Feature space optimization for brain tissue classification in non-contrast computed tomography images

In this work, a classification problem focused on brain tissue types in non-contrasted computed tomography images is explored. A group of characteristics is proposed in the spatial domain, and block model for feature selection is added to reduce the dimension of the problem. A stochastic search technique was used by means of genetic algorithms. This scheme presents a 61.9% reduction of the feature space. Finally, a comparative analysis performing classification tests under supervised learning for various classifiers on the set of samples. The full feature space presents an average Acc of 98.1± 0.81%, for FS
0
, 97.82 ± 0.46%, and considering only average intensity 97.33 ± 0.67%. This analysis shows statistical evidence to affirm that the characteristic spaces affect the classification performance.