A Tree - structured Approach to Medical Diagnosis

A general approach for the generation of structured knowledge by integrating tree-structured and connec-tionist mechanisms is proposed. We present our main ideas concerning the amalgamation of symbolic tree structures and topology preserving neural networks. Goals of this research are to develop a model that is explainable in symbolic terms and supports direct interaction with the user as well as being trainable and robust like connectionist systems. As a report on the progress of our work, we introduce a dynamic hierarchical neural architecture that provides topology preservation and structural interpretability. It consists of a number of Kohonen chains which grow by inserting neurons according to a quantization error criterion. The hierarchical structure is developed mainly by detecting topological defects in a chain and resolving them by splitting of the chain. Neighboring chains remain connected and this topology preserving feature provides the interpretability of the network structure. We evaluate the topologic eeect by using a known measure for the topology preservation and give a simple example for the training result in two dimensions. We then demonstrate the capabilities of our approach in the medical diagnosis of Ulnar nerve lesions, which lead to motoric dysfunction of the ring and the little nger. Visual inspection by medical experts is supposed to be supported by an automatic diagnosis method that is based on the experience of many other previous cases. It is shown how our network model can be used to reach the goals of medical diagnosis such as distinction of diierent dysfunctions or judgement of success after surgical reconstruction. We present the current state of our research and outline future work.