BEST IN PHYSICS (IMAGING): Deep Learning-Based Motion Compensation for 4D-CBCT Reconstruction
Z Zhang1*, J Liu2, D Yang3, U Kamilov2, G Hugo1, (1) Washington University School of Medicine in St. Louis, St. Louis, MO, (2) Washington University in St. Louis, St. Louis, MO, (3) Duke University, Chapel Hill, NC
Presentations
TH-E-201-3 (Thursday, 7/14/2022) 1:00 PM - 2:00 PM [Eastern Time (GMT-4)]
Room 201
Purpose: Motion-compensated (MoCo) reconstruction shows great promise in improving 4D-CBCT quality, but MoCo will be more accurate with motion information at 4D-CBCT imaging time. Such ‘data-driven’ approaches are however hampered by the quality of the initial 4D-CBCT images used to build the motion model. The goal of this work is to improve the quality of reconstructed 4D-CBCT images through deep learning-generated motion models for MoCo.
Methods: A 3D artifact-removal convolutional neural network (CNN) was trained to improve conventional phase-correlated Feldkamp–Davis–Kress (PCF) reconstructions by reducing undersampling-induced streaking artifacts while maintaining motion information. The CNN-generated artifact-mitigated phase images (CNN only) were used to estimate a high-quality motion model for MoCo reconstruction (CNN+MoCo).The proposed method was evaluated using in-vivo patient datasets, a XCAT motion phantom dataset, and the publicly available SPARE challenge dataset. The quality of reconstructed 4D-CBCT images was quantitatively assessed through root-mean-square-error (RMSE) and normalized cross-correlation (NCC) metrics using phantom and SPARE datasets.
Results: The trained CNN effectively reduced the streaking artifacts of PCF images for all datasets. More detailed structures can be recovered using the proposed CNN+MoCo reconstruction. Phantom experiments showed that the accuracy of estimated motion model using CNN only images was greatly improved over PCF. CNN+MoCo showed lower RMSE and higher NCC compared to PCF, CNN only, and conventional MoCo. For the SPARE dataset, the RMSE in body region for PCF, CNN only, conventional MoCo and CNN+MoCo were 0.0038, 0.0029, 0.0026 and 0.0023. Corresponding NCC were 0.89, 0.94, 0.93 and 0.95.
Conclusion: CNN-based artifact removal can generate high-quality CBCT phase images for accurate motion modeling and improve the quality of MoCo reconstructed 4D-CBCT images.
Funding Support, Disclosures, and Conflict of Interest: Washington University receives research support from Varian Medical Systems unrelated to the present work.
Keywords
Cone-beam CT, Motion Artifacts, Reconstruction
Taxonomy
Contact Email