The purpose of this study is to develop automatic algorithms for the segmentation phase of radiotherapy treatment planning. We develop new image processing techniques that are based on solving a partial diferential equation for the evolution of the curve that identifies the segmented organ. The velocity function is based on the piecewise Mumford-Shah functional. Our method incorporates information about the target organ into classical segmentation algorithms. This information, which is given in terms of a three- dimensional wireframe representation of the organ, serves as an initial guess for the segmentation algorithm. We check the performance of the new algorithm on eight data sets of three diferent organs: rectum, bladder, and kidney. The results of the automatic segmentation were compared with a manual seg- mentation of each data set by radiation oncology faculty and residents. The quality of the automatic segmentation was measured with the k-statistics", and with a count of over- and undersegmented frames, and was shown in most cases to be very close to the manual segmentation of the same data. A typical segmentation of an organ with sixty slices takes less than ten seconds on a Pentium IV laptop.

Purpose/Objective: The implementation of three-dimensional radiation therapy depends on the segmentation of tumor volumes and normal anatomical structures. However, the segmentation procedure as currently implemented is a subjective and time-consuming part of the treatment planning process. The purpose of this work was to evaluate a new image processing technique that was developed by us that is based on solving partial differential equations (PDEs) for automating the segmentation phase of radiotherapy treatment planning. Materials/Methods: Using the AcQsim (Philips Medical Systems) manual segmentation software, 8 CT studies were segmented manually by Radiation Oncology faculty and residents. These studies visualized the bladder, the rectum and the kidneys in their entirety. The consistency of the segmentations from plane to plane was verified by viewing surface renderings of the segmented structures from several angles. We then applied the level-set based segmentation algorithm to the same data sets and compared the manually segmented images with the level-set results. A segmentation and visualization computer framework we called VolVisT was written to carry out this study. The level-set segmentation was accomplished in two steps: First, a radiation oncologist selected a wire-frame model from a library within VolVisT that most closely approximated the target organ of interest. Then they manually translated, rotated and modified the wire-frame with a deformable warping tool in three dimensions. VolVisT provided the operator with a visualization through the wire-frame of a three-dimensional rendering of selectable axial, saggital, and transverse planes through the CT data set. Even with this optimized deformable warping of the wire-frame, the enclosed volume differed significantly from the manual segmentation of the target organ. In the second step, the segmentation was completed without further human intervention with the aid of a geometric PDE-based image segmentation technique. First the target data set was preprocessed with a nonlinear diffusion algorithm. Then a level-set was evolved within the processed data set under a velocity field Sn to satisfy a nonlinear partial differential equation. The velocity in the level-set approach was defined using the Mumford-Shah function. In addition, the evolution was constrained to remain within the wire-frame. These manually initiated level-set segmentations were compared with the totally manual segmentations using the Kappa statistic. Each voxel in a pair of segmentations A and B was classified as being in A and B, or in neither A nor B, or in A but not B, or in B but not A. Kappa uses the overall proportion of agreement in relation to the proportion of expected chance. The computed values lie in the range between 0 and 1. A value of 1 indicates that the volumes are identical. It is conventional to interpret Kappa values between 0.80 and 1 to indicate excellent correlation, values between 0.4 and 0.80 to indicate a fair to good correlation, and values smaller than 0.4 to indicate a poor correlation between the volumes. Results: The table below shows the Kappa values that were calculated in this study. The table shows an excellent correlation between the level set approach and the manual segmentations in 7 of the 8 cases. Only one case (#4) resulted in a fair to good correlation. Using an operator placed wire-frame defines a constraining volume that addresses cases in which the organ to be segmented can be in close proximity to other features and in gray scale regions where the boundary between the features is not clearly defined. Using the combination of nonlinear diffusion preprocessing and level-set evolution appears to be a robust technique in regions where some discernable boundary exists. Conclusions: Our results demonstrate that the approach we developed and investigated produces segmentations that correlate excellently with manual segmentations.