Robust estimator framework in diffusion tensor imaging

Diffusion of water molecules in the human brain tissue has strong similarities with diffusion in porous media. It is affected by different factors such as restrictions and compartmentalization, interaction with membrane walls, strong anisotropy imposed by cellular microstructure, etc. However, multiple artefacts abound in in vivo measurements either from subject motions, such as cardiac pulsation, bulk head motion, respiratory motion, and involuntary tics and tremor, or hardware related problems, such as table vibrations, etc. All these artefacts can substantially degrade the resulting images and render postprocessing diffusion analysis difficult or even impossible. In order to overcome these problems, we have developed a robust and efficient approach based on the least trimmed squares algorithm that works well with severely degraded datasets with low signal-to-noise ratio. This approach has been compared with other diffusion imaging post-processing algorithms using simulations and in vivo experiments. We demonstrate that the least trimmed squares algorithm can be easily adopted for multiple non-Gaussian diffusion models such as the biexponential model. The developed approach is shown to exhibit a high efficiency and accuracy and can, in principle, be exploited in other diffusion studies where artefact/outlier suppression is demanded.