| dc.description.abstract | Segmentation, or delineation of the boundaries of a region of interest, is an integral part of implementing intraoperative image guidance for kidney tumor resection. Results are affected by the kidney's physiology and pathology as seen in 3-D image data sets, as well as by the methods guiding contour growth. This work explores the variables involved in using level set methods to segment the kidney from computed tomography (CT) images. Multiple level set classes found in the Insight Toolkit were utilized to build a single, semi-automatic segmentation algorithm. This algorithm takes seed points and the image's contrast state as user input and functions independently thereafter. Comparison of the semi-automatic algorithm to an expert's hand-delineation of boundaries, hereafter "handsegmentation," showed that the algorithm performed well both for the images used in its creation and for new image sets. The algorithm also showed lower variability between raters than did handsegmentation. The automatic method's ability to function in a realistic image guidance situation was also evaluated. For three open kidney surgical cases, intraoperative laser range scans were registered to surfaces generated by both handsegmentation and the semi-automatic algorithm. Mean closest point distances between these registered surfaces as well as visual inspection of the distribution of closest point distances showed that the semi-automatic method provided a surface for registration which was comparable to handsegmentation. The inverse of each resultant transformation from these registrations was applied to CT image points, and variability introduced by the different transformations was found to be low, supporting the comparability of the autosegmentation to handsegmentation. | |
