Segmentation Principles and Basic Techniques

Introduction

Image segmentation aims to generate pixel groups from an image that represent parts of depicted objects. In medical imaging, it often delineates specific structures and parts of classification.
Segmentation strategies in medical imaging combine data knowledge (assumptions about continuity, homogeneity, and local smoothness) with domain knowledge (information about the objects to be delineated).
The primary purpose of segmentation is to create semantic entities, assigning each pixel to exactly one segment.
A uniform segmentation criterion for all objects is difficult to find, and a general solution for segmentation tasks is very difficult.
Medical image represents the measurement of a diagnostically relevant entity, but external influences add object- or location-dependent components.
Data-driven segmentation often results in a collection of regions because measured values are not unique for a specific object class.

Foreground Segmentation: Focuses on a single object, creating good partitioning of foreground objects while background quality is irrelevant; requires model knowledge and sometimes user input to separate foreground from background.
Hierarchical Segmentation: Applies a multiresolution concept for gradual refinement; initial segmentation creates smaller segments (oversegmentation), which are then merged into larger segments at the next level.
Multilayer Segmentation: Assumes a common segmentation criterion exists but its scale varies; segmentation occurs at different scales producing layers, requiring later analysis to estimate local scales and patch segments.
Segmentation and classification in medical imaging often mix, as a segmentation criterion on pixel values can assign class membership to a segment.
Segmentation is essential for computer-assisted analysis of medical images.

Spatial and Temporal Continuities: Continuity in space and time are key properties for segmentation.
- Spatial Continuity: Partitions a 2D or 3D image, ensuring homogeneity within segments exceeds that between adjacent segments.
- Temporal Continuity: Treats time as a fourth dimension, using segmentation results from one time step to constrain or initialize the next.
The result depends on the initial segmentation, emphasizing the importance of a good initialization.
Segmentation as Registration: Imposing continuity constraints in (n-1)-dimensional space reduces n-dimensional segmentation to a registration task, assuming a one-to-one correspondence between segments.

Spatial and temporal continuity can be characterized by homogeneous local appearance, implying homogeneity of intensity within a segment.
Intensity-based segmentation schemes are popular because pixel or voxel intensities of a structure vary little throughout the segment.
Noise: Often modeled as Gaussian with zero mean and variance according to the SNR; can be reduced during preprocessing or within segmentation using a multi-resolution strategy like a Gaussian pyramid.
Multiresolution Approach: Gaussian pyramid creates images at different resolutions via repeated low-pass filtering and subsampling.
Shading: Can influence intensity-based segmentation, stemming from image acquisition; if less pronounced than segment homogeneity, it can be removed during preprocessing; boundary criteria can sometimes resolve shading problems.
Region-based and Edge-based Segmentations: Boundaries are localized by tracking zero crossings of the second derivative or computing local maxima of the gradient length.
Gradient-based segmentation is simplest when the gradient length is approximately equal or all nonzero gradient lengths are from segment boundaries.
Edge-based segmentation is more sensitive to noise compared to region-based segmentation.

Texture-based segmentation assumes at least two different textures to be separated.
Texture computation becomes unreliable at unknown segment boundaries if pixels from multiple segments contribute to the texture measure.
Texture Computation Strategies:
1. Choose texture measures requiring few pixels for reliable features.
2. Carry out segmentation iteratively.
3. Use a multi-resolution framework.
4. Precede texture-based segmentation with an initial homogeneity-based segmentation.

Detecting an object requires additional information about what to segment.
Further knowledge may be entered through an explicit description of boundary properties or about a foreground object.
Domain Knowledge Information:
- Appearance of boundaries between segments.
- Location of an object within an image.
- Orientation and size of the object.
- Spatial relationships of the object.
- Shape and appearance of the object.
Domain Knowledge for Describing a Foreground Object:
- Discriminative: Object and background must have different properties.
- Generalizable: The property must be true for all possible instances.
- Efficiently computable with sufficient reliability.

Domain knowledge may be introduced via:
- An adjustable model included in the segmentation method.
- Interactively at run-time.
Descriptions of Domain Knowledge Types:
1. A parameterized description.
2. A sampled description.
3. An implicit description.

Incorporating domain knowledge must include known variation among instances of a class.
Information about acceptable variance may be obtained from expert information or training.
Many implicit knowledge representations use simple assumptions about the local smoothness of object boundaries.

Interactive incorporation of domain knowledge is flexible because an expert user decides on the necessary input at segmentation time.
Interaction at runtime delivers cues about potential input errors.
Interactive input may happen directly on the image or through adjustment of segmentation parameters.
Different Kinds of Interaction During Image Analysis:
- In a priori parameterization, the user enters parameter values.
- Through segmentation guidance, the user supports the segmentation.
- Feedback happens after segmentation.
- Correction changes the segmentation result.
- Confirmation is the process by which the user accepts or rejects a result.

Relies on user guidance to outline foreground structure boundaries.
Interactive segment delineation by an expert often provides a reference for segmentation schemes.
The underlying assumption is that a domain expert provides proper model knowledge.
A problem of interactive segmentation is that a user may not possess all necessary information or apply it correctly.
Low-Level Techniques for Interaction:
- Boundaries highlighted by displaying the intensity gradient.
- User provides a few boundary points connected by line segments.
- User input is corrected automatically.
- User corrects a delineated boundary.