Efficient 3D medical image segmentation

Abstract

3D Medical imaging techniques have become extremely important tools in patient diagnosis. However, they produce large amounts of data that is difficult to interpret, and can currently only be analysed by highly trained people. Datasets are large – the female Visible Human dataset is around 40 Gb in size. Processing any dataset of this size will obviously be computationally demanding. Currently segmentation of images is a predominantly manual process. Tools that are available allow segmentation to be done on a slice-by-slice basis, often using a flood-fill or region growing approach based on colour or texture space. This report outlines research into an automated texture based segmentation technique. The research compared the effectiveness of using simple and energy efficient DCT (Discrete Cosine Transform) and Haar transforms (in both 2D and 3D forms) as a description of texture at each location within an image. This description was initially used as a vector in feature space, allowing segmentation to be carried out using a Gaussian Mixture Model and some post processing techniques. The transforms were then extended to make them independent of variations in intensity, a common issue in medical imaging. However, although now robust to intensity variations, the results were not of sufficient quality to be useful in a real application. To improve the quality of results, a model based approach based on an AAM (Active Appearance Model) was considered. A traditional AAM uses an intensity based appearance model, which while less computationally demanding than a more complex texture based appearance model, can give poor results when subjected to intensity variations. When complex texture descriptions are used to create the appearance model results are much improved, but this is at the expense of run time, which can make the techniques less practical. A novel combination of mDCT (modified DCT, which is intensity invariant) and an AAM was implemented and tested. When presented with 3D volumes which had been subjected to intensity variations this was seen to generate much better results than a traditional AAM, while maintaining a practical run time. Using this approach the time taken to carry out segmentations was less than 10 minutes (when run in Matlab on a typical datacentre based Linux machine). This showed the process to be practical in terms of quality of results, run time and energy efficiency.