Research On The Application Of Neural Networks In Medical Diagnosis

Neural Networks is a cross-discipline which integrates neural science, information science, computer science, and it has been developing rapidly in recent years. Neural Networks is an information processing system through abstracting, simplifying, simulating the structure, function and other aspects of biological Neural Networks theory. The application of Neural Networks has penetrated into various fields and made an impressive progress. With the advancement of various cutting-edge technologies, modern medicine benefits significantly and develops rapidly. With its unique advantages, Neural Networks has been applied to several areas of medical diagnosis, such as signal processing, feature extraction, pattern recognition, etc. Satisfied performance has been achieved, which arouses the attention of many researchers. As a result, Neural Networks will have a bright prospect in the field of medical diagnosis.At present, BP Neural Networks is a kind of Neural Networks which has been widely used in the medical diagnosis. However, the learning algorithm of BP Neural Networks in the medical diagnosis still has shortages, such as the slow speed of convergence and the difficulty of determining the number of hider layer units. In this paper, these problems are systematically analyzed and discussed in detail. Furthermore, the solutions are put forward pertinently and applied to some medical diagnosis examples.The main work and conclusions of this paper are summarized as follows:(1) On the base of comprehensively analyzing Neural Networks, medical diagnosis and the application of Neural Networks in the field of medical diagnosis, the most widely used BP Neural Networks in medical diagnosis is studied. The structure and learning rules of BP Neural Networks are analyzed in detail. To solve the slow convergence speed and ambiguous defects in the hider layer units, several improved Neural Networks models are analyzed and summarized. Furthermore, a method of determining the best number of hider layer units is proposed, which is named boundary numbers limiting the number of hider layer units. This method is proved to be feasible and superior, which could optimizes the structure of BP Neural Networks effectively.(2) Based on analyzing and summarizing the modularized network, the optimized structure of BP Neural Networks and the modularized network are applied to the practical example of medical diagnosis - the diagnosis of Parkinson's disease (PD). The respective BP diagnosis network is established on classifying the diagnosis indicators of Parkinson's disease. Firstly, we calculate each kind of diagnosis network, and then integrate all the results to get a final judgment. Compared with the conventional network, our results show that the correct detection rate of Parkinson's patients and non-Parkinson's patients is increased, and the misdiagnosis rate of Parkinson's patients and non-Parkinson's patients is reduced in test samples.Two typical innovations in this paper are presented below:(1) To determine the number of BP Neural Networks hider layer units, a method of determining the best number of hider layer is proposed, which is named boundary numbers limiting the number of hider layer units. This method optimizes the structure of BP Neural Networks effectively and it is proven to be feasible and superior.(2) The method of modularizing network is applied to medical diagnosis model of Parkinson's disease for the first time. With further optimization by the method of boundary numbers limiting the number of hider layer units, the modularized network is proven to be feasible in practical applications.