Integration of the Gibbs prior model, the marching cubes method and the deformable model in medical image segmentation

Medical image segmentation is the process that defines the region of interest in the image volume. It is the basis of high-level image analysis such as registration, motion analyses. However, as of today, it still remains an open problem. Classical segmentation methods such as region-based methods and boundary-based methods cannot make full use of the information provided by the image. And as the 3D imaging technique develops, the data size of the medical image volume also increases, which raises the difficulty to segment objects in real time. The first contribution of this dissertation is a general hybrid segmentation framework which combines the Gibbs Prior model (basically a region based method) and the deformable model (a boundary-based method). Although MRF based model and deformable models have been applied to segmentation for a long time, it is the first trial to combine these two methods in a hybrid framework. The segmentation is now performed by iteratively applying the Gibbs Prior model and deformable model onto the image. The Gibbs Prior model will provide binary region estimations for the deformable model to fit with. The deformable model can refine the segmentation result and update the parameters of the Gibbs Prior model for the next iteration. The second contribution of the dissertation is that we make the 3D segmentation more efficient by insert the Marching Cubes method into the hybrid framework. The Marching Cubes method can convert a 3D binary mask into an approximate deformable mesh surface. The deformable model can use the result of marching Cubes method as the initial surface instead of expand from inside of the object. The usage of the Marching Cubes method can greatly speed up the segmentation process and reduce the topology mistakes on the deformable surface. The hybrid segmentation framework has been applied to segment objects clinical images. The experimental data includes MRI (T1, T2, PD), CT, X-ray, Ultra-Sound images. 17 Brains (2 normal, 15 with lesion) have been segmented by the hybrid framework and results have been evaluated using expert manual segmentation as the ground truth. The validation shows that high quality results are achieved with relatively efficient time cost, which prove that the hybrid segmentation methods may have further clinical usage