Medical image understanding: Segmentation, modeling and representation

Medical image understanding is becoming increasingly important in clinical applications such as quantification of tissue pathology, surgery simulation, and radiotherapy planning. In this thesis we develop advanced image analysis techniques for medical image understanding by segmentation, modeling, and representation.; A new segmentation scheme by integrating region and edge information is first developed to segment images with complex structures such as MR brain images. A sophisticated region merging procedure is introduced to obtain a reliable and meaningful region information to further combine region and edge information. The region merge score takes into account the grey-level similarity, size constraint, and region connectivity, so that the algorithm has the following tendencies: regions with similar grey-level intensity are merged, larger regions are preferred over smaller ones, and regions with a small common boundary should not be merged. Edge information is then used to eliminate false boundaries, and active contour models are used to modify region contours to conform to the gradient information of the image. The experimental results show that segmentation by integrating region and edge information results in a highly reliable segmentation result for MR brain image segmentation and tissue extraction.; A deformable surface-spine model is then developed to reconstruct surfaces of human organs such as the prostate. The deformable model is dynamic and governed by Lagrangian motion equation. In order to solve the dynamic (Lagrangian) equation, a finite-element implementation is developed, where 9 degree-of-freedom (dof) triangular elements and 4-dof spine elements are constructed. The deformable surface-spine model is successfully applied to reconstruction of surgical prostate models. The experimental results show that the deformable surface-spine models not only greatly improve the smoothness and consistency of the reconstructed surfaces, but also respond dynamically to the external forces which is one of the main requirements in surgery simulation and planning. Further, with such accurate reconstructed 3-D prostate models, we implement a prostate needle biopsy simulation system to validate current needle biopsy techniques for prostate cancer research.; Finally, a feature-based deformable model is developed to represent surfaces with discontinuities. By using finite-element method, a surface representation scheme that can incorporate surface discontinuities is introduced. A formulation for finite-element representation that can handle discontinuities is then developed, which is the key step in developing our feature-based deformable model. By incorporating the discontinuity component into a finite-element representation, 9-dof triangular elements are constructed with continuous and discontinuous stretching and bending energies. The feature-based deformable model is governed by Lagrangian dynamic equation, and an iterative algorithm is developed to find the final surface shape with the minimum energy of the model. The problem of surface discontinuity detection is also discussed and a general framework is developed by using Canny's edge detector in depth, normal, and curvature domains respectively.