Physicians need axial Computed Tomography (CT) scans of the brain skull to detect early brain deformation. A case of brain atrophy can be recognised by an increase of the intra-cranial ventricles, referred to as liquor rooms. The precise volumetric proportion between brain and liquor values has to be calculated by means of a similarity-based segmentation algorithm. Via a DICOM standard interface, the patient's image slices are ported to the parallel system to pre-process, segment and label every image into brain and liquor. The HP-ISIS system depends on sophisticated case based reasoning, performed via special fast search algorithms, to discover similar cases in a large database of reference cases. The parameters of the closest case are selected for segmentation and measurement of the patient image to calculate the total liquor/brain ratio.
HP-ISIS takes the calculation and analysis time of CT images down from 40-50 minutes to nearly real time. The results of automatic brain/liquor ratio determination are available in 1-2 minutes and this is unique in the clinical environment. The Martin-Luther-Universität Halle has compared the automatically labelled images with manually labelled ones and indicated an average error rate of 2.5%. Another study, involving two physicians and 600 images, showed that the HP-ISIS algorithm could label more liquor than the human expert does. It is possible to use the system as a Computer Assisted Measurement Device. The physician can patch the automated detected areas in slices with nose and ear areas, if he does not agree with the segmentation.
Initially, the project partners implemented the algorithm on a four-processor parallel Parsytec workstation installed at the University Hospital of Leipzig. For reasons of practical use, the HP-ISIS team later on decided to couple PCs through 100Mb/s Ethernet and to design a Windows version of PVM, Parallel Virtual Machine, a high level communication software. The Unix platform is still preferred but the remote user interface permits an easy and convenient operation from any PC in the local network. As such, the target system for demonstrations and customers consists of a PPro-class PC with at least two processors and Windows NT as the operating system. This lighter version can be coupled to patient monitoring software, designed by MediSYS, one of the consortium members, to turn it into a bookkeeping and analysis tool for patient treatment monitoring.
At the Technical University of Delft, 3D segmentation has been experimented for yet a greater accuracy but the low-level image processing on an Ethernet network has led to surprising congestion, because multiple volume elements are transferred over the network at the same time. This procedure has been abandoned for the time being in favour of the existing 2D method. Together with the Institut für Bildverarbeitung und angewandte Informatik and the Department of Medical Computer Science, Statistics and Epidemology at the University of Leipzig, TU-Delft has combined within HP-ISIS its advanced low-level image processing software Scil-Image, with the assets of high level image processing, case based reasoning, and the approach towards parallel computer programming for real time image processing.
After the clinical validation of the parallel implementation of the similarity-based image segmentation algorithm for brain-liquor determination has been completed, the consortium plans the production of a marketable kit for daily use in hospitals all over Europe. A full report of the consortium's activities is available at the home page of the HP-ISIS project.