Medical image segmentation based on Dirichlet energies and priors
Ang Li · The Sydney eScholarship Repository (The University of Sydney) · 2014
As the first and fundamental step of the computer-aided medical image analysis, automated segmentation methods have been extensively investigated due to their efficiency and reproducibility. However, image artifacts including noise, intensity inhomogeneities and missing/blurred boundaries pose a significant challenge for the computer-aided segmentation. In this thesis, we proposed three methods to improve segmentation accuracy and robustness when images present blurred boundaries and intensity inhomogeneities. On the basis of image prior knowledge and statistical analysis, our contributions to medical image segmentation can be categorized threefold. The first proposed method reformulates graph weights and takes into account both global probability knowledge and local intensity. With the prior knowledge guidance, our method automatically initializes the target and background to avoid the initialization bias in the random walker method. Secondly, a combinatorial Bayesian and Dirichlet model that utilizes more comprehensive prior knowledge is proposed to better handle the inhomogeneity and shape irregularity in medical images. We formulated a new Dirichlet graph energy function taking into account both probabilistic sub-graph features and image intensity. Thirdly, we proposed a new Statistical and Dirichlet Integral framework, including a generalized energy formulation harnessing statistical intensity approximation and a Gaussian mixture model term approximating the probability density distribution of the foreground and the background. We conducted thorough experimental validations of proposed methods on public datasets. We also compared our methods with other state-of-the-art segmentation algorithms, including graph cut and random walker. The experimental assessment and comparison have demonstrated that our proposed methods were able to segment the regions of interest with better accuracy and robustness compared to the benchmark methods.