Training An Unsupervised Deformable Image Registration Model with Thoracic Fast, Low-Dose Helical CT Acquisitions
R Savjani1,3*, P Paysan2, D O'Connell1, J van Heteren3, S Scheib2, D Low1, (1) University of California Los Angeles, Los Angeles, CA (2) Varian Medical Systems Imaging Laboratory, Baden - Daettwil, CH (2) Varian Medical Systems Imaging Laboratory, Baden - Palo Alto, CA (3)
Presentations
MO-IePD-TRACK 4-5 (Monday, 7/26/2021) 5:30 PM - 6:00 PM [Eastern Time (GMT-4)]
Purpose: Monitoring respiratory motion is critical while patients are undergoing thoracic radiotherapy. Low-dose, fast helical CT scans can capture the variability in respiratory motion while patients are free breathing. These scans have been used previously to fit a model (5DCT) to derive CT volumes from a baseline CT volume, as well as the tidal volume and air flow derived from an external surrogate marker. Here, we leverage these data to apply an unsupervised deep-learning based deformable image registration algorithm (VoxelMorph) to build towards a motion management model.
Methods: We obtained data from 32 retrospective patients (27 for training, 5 for validation) each scanned with 25 fast, low-dose helical CT acquisitions. A VoxelMorph model was trained at 2-mm resolution using mean square error for image similarity loss with regularization. A random pair of images within each patient served as input, training randomly across patients for a total of 25,000 pairs. The generated deformation vector fields from the model were used to warp each structure to a selected reference volume for each patient. Dice scores between the warped structure and an autosegmentation were computed and also compared to a state-of-the-art algorithm, deeds.
Results: Median Dice scores were higher than that of deeds on the 5 independent test sets for 19 of 20 structures (stomach was lower). Median Dice scores ranged from 76.2% in the duodenum to >95% in the right and left lungs. On inference, deformations could be generated on the 5DCT VoxelMorph model in less than 1 second on a GPU, whereas deformations for deeds took approximately 10 minutes on CPUs.
Conclusion: Fast, helical free breathing scanning provides a rich dataset to train deep-learning based image deformation models. Together, this framework is being developed towards developing a rapid motion model to help predict respiratory motion in real-time for radiotherapy.
Funding Support, Disclosures, and Conflict of Interest: ASTRO-Varian Research Training Fellowship awarded to Ricky Savjani to conduct this research in collaboration with Varian Medical Systems and UCLA.
ePosters
Keywords
Registration, MOSFET Detector, Motion Artifacts
Taxonomy
Contact Email