Features

FARSIGHT contains a rich and constantly growing library of software routines to compute diverse measurements from images. Measurements are also commonly referred to as "features" in the image analysis community. Broadly speaking, there are two classes of features in FARSIGHT:

1. Intrinsic Features: Intrinsic features are measurements that quantify aspects of the morphology and/or appearance of objects in a single image channel. They can be computed based on a segmentation (object delineation) that results in a set of pixel/voxel labels. One must pay close attention to the units of the features. Also keep in mind the possibility of non-isotropic images - this is a common occurrence in biological microscopy. Usually, the axial dimension of a voxel is much greater than its lateral dimensions. To obtain feature measurements in physical units, one must scale pixel/voxel based feature values appropriately. The intrinsic measurements of each object are directly determined by its morphological class. Some examples are noted in the table below.

Object Morphology	Examples of Intrinsic Measurements
Blobs	location, diameter, volume, shape factor, surface area, eccentricity, texture
Tubes	centerline, surface locations, local diameter, branch/crossover points, orientation
Shells	thickness, shape factors, brightness
Plate/Laminae	thickness, surface area, brightness
Man-made objects	location, pose
Foci	location, brightness, diameter
Cloud	Brightness, texture

The goal of FARSIGHT is to have a sufficiently rich, and well-documented library of intrinsic features available to the user for each morphological category. Another goal is to incorporate appropriate corrections when possible. For instance, most intensity/brightness measurements require a background correction step. Careful documentation of each feature, and clarifying its properties such as the units and invariance are an important goal.

2. Associative Features: Associative features quantify associations/relationships between objects. They can be computed once objects have been delineated. Some associative features involve objects and image data. Others relate objects to other objects. Associative features are quite diverse, and there are a combinatorially large number of possible associations. However, a much smaller number of them are are interesting and/or useful for a given investigation. The goal of FARSIGHT is to have a versatile set of 'programmable' computational methods available to the user. These are based on the simplest notions such as spatial proximity, neighborhood, and adjacency.

Contents 1 Intrinsic Features for Blobs 1.1 Volume 1.2 Integrated Intensity 1.3 Centroid (unweighted and weighted) 1.4 Axes Lengths 1.5 Eccentricity 1.6 Elongation 1.7 Orientation 1.8 Bounding Box Volume 1.9 Oriented Bounding Box Volume 1.10 Sum 1.11 Mean 1.12 Median 1.13 Minimum 1.14 Maximum 1.15 Sigma 1.16 Variance 1.17 Radius Variation 1.18 Skew 1.19 Energy 1.20 Entropy 1.21 Surface Gradient 1.22 Interior Gradient 1.23 Interior Intensity 1.24 Surface Intensity 1.25 Intensity Ratio 1.26 Shared Boundary 1.27 Surface Area 1.28 Shape 2 External Links

Intrinsic Features for Blobs

These features can be calculated with two input images (Data Image and Label Image). They are most commonly used for blob-like regions, such as cell nuclei.

Volume

Number of voxels in the object. (Dirk)

Integrated Intensity

Sum of the intensities of all voxels in the object. (Dirk)

Centroid (unweighted and weighted)

The unweighted centroid calculates the center of the object. The weighted centroid uses the image intensity values to calculate the intensity center of the object. (Dirk)

Axes Lengths

The length of the axes of the ND hyper-ellipsoid fit to the object. (Dirk)

Eccentricity

Ratio of the distance between the foci of the best-fit hyper-ellipsoid to the length of its major axis. (2D) (Dirk)

Elongation

Ratio of the major axis length to minor axis length of the best-fit hyper-ellipsoid. (2D) (Dirk)

Orientation

Angle between the major axis of the best-fit hyper-ellipsoid and origin. (2D) (Dirk)

Bounding Box Volume

Number of voxels in the bounding box of the object. (Dirk)

Oriented Bounding Box Volume

Number of voxels in the oriented bounding box of the object. The oriented bounding box is defined as the bounding box aligned along the axes of the object. (Dirk)

Sum

Same as integrated intensity (ITK statistics)

Mean

Average intensity of voxels in the object (ITK statistics)

Median

Middle intensity of voxels in the object (ITK statistics)

Minimum

Minimum intensity of voxels in the object (ITK statistics)

Maximum

Maximum intensity of voxels in the object (ITK statistics)

Sigma

Standard deviation of intensity of voxels in the object (ITK statistics)

Variance

Variance of intensity of voxels in the object (ITK statistics)

Radius Variation

Standard deviation of distance from surface voxels to centroid (Isaac)

Skew

Skew of the normalized intensity histogram (Isaac) (See Supplement B2)

Energy

Energy of the normalized intensity histogram (Isaac) (See Supplement B2)

Entropy

Entropy of the normalized intensity histogram (Isaac) (See Supplement B2)

Surface Gradient

Average of surface gradients (Isaac)

Interior Gradient

Average of interior gradients (Isaac)

Interior Intensity

Average of interior intensities (Isaac)

Surface Intensity

Average of surface intensities (Isaac)

Intensity Ratio

Ratio of surface intensity to interior intensity (Isaac)

Shared Boundary

Ratio of surface area that touches another object to total surface area (Isaac)

Surface Area

Number of voxels on surface of the object (Isaac)

Shape

Ratio of surface voxels to total voxels - compactness or thinness of object (Isaac) (See Supplement B2)

External Links

itkLabelGeometryImageFilter

Kitware Source Newsletter