Deformable Simplex Meshes Based 3d Medical Image Segmentation

With the rapid advances in medical imaging technology, computed tomography(CT),magnetic resonance imaging(MRI), ultrasound and other imaging modalities have beenwidely used within the clinical environment such as computer-aided diagnosis, surgicalplanning and simulation and radiotherapy planning.Three-dimensional(3D) visualization of the organs within medical image sequencesallows physicians to virtually interact with anatomical structures and learn important diag-nosis and treatment information. While the precise segmentation of organs from medicalimage sequences form the basis of effective 3D visualization.Deformable models have been widely used in medical image analysis, especially inimage segmentation. Deformable models are curves or surfaces that deform under theconstraints of all kinds of internal and external energy in order to segment the object ofinterest.There are two widely used evolution methods for deformable models: greedy algo-rithm which minimizes an energy functional and physics-based method. It has alreadyproven that the greedy algorithm outperformed physics-based method both in computa-tion cost and segmentation accuracy. Simplex meshes are efficient and versatile surfacerepresentation for physics-based 3D deformable models. In order to fully take advantageof the benefits of the greedy algorithm, a new greedy algorithm based deformable simplexmeshes is proposed.Generalized gradient vector ?ow(GGVF) field is a widely used classical externalforce for physics-based deformable models. The traditional external energy(i.e., imageintensity and gradient magnitude) for deformable models has many problems, such aslimited capture range, sensitivity to initialization and noise and poor convergence toboundary concavities. In order to overcome the main issues of various external energyfor deformable models, the GGVF field is adapted to GGVF external energy and appliedto the greedy algorithm.In order to demonstrate the performance and accuracy of this new proposed 3Dsegmentation method, qualitative analysis of a synthetic planet data set and clinical 3Dimages, and quantitative analysis of a synthetic sphere data set and the same data set with additional gaussian noise are conducted. The experimental results showed that themean radial error(MRE) for the synthetic sphere data set and the same data set with addi-tional gaussian noise was 0.55 pixels and 0.59 pixels, respectively; the new proposed 3Dsegmentation method can successfully segment different kinds of 3D data sets, and canachieve sub-voxel accuracy.