The fundamental objective of our proposed research project is to approach biomedical image interpretation from a very new perspective: that of knowledge-intensive experimental design of the segmentation process itself. We use methods of artificial intelligence, specifically of knowledge representation, diagnostic decision-making, planning and learning, to carry out our objectives. Our central hypothesis is simple: to make significant progress in automating image recognition and measurement tasks we need to treat recognition problems at the level of experimental design, so the best solutions to various types of imaging problems can be derived by a process of explicit specification, testing, and evaluation of different segmentation strategies. We have already built a preliminary prototype of the proposed system, and have tested it on brain lesion recognition problems from multimodality magnetic resonance imaging (MRI). We are now proposing to test both the methodological and practical assumptions underlying the system. We will concentrate on automatic segmentation and interpretation techniques for individual and serial MRI examinations, which will be applied to automatically quantitate CNS changes in patients with tumors, AIDS-related lesions, MS lesions, and other conditions. The significance of this research for MR image interpretation lies in its ability to provide both the clinical researcher and the laboratory investigator the tools needed to carry out their work more efficiently and effectively. In the clinical case we are focusing on the assessment of volume changes in AIDS-related and other lesions to quantitate their response to treatment, and in an industrial laboratory application the quantitation of lesion volumes is also critical in assessing the effectiveness of drugs undergoing testing. In both cases there is a clear potential contribution to biomedical knowledge and future health care.
