FeatureSpaceMatrix

MLModule

genre	FiberClustering
authors	Jan Klein, Philip Bittihn, Peter Ledochowitsch
package	FMEwork/Release
dll	MLFiberClustering
definition	MLFiberClustering.def
keywords	DTI, fiber, clustering, feature, space, matrix

Purpose

The module FeatureSpaceMatrix calculates the feature space matrix from a fiber set which may be used for fiber clustering.

Windows

Default Panel

Input Fields

inputFiberSet

name: inputFiberSet, type: MLBase

A FiberSet or FiberSetContainer.

Output Fields

outFeatureMatrix

name: outFeatureMatrix, type: MLBase

The feature space matrix with dimension numObjects x numFeatures.

outFeatureWeightVector

name: outFeatureWeightVector, type: MLBase

Feature weights used in ClusteringAlgorithms stored as a vector with dimension (numFeatures x numSegments).

Parameter Fields

Field Index

&Apply: Trigger	EY: Float
Cov XX: Float	EZ: Float
Cov XY: Float	Feature Space: Enum
Cov XZ: Float	Number of Segments: Integer
Cov YY: Float	XMean Direction: Float
Cov YZ: Float	YMean Direction: Float
Cov ZZ: Float	ZMean Direction: Float
EX: Float

Visible Fields

Feature Space

name: featureSpace, type: Enum, default: Standard

Defines the feature space in which to work in.

Values:

Title	Name	Description
Standard	Standard	Feature space corresponding to input fibers and selected feature weights.
Feature Space 1	Feature Space 1	Not implemented: arbitrary feature spaces could be added in the future.

Number of Segments

name: numSegments, type: Integer, default: 1

Sets the number of target segments.

Each fiber is partitioned into the desired number of segments, independently of its length. Then each segment is treated as a short fiber and the whole feature space is calculated and the feature spaces combined (numFeatures x numSegments resulting columns).

However this procedure does not yield great results: the algorithms (in ClusteringAlgorithms), which compute the affinity matrix , sometimes end up comparing two fibers starting from opposite ends. In addition a high dimensional feature space usually leads to many weights being of the same order of magnitude and therefore to a non-sparse affinity matrix and ambiguous clustering.

XMean Direction

name: xMeanDirection, type: Float, default: 0

Sets the weighting of mean x-components.

YMean Direction

name: yMeanDirection, type: Float, default: 0

Sets the weighting of mean y-components.

ZMean Direction

name: zMeanDirection, type: Float, default: 0

Sets the weighting of mean z-components.

Cov XX

name: covXX, type: Float, default: 1

Sets the weighting of cov(xx).

Cov YY

name: covYY, type: Float, default: 1

Sets the weighting of cov(yy).

Cov ZZ

name: covZZ, type: Float, default: 1

Sets the weighting of cov(zz).

Cov XY

name: covXY, type: Float, default: 1

Sets the weighting of cov(xy).

Cov XZ

name: covXZ, type: Float, default: 1

Sets the weighting of cov(xz).

Cov YZ

name: covYZ, type: Float, default: 1

Sets the weighting of cov(yz).

EX

name: eX, type: Float, default: 1

Sets the weighting of ex.

EY

name: eY, type: Float, default: 1

Sets the weighting of ey.

EZ

name: eZ, type: Float, default: 1

Sets the weighting of ez.

&Apply

name: apply, type: Trigger

If pressed, the module computes anew.