Research On Stem Cell Mechanism-based Evolutionary Neural Network & Its Application

Back propagation neural network (BPN) is a self-adaptive, self-learning and non-linear algorithm and it is valuable for solving non-linear and complex problems. However, it easily traps into a local minimum and converges slowly and its generalization ability is poor. Starting from training samples and training algorithm, this paper explores optimization of BPN. On that basis, it is investigated to establish the diagnosis models for distinguishing nature of stomach diseases.First splitting method of training set is investigated based on the similarity to mean vector in this paper. In general, the total samples will be fixed after all data have been collected. So, the quality of training sample can be improved only by its extraction method. This paper proposes a new splitting method of sample set based on the similarity to mean vector (SSSMV) to improve the quality of training sample. SSSMV represents each sample as a vector and establishes a virtual vector, named mean vector, which is composed of mean value. Then the similarity of each sample to mean vector is calculated. At last, samples are extracted to build a training set based on their similarity and category. A simulation study is taken using UCI data sets. Hotelling T2 test shows that there is no significant difference between the training set and the test set from SSSMV and it is better than random splitting samples (RSS). The sample sets from SSSMV and RSS are used for training BPN. The results show that the neural network from SSSMV has better generalization ability, its error difference and accuracy difference between training set and test set is smaller than that of the neural network from RSS.Then this study explores optimization of training algorithm of BPN based on stem cell mechanism in cytology. Genetic algorithm-based BPN (GABPN) is considered as one of the most promising optimization methods of BPN. It can make a network for the global convergence, but its convergence speed is slower. In addition, more parameters are added into and setting properly these parameters becomes more difficult. Therefore, learning from stem cell theory in cytology, this study explores optimization of BPN and proposes stem cell-based evolutionary BPN (SCEBPN). The main algorithm and evolutionary framework of SCEBPN is established drawing on such mechanisms as reproduction, transplantation and differentiation of stem cell, replacement and apoptosis of aged cell. Its operators are designed, which include reproduction, transplantation and differentiation of stem cell node, replacement and apoptosis of aged node. In SCEBPN, selection operation is used, but such operations as encoding and decoding, crossover, mutation and so on are leaved out. So its computational complexity drops greatly down. Simulation results show that SCEBPN can not only ensure a BPN towards global convergence, but also it can converge faster than GABPN and BPN. These results confirm that SCEBPN is an excellent algorithm for improving a BPN.Finally, it is investigated to establish the diagnosis models of stomach diseases by ultrasonography using SCEBPN. Ultrasound diagnosis of stomach diseases is selected as subject and data about stomach diseases are collected by color ultrasound. The collected data is split into training set and test set. The diagnosis model of stomach diseases is successfully built using SCEBPN and it is trained by training set. It solves the problems of multiple linear regression analysis, BPN and GABPN for designing diagnosis models. The models are tested using test set. The SCEBPN model can effectively identify the nature of stomach diseases, its accuracy is as high as 94.7% and its area under curve is 0.936 in receiver operating characteristic analysis. These Experimental results confirm that SCEBPN is superior to multiple linear regression analysis, BPN and GABPN, it can be used for establishing a diagnosis model of diseases and it has a good practicability.In summary, this paper explores optimization problem about BPN from the aspects of training sample and training algorithm. A new extraction method of training samples is put forward based on the similarity to the mean vector and a new training algorithm based on stem cell mechanism is proposed to optimize BPN. On that basis, it is studied to establish the diagnosis models of stomach diseases by ultrasonography and it is confirmed that SCEBPN is of great value in this practical application.