The Department of Mathematics at the University of Joensuu has been involved in a programme on medical image processing as an extension to the Euromed project. Mr. Olavi Kelle specified some particular issues with relation to remote brain image segmentation during the last ITIS-ITAB event. The brain is a complex structure and there is a great medical need to assign the various brain tissues to different classes of white matter and grey matter, in order to predict pathologies, such as tumours and necrosis. The method of segmentation used by the Finnish team is based upon the Markov Random Field (MRF) modelling or maximum a posteriori (MAP) probability. Typical for the application is that the result of the segmentation can be visualized as a Virtual Reality Modelling Language (VRML) world.
Brain volume estimation is a very useful tool for the diagnosis of diseases, for the analysis of cancer, and in surgery planning. The team has several segmentation methods under research, relating to pixel-based, edge-based, region-based, and neighbourhood-based applications. Unfortunately, all these methods have their shortcomings. In order to combine the surface data with the voxel data, a prototype system was developed to simulate the parts of the brain. A Web-based application was built to connect the system to the medical database, as to retrieve the patient and the study. In the case of 3D data, it is possible to select the area of interest, the different views, the image processing function and the parameters to generate the segmentation.
After the segmentation has been calculated, it can be viewed with the use of java-applets, in order to check its correctness. Unique to this system is that the data can be post-processed and transformed into VRML worlds for online high level interactivity. To visualize the images, the team has used standard browsing software and VRML 2.0 language. A fully automatic segmentation that is completely exact is hardly achievable. Mr. Kelle nonetheless believes in mathematical segmentation methods. Even if the information is not fully complete, it can be of great help to the medical professionals. The segmented data can be post-processed manually, in order to improve the segmentation algorithms.
The entire procedure can be divided into the three steps of pre-processing, processing, and post-processing, ranging from classification over the exact determination of the homogeneity criterion to the final visualization. The system is suitable for multi-modality image segmentation. Mr. Kelle equally presented an overview of the various algorithms that are applied to perform accurate segmentation of 3D brain images. In the future, the team aims to use standard frame atlases as to refine the mathematical calculations.