Y Zhao*, X Chen, C Yu, L Court, B McDonald, T Pan, H Wang, X Wang, J Phan, J Yang, UT MD Anderson Cancer Center, Houston, TX

• 
  Y Zhao

• 
  
• 
  
• 
  
• 
  
• 
  
• 
  
• 
  

SU-H400-IePD-F8-5 (Sunday, 7/10/2022) 4:00 PM - 4:30 PM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 8

Purpose: To improve the accuracy and speed of CT-MR and MR-MR registration for head-and-neck MR-Linac treatments in clinic through an automated deformable image registration (DIR) framework.

Methods: We developed a hierarchical registration framework. Following a whole-volume rigid registration, the input images were divided into overlapping patches. Then a patch-based rigid registration was applied to achieve sufficient local alignment for subsequent DIR. We developed a ViT-Morph model, a combination of a convolutional neural network (CNN) and the Vision Transformer (ViT), for the patch-based DIR. The modality independent neighborhood descriptor (MIND) was adopted in our model as the similarity metric to account for both inter-modality and intra-modality registration. The CT-MR and MR-MR DIR models were trained with 236 CT-MR and 212 MR-MR image pairs from 31 patients, respectively, and both were tested with 24 image pairs (CT-MR and MR-MR) from another 6 patients. The registration performance was evaluated with 7 manually contoured organs (brainstem, spinal cord, mandible, left/right parotids, left/right submandibular glands) using Dice similarity coefficient (DSC) and mean surface distance (MSD) by comparing with the popular deep learning-based DIR framework, Voxelmorph, using MIND as the similarity metric.

Results: The average DSC and MSD calculated over all organs between the deformed contours and the reference manual contours were 0.76±0.05/1.91±0.47mm for CT-MR registration and 0.86±0.06/0.92±0.34mm for MR-MR registration, respectively. Our method outperformed VoxelMorph by 6% for CT-CT registration, and 4% for MR-MR registration based on DSC measurements. Among all organs, right parotid achieved the best results (0.81±0.03/1.99±0.47mm) while mandible had the worst (0.71±0.04/1.87±0.27mm) in CT-MR registration; and brainstem had the best results (0.93±0.02/0.63±0.18mm) while right submandibular gland had the worst (0.84±0.05/0.98±0.33mm) in MR-MR registration.

Conclusion: We developed a fully automated hierarchical registration framework that achieved notably improved DIR accuracy of both CT-MR and MR-MR registration for head-and-neck MR-guided adaptive radiotherapy.

Keywords

Registration, MRI, Image-guided Therapy

Taxonomy

IM/TH- Image Registration: Multi-modality registration

Contact Email