Program Information
Investigation of a Pelvic Bone Shape Model in Support of Bone Classification for Synthetic CT Generation
L Liu*, Y Cao , J Fessler , J Balter , Univ Michigan, Ann Arbor, MI
Presentations
SU-F-303-14 (Sunday, July 12, 2015) 4:00 PM - 6:00 PM Room: 303
Purpose: Most synthetic CT methods require user intervention to aid in bone localization in the pelvis. This study investigates an automated method for identifying bone regions to avoid confounding their classification with that of transient air locations in the pelvis.
Methods:The space that encapsulated the pelvic bones and excluded air was modeled as a combination of eigenmodes of a shape atlas. Under IRB-approval, CT data from 20 patients (8 with matching MRI image volumes) were investigated. The atlas was derived from intensity-based deformable alignment of the pelvic CT image volumes. Eigenvectors were calculated via Principle Component Analysis (PCA) of the B-spline coefficients that optimally aligned a randomly selected member of the population to all other members. Eigenmodes of this atlas model were then fit to candidate MR image volumes via a cost function that minimized the mean square error between the deformed atlas MRI and the target MRI image volumes. Using these eigenmodes permitted an efficient search through a very small number of parameters to estimate pelvic shape. The accuracy of pelvic bone identification was tested in the patient population using a leave-one-out cross validation test.
Results:Using the first eigenmode only, the percentage of correctly identified bone voxels is 84% with an average computational time of 69 seconds, as compared to 86% when fitting the model to the source CT images by maximizing the percentage of identified bone voxels. Including the second and third eigenmodes to the model results in an improvement that is less than 2%. Dilating the bone-containing space by 5mm improves the overlap with true bone to 96%, while the percentage of air voxels mislabeled as bone remains below 0.6%.
Conclusion:Our result suggests using one leading eigenmode of pelvic bone shape variation achieves a sufficiently accurate model to support atlas-based synthetic CT generation in the pelvis.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by NIH R01 EB016079
Contact Email: