An optimized feature detector for markerless motion tracking in motion-compensated neuroimaging

Abstract

Head movements during PET and MRI scans can have a detrimental effect on image quality and quantitative measurements. For both modalities, motion correction methods exist that rely on accurate characterization of head motion. In the case of prospective correction in MRI, the motion estimates also need to be delivered in real-time. Motion tracking methods that rely on attached markers are susceptible to decoupling of the head and marker, hinder clinical workflow, and have line of sight issues due to the geometry of the bore and headcoil. In this study, we aim to optimize a methodology that measures head motion by detecting and tracking SIFT features native to the forehead. These features can be extracted and described in many ways, with different algorithms offering varying levels of computational efficiency and robustness to scene changes. A phantom study was performed to assess the accuracy and speed performance of five different feature detectors: SIFT, SURF, ORB, BRISK and AKAZE. Except for ORB, position estimates obtained using the different feature detectors showed similar agreement (error <;0.4 mm) with the ground-truth robot measurements. Processing time varied, with SURF, BRISK and AKAZE offering a substantial speed increase over SIFT while maintaining similar accuracy. We conclude that SURF, BRISK and AKAZE appear to be suitable alternative feature detectors to SIFT for prospective motion correction in MRI and MRI-PET.