Meeting Banner
Abstract #3576

The Parallel Kalman Filter: An Efficient Tool to Deal with Real-Time Non Central χ Noise Correction

Veronique Brion1, Olivier Riff1, Maxime Descoteaux2, Jean-Franois Mangin1, Denis Le Bihan1, Cyril Poupon1, Fabrice Poupon1

1NeuroSpin, CEA/IBM, Gif-sur-Yvette, France; 2Sherbrooke University, Sherbrooke, Canada

This abstract proposes a novel real-time non central χ noise correction method for diffusion-weighted MR data that are known to be particularly sensitive to noise, as the diffusion indicator in the tissues corresponds to a signal loss. The technique is based on a Parallel Kalman Filter which is well adapted for non-Gaussian noise distributions, and which is as suitable for real time purposes as the standard Kalman filter. The results on simulated and real HARDI data show that it outperforms the standard Kalman Filter approach since non-Gaussian noise distributions are directly embedded in the process through their Gaussian mixture approximation.

Abstract PDF E-Poster Video

Keywords

account accuracy acquisitions adapted advised already analytical angular applied approximated architecture artificially assumes available ball basis better bias central channels characterizing cluster coefficients collapsed compatible completely confirmed constrains corrected correction correctly corrects correlations corrupted coupled deal decomposed decreased defines deliver density depicts derived detailed developed deviation diffusion directly distribution distributions effective efficient embedded embeds empirical employ employed environment error estimation estimator exists extended feedback fern field filter flaw free furthermore gained give gives global highest highly huge hypothesis idea improvement incremental injected introduced iteration keeps lead leaves like linear loop material measured media mixture model moreover neglects neighborhood nevertheless noise optimal orientation overcome panel parallel parallelization partial particular pixel position posteriori prevails problem processing produce propose proposed real really reduction refined regularization rely relying repetition representation retrieved robotic shot simplify simulated smoothing spatially spin square structures suitable systems take takes term tool towards trio upon view volume whereas whose workstation yielding