Program Information
A Generation Methodology for Numerical Phantoms with Statistically Relevant Variability of Geometric and Physical Properties
S Dolly1*, E Ehler1, Y Lou2, M Anastasio2, H Li3, (1) University of Minnesota, Minneapolis, MN, (2) Washington University in St. Louis, Saint Louis, MO, (3) Washington University School of Medicine, Saint Louis, MO
Presentations
TH-CD-601-1 (Thursday, August 3, 2017) 10:00 AM - 12:00 PM Room: 601
Purpose: Numerical phantoms enable ground-truth assessment, the validity of which depends heavily on the realism of the phantoms. Realistic phantoms should model both geometric and physical properties of human anatomy. However, most currently available phantoms lack the ability to realistically model both property types. In this study, a new numerical phantom generation methodology was proposed which quantifies statistical variations of both geometric and physical properties, by learning these variations from a patient population via machine learning techniques. An ensemble of head-and-neck computed tomography (CT) phantoms was created to demonstrate this methodology and its applications in radiation therapy.
Methods: Anonymized CT images and organ segmentation data were extracted from a database of patients treated with radiation therapy for head-and-neck cancer. First, geometric attribute distribution (GAD) models were created to quantify the inter- and intra-organ geometric variations, in terms of position and shape, by performing principal component analysis (PCA) on the organ segmentation data. Then, a newly proposed physical attribute distribution (PAD) model was constructed to quantify photon attenuation coefficient variations using PCA and the CT voxel information from the training data. Both the GAD and PAD models were employed to generate new numerical phantoms that resemble the statistical variations of the training data. X-ray projections of the generated phantoms were simulated, and then reconstructed to produce CT images having similar shape and interior texture variations as realistically acquired CT scans.
Results: An ensemble of head-and-neck phantoms was generated using the methodology, with each phantom exhibiting realistic organ position and shape. The computer-generated CT images demonstrated similar CT number distributions (mean and standard deviation) as the training data.
Conclusion: A novel methodology was developed to generate numerical phantoms that are realistic in terms of both anatomic geometry and physical properties, and can be applied for meaningful, ground-truth patient simulation studies.
Contact Email: