An assessment of two methods for generating automatic regions of interest.

Abstract

Two fully automatic methods for generating regions of interest (ROIs) for nuclear medicine images are described and assessed. One of these, involving registration of a previously defined ROI onto a new image, uses spatial information and is appropriate for two- and three-dimensional images which may be static or dynamic. The other method is based on artificial neural networks and uses temporal information. It is appropriate for dynamic images only. The registration method has been tested using 10 pairs of stress and redistribution images obtained from cardiac perfusion SPET. Regions of interest of the left ventricular muscle, defined on the stress images, were registered onto the redistribution images, where they were compared with reproducibility of manually drawn ROIs. Both methods were tested on 17 99Tcm-MAG3 kidney dynamic studies, where the original ROIs corresponding to both kidneys and the bladder were defined using the COST B2 hybrid phantom. Our results indicate that neither method is as reliable as having ROIs redrawn by the operator, although there are indications that an artificial neural network which combines the use of the spatial and temporal information could prove useful for dynamic studies.