## Abstract / Description of output

Introduction

A critical step in the calculation of imaging biomarkers derived from functional, diffusion, perfusion and permeability MRI is the registration of the component

volumes that constitute these datasets. This step, which in its simplest form provides a rigid (6 d.o.f.) or full affine (12 d.o.f.) transformation of each volume to a predefined reference volume, is required to remove bulk patient motion and/or artifacts such as eddy current induced distortions so that each voxel represents spatially consistent tissue information across the acquisition. Such MR datasets typically comprise many tens of volumes, and contain up to 100 individual images, registration of which leads to a significant computational overhead in the processing pipeline; one which needs to be reduced if results are to be presented to clinicians in an acceptable time. In this abstract we present initial results from the application of a novel registration method based on the ‘Massively Optimized Parameter Estimation and Data compression’ (MOPED) algorithm, developed in the field of astronomy [1], which has the potential to reduce significantly the time taken to align high dimensional MRI data. The MOPED approach works by enabling very fast calculations of the likelihood of a given dataset. The speed-up is attained by a carefully designed data compression step, via construction of a set of optimized weighting vectors (y-vectors), which are designed to retain as much information as possible.

Given a small set of constraints, well met in the case of medical image registration, this compression step allows results of the same accuracy as would be obtained using the full dataset, but with several million times fewer calculations. The core algorithm also has the remarkable property that the calculation time is proportional to the number of parameters, for example 12 for a full 3D affine transformation, rather than the number of voxels, thereby allowing easy scaling from low to high resolution. To show the potential of this method, we present timings and χ2 maps obtained for MOPED, and for comparison FLIRT (FMRIB, Oxford, UK; http://www.fmrib.ox.ac.uk), when registering brain volumes acquired with different contrasts and acquisition matrices.

A critical step in the calculation of imaging biomarkers derived from functional, diffusion, perfusion and permeability MRI is the registration of the component

volumes that constitute these datasets. This step, which in its simplest form provides a rigid (6 d.o.f.) or full affine (12 d.o.f.) transformation of each volume to a predefined reference volume, is required to remove bulk patient motion and/or artifacts such as eddy current induced distortions so that each voxel represents spatially consistent tissue information across the acquisition. Such MR datasets typically comprise many tens of volumes, and contain up to 100 individual images, registration of which leads to a significant computational overhead in the processing pipeline; one which needs to be reduced if results are to be presented to clinicians in an acceptable time. In this abstract we present initial results from the application of a novel registration method based on the ‘Massively Optimized Parameter Estimation and Data compression’ (MOPED) algorithm, developed in the field of astronomy [1], which has the potential to reduce significantly the time taken to align high dimensional MRI data. The MOPED approach works by enabling very fast calculations of the likelihood of a given dataset. The speed-up is attained by a carefully designed data compression step, via construction of a set of optimized weighting vectors (y-vectors), which are designed to retain as much information as possible.

Given a small set of constraints, well met in the case of medical image registration, this compression step allows results of the same accuracy as would be obtained using the full dataset, but with several million times fewer calculations. The core algorithm also has the remarkable property that the calculation time is proportional to the number of parameters, for example 12 for a full 3D affine transformation, rather than the number of voxels, thereby allowing easy scaling from low to high resolution. To show the potential of this method, we present timings and χ2 maps obtained for MOPED, and for comparison FLIRT (FMRIB, Oxford, UK; http://www.fmrib.ox.ac.uk), when registering brain volumes acquired with different contrasts and acquisition matrices.

Original language | English |
---|---|

Title of host publication | Proceedings of the International Society for Magnetic Resonance in Medicine |

Publisher | The International Society for Magnetic Resonance in Medicine |

Pages | 671 |

Number of pages | 1 |

Publication status | Published - 1 May 2010 |