Volume Based Evaluation of Geometric Distortion in Magnetic Resonance Imaging Using 3D Template Matching
K Huang1*, Y Cao1, U Baharom2, J Balter1, (1) University of Michigan, Ann Arbor, MI, (2) Integrated Medical Technologies, Troy, NYMO-B-Salon EF-7 Monday 10:00:00 AM - 12:00:00 PM Room: Salon EF
Purpose: Treatment planning with solely MR images is drawing increased interest in radiation therapy. One of the major issues potentially limiting such use is the possibility of geometric distortion inherent in MR images. Previously the evaluation of MR distortion has been either done in 2D or in 3D using 2D images in AP, SI, and LR planes. We proposed a method to determine the distortion of MR image volume in a real 3D approach.
Methods: A large phantom with 3D array of spheres filled with contrast was scanned with a 3T MRI simulator (Siemens Skyra) using T1 3D-Vibe protocol (TR 4.39ms and TE 2.03ms). 3D template matching using templates of sphere and half spheres with different orientations were used to locate the sphere centers. The normalized cross-correlation (NCC) value measures how well the image pattern is matched with templates. The half-sphere templates are used when the NCC value is below threshold. To remove the shift of the sphere centers induced by the physical rotation and translation of the phantom, the Procrustes method was applied. Then the distortion at each sphere center is calculated and the distortion map of the whole image space was generated by interpolation.
Results: Among the half-sphere templates used in matching, down half-sphere was most frequently used (83%), followed by left/right half-sphere (8% each). Among all the sphere centers, 80% have distortion less than 1mm, 14% have 1~1.5mm distortion, 4.4% have 1.5~2mm distortion, and 1.6% have distortion more than 2mm. The coordinates of the location closest to center with >2mm distortion is at (11.0, -17.8, -4.7).
Conclusions: This work successfully presented 3D template matching is effective to locate the coordinates of a sphere center in 3D simultaneously. With this method, a high fidelity MR distortion map was able to be generated in less than a minute.