SEGMENTATION OF BRAIN CT IMAGES USING THE CONCEPT OF REGION GROWING

A method is described for extracting and isolating cerebrospinal fluid and tissue areas of brain images obtained with computed tomography. The classification of the pixels into components is based on region growing and nearest neighbor principles. To aid the performance of this method, the algorithm utilizes a priori information on the anatomic composition of the brain, and reduces the 'cupping effect' in the CT image that is attributed to beam hardening artifacts. In order to avoid subjectivity, the performance of the algorithm was tested by superimposing five computer-simulated circular lesions on different areas of the original CT scans, 8 mm thick. These images were taken at different levels in the brain, thereby accommodating different anatomy as well as the apical artifact of CT scanning. In this exploratory investigation, the false negative error of segmentation for lesions having diameter of 20 pixels was found in the order of 25% at an estimated partial volume (PV) effect of 50% that decrease further to about 5% for a PV of 80%. At that point the false positive error becomes the dominant error in the analysis.