Interactive Brain Tumor Segmentation with Inclusion Constraints

This thesis proposes an improved interactive brain tumor segmentation method based on graph cuts, which is an efficient global optimization framework for image segmentation, and star shape, which is a general segmentation shape prior with minimal user assistance. Our improvements lie in volume ballooning, compactness measure and inclusion constraints. Volume ballooning is incorporated to help to ”balloon” segmentation for situations where the foreground and background have similar appearance models and changing relative weight between appearance model and smoothness term cannot help to achieve an accurate segmentation. We search different ballooning parameters for different slices since an appropriate ballooning force may vary between slices. As the evaluation for ”goodness of segmentation” in parameter searching, two new compactness measures are introduced, ellipse fitting and convexity deviation. Ellipse fitting is a measure of compactness based on the deviation from an ellipse of best fit, which prefers segmentation with an ellipse shape. And convexity deviation is a more strict measure for preferring convex segmentation. It uses the number of convexity violation pixels as the measure for compactness. Inclusion constraints is added between slices to avoid side slice segmentation larger than the middle slice problem. The inclusion constraints consist of mask inclusion, which is implemented by an unary term in graph cuts, and pairwise inclusion, which is implemented by a pairwise term. Margin is allowed in inclusion so that the inclusion region is enlarged. With all these improvements, the final result is promising. The best performance for our dataset is 88% compared to the previous system in [25] that achieved 87%.