas and prostate cancer
• Randomized trials proving the superiority of heavy charged particles in comparison to photon IMRT and protons required – Randomized trial of proton vs. Carbon ion RT for chordomas and
chondrosarcomas of the skull base
• Radiobiologic research will enable better exploitation of the advantages of carbon ion RT in future trials
• Physics Research: e.g.High Precision Treatment of Moving Targets
What we need to do.....
24
HIT
HEIDELBERG ION-BEAM THERAPY
CENTER
• began patient treatment in Nov. 2009
• main focus:
– clinical studies to evaluate the benefits of ion therapy for several indications
• ULICE project
(Union of Light Ions Centers in Europe)
– development of a database with transnational access
– platform for international clinical multicenter studies
– accessible by external/internal oncologists, physicists, researchers
25
Establish a common database for hadrontherapy.
- few centers have treated patients with particle therapy
- heterogeneous concepts of dose prescription, delivery
- hetereogeneous studies and treatment protocols
Exchange of
clinical
experience
Harmonized
study and
treatment
concept
Transfer of
know-how
Standards and
basic (biologic)
data ULIC
E D
ATA
BA
SE
ULIC
E D
ATA
BA
SE
26
METHODS
• Approach:
– central, web-based system
– interfaces to the mandatory existing information systems of the hospital
– avoidance of redundant entry of any patient/clinical data
– study-specific modules, which can easily be adapted
– implementation of security and data protection measures to fulfill legal
requirements
Kessel K., ..., Combs SE, Radiat Oncol, 7, 115 (2012).
27
GENERAL PRINCIPLES AND ARCHITECTURE
• database with standard interfaces of the PACS world
– storage of all data in an SQL database (DICOM data model)
– possibility to dynamically extend with additional data structures
– interfaces to process and import information from
• DICOM data
• HL7 messages
– Java applet for
• manual import of data by web-upload
• receiving and sending data with DICOM C-Store
– the underlying components are compliant with the IHE Framework
• telemedicine record
– combines the functionality of
• a medical record with
• a professional DICOM RT viewer (Class IIb)
– allows us to exchange, store, process and visualize text data, all types of
DICOM images, and any other multimedia data
– with extensions for study documentation
Kessel K., ..., Combs SE, Radiat Oncol, 7, 115 (2012).
28
INTEGRATION OF OTHER INFORMATION
SYSTEMS
Kessel K., ..., Combs SE, Radiat Oncol
29
SECURITY CONCEPT
• the system uses the encrypted https protocol to exchange data between
the central server and other systems
• users need an account and password to access the system
• a roles-and-rights concept enables a very detailed configuration for each
study or client
• (basic patient) data can be pseudonymized
– the original patient information is kept in a separate database
– when a user has the right to view all data of a patient, the data is de-
pseudonymized “on the fly”
• external users
– communication via application gateway in the demilitarized zone (DMZ)
– client certificates to verify authorized browsers
Kessel K., ..., Combs SE, Radiat Oncol, 7, 115 (2012).
30
RESULTS
• documentation began May 2011 and we documented 900 patients
• all access to the system takes place over the Internet
– no software needs to be installed (except a JRE)
• exchange and storage of various DICOM RT data
• visualization with the DICOM RT ion viewer
– examination and review of radiation plans from every computer in the
hospital without having to go to a PACS or a radiation therapy planning
station
• with our web-based approach, we have succeeded in developing a
database for multicenter studies within the ULICE project
• future prospects: extend automatic and electronic study analyses
Kessel K., ..., Combs SE, Radiat Oncol, 7, 115 (2012).
31
More to Integrate
Andre Dekker, PhD
MAASTRO Knowledge Engineering
32
How to learn better?
• AUC of 0.85 is needed
• Survival model AUC 0.75
• More patients
• More variables
• Diversity
33
More variables from a simple CT
34
Biomarkers
General
Radiomics
(n = 150)
LOO-AUC 0.75
Selected image features: CT Range,
Cluster Shade, Run Percentage
LOO-AUC 0.76
Selected biomarkers: IL6, IL8, CEA
LOO-AUC = 0.72:
Gender, WHO, FEV1, Positive lymph
nodes,
Tumor Volume
35
Biomarker: IL6, IL8, CEA
General: Gender, WHO-PS, FEV1, Positive lymph nodes, Tumor Volume
Radiomics: Range, Run Length, Run Percentage
General Biomarkers
n = 131
Radiomics
AUC 0.87
36
DATA ANALYSIS
37
DATA ANALYSIS
- Evidence Based Radiation
Therapy Quality Assurance
38
DATA ANALYSIS
- Evidence Based Radiation
Therapy Quality Assurance
(Structure Definition)
39
Critical Impact of Radiotherapy
Protocol Compliance and
Quality TROG 02.02
L. J. Peters et al, Journal of Clinical Oncology, vol. 28, Number 18, June 2010
40
Failure To Adhere To Protocol
Associated With Decreased
Survival: RTOG 9704
R. Abrams et al, Int. J. Radiation Oncology Biol. Phys., Vol. 82, No. 2, pp. 809–816, 2012
41
Target Defined from Multiple
Institutions
S. Kong, Y. Xiao, M. Machtay, et. Al., A “Dry-Run” Study for RTOG1106/ACRIN6697: A Randomized Phase
II Trial of Using During-Treatment FDG-PET and Modern Technology to Individualize Adaptive Radiation
Therapy in Stage III NSCLC ,IASLC, 2011
42
Target Inter-Observer
Variability
Pre-GTV Statistics
43
OAR Variation
Th
e maxim
um
volu
me o
f
bra
ch (b
rach
ial p
lexu
s) is up
to 4
-5 fo
ld o
f the m
inim
um
.
44
Dry Run for Segmentation and
Plan Evaluation – 1106 Example
GTV contours from different institutions. Red thick line represents the consensus contour.
(a) Case1, (b) Case2.
45
Oral Presentation AAPM 2012
46
RTOG 0617, NCCTG N0628,CALGB 30609
Conventional vs. High Dose RT
R
A
N
D
O
M
I
Z
E
RT: 60 Gy
Paclitaxel
Carboplatin +/-
Cetuximab
RT: 74 Gy
Paclitaxel
Carboplatin +/-
Cetuximab
Paclitaxel
Carboplatin X 2
+/- Cetuximab
J. Bradley et al, ASTRO 2011
47
Overall Survival
Ove
rall
Su
rviv
al (%
)
0
25
50
75
100
Months since Randomization
0 3 6 9 12
*One-sided p-value, left tail
Patients at Risk60 Gy74 Gy
213204
190175
149137
124116
104 93
Dead
5870
Total
213204
HR=1.45 (1.02, 2.05) p*=0.02
60 Gy74 Gy
48
DATA ANALYSIS
- Evidence Based Radiation
Therapy Quality Assurance
(Image Guided Radiotherapy)
49
IGRT Data Submission
Components
50
Variations Between Systems
Y. Cui (Xiao) et al, Int. J. Radiation Oncology Biol. Phys., Vol. 81, No. 1, pp. 305–312, 2011
51
IGRT Credentialing for RTOG
Protocols
52
IGRT Variations
Y. Cui (Xiao) et al, Implementation of Remote 3D IGRT QA for RTOG Clinical Trials,
Int. J. Radiation Oncology Biol. Phys., In Press
53
DATA ANALYSIS
- Evidence Based Radiation
Therapy Quality Assurance
(Intensity Modulated
Radiotherapy Planning)
DUKE University Radiation Oncology
Establishing knowledge-based
models for IMRT planning
quality evaluation
Q. Jackie Wu
Yaorong Ge
Jan 2012
55
Case1 Right Parotid Case 2 Left Parotid
0 20 40 60 80 1000
20
40
60
80
100
Percent dose (%)
Perc
ent volu
me (
%)
Case 1 R Parotid Case 2 L Parotid
Parotid DVH
56
Knowledge Models For IMRT
Planning QA
• Understand the patient anatomical, physiological
factors that influence dose coverage
• Quantify their individual influence via mathematical
modeling and machine learning
• Codify treatment planning experience and guidelines
using knowledge engineering
• Model Use these factors to model the inter-patient
variation and to provide “peer-review” type guidance
57
0 50 100 0
50
100
Case 1
50 100
Case 2
50 100
Case 3
50 100
Case 4
50 100
Case 5
50 100
Case 6
0 50 100 0
50
100
Case 7
50 100
Case 8
50 100
Case 9
50 100
Case 10
50 100
Case 11
50 100
Case 12
0 50 100 0
50
100
Case 13
50 100
Case 14
50 100
Case 15
50 100
Case 16
50 100
Case 17
50 100
Case 18
0 50 100 0
50
100
Case 19
50 100
Case 20
50 100
Case 21
50 100
Case 22
50 100
Case 23
50 100
Case 24
Example Of Parotid Sparing
Modeling
Vo
lum
e (
%)
Dose (%)
Vo
lum
e (
%)
Vo
lum
e (
%)
Vo
lum
e (
%)
Dose (%)
Dose (%)
Dose (%)
Actual Plan DVH
Modeled Lower
and Upper
Bound of DVH
58
Moving Forward
• Bigger Database
– Knowledge and experience embedded in the IMRT
cases
– Collaboration with RTOG who has large case pools
• Comprehensive Knowledge Base
– All types of knowledge sources
– Predictive models
– Ontological framework
• Intuitive and Integrated Decision Support
– Minimize cognitive burden
59
Intuitive and Integrative Decision
Support
Prediction QUANTEC Guidelines
Gui
60
Experience-based dose volume
histogram prediction in IMRT: A new
QC paradigm
Kevin L. Moore, Ph.D., DABR
61 KL Moore et al, "Experience-based QC of IMRT planning" IJROBP 82, 545 (2011)
Introduction
• Current TPSs cannot guarantee that IMRT plans are fully optimal, and applying quality
control methods to the treatment planning process can have a very large impact
• The degree to which plan quality variations compromise patient care is unknown
• Investigations (WashU, JHU, ROR) have shown that plan quality variations can be
quite sizeable, vastly exceeding other uncertainties and putting patients that should
have had a low risk of complications into a high-risk category
• Quantitative assessment of PTV metrics is easy, normal tissue sparing is not.
Anatomical variations confound the plan evaluation process because expected values
are (as yet) not available.
62
Predicting DVHs with universal basis function evolution
• With universal fitting function and parameter fits pjq(k), which are derived from all N
training patients, we have the desired result:
• This equation takes as input geometric variables and the derived fitting parameters
and outputs a predicted DVH for an organ based only on the input structure set SSij
63
Results: rectum and bladder predictions (training pts)
N=20 training patients
64
Results: rectum and bladder predictions (“new” pts)
M=20 validation patients
65
Quantifying plan quality variations in RTOG trials
• These data represent an unique and nearly ideal test bed to not only quantify plan
quality variability but, due to the associated outcomes data with the treatment plans, it
can address its clinical importance as well
• Employ modeling toolkit as
secondary study on closed
RTOG protocol (e.g. 01-26) to
quantify the clinical
importance of IMRT plan
variability
• Plan quality variations in a large-scale, multi-
institutional trial will serve as an excellent
surrogate for the variability in plan quality
across IMRT practitioners
66
DATA ANALYSIS
- Analytical Algorithms
67
Statistical inference
Frequentist inference Bayesian inferenceProbability: the proportion of times that an event would occur in a large number of similar repeated trials.
Model parameters are fixed, use the observed data to make Inference about parameters, e.g. Maximum Likelihood Estimation, confidence intervals and P-values
Probability describes degree of belief. It reflects one’s strength of belief that the proposition is true. Bayesian inference inherently embraces a subjective notion of probability.
Start with a prior belief about the likely values of model parameters, then use observed data to modify these parameters, i.e., deriving posterior probability distribution.
The Dempster-Shafer theory is an extension of the Bayesian inference.
Wenzhou Chen, Yunfeng Cui, Yanyan He, Yan Yu, James Galvin, Yousuff M. Hussaini, Ying Xiao, “Application of Dempster-Shafer Theory in Dose Response Outcome Analysis”, Physics in Medicine and Biology, In Press
68
Probability, Belief & Plausibility of
RP
69
LKB parameters from Dempster–
Shafer theory and other references
70
FUTURE DIRECTIONS
Introducing A New Organizational Structure
NCI Clinical Trials Network
72
Institution Patient data
ACR
Cloud IROC
QA Medidata
Rave
Study Group
Patient data
IROC
Study Groups (second analysis, outcomes,
publications etc.)
QA
QA
Imaging
Studies
Data and QA Process Flow
(receipt, QA, storage)
73