Markov random field-Gibbs context-dependent volumetric image classification on a small scale cluster computing system

Date of Completion

January 1998


Health Sciences, Medicine and Surgery|Health Sciences, Radiology|Computer Science




The research work focuses on the development of a volumetric context-dependent classification technique and on a system architecture for image classification on a small-scale cluster computer. Future medical imaging systems will provide the medical practitioner with unique capabilities for rapidly analyzing a patient, however such systems will be required to automatically analyze and interpret massive patient datasets in real-time. Both effective image classification and a viable parallel-processing platform are viewed as essential to the successful development of these future systems.^ In this research effort a context-dependent classification technique based upon a MRF model is developed along with several novel extensions to the MRF-Gibbs image classification approach. MRF models allow for spatial relationships between locations in an image to be used in the classification process and are viewed as more effective then context-independent approaches, however they have been traditionally applied to small planar images. Here a MRF-Gibbs image model is developed for volumetric datasets and is demonstrated on a massively large dataset. The novel extensions presented consist of an intelligent processing reduction technique along with methods for incorporating spatial consistency into the MRF-Gibbs classification process.^ A robust system architecture for image classification and medical imaging on a cluster is also presented. Cluster computing has recently experienced a surge in popularity, since many computer engineers viewed them as an effective parallel-processing platform due to their low cost, fault-tolerance, and wide support. Classification application performance results using the proposed system architecture on a small-scale cluster, along with classification accuracy results of the MRF-Gibbs method are also presented.^ In summary, the research described in this document represents a meaningful step towards obtaining real-time computational intensive applications for medical imaging and for extending the technology of image classification. The research specified represents a meaningful contribution in numerous areas of computer science (parallel processing, cluster computing, distributed processing, software engineering) as well as medical imaging itself. ^