“Big Data” in Radiation
Oncology
Hao Howard Zhang, PhD and Warren D. D’Souza, PhD, MBA
Department of Radiation Oncology
Disclosures
Research funding from
NIH/NCI
Varian Medical Systems (Palo Alto, CA)
Philips Healthcare (Cleveland, OH)
New Economics of Cancer Care
• ASTRO considering to even out variations in care delivery
with a revamped accreditation program
• Essential value-based purchasing components include
physician education, implementation of quality metrics and
deployment of
big data in decision making
(Health Imaging, Lisa Fratt, May 2013)What is “Big Data” ?
Big data
is the term for a collection of data sets so large
and complex that it becomes difficult to process using
on-hand database management tools or traditional data
processing applications. The challenges include
capture, curation, storage, search, sharing, transfer,
analysis, and visualization. Big data sizes are a
constantly moving target, as of 2012 ranging from a few
dozen terabytes to many petabytes of data in a single
data set.
“Big Data, for better or worse: 90% of world’s data
generated over last two years”
Big Data in Medicine
• Quality
• The granddaddy of medical databases is The
Society of Thoracic Surgeons STS National
Database, launched in 1989 as an initiative for
quality improvement and patient safety among
cardiothoracic surgeons.
• Outcomes
• Appropriate use ~ value-based reimbursement.
Currently, fee-for-service reimbursement
incentivizes physicians to treat every patient. In
some cases, the patient can be served by other
options.
Challenges with Big Data
• Data Provenance
• Source and reliability of data
• Data Aggregation
• Combining data from multiple sources and across
institutions
• Data Interpretation
National Radiation Oncology
Registry
RT Data Mining Infrastructure
Oncospace
IHE-RO
• IHE-RO is an initiative that helps to ensure a safe,
efficient radiation oncology practice by improving
system to system connections.
• Image-based (3-dimensional) radiation therapy
treatment planning.
• Exchanging and storing image registration,
radiation therapy structure sets, radiation therapy
doses.
• Exchange of data required to perform sophisticated
treatment planning.
National Cancer Informatics
Program
The Old caBIG initiative
• Open development of informatics capabilities for research
• Access to well-described data to facilitate integrative
cancer research
• Provides informatics infrastructure and standards to
improve interoperability between information systems
• Foster collaborative relationships among researchers
across the basic, translational, and clinical continuum
• Training next generation of biomedical investigators
Radiation Oncology Big Data
• Electronic health records
• Demographics
• Treatment delivery modality
• Diagnostic Imaging
• Anatomical and functional
• Treatment Planning
• Beam/plan parameters, DVH parameters
• Additional Imaging
• kV, Simulation CT
• Outcomes
• Tumor response, Toxicity/Complications
• Blood and tissue samples
Optimal Treatment Strategy
Lack of apriori Knowledge
The inverse planning does not quite listen
to me!
Keep trying and …
Hope !!!
I want:
Prostate dose 70Gy Seminal vesicles 50-70 Gy
Bladder max 60Gy Rectum max 60Gy
…….. …..
Plan Optimality-Practicality
Tradeoff
Treatment Plan Quality – Overlap
Volume Histogram
Wu et al. Patient driven-geometry information retrieval for IMRT treatment plan quality control. Med Phys 36, 5497 (2009).
Knowledge-Based Replanning
Wu et al. Patient driven-geometry information retrieval for IMRT treatment plan quality control. Med Phys 36, 5497 (2009).
Model-Based OAR Sparing
Moore et al. Experience-based quality control of clinical intensity-modulated radiotherapy planning. Int J Radiat Oncol Biol Phys 81, 545 (2011).
Predicting Treatment Plan Output
Left Parotid
Right Parotid
Zhang et al. Modeling plan-related clinical
complications using machine learning tools in a multi-plan IMRT framework. Int J Radiat Oncol Biol Phys 74, 1617-26 (2009)
Decision Features
• Decision making in radiotherapy
• Clinical Features such as patient performance status,
organ function, grade and extent of tumor (TNM
system)
• Toxicity measurements and scoring based on
validated scoring systems
• Spatial and temporal distribution of radiotherapy dose
• Additional therapies such as chemotherapy
(sequential or concurrent), targeted agents and
surgery
• Imaging features, including size, volume and more
Clinical Decision Support System
Lambin et al. Predicting outcomes in
radiation oncology – multifactorial decision support systems.
MAASTRO Clinic Larynx Data
Egelmeer 2011Local Control Nomogram
• Prognostic factors for overall survival were low hemoglobin level, male sex,
high T-status, presence of nodal involvement, older age, lower EQD2T, and non-glottic tumor
• Unfavorable prognostic factors for local control were low hemoglobin level,
Overall Survival
Egelmeer et al. Development and
Validation of a nomogram for prediction And local control in laryngeal
Carcinoma patients treated with Radiotherapy alone: a cohort study
Based on 994 patients. Radiother Oncol 100, 108-15 (2011)
RTOG 9311
• 163 patients entered in a Phase I/II 3-D radiation therapy dose escalation trial.
• Patients stratified for different radiation treatment levels depending on their V20 (the percentage of their total lung volume that would receive in excess of 20 Gy with the treatment plan).
• Patients with a V20<25% (Group 1) received 70.9 Gy/33 fractions, 77.4Gy/36 fractions, 83.8 Gy/39 fractions and 90.3 Gy/42 fractions successively.
• Patients with a V20 of 25-37% (Group 2) received 70.9 Gy and 77.4 Gy successively.
• The treatment arm for patients with a V20>37% (Group 3) closed early secondary to poor accrual(2 patients) and the perception of excessive risk for the development of pneumonitis.
• Patients were allowed to receive neoadjuvant chemotherapy before radiation therapy only, but not concurrently.
Radiation Pneumonitis Events
Netherlands Cancer Institute Data
• Patients
– 81 patients (41 with malignant lymphoma and 40 breast cancer); – Mean age = 41 years (range = 18-74 years); 25 men and 56
women; 26 smokers
• Malignant lymphoma
– Radiation alone (n = 18) – Chemotherapy + radiation (n = 23)• Breast
– Radiation alone (n = 5) – Radiation + Chemotherapy (n = 24) – Radiation + tamoxifen (n = 11)Theuws et al. Prediction of overall pulmonary function loss in relation to the 3-D dose distribution for patient with breast cancer and malignant lymphoma. Radiother Oncol 1998;49:233-43.
Treatment Modality Influence
Theuws et al. Prediction of overall pulmonary function loss in relation to the 3-D
dose distribution for patient with breast cancer and malignant lymphoma. Radiother Oncol 1998;49:233-43.
circles – Breast; squares – Lymphoma
Smoking Status
circles – Breast; squares – Lymphoma
open symbols – radiation alone; closed symbols – radiation + chemotherapy
Theuws et al. Prediction of overall pulmonary function loss in relation to the 3-D dose distribution for patients with breast cancer and malignant lymphoma. Radiother Oncol 1998;49:233-43.
Duke Radiation Pneumonitis
234 patients; 34 patients with
Grade 2+ pneumonitis;
70% of patients treated at 1.8–2.0 Gy/fraction,
once daily; remaining treated at 1.25 Gy/fraction
To CTV and 1.6 Gy/fraction to GTV, twice daily
27 non-dose factors (biological, clinical and
other factors)
Chen et al. Using patient data similarities to predict radiation
Radiation Pneumonitis Models
Dose factors
Dose + Clinical factors
Wisdom of Crowds
Das et al. Combining multiple models to generate consensus: Application to radiation-induced pneumonitis prediction. Med Phys 35, 5098-109 (2008).
Probability of Radiation
Pneumonitis
Das et al. Combining multiple models to generate consensus: Application to radiation-induced pneumonitis prediction. Med Phys 35, 5098-109 (2008).
Challenges with Outcomes
Analysis
• Quantitative Analysis of Normal Tissue Effects in the
Clinic (QUANTEC)
• Current state of dose response knowledge
• Challenge in generalizing recommendations
• Difficulty in determining actual dose to patients, consistency
of structure delineation, outcome scoring, heterogeneity in institutional treatment delivery practices
• Inferences about treatment plan quality
Tan et al. 2013. Int J Radiat Oncol Biol Phys 85: 1375-82.
FDG-PET Features & Pathologic
Tumor Response
A new SUV intensity feature - Skewness pre-CRT
• Top: responder, more skewed (fewer
higher SUVs)
• Bottom: non-responder, less skewed
(more higher SUVs)
Three texture features post-CRT – Inertia, Correlation, and Cluster Prominence
• Top: responder, homogeneous FDG
uptake post-CRT
• Bottom: non-responder, heterogeneous
Tan et al. 2013. Int J Radiat Oncol Biol Phys 85: 1375-82.
Spatial-Temporal FDG-PET
Zhang et al. 2014. Int J Radiat Oncol Biol Phys 88: 195-203
Therapy Response Prediction
• Response of 20 patients with esophageal cancer to chemoradiotherapy (CRT)
• SVM model with selected features from all feature groups: AUC = 1.0, sensitivity = 100%, specificity = 100%
• Models with conventional PET/CT response measures or clinical parameters: AUCs < 0.75
• Multi-institution data (Univ of Maryland, Wake Forest Univ, Oregon Health & Science Univ)