Model based interpretation of magnetic resonance human brain scans
Ioannis Kapouleas · 1990
A new approach to automating radiologic diagnosis is described and tested in a system that locates multiple sclerosis lesions in magnetic resonance human brain images. No automatic systems are available now for this or for analogous diagnostic tasks in radiology. The originality of the approach consists of the combination of low-level computer vision methods with domain-specific geometric models of human brain anatomy. New low-level methods of high reliability have been developed, as well as a geometric modeling method which tolerates the variability of biological structures. These methods have been tested on data from 17 patients with Multiple Sclerosis (1132 images) with excellent results. The approach and several of the methods developed for this application could be used for other tasks in radiology and probably other imaging tasks as well. The new low level vision methods identify successively the brain mass in the images, locate suspected lesions, and some normal structures. Other low level methods eliminate the majority of false positive lesions from the previous steps and locate landmarks such as the interhemispherical fissure. These methods take advantage of the special characteristics of tomographic images. A modeling method that employs B-spline surfaces has been developed to model the surfaces of the organs in a human brain in 3D. This method allows the model surfaces to be deformed in order to fit each individual patient's brain, and also allows the proportional deformation of the shapes of difficult-to-identify regions according the the deformation of easier-to-identify regions. The system calculates the appropriate position, orientation and proportional deformation for the model by using landmarks within the patient's brain, and the moment of inertia method. The system performs well because it is (a) working with relatively well-defined real images which differ from each other in non-trivial way, but are still very similar (b) employing an approach of stepwise refinement, where the steps are chosen so that the easiest-to-detect features are found at each step, and then are used to calibrate the application of the next step, until the desired useful features are found. Finally, extensive testing has been performed to ensure that the approach really works.