An adaptive level set method for interactive segmentation of intracranial tumors

Abstract

Meaningful segmentation of intracranial lesions can be of assistance for planning open navigated microneurosurgical procedures, as well as for radiotherapy. Meaningful segmentation, however, may be hampered by lack of computational power. The respective segmentation method should be based on state-of-the-art mathematical tools, and it should be suitable for real applications. A three-dimensional computational method for interactive segmentation of intracranial tumors is presented. It is based on a front propagation method, in which the evolving front gradually approaches the boundary of a given segment. It generates and remembers the entire evolution of the interface. The segment boundary is chosen from a one parameter family. User interaction is realized by selecting "seed points" inside the object/lesion. External evolution velocity regulates the segmentation process, while approaching the boundary. Adaptively resolved grids ensure computational efficiency for larger segments. The resolution is steered by an image-based indicator, which allows coarse representation of the solution in low-frequency regions, but high resolution along suspected edges of the image. Model-based segmentation was performed on the imaging data of n = 12 patients and the results compared with manual segmentation of the same tumors. The method allowed for basic segmentation in all tumors <3 minutes. This increased 2-4 fold in four irregular tumors, where discrepancies existed in comparison with manually performed segmentation. The implicit formulations of this method establish methodical and topological flexibility in three dimensions. It is thus suitable for the segmentation of objects with non-sharp boundaries such as intracranial tumors.