Ivanovska,T., Linsen,L.
Abstract:
Most medical scanning techniques generate scalar fields, for which a large range of segmentation algorithms exists. Some scanning techniques like cryosections, however,generate color data typically stored in RGB format. Since standard segmentation algorithms such as isosurface extraction, level-set and region growing methods all have their advantages and drawbacks and many extensions and specializations of the algorithms have been developed to solve specific problems, one would need to generalize all these approaches to color data to have the full range of algorithmic solutions at hand. A more viable way to proceed is to convert the color data field to a scalar field in a preprocessing step, which allows for the direct application of all above-mentioned segmentation approaches. However, color-to-scalar conversion can lead to a loss in information. In particular, different colors with the same luminance are often mapped to the same scalar value and segmentation algorithms would fail in distinguishing the respective regions. We propose a procedure that converts color to scalar data while preserving the properties that are important for segmentation purposes.
We first convert the colors from RGB to Lab color space, which separates the luminance channel from the chrominance channels and distributes the chrominance with respect to human perception. Then, we cluster the colors present in the data using a genetic algorithm or an axes-aligned binary space-partitioning approach. We compare our approaches to median cut, k-means, and c-means approaches and discuss the advantages and drawbacks. The number of clusters is given by the number of distinguishable scalar values in the resulting data set. In order to assign to each cluster an appropriate scalar value, we use the ideas of the recently presented Color2Gray algorithm and generalize it for application to volume data. The Color2Gray algorithm is based on computing differences in the Lab color space.
The Color2Gray algorithm in its originally proposed form is too inefficient to be applied to volume data, but a restructuring of the algorithm coupled with a prior clusterization step allows us to apply the algorithm even to large volume data. We segment the resulting scalar field using standard segmentation algorithms and discuss our results in comparison to standard conversion results.