A RULE-BASED METHOD FOR CONTEXT-SENSITIVE THRESHOLD SEGMENTATION IN SPECT USING SIMULATION

Robust techniques for automatic or semi-automatic segmentation of obje cts in single photon emission computed tomography (SPECT) are still th e subject of development. This paper describes a threshold based metho d which uses empirical rules derived from analysis of computer simulat ed images of a large number of objects. The use of simulation allowed the factors affecting the threshold which correctly segmented objects to be investigated systematically. Rules could then be derived from th ese data to define the threshold in any particular context. The techni que operated iteratively and calculated local context sensitive thresh olds along radial profiles from the centre of gravity of the object. I t was evaluated in a further series of simulated objects and in human studies, and compared to the use of a global fixed threshold. The meth od was capable of improving accuracy of segmentation and volume assess ment compared to the global threshold technique. The improvements were greater for small volumes, shapes with large surface area to volume r atio, variable surrounding activity and non-uniform distributions. The method was applied successfully to simulated objects and human studie s and is considered to be a significant advance on global fixed thresh old techniques.