Approximation Degrees in Decision Reduct-Based MRI Segmentation

Segmentation of magnetic resonance images (MRI) is a process of assigning the tissue class labels to voxels. One of the main sources of segmentation error is the partial volume effect (PVE), which occurs most often with low resolution images. Indeed, for large voxels, the probability of a voxel containing multiple tissue classes increases. We have utilized a classification approach based on the attribute reduction, derived from the data mining paradigm of the theory of rough sets. An approximate reduct is an irreducible subset of features, which enables to classify decision concepts with a satisfactory degree of accuracy in the training data. The ensembles of the best found reducts trained for appropriate approximation degrees are applied to segmentation of previously unseen (parts of) images. One of the challenges is to adjust the approximation level during the training phase to obtain the best classification results for new cases. In this paper, it is proved experimentally that the choice of approximation level, consequently related to generality of classification rules induced by reducts, should correspond to expected quality of images. We show that when dealing with noisy images, or images with high PVE level, better results are obtained with higher degrees of approximation.