Click here to

Session: Multi-Disciplinary: Segmentation III [Return to Session]

Segmentation by Individualized Registration (SIR) for CBCT-Based Adaptive Radiation Therapy

X Liang1*, J Chun2, H Morgan1, T Bai1, D Nguyen1, J Park1, S Jiang1, (1) Medical Artificial Intelligence and Automation (MAIA) Laboratory, Department of Radiation Oncology, University of Texas Southwestern Medical Center, Dallas, TX, (2) Department of Radiation Oncology, Yonsei University, Seoul, KR


TH-IePD-TRACK 3-4 (Thursday, 7/29/2021) 3:00 PM - 3:30 PM [Eastern Time (GMT-4)]

Purpose: The purpose of this study is to develop a method that can automatically generate segmentations on cone-beam CTs (CBCT) for head and neck adaptive radiation therapy (ART). Due to the prevalence of artifacts and truncations on CBCTs, directly contouring on CBCTs is very challenging. Thus, we propose to utilize a learning based deformable image registration method and contour propagation to get updated contours on CBCTs, where expert-drawn contours on the planning CT (pCT) can serve as prior knowledge.

Methods: Different from other deep learning based image registration methods which are generally trained on a large datasets and tested on separate data, our proposed method uses deep learning architectures as a function f taking only a pair of moving and fixed images (pCT and CBCT in our case) as input and yielding warped moving images as outputs. The function f is composed of two parts: deformation vector field (DVF) generation and spatial transformer. DVF generation can predict a DVF from a pair of pCT and CBCT, then spatial transformer can warp pCT to deformed pCT (dpCT) based on predicted DVF. The function f can be learned by minimizing the cost function of image similarity between fixed and warped images plus regularization terms.

Results: When compared with manual segmentations by a radiation therapy expert, the average dice coefficients of organ-at-risks and target by Elastix, 3DSlicer Bspline deformable registration, 3DSlicer demon deformable registration, FAIM, Voxelmorph, 5-cascaded Voxelmorph, VTN, and 10-cascaded VTN are 0.84, 0.83, 0.80, 0.83, 0.80, 0.84, 0.81, and 0.84 respectively. However, the average dice coefficients of Voxelmorph, 5-cascaded Voxelmorph, VTN, and 10-cascaded VTN using our proposed training strategy can reach up to 0.83, 0.85, 0.85, and 0.86.

Conclusion: Our proposed method outperforms both the state-of-the-art learning based and traditional based image registration methods.

Funding Support, Disclosures, and Conflict of Interest: This study is funded by Varian Medical Systems Inc.



    Cone-beam CT, Segmentation, Registration


    IM/TH- Image Registration Techniques: Nonrigid registration (other)

    Contact Email