Interactive 3D Segmentation as an Example for Medical Visual Computing

Segmentation of medical volume data sets (i.e., partitioning images into a set of disjoint regions representing different semantic objects) is an important research topic due to its large number of potential clinical applications. In order to get accepted by physicians and radiologists a generic, interactive 3D segmentation algorithm has to be simple-to-use, accurate, and show immediate feedback to the user. In this work we present a novel 3D segmentation paradigm that effectively combines interaction, segmentation and volumetric visualization in a single framework integrated on a modern graphics processing unit (GPU). This is an example of the fruitful combination of computer graphics and computer vision, a field nowadays called visual computing. Direct interaction with a large volumetric data set using 2D and 3D painting elements is combined with a segmentation algorithm formulated as a convex energy minimization. This globally optimal segmentation result and its evolution over time is continuously visualized by means of a hardware accelerated volume rendering along with the original data. By implementing all of these components on a GPU, a highly responsive interactive 3D segmentation system requiring minimal user interaction is achieved. We demonstrate quantitative and qualitative results of our novel approach on liver and liver tumor segmentation data where a manual ground truth is available.