This paper presents the main results achieved by the COVIRA consortium project during a 15-months period under the AIM Exploratory Action programme. The prime result of the project is a specification of a system for knowledge based interpretation of cranial MR images as a means of computer assistance in diagnostic radiology, radiation therapy planning and stereotactic neurosurgery. This specification is based on demonstrated results from prototypical implementations and an analysis of the state-of-the-art and of the clinical requirements.

Image segmentation and interpretation results were obtained for a set of test images of tumor patients selected by medical experts from the above fields. For each patient slice, two MR spin echo images are available.

For image segmentation, results are presented for three schemes: one using a Canny edge detector followed by a smooth patch fitting for region finding, another using a region detection according to Nagao/Matsuyama and a Marr-Hildreth edge detector based on one of the echoes to guide region merging, and a third scheme using a multiscale approach to edge detection.

The image interpretation results presented are based on a case model representation of clinical, anatomical, MR-Physics, and tissue parameter knowledge. Results are presented for two schemes: one using a fuzzy clustering approach providing a fuzzy segmentation and a subsequent fuzzy relational matching step, and another using a blackboard approach based on the classical low-, middle- and high-level of processing, allowing for a dynamic control with feedback and backtracking mechanisms.

The results were comparatively evaluated by medical experts based on criteria of clinical usefulness.