Automatic segmentation of the heart in 3D MR images

This paper presents an approach to a fully automated system for segmenting the heart in 3D MR images of the torso. The system initially processes one axial slice containing a cross-section of the heart with significant sectional area. The system then hierarchically segments this image by first separating the torso from the background and then the lung fields from the torso. The heart location is estimated from the lung fields. This estimate is used as the region of interest for finding the edges of the heart in the image. The region is refined to straddle the detected edge, and is then projected onto adjacent images to provide a starting point for further segmentation. All thresholds are determined from the image data. Anatomical knowledge of the locations and relative sizes of the various structures is used to guide segmentation. The approach attempts to use easily-detected structures to constrain the search space for structures which are more difficult to determine. The algorithms also use the structures detected in a given image to constrain the search space in adjacent images, thereby reflecting the inherent spatial coherence of the heart