Novel seed determination method based parallel geodesic distance computation algorithm for CT lung segmentation
The segmentation of anatomical structures is one the most fundamental and challenging problems in image based medical diagnosis. This problem is further complicated because different modalities require specific segmentation techniques. It is also the case that various techniques may be required even for the same modality depending on the nature of the problem. Geodesic segmentation techniques take into account the geometrical characteristics of the surfaces analyzed. The flexibility of geodesic methods makes them particularly suitable for higher dimensional image analysis. In this thesis, a novel lung segmentation technique is proposed by determining seed points in the image domain to compute geodesic distance in parallel locally around each seed point. The local segmentation accuracy improvement is achieved by specifying local tolerance value for region around each seed point. The seed region can be determined either using a manually selected trial seed region or it can be automatic using a priori Hounsfield Units for specific tissues. The technique reduces the computational domain over the image over which the geodesic distance is computed in parallel over small pixel blocks around the seed points. The GPU based parallel implementation has potential to extend to higher dimensional and multi-modal images. The proposed technique is subjectively and objectively evalutated and found to be in agreement with popular level set curve evolution techniques.