Peter James Elliott

Interactive image segmentation for radiation treatment planning

COVIRA (COmputer VIsion in RAdiology) is a project in the European Communitys Advanced Informatics in Medicine program. The goal is to improve the diagnosis and planning of treatment (radiotherapy) for patients with brain tumors and other diseases. The aim of radiotherapy is to provide a high dose of radiation to a tumor while sparing as much as possible of the surrounding healthy tissue. A necessary first step is defining the target volume and organs at risk by manually outlining the required contours on magnetic resonance or computed tomography scans. For a full three-dimensional plan this is time-consuming, as 40 or more scans are used. Computer image segmentation speeds up the process, and a method that combines information from edge and region detectors is described. Since this method is not able to completely meet the clinical requirements, an interactive image segmentation algorithm has been developed that enables the operator to employ clinical judgment. Probabilities are assigned to edges and regions and presented to the user as a hierarchy of segmentations. The approach is being subjected to extensive clinical evaluation, using pilot applications running on IBM RISC System/6000 workstations.

An object-oriented system for 3D medical image analysis

As part of a European Commission-funded research project on medical image analysis, we have developed a system aimed at solving a real clinical problem: the outlining of the target volume and organs at risk for three-dimensional conformal radiation treatment planning. The clinical requirement is to make this process less tedious and time-consuming. We show how object-oriented design techniques can be used to good effect in a collaborative project such as this, so that image segmentation and other algorithms from a number of different partners in the consortium can be integrated into one system, for comparative evaluation by the clinical partners. We describe in some detail two of these algorithms, one a straightforward region and volume grower, in which particular attention has been paid to adapting the user interface to suit clinical needs, and the other an interactive tool using b-spline surface patches for direct three-dimensional segmentation. We report on the clinical feedback we have received, which indicates that the most technically advanced algorithms are not necessarily the most useful from a clinicians point of view.

Interactive Image Segmentation Applied to CT and MR Images

In radiation treatment planning, it is necessary to mark out selected parts of CT or MR images which are clinically relevant, before designing the treatment plan. This process is usually done manually with a mouse, and for a typical data set of some 40 slices it can be a time consuming and tedious process. Automatic image segmentation techniques have been investigated, but are of little use in this field since it is difficult to incorporate the relevant clinical knowledge. We describe here two methods: Interactive Volume Growing and Hierarchical Probabilistic Segmentation, which assist the clinical user in outlining relevant volumes whilst allowing him full control of the process. Both methods have been integrated into one of the pilot systems developed as part of the COVIRA project.

Automatic Location of the Eyes in CT Images of the Head

A technique for locating the eyes in Computed Tomography (CT) brain scan data is described. The objective is to automatically localise the eyes for protection during radiotherapy planning. The image feature that is exploited is the circularity of the eyes. After data preprocessing to remove unwanted features in the images, signature analysis is performed to locate areas of interest. By applying the Canny edge detector to these areas, data is further reduced to the significant edge fragments. The Hough Transform is then applied to estimate radii and centres of the CT sections through the eyes. The Converging Squares algorithm is used as an efficient and robust method to search the Hough Transform parameter space. The results are processed by the hypothesis generation stage which clusters them according to the x,y,z coordinates of the suggested centres. The ISODATA algorithm is used for clustering. The hypotheses are assessed and sorted and the most valid hypothesis is selected and refined using a second Hough Transform. After the rejection of the invalid members of the hypothesis cluster, an ellipsoid is fitted to the new cluster centre and the results are drawn on the data. The method is fast and robust. The method was tested using five different, data sets and it performed well oil all of them.

Segmenting Cranial MR Images using Edge-Region Integration

As part of the EEC funded AIM project COVIRA to improve visualisation of MR data for radiology, radiation therapy planning and neurosurgery, a data driven method for image segmentation has been developed. Its results have been compared to ideal data prepared manually by a group of radiologists. There was found to be an overlap between the manually and automatically produced results of 84% to 93% and an evaluation showed the usefulness of the overlapping edges as high.