Room: Exhibit Hall | Forum 6
Purpose: To develop a deformable image registration (DIR) framework for large neck motion in multimodal images using a neural network to create a MR-derived synthetic CT.
Methods: 25 aligned MR-CT pairs were acquired. 19 pairs were used for training a generative adversarial network to create MR-derived synthetic CTâ€™s. The remaining 6 patients were used for registration and evaluation. The synthetic CTâ€™s replaced the MRâ€™s in the DIR to create a pseudo monomodality registration. Each patientâ€™s MR was registered with a CT in a different head pose, in order to challenge the registration. The registration used a multi-resolution b-spline approach with mutual information. The direct (using MR) and synthetic CT based registration were performed in both directions (MR-CT and CT-MR), and were evaluated using the mean distance to agreement (MDA) in spinal cord contours, 11 landmarks, and the Jacobian determinant between the target and deformed source images.
Results: Replacing the MR with the synthetic CT led to an improvement in registration, with greater improvement in the MR-to-CT direction. The average landmark error decreased from 11mm to 6mm and 8mm to 7mm in the MR-to-CT and CT-to-MR directions, respectively. The Jacobian determinant was never negative (no anatomically infeasible folding). The average spinal cord MDA improvement was 3mm and 1mm for the MR-to-CT and CT-to-MR directions, respectively. There was larger relative improvement when the initial misalignment was worse between the MR and CT images.
Conclusion: We showed that replacing the MR with a synthetic CT in multimodal deformable image registration offers improvement to directly registering the MR and CT. We showed that this method works even in the presence of different head poses. Future work will include more patients and different anatomical sites.
Funding Support, Disclosures, and Conflict of Interest: NIH R44CA183390, NIH R01CA188300