(:mini m1=500x0x0x0x100 m2=200x0x0x0x100 :)
Semi-automatic Segmentation
Introduction
Segmentation is the problem of delineating structures of interest in an image from surrounding structures. Semi-automatic segmentation has always been a mainstay of ITK-SNAP, but with the release of version 3.4, semi-automatic segmentation has been made dramatically more powerful for complex segmentation problems, including problems where the structure of interest looks very similar to some of the adjacent structures.
Semi-automatic segmentation in ITK-SNAP consists of four stages:
- Region of interest selection: In this stage we define the region of interest (ROI) that contains our structure of interest. An ROI is just a 3D rectangular region outlined in red in the figure below. All subsequent segmentation operations are applied to this ROI, rather than the whole image. This makes the segmentation go faster and uses less memory.
Mini2:roisel.png
- Pre-segmentation: In this stage, we transform the anatomical image (or multiple anatomical images) to a new synthetic image called the speed image. The speed image has values in the range -1 to 1. The goal of pre-segmentation is to make the speed values close to 1 inside of the structure of interest, and close to -1 for other parts of the image.
Mini1:preseg.png
The example of pre-segmentation above transforms three anatomical images from a brain tumor MRI into a speed image, with the voxels inside the tumor having positive (white) speed values, and non-tumor voxels having negative (blue) speed values.
There are four pre-segmentation modes available in ITK-SNAP:
- Thresholding: Voxels with intensity inside of a user-specified range are mapped to positive speed values. This is the simplest pre-segmentation method suitable for easy tasks.
- Classification: User draws examples of two or more tissue classes in the image, and a machine learning algorithm uses these examples to assign speed values to the rest of the image. This is the most powerful pre-segmentation method in ITK-SNAP.
- Clustering: The image is automatically partitioned into two or more tissue classes by clustering voxels with similar intensities together. This approach requires the least user input, but is not always reliable.
- Edge attraction: This approach finds and accentuates edges in the image. The speed image is close to 1 away from edges and close to 0 at edges. Edge attraction is useful when image intensity within the structure of interest varies.
- Initialization: In this stage, we place one or more seeds inside of the structure of interest. These seeds, shown in red below, are spheres that will be grown to form the segmentation of the structure of interest in the next phase.
Mini2:seeds.png
- Evolution: In this stage, the seeds evolve, expanding over the regions of the positive portions of the speed image, and contracting over the negative regions.
Mini1:evol.png
Unlike the pre-segmentation phase, which labels individual voxels as likely to be the object of interest or not, the evolution phase produces a contiguous object or set of objects based on the seeds placed by the user.
Simple Segmentation using Thresholding
