Master of Science
In this thesis, we propose a brain tumor segmentation system that requires only 4 clicks from users to specify a tight bounding box that completely contains the tumor.
We convert the segmentation problem to an energy minimization problem. We utilize the basic energy function that combines intensity appearance and boundary smoothness. Global and local appearance models are experimented and compared in our work.
The basic energy function does not assume any shape prior and thus leads to unrealistic shapes. We take the advantage of the fact that most of the tumors are approximately convex in shape and incorporate the star shape prior to prohibit unlikely segmentations.
Another problem with the basic energy function is the undersegmentation problem. With the bounding box provided by the user, we are able to have a rough idea of the tumor size. Therefore, to encourage the segmentation to be a certain size, we add volumetric bias to our energy, which helps solve this problem.
We also try to model the tumor as multi-region object where regions have distinct appearance. Specifically, we incorporate interior+exterior model for the tumor into our energy function.
Our final result is promising in terms of f-measure. Our best performance for 88 volumes is 87% using volumetric ballooning.
Liu, Liqun, "Brain tumor segmentation with minimal user assistance" (2014). Electronic Thesis and Dissertation Repository. 1867.