From Slicer Wiki
Jump to: navigation, search
Home < Modules:UnbiasedNonLocalMeans-Documentation-3.4

Return to Slicer 3.4 Documentation

Gallery of New Features

Module Name


General Information

Module Type & Category

Type: Interactive

Category: CLI/DiffusionApplications

Authors, Collaborators & Contact

  • Author: Antonio Tristán Vega , Santiago Aja Fernández
  • Contact:

Module Description

Filters a set of diffusion weighted images using Unbiased Non Local Means for Rician noise. It exploits not only the spatial redundancy, but the redundancy in similar gradient directions as well; it takes into account the N closest gradient directions to the direction being processed (a maximum of 5 gradient directions is allowed to keep a reasonable computational load, since we do not use neither similarity maps nor block-wise implementation). The noise parameter is automatically estimated. A complete description of the algorithm may be found in "DWI filtering using joint information for DTI and HARDI", by Antonio Tristan Vega and Santiago Aja-Fernandez (under review).


Examples, Use Cases & Tutorials

Quick Tour of Features and Use

It is very easy to use it. Just select a DWI, set the parameters (if you really need it), and you're ready to go.

  • Input DWI Volume: set the DWI volume
  • Output DWI Volume: the filtered DWI volume
  • Estimation radius: This is the 3D radius of the neighborhood used for noise estimation. Noise power is estimated as the mode of the histogram of local variances
  • Search radius: This filter works by computing the weighted average of pixels in a large neighborhood. The search radius is the 3D radius of that neighborhood.
  • Comparison radius: The weights of the average are computed as the negative exponential of the (normalized) distance between two neighborhoods which have a 3D radius of this size.
  • h: The normalization constant for the distance between neighborhoods. This parameter is related to the noise power, and should be in the range 0.8->1.2. Higher values produce more blurring (and more noise reduction), while lower values better preserve edges.
  • Number of neighborhood gradients: This filter works gathering joint information from the N closest gradient directions to the one under study. This parameter is N. It is strongly recommended to keep this parameter less than 5 to keep a reasonable computational load. Even so, this filter is extremely slow.



Volumes. Needed to load DWI volumes

Known bugs

Usability issues

This filter is painfully slow. May take hours.

Source code & documentation

More Information


Antonio Tristan Vega, Santiago Aja Fernandez. University of Valladolid (SPAIN). Partially founded by grant number TEC2007-67073/TCM from the Comision Interministerial de Ciencia y Tecnologia (Spain).