Program Information
Validation of Image Registration Methods for Brain Magnetic Resonance Imaging
J Lin1 , D Fuentes1 , A Chandler2 , J Hazle1 , D Schellingerhout1 , C MacLellan1*, (1) The University of Texas MD Anderson Cancer Center, Houston, Texas, (2) GE Healthcare, Waukesha, WI
Presentations
TU-H-CAMPUS-IeP3-3 (Tuesday, August 2, 2016) 5:30 PM - 6:00 PM Room: ePoster Theater
Purpose: To assess the performance of both commercial and open-source co-registration solutions as applied to intra-subject, multi-sequence, magnetic resonance (MR) images of the brain.
Methods: Twenty (20) patients were imaged on clinical 3.0T MR scanners to obtain T2-weighted (T2W), fluid attenuation inversion recovery (FLAIR), susceptibility weighted angiography (SWAN), and T1 post-contrast (T1C) image sequences. Fiducial landmark sites (n=15 per sequence, 4 sequences per patient, 1200 total planned for 4 sequences and 20 patients, 1175 total realized) were specified throughout these image volumes to define identical locations across sequences. Both commercial (General Electric [GE] VolumeViewer) and open-source software (Advanced Normalization Tools [ANTs]) registration solutions were applied using the T2W sequence as the fixed reference. Landmark and image similarity (cross-correlation [CC], mutual information [MI]) based registrations were performed for all image pairs. Rigid (6DOF) and affine (12DOF) transformations were considered. The Euclidean Target-to-Registration Error (TRE) was calculated at all landmarks of each image pair.
Results: Prior to registration, TRE values for FLAIR, SWAN, and T1C were 2.07 ± 0.55 mm, 2.63 ± 0.62 mm and 3.65 ± 2.00 mm, respectively. Post-registration, the best (smallest) average TRE values for FLAIR, SWAN, and T1C were 1.55 ± 0.46 mm (rigid MI), 1.34 ± 0.23 mm (affine MI) and 1.06 ± 0.16 mm (GE), respectively.
Conclusion: This study presents a methodology to quantify the registration accuracy of commercial algorithms (whose figures of merit are often not publically available, despite clinical use), and compares that accuracy to open-source alternatives. All sequences, on average, were improved by spatial registrations that corrected for patient motion, and such motion itself was found to increase with time spent in the MR scanner. The neuroanatomical information encoded in these landmarks, as placed on images with different contrast mechanisms, collectively represents a comprehensive dataset for quantitative evaluation of clinically-used registration software.
Funding Support, Disclosures, and Conflict of Interest: A.C. is an employee of GE Healthcare. This research was supported in part by the National Cancer Institute (Cancer Center Support Grant CA016672 and Training Grant 5T32CA119930). J.S.L. acknowledges support from the Baylor College of Medicine Medical Scientist Training Program and the Cullen Trust for Higher Education Physician/Scientist Fellowship Program.
Contact Email: