Deep Learning-Based Simultaneous Multi-Phase Deformable Image Registration of Sparse 4D-CBCT
I Herzig1*, P Paysan2, S Scheib2, F-P Schilling3, J Montoya3, M Amirian3, T Stadelmann3, P Eggenberger1, R Fuechslin1, L Lichtensteiger1, (1) Zurich University of Applied Sciences ZHAW, Institute for Applied Mathematics and Physics IAMP, Winterthur, CH (2) Varian Medical Systems Imaging Laboratory, Daettwil AG, CH (3) Zurich University of Applied Sciences ZHAW, Centre for Artificial Intelligence CAI, Winterthur, CH
Presentations
MO-H345-IePD-F2-4 (Monday, 7/11/2022) 3:45 PM - 4:15 PM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 2
Purpose: Respiratory gated 4D-CBCT suffers from sparseness artefacts caused by the limited number of projections available for each respiratory phase/amplitude. These artefacts severely impact deformable image registration methods used to extract motion information. We use deep learning-based methods to predict displacement vector-fields (DVF) from sparse 4D-CBCT images to alleviate the impacts of sparseness artefacts.
Methods: We trained U-Net-type convolutional neural network models to predict multiple (10) DVFs in a single forward pass given multiple sparse, gated CBCT and an optional artefact-free reference image as inputs. The predicted DVFs are used to warp the reference image to the different motion states, resulting in an artefact-free image for each state. The supervised training uses data generated by a motion simulation framework. The training dataset consists of 560 simulated 4D-CBCT images of 56 different patients; the generated data include fully sampled ground-truth images that are used to train the network. We compare the results of our method to pairwise image registration (reference image to single sparse image) using a) the deeds algorithm and b) VoxelMorph with image pair inputs.
Results: We show that our method clearly outperforms pairwise registration using the deeds algorithm alone. PSNR improved from 25.8 to 46.4, SSIM from 0.9296 to 0.9999. In addition, the runtime of our learning-based method is orders of magnitude shorter (2 seconds instead of 10 minutes). Our results also indicate slightly improved performance compared to pairwise registration (delta-PSNR=1.2). We also trained a model that does not require the artefact-free reference image (which is usually not available) during inference demonstrating only marginally compromised results (delta-PSNR=-0.8).
Conclusion: To the best of our knowledge, this is the first time CNNs are used to predict multi-phase DVFs in a single forward pass. This enables novel applications such as 4D-auto-segmentation, motion compensated image reconstruction, motion analyses, and patient motion modeling.
Funding Support, Disclosures, and Conflict of Interest: Co-financed by Innosuisse, grant no. 35244.1 IP-LS. S.S. and P.P. are full-time employees of Varian Medical Systems Imaging Laboratory GmbH
Keywords
Cone-beam CT, Deformation, Registration
Taxonomy
Contact Email