(Tuesday, 7/16/2019) 10:30 AM - 11:00 AM
Room: Exhibit Hall | Forum 9
Purpose: To compare correlation-based tracking of gradient features against features selected by deep neural network (DNN), with the goal of identifying which image features are good for kV X-ray-based soft tissue monitoring during free-breathing MV radiation.
Methods: Test images and corresponding ground truth data were obtained from XCAT simulations applied to the abdominal area in XCAT phantom volumes. Realistic x-ray projections, containing kV and other noise sources, were generated using the iTools software package (Varian, Palo Alto CA) to modify the XCAT output. Short-arc CBCT images with different arc lengths and imaging angles were then reconstructed from the projections using the FDK algorithm. Finally, central slices of volumes in each simulated motion cycle were concatenated to create a 2D sequence of images with known organ positions. Image features generated from DNN and from traditional image filters were used as input to correlation-based tracking algorithms.
Results: Tracking accuracy was evaluated in terms of success rate (at 3 mm precision) and Dice coefficients. The success rate and averaged Dice for tracking with DNN features were 72% and 0.80, and for the non-DNN features were 78% and 0.85. Processing speeds for 3D volumes were equivalent to 0.75 fps for DNN and 2.1 fps for non-DNN. This showed that using non-DNN features is accurate, robust, and promising for 3D applications. Further analysis demonstrated that using DNN features for high-accuracy soft tissue tracking has its limitations they are not shift invariant.
Conclusion: This study showed that correlation-based algorithms using traditional image features continue to be promising for abdominal soft tissue monitoring.
Funding Support, Disclosures, and Conflict of Interest: All authors are employees of Varian Medical Systems, Inc.
Keywords
Image Correlation, Organ Motion, Cone-beam CT
Taxonomy
IM- Cone Beam CT: Machine learning, computer vision
Contact Email
