MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Estimating Target Registration Error for Automated Deformable Image Registration QA

J Sage*, D Boukerroui, M Gooding, Mirada Medical, Denver, CO

Presentations

(Thursday, 7/16/2020) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]

Room: Track 5

Purpose: To validate a deformable image registration (DIR) QA tool using local similar points (LSP) against the POPI standard.

Methods: Image features were detected in the 10% and 50% phase images of the POPI data using a volumetric salient feature detector. The region around each feature was encoded using a binary descriptor. Features within each 3D image were matched to the most similar feature on the other image within a geometric distance threshold using the test DIR to relate the image geometries. Matches were rejected if not reciprocal or if not unique within the distance range. TRE was measured using the manual POPI landmarks for the undeformed data, the POPI Demons registration and optical flow registration (Mirada Medical Ltd). TRE was also estimated using the LSP selected points, both at the position of POPI landmarks and at locations on a fixed grid across the image.

Results: Using the POPI landmarks the mean(standard deviation) TRE in mm for the undeformed, Demons and optical flow cases were 5.8 (2.6), 1.3 (0.5) and 1.1 (0.5) respectively. Using LSP, 5000 matched features were detected. At the POPI landmark positions the TRE values estimated were 5.8 (2.3), 1.0 (0.4) and 0.9 (0.5). For the regular grid points the TRE values estimated using LSP were 4.0 (2.9), 1.9 (1.4) and 1.8(1.3) respectively.

Conclusion: LSP produced estimations of registration error in close agreement with errors from manual landmark positions. Errors estimated at points away from the manual landmarks are greater. Manual landmarks will be self-selected as well-defined points where registration algorithms might be expected to perform well. LSP shows potential as a tool to automate the patient specific QA of deformable registration.

Funding Support, Disclosures, and Conflict of Interest: Funded by Innovate UK, Grant no. 105205

Keywords

Registration, Deformation, Quality Assurance

Taxonomy

IM/TH- Image Registration: CT

Contact Email