Automatic segmentation and 3D rendering of liver veins from ultrasonic images

Abstract

In this thesis, automatic segmentation of liver veins from ultrasound slices is implemented and further rendered into 3D image to assist the medical doctor to more accurately apply any study intended on the liver. The main part of this study is the texture analysis since it is the hardest problem in common. As a start, 4 different techniques in texture analysis have been introduced, the Spatial Gray-level Dependence Matrices (SGLDM), The Fourier Power Spectrum (FPS), The Gray- Level Difference Statistics (GLD), and Law's Texture Energy Measure (TEM). Each of these methods have been discussed in detail in the thesis to show its power and weakness points. From the viewpoint of speed and accuracy of classification, it was found that these features do not perform well enough, either consuming much time or producing low classification rate. The Multi-resolution Fractal Features is also explained and discussed to show its ability to read important measurement in ultrasound, which are granularity, regularity, and roughness. This method is based upon the concepts of Multiple Resolution Imagery and Fractional Brownian Motion and it was adopted in this research. Clustering methods (Hierarchical and non-hierarchical) have been introduced and explained, the Vector Quantizer method which is a member of the hierarchical family, is recommended for its ability to simplify the data to be clustered without losing essential information. Then, labeling techniques which is known for its simplicity and speed, P and 4-connected neighbors, and P and 8-connected neighbors, are also introduced and discussed. Finally, the segmented veins in each slice are rendered to be displayed in 3D form using Ray Casting technique. A real time implementation of the proposed algorithms is performed on a pentium II 400-MHz, 64 Mb RAM, and 4 Mb SVGA Card, with an acceptable processing time which makes the proposed techniques acceptable from all aspects.