Semi-Automatic Rough Classification of Multichannel Medical Imaging Data

Rough set theory is an approach to handle vagueness or uncertainty. We propose methods that apply rough set theory in the context of segmentation (or partitioning) of multichannel medical imaging data. We put this approach into a semi-automatic framework, where the user specifies the classes in the data by selecting respective regions in 2D slices. Rough set theory provides means to compute lower and upper approximations of the classes. The boundary region between the lower and the upper approximations represents the uncertainty of the classification.We present an approach to automatically compute segmentation rules from the rough set classification using a k-means approach. The rule generation removes redundancies, which allows us to enhance the original feature space attributes with a number of further feature and object space attributes. The rules can be transferred from one 2D slice to the entire 3D data set to produce a 3D segmentation result. The result can be refined by the user by interactively adding more samples (from the same or other 2D slices) to the respective classes. Our system allows for a visualization of both the segmentation result and the uncertainty of the individual class representations. The methods can be applied to single- as well as multichannel (or multimodal) imaging data. As a proof of concept, we applied it to medical imaging data with RGB color channels.