Genetic algorithms and deformable geometric models for anatomical object recognition

We examine a method of locating an extended anatomical structure within the human brain from evidence provided by 3D magnetic resonance (MR) images. The problem that we deal with is the determination of the location, size, orientation and shape of a major portion of the human brain stem: an extended anatomical structure of the middle and lower brain. This object is not easily visualizable and cannot be extracted with traditional intermediate level segmentation techniques because only its lower part (cortico-spinal tract) and upper part (mesencephalon) are well defined. A substantial length is completely connected with the rest of the brain by the right and left superior and inferior cerebellar peduncles and the central tegmental tract, having almost the same image intensity, making our approach especially advantageous over other segmentation techniques. The search problem can be considered as a one of large scale optimisation and we describe a genetic algorithm (GA) based method for its solution. Finally we briefly describe recent extensions to our approach that allow the GA-based system to be used for arbitrarily shaped objects.<>