Tissue Segmentation

Tissue Segmentation
This plugin segments a brain MR volume into tissue classes and returns fuzzy memberships and hard classification. The algorithms currently available are FANTASM [1], K-means [2] and EM segmentation [3] and its variations. If the input image is skull-stripped, with the default parameters, the CSF, GM and WM memberships are generated.
Input and output

Input

The raw MR volume.

Output

All result images: Hard segmentation, fuzzy memberships and inhomogeneity field.

Hard segmentation: Label image with labels ranging from 0 to number of classes.

Fuzzy segmentation: Membership functions.

Both Fuzzy and Hard segmentation.

Parameters

Number of classes

Number of tissue classes to segment the volume

Algorithm:

FANTASM: Fuzzy and noise tolerant adaptive segmentation method. This algorithm generates an N-Class segmentation of an input MRI image. The core of the algorithm is an iterative approach that estimates the fuzzy classification of a tissue class. The approach is an adaptive Fuzzy C-Means algorithm (FCM) which can place additional constraints on the membership functions that force spatial smoothness. It can additionally compute an inhomogeneity field.

K means : Segments the image into K classes and returns the class labels.

EM : Expectation-maximization based segmentation using a Gaussian mixture model with equal variances and equal priors.

EMp : Expectation-maximization based segmentation using a Gaussian mixture model with equal variances and unequal priors.

EMpv : Expectation-maximization based segmentation using a Gaussian mixture model with unequal variances and unequal priors.

Correct Inhomogeneity

Inhomogeneity correction using a penalization on the first and second derivative of an inhomogeneity field, see [1] for more details.

Background Cropping

If this is true, the image is cropped to remove the background, depending on the background threshold. The subsequent computations are done on the cropped image. This is used mostly for memory efficiency.

Smoothing Parameter

The range is [0,1]. It controls the spatial smoothness of the resulting segmentation. Larger value means more spatial smoothness.

Maximum Difference

It is a termination criteria for the algorithm. The algorithm converges if the maximum difference between the memberships of subsequent iterations is smaller than this value. Setting this value too low might increase in runtime without significant improvement in the result.

Maximum Iteration

Another termination criteria for the algorithm. If the number of iteration becomes larger than this value, the algorithm stops.

Initialization

This is the initialization of the class centroids.

Range: The initial centroids are chosen based on an equal division of the image range. If the image has too many outliers, this is probably not a good idea. Use manual initialization instead.

Mode: The initial centroids are chosen as the modes of the image histogram. If the image histogram does not contain enough mode, or there is too much of partial voluming so that the modes are not very distinct, the initialization could be very wrong which will trap the algorithm in a local minima. Use manual initialization instead.

Manual: Manually input the initial centroids.

Masking Mode

One

All

Inhomogeneity Field Degree

Inhomogeneity is corrected based on a spline based field. This parameter denotes the degree of the spline.

Inhomogeneity Mode

Image : The inhomogeneity field is multiplied with the image intensity values in the objective function. See [1] for the objective function.

Centroid: The inhomogeneity field is multiplied with the image centroids in the objective function.

Separate: *Broken*, DO NOT USE.

Example Usage

Example input image

Hard Segmentation

CSF Memberships

GM Memberships

WM Membershipss

Inhomogeneity field

References

[1] D.L. Pham, "Spatial Models for Fuzzy Clustering," Computer Vision and Image Understanding, vol. 84, no. 2, pp. 285-297, 2001.

[2] K Means Clustering, "Wikipedia."

[3] A. P. Dempster, N. M. Laird and D. B. Rubin, "Maximum Likelihood from Incomplete Data via the EM Algorithm", Journal of the Royal Statistical Society. Series B (Methodological) vol. 39, no. 1, pp. 1-38. 1977.

Input	The raw MR volume.
Output	All result images: Hard segmentation, fuzzy memberships and inhomogeneity field. Hard segmentation: Label image with labels ranging from 0 to number of classes. Fuzzy segmentation: Membership functions. Both Fuzzy and Hard segmentation.

Number of classes	Number of tissue classes to segment the volume
Algorithm:	FANTASM: Fuzzy and noise tolerant adaptive segmentation method. This algorithm generates an N-Class segmentation of an input MRI image. The core of the algorithm is an iterative approach that estimates the fuzzy classification of a tissue class. The approach is an adaptive Fuzzy C-Means algorithm (FCM) which can place additional constraints on the membership functions that force spatial smoothness. It can additionally compute an inhomogeneity field. K means : Segments the image into K classes and returns the class labels. EM : Expectation-maximization based segmentation using a Gaussian mixture model with equal variances and equal priors. EMp : Expectation-maximization based segmentation using a Gaussian mixture model with equal variances and unequal priors. EMpv : Expectation-maximization based segmentation using a Gaussian mixture model with unequal variances and unequal priors.
Correct Inhomogeneity	Inhomogeneity correction using a penalization on the first and second derivative of an inhomogeneity field, see [1] for more details.
Background Cropping	If this is true, the image is cropped to remove the background, depending on the background threshold. The subsequent computations are done on the cropped image. This is used mostly for memory efficiency.
Smoothing Parameter	The range is [0,1]. It controls the spatial smoothness of the resulting segmentation. Larger value means more spatial smoothness.
Maximum Difference	It is a termination criteria for the algorithm. The algorithm converges if the maximum difference between the memberships of subsequent iterations is smaller than this value. Setting this value too low might increase in runtime without significant improvement in the result.
Maximum Iteration	Another termination criteria for the algorithm. If the number of iteration becomes larger than this value, the algorithm stops.
Initialization	This is the initialization of the class centroids. Range: The initial centroids are chosen based on an equal division of the image range. If the image has too many outliers, this is probably not a good idea. Use manual initialization instead. Mode: The initial centroids are chosen as the modes of the image histogram. If the image histogram does not contain enough mode, or there is too much of partial voluming so that the modes are not very distinct, the initialization could be very wrong which will trap the algorithm in a local minima. Use manual initialization instead. Manual: Manually input the initial centroids.
Masking Mode	One All
Inhomogeneity Field Degree	Inhomogeneity is corrected based on a spline based field. This parameter denotes the degree of the spline.
Inhomogeneity Mode	Image : The inhomogeneity field is multiplied with the image intensity values in the objective function. See [1] for the objective function. Centroid: The inhomogeneity field is multiplied with the image centroids in the objective function. Separate: Broken, DO NOT USE.

Example input image
Hard Segmentation
CSF Memberships
GM Memberships
WM Membershipss
Inhomogeneity field

Example Usage Example input image Hard Segmentation CSF Memberships GM Memberships WM Membershipss Inhomogeneity field

References

Example Usage

Example input image

Hard Segmentation

CSF Memberships

GM Memberships

WM Membershipss

Inhomogeneity field