Because of the lack of a "gold standard'' against which classification algorithms can be measured, it is difficult to compare the technique described here with others. Each technique presents a set of results from some application area, and so anecdotal comparisons can be made, but quantitative comparisons require reimplementing other algorithms. Work in generating a standard would greatly assist in the search for effective and accurate classification techniques. The voxel histogram technique appears to achieve a given level of accuracy with fewer vector elements than the eigenimages of Windham et al.  or the classification results of Choi et al. , which use three-valued data. Their results are visually similar to the voxel histogram results and underscore the need for quantitative comparison. Because neighboring sample values are interpolated, a given accuracy can be achieved with two-valued or even scalar data, while their technique is likely to require more vector components. Kao et al.  show good results for a human brain dataset, but their technique may be less robust in the presence of material mixture signatures that overlap, a situation their examples do not include.
Was this article helpful?