kowaliw dot ca :: The Unconstrained Automated Generation of Cell Image Features for Medical Diagnosis

The Unconstrained Automated Generation of Cell Image Features for Medical Diagnosis

T. Kowaliw and W. Banzhaf
to appear at ACM GECCO 2012
paper preprint (PDF)
© ACM, (2012). This is the author's version of the work. It is posted here by permission of ACM for your personal use. Not for redistribution. The definitive version was published in GECCO-2012.
definitive version
BibTEX

For an introductory text, see the TEF project page. See also our source code for the 2D neighbourhoods.

Abstract

An extension to a non-linear offline method for generating features for image recognition is introduced. It aims at generating low-level features automatically when provided with some arbitrary image database. First, a general representation of prioritized pixel- neighbourhoods is described. Next, genetic programming is used to specify functions on those representations. The result is a set of transformations on the space of grayscale images. These transforms are utilized as a step in a classification process, and evolved in an evolutionary algorithm. The technique is shown to match the efficiency of the state-of-the-art on a medical image classification task. Further, the approach is shown to self-select an appropriate solution structure and complexity. Finally, we show that competitive co-evolution is a viable means of combating over-fitting. It is concluded that the technique generally shows good promise for the creation of novel image features in situations where pixel-level features are complex or unknown, such as medical images.