A comparative study of segmentation algorithms applied to 2- and 3- dimensional medical images
A method that enables discriminating between CSF-grey matter edges and grey-white matter edges separately has been suggested. It was obvious that edges from this method are more complete that those resolved by the original method and have fewer artifacts. Some edges that were undetected before, are now detected because they do not have any influence from stronger nearby edges. Texture noise is also suppressed and this allows us to work at higher space scales. These 3D edge detection methods proved to be superior to the equivalent 2D methods because they can calculate the gradient more accurately and the edges detected have better continuity which is uniformly preserved along all three directions. The split and merge technique was the second method that has been examined. The existing algorithms need data structures that have dimensions that are powers of two (quadtrees). Such a 3D method would not a practical for volume analysis because of memory limitations. For example, a 256x256x256 array of bytes is about 17Mbytes and since the method requires about 14 bytes per voxel, memory sizes that computers usually have are exceeded. In order to solve this problem an algorithm that applies a split and merge technique on non cubic datasets has been developed. Along the x,y axes, the data must have dimensions that are powers of 2 but along the z axis it is possible have any dimension that may meet the current memory limits. The method consist of three main steps a) splitting of an initial cutset, b) merging and c) grouping of the resulting nodes. An extra boundary elimination step is necessary to reduce the number of the resulting regions. The original method is controlled mainly by a parameter ε that is kept constant during the process. Currently, improvements that could be achieved introducing a level of optimisation during the grouping step are being examined. Here, the grouping is done in a way that stimulates the formation of a crystal during anealing by a progressive increase (relaxing) of the parameter ε. Such method has given different results from a method that consist of a split and merge step with ε = ε1 and a step of grouping with constant ε = ε1 and a step of grouping with constant ε = ε2 > ε1. At the moment, it has been difficult to establish quantitative ways of measuring any level of improving since there is no objective segmentation to compare with. So far, the method has processed adequately up to a block of 32 56c256 sized slice and can produce 3D objects representing regions like the ventricles, the white or the grey matter.