Segmentation accuracy in magnetic resonance angiography images

Today a large amount of detailed information can be obtained from a single Magnetic Resonance Angiography image. Review of the source data can be tedious given the low density of information in each viewable image. One method of display is the Maximum Intensity Projection (MIP) algorithm. The information density is increased as one easily viewable image is created; however a loss of information results. Surface rendering detects and displays the surfaces of the tissues in the image. Volume rendering processes the information from the volume of the tissue for display. These rendering algorithms give more information than the MIP but increase the processing time. Segmentation separates the tissues into groups and uses one group, decreasing the amount of data to be processed and the time required for display. Accurate segmentation also provides a data set that can be displayed with greater signal difference to noise ratio (SDNR) and on which data fitting can be performed. The Z-Buffer Segmentation (ZBS algorithm gives an image that appears to contain all the information present in the original data. A truth standard is needed to quantitatively measure the algorithms performance. Computer simulated images of blood vessels were generated to serve as truth. The performance of the algorithm was measured and compared with the Recursive ZBS algorithm and a simple thresholding algorithm. Noise was added to each of the compute simulated images. The images were segmented while the parameters of the algorithm were varied. The ZBS algorithm gave and accurate segmentation (98% accurate) when the vessel diameter was 3mm and the SDNR was 5. The Recursive ZBS algorithm was shown to decrease the algorithms dependency on the three-dimensional region growing threshold, but did not increase the maximum accuracy obtainable with the ZBS algorithm. Both the Recursive ZBS and ZBS algorithm showed a greater accuracy than the simple thresholding algorithm. The ZBS algorithm gives an accurate segmentation and the use of computer simulated images as truth can be used to evaluated performance quantitatively. The evaluation of the algorithms gives insights into possible future improvements.