Mixed-integer evolution strategies for parameter optimization and their applications to medical image analysis
05 / 2005 - 10 / 2009
Nederlandse Organisatie voor Wetenschappelijk Onderzoek (NWO)
Over the past several years we have successfully developed a multi-agent image interpretation system with software agents functioning as (segmentation) experts on a particular image object. These agents collectively aim at a consistent image interpretation result through communication, collaboration, and the resolution of conflicts. We tested this approach on medical images and demonstrated that multi-agent image segmentation was more reliable and could be applied to a larger range of problems than could be achieved otherwise. Agent knowledge (about 500 rules) is thereby mostly generic and modular. Only approximately 5% of these knowledge rules is application specific and describes (1) high-level common sense knowledge about the application area, and (2) very specific knowledge concerning which segmentation algorithm and parameters to use under various conditions and how to interpret the consequent (lack of) results. However, this second knowledge category is very difficult to specify and does not easily transfer to another image interpretation domain. Effectively it causes a 'knowledge gap' that needs to be bridged for each new application area. This hampers the general applicability of multi-agent image processing, which is otherwise very promising and highly innovative. To create a truly generic system, we propose to enable agents to bridge this 'knowledge gap' automatically by learning such knowledge through a hybrid combination of reinforcement learning and evolutionary algorithms. Goal and challenge in this multi-agent environment with inter-agent dependencies is to make agents automatically learn which image interpretation strategy and parameters are most appropriate and how to interpret the consequent results.