2982 - TetraMorph: Whole-Body, Patient-Specific Mesh Phantoms - Applications to RPT and EBRT Dosimetry
Presenter(s)
R. J. Dawson1, L. M. Carter2, C. Choi1, B. Shin1, C. Huesa-Berral3,4, A. Bertolet3,4, A. Kesner2, H. Paganetti5, and W. Bolch1; 1University of Florida, Gainesville, FL, 2Department of Medical Physics, Memorial Sloan Kettering Cancer Center, New York, New York, New York, NY, 3Massachusetts General Hospital, Boston, MA, 4Harvard Medical School, Boston, MA, 5Department of Radiation Oncology, Massachusetts General Hospital/Mass General Brigham and Harvard Medical School, Boston, MA
Purpose/Objective(s): To develop an automated workflow for generation of patient-specific tetrahedral mesh phantoms based on input anatomical images, with applications to radiopharmaceutical and external beam radiotherapy dosimetry. Phantoms produced using this method were made to be consistent with tissue compositions, densities, and tag identifiers present in the ICRP-145 mesh-type reference computational phantoms (MRCPs), and, importantly, preserve the finely-detailed radiosensitive anatomical regions defined in the MRCPs.
Materials/Methods: From the MRCPs, the University of Florida, in collaboration with Memorial Sloan Kettering, developed the largest mesh-type computational phantom library to date; this library is parameterized to reflect anthropometric characteristics of the North American population and is collectively termed the UF/MSK phantom library. The following procedure was applied to CT images (acquired through an IRB protocol at UF) for a cohort of pediatric patients (medulloblastoma via craniospinal irradiation) and adult patients (Hodgkin’s lymphoma) treated with pencil beam scanning proton therapy and 3D conformal radiation therapy. For each patient image, a sex-, age-, height-, and weight-matched phantom was selected from the library as the starting anatomical model and subsequently refined through structure-specific contour matching as described below. A pretrained deep learning-based autosegmentation algorithm (TotalSegmentator, based on the nnU-Net framework) was used to contour the patient image without human intervention. Corresponding contours were generated within the mesh phantom which then served as moving images for a series of deformable image registration steps, with the autosegmented contours as fixed images. Deformation vector fields were then applied to the nodes of the tetrahedral mesh phantoms, and any induced geometric defects were repaired. This process preserved key geometric information present in the patient image, including organ/tissue morphometries, volumes, relative centroid positions, and maintained accuracy of major organ contours as evaluated by Dice similarity coefficients and Hausdorff distances relative to the ground truth segmentation.
Results: The phantom generation method described was able to produce a whole-body tetrahedral mesh model for any arbitrary input CT image. Phantoms created with TetraMorph were suitable for Monte Carlo radiation transport simulations. Images from the SNMMI Lu-177 Dosimetry Challenge 2021 were used to demonstrate the utility and accuracy of the approach.
Conclusion: Computational human phantoms produced by merging patient volume images with patient-matched tetrahedral mesh phantoms can be used to compute organ doses following radiotherapy and radiation-based medical imaging. This work has potential for clinical implementation, medical record supplementation, prospective EBRT TPS and RPT dose optimization, and future epidemiological studies.