Program Information
Kernel Density Estimation of Tumor Position Undergoing Respiration Via Surface Monitoring
M Tajdini1*, P Sabouri2 , Y Ghanbari3 , D Ruan4 , A Sawant5 , (1) ,,,(3) University of Maryland School of Medicine, Baltimore, MD, (4) UCLA School of Medicine, Los Angeles, CA, (5) University of Maryland School of Medicine, Baltimore, MD
Presentations
TH-AB-205-12 (Thursday, August 3, 2017) 7:30 AM - 9:30 AM Room: 205
Purpose: Accurately locating highly-mobile thoracic and abdominal tumors is a challenging task due to the complexity of respiratory motion. Here, we develop a principal component analysis-based kernel density estimation (PCA-KDE) technique to estimate tumor position. The PCA-KDE characterizes the joint probability distribution of response (tumor target) and covariate (the patient’s body surface) variables, taking into account the intrinsic uncertainties of respiratory motion.
Methods: 4DCT data from eight patients were retrospectively studied. For each patient, a reference phase was selected (end-of-inhale) and all the remaining phases were registered using the open-source deformable image registration software Elastix. Subsequently, the displacement trajectories of points on the patient surface and on the tumor were derived. For every target point on the boundary of the tumor, the KDE characterized its joint probability distribution as a superposition of Gaussian components, each centered at a sampled value and weighted according to how close the external surrogate at each phase was to the training samples at other phases. The PCA was also performed to construct a lower-dimensional manifold of training space.
Results: The entire shape of the tumor including its centroid position was estimated at all 10 phases of 4DCT, with RMSE between the actual and estimated values as low as 0.73 mm and maximum error up to 1.58 mm, allowing us to capture the complete motion of the tumor.
Conclusion: We successfully applied the PCA-KDE method for thoracic and abdominal tumor position estimation. Our results showed that there was strong correlation between the motion of the surface and that of the tumor. In contrast to current “single-point” (typically centroid) position estimation techniques, our developed technique may be used to create external surrogate-based tumor motion models to estimate the entire shape of the tumor and capture its complex motions such as rotation and deformation.
Funding Support, Disclosures, and Conflict of Interest: This work was supported through research funding from the National Institutes of Health (NIH-R01- CA169102) and Varian Medical Systems.
Contact Email: