Feature Selection for Self-Organizing Feature Map Neural Networks With Applications in Medical Image Segmentation


This thesis presents a novel feature selection algorithm for medical image segmentation. A multiobjective optimization genetic algorithm is used to search among the candidate features to find an optimal subset which results in the highest segmentation. A self-organizing feature map serves as the classifier such that the fitness of the genetic algorithm is determined by two quality measures of the map, quantization error and topology preservation. The algorithm is applied to a 3D simulation model of the human brain and six MRI data sets, and shows promising results in comparison with using principal component analysis as the basis for feature selection. This indicates that tailoring a self-organizing feature map to a specific subset of features has the potential to increase the segmentation accuracy of medical images.