Medical Image Processing: Feature Recognition and Classification

Feature Recognition and Classification

• Feature recognition is necessary to reduce the dimensionality of the classification task by measuring essential properties or features of objects.

Features

• Features are higher-level representations of structure and shape.
• Features should be chosen to preserve the information that is important to the particular task at hand.
• Two categories of features:
  • Describing the contents of objects (e.g., mean pixel value, standard deviation, contrast, entropy, texture)
  • Describing the shape of objects (e.g., size, perimeter, circularity, skeleton, Euler number)

Image Representation and Feature Extraction

• Representation makes object information more accessible for computer-interpretation.
• Description quantifies the representation of the object.
• Feature extraction detects and represents features of interest within an image for further processing.

Why Features?

• Features are used for diagnosis, radiation therapy, and multi-modality registration of images.
• Features are extracted to identify treatment areas and boundaries.

Features Characteristics

• Features should be:
  • Robust (invariant to translation, orientation, scale, and illumination)
  • Discriminating (different values for objects in different classes)
  • Reliable (similar values for objects of the same class)
  • Independent (uncorrelated features)
  • Simple to extract for large-scale screening.

Feature Vector

• A feature vector or pattern vector contains the measured features for a particular object or region.
• The feature space is n-dimensional, with each feature constituting a dimension.
• In classification, the goal is to assign each feature vector to one of a set of classes.

Types of Representation

• Two types of representation:
  • Boundary (External characteristics)
  • Region (Internal characteristics)

Boundary Descriptors Techniques

• Chain Code:
  • A boundary representation technique that represents a contour as a sequence of straight line segments of specified length and direction.
  • The simplest chain code mechanism allocates numbers based on directions, reducing the boundary of an object to a sequence of numbers.
  • Chain code provides good compression of boundary description but is sensitive to noise and not rotation and scale invariant.