Rapid Mode Estimation for 3D Brain MRI Tumor Segmentation

In this work we develop a method for the efficient automated segmentation of brain tumors by developing a rapid initialization method. Brain tumor segmentation is crucial for brain tumor resection planning, and a high-quality initialization may have a significant impact on segmentation quality. The main contribution of our work is an efficient method to initialize the segmentation by casting it as nonparametric density mode estimation, and developing a Branch and Bound-based method to efficiently find the mode (maximum) of the density function. Our technique is exact, has guaranteed convergence to the global optimum, and scales logarithmically in the volume dimensions by virtue of recursively subdividing the search space through Branch-and-Bound. Our method employs the Dual Tree data structure originally developed for nonparametric density estimation, and recently used for object detection with branch-and-bound. In this work we ‘close the loop’, and use the Dual Tree data structure for finding the mode of a density. This estimated mode provides our system with an initial tumor hypothesis which is then refined by graph-cuts to provide a sharper outline of the tumor area. We demonstrate a 12-fold acceleration with respect to a standard mean-shift implementation, allowing us to accelerate tumor detection to a level that would facilitate high-quality brain tumor resection planning.