Robust Regularization for the Estimation of Intra-Voxel Axon Fiber Orientations Also presented at MMBIA Anchorage, Jun, 2008 Alonso Ramirez-Manzanares (PICSL) Hui Zhang (PICSL) Mariano Rivera (CIMAT) James C. Gee (PICSL)
Overview Motivation Statement of the problem Our Proposal Results for in-vivo human data Results for synthetic data Conclusions
MOTIVATION
Motivation (1/3): Intra-voxel fiber orientations. Behrens et al, Neuroimage'07 “We detect complex fibre architecture in approximately a third of voxels with an FA greater than 0.1” DT Multi-DTs
Motivation (2/3): The noisy orientations Because of: - acquisition noise - a reduced # of diffusion encoding orientations (clinical applications) - patient movement
Motivation(3/3): Data averaging and Spatial integration From web site: In our case:
THE PROBLEM
The Problem (1/2): The spatial regularization of directional fields This is a well- known task, for instance, in Optical Flow computation.
The Problem (2/2): The spatial regularization of multi-fiber orientation fields. Problems: a) The need to regularize orientations (not directions) b) The need to match orientations c) The need to use indicator variables of the number of bundles d) The subtle axon fiber structures
OUR PROPOSAL
Proposal(1/5): Observation model, Diffusion Basis Function (DBF) approach. Ramirez-Manzanares et al. IEEE-TMI '07 Tuch et al, MRM '02 Tensor Basis DBFs
Proposal(2/5): The robust spatial regularization term Inspired in statistical robust regression
Proposal(3/5): The robust spatial regularization term Indicator variables Robust Weights Robust regularization
Proposal(4/5): Single DT as a diffusivity profile constraint Plausible Implausible solution solution
Proposal(5/5): The Integration of terms and methods Data and contrast term Ramirez-Manzanares et al TMI’07
RESULTS FOR IN-VIVO HUMAN DATA
Results(1/6): In-vivo human data (b=1000, 60 DWI) Non-Regularized Robust Regularized
Results(2/6): In-vivo human data, a closer view Non-Regularized Robust Regularized DT
Results(3/6): In-vivo human data, a closer view Non-Regularized Robust Regularized DT
RESULTS FOR SYNTHETIC DATA
Results(4/6): Synthetic data Robust Regularized Iteration 1 Robust weights Iteration 1 Robust weights Iteration 3 Non Regularized DT Robust Regularized Iteration 3
Results(5/6): Realistic/complex synthetic data Ground truth non-Regularized SNR=20 DT Iteration 1 Iteration 2 Iteration 3
Results(6/7): Synthetic data, quantitative results
Results(7/7): Comparison: Robust vs. non-Robust Non-Robust Regularized Ramirez-Manzanares et al TMI’07 Robust Regularized
Conclusions
Thank you for your attention! Questions?