Abstract
Purpose: To register CT to MR images of the antebrachial region, with large deformations or motion between scans, by use of rigid body registration and subsequent interpolation of soft tissue.
Methods: A workflow was designed in which deformable BSpline registrations were initialized using rigid body transformations for bone tissue, and a dual quaternion based interpolation scheme for soft tissue. Rigid body transformations were calculated using segmentations of the humerus, radius and ulna bones. These were registered using an Iterative Closest Point (ICP) algorithm. CT and MR images of 18 in vivo scanned arms were registered using this workflow, and compared to a traditional workflow consisting of rigid and deformable BSpline registration in elastix. The registrations were evaluated quantitatively by segmenting water, fat and bone tissue on both CT and MR in order to calculate Dice similarity coefficient.
Results: Preliminary results of one registered arm using the proposed workflow showed a Dice coefficient of 0.79, 0.68 and 0.79 for the water, fat and bone tissue respectively, averaging 0.75 over all tissue. For comparison, Dice coefficients using the traditional workflow were 0.74, 0.68 and 0.42, averaging 0.62.
Conclusion: The preliminary results showed in increase in Dice coefficient when using the proposed method. Due to the lack of a golden standard for the segmentation of the CT and MR images the results could only be interpreted relative to each other. Qualitative inspection of the results of the proposed workflow indicate regions with considerable registration errors at locations that are distant from the bone. Registration errors also occur close to the edge of the image, caused by a difference in field of view between the CT and MR images. However, tissue close to the bone is often well registered. Appropriate cropping of the images could improve the results of the registration.
Methods: A workflow was designed in which deformable BSpline registrations were initialized using rigid body transformations for bone tissue, and a dual quaternion based interpolation scheme for soft tissue. Rigid body transformations were calculated using segmentations of the humerus, radius and ulna bones. These were registered using an Iterative Closest Point (ICP) algorithm. CT and MR images of 18 in vivo scanned arms were registered using this workflow, and compared to a traditional workflow consisting of rigid and deformable BSpline registration in elastix. The registrations were evaluated quantitatively by segmenting water, fat and bone tissue on both CT and MR in order to calculate Dice similarity coefficient.
Results: Preliminary results of one registered arm using the proposed workflow showed a Dice coefficient of 0.79, 0.68 and 0.79 for the water, fat and bone tissue respectively, averaging 0.75 over all tissue. For comparison, Dice coefficients using the traditional workflow were 0.74, 0.68 and 0.42, averaging 0.62.
Conclusion: The preliminary results showed in increase in Dice coefficient when using the proposed method. Due to the lack of a golden standard for the segmentation of the CT and MR images the results could only be interpreted relative to each other. Qualitative inspection of the results of the proposed workflow indicate regions with considerable registration errors at locations that are distant from the bone. Registration errors also occur close to the edge of the image, caused by a difference in field of view between the CT and MR images. However, tissue close to the bone is often well registered. Appropriate cropping of the images could improve the results of the registration.
| Original language | English |
|---|---|
| Publication status | Published - 25 Jan 2019 |
| Event | Dutch Biomedical Engineering Conference - Egmond aan Zee, Netherlands Duration: 24 Jan 2019 → 25 Jan 2019 |
Conference
| Conference | Dutch Biomedical Engineering Conference |
|---|---|
| Country/Territory | Netherlands |
| City | Egmond aan Zee |
| Period | 24/01/19 → 25/01/19 |