Program Information
A Real-Time 3D Ultrasound Tracking Algorithm with Simultaneous Templates for Use in Radiotherapy
A Shepard1*, B Wang2 , D Mills2 , R Darrow2 , H Chan2 , T Foo2 , B Bednarz1 , (1) University of Wisconsin, Madison, WI, (2) GE Global Research Center, Niskayuna, NY
Presentations
SU-E-601-4 (Sunday, July 30, 2017) 1:00 PM - 1:55 PM Room: 601
Purpose: To develop and evaluate a novel real-time 3D ultrasound tracking algorithm for the localization of key structures that aid in the identification of PTV motion throughout radiotherapy treatments.
Methods: A learning-based 3D tracking algorithm was developed using the principles of block-matching with normalized cross-correlation as a similarity metric. As each new frame is acquired, the algorithm uses globalized block-matching to determine gross displacements of the entire feature, and localized block-matching at strategic locations along the boundary of the target to account for local deformations. The localized matching is performed through simultaneous use of templates from both the previous volume and a training set volume to provide both temporal (previous) and shape (training) constancy to the tracked contour. A rolling training set is obtained during tracking by storing final contours for the first several volumes to aid in the accurate tracking of subsequent volumes. GPU computing is used to address computationally expensive aspects of implementation. The algorithm was tested on three ultrasound image sequences of the liver ranging from 53-58 seconds (497-1008 frames). For each sequence, tracking was performed on a vessel bifurcation and the tracking performance was characterized relative to manual annotations on 10% of the frames. Annotations were performed by three users and the mean result was taken as ground truth.
Results: A mean tracking error of 1.31 mm was obtained using a constant set of parameters for all sequences. The average analysis rate was 16.6 frames per second (fps) during training set accumulation and 18.1 fps post-accumulation with image sets preloaded into memory. The analysis rate exceeded the imaging rate for all sequences.
Conclusion: The algorithm displayed accuracies sufficient for tracking in radiotherapy while processing 3D ultrasound volumes at rates exceeding the 4D ultrasound imaging rates, making it a viable option for potential use in radiotherapy guidance.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by NIH grant R01CA190298
Contact Email: