Smart Pathological Brain Detection System By Predator-prey Particle Swarm Optimization And Single-hidden Layer Neural-Network

(1) Objective: This thesis introduced the background and significance of computer-aided diagnosis and then made a brief introduction to the development of magnetic resonance (MR) image in China and abroad. The smart pathological brain detection system (SPBD) proposed in this thesis is a computer-aided diagnosis system.The research of artificial intelligence algorithm will help improve the efficiency and accuracy of detection classification, which is of great significance in the field of brain detection. An idea of neural network learning algorithm combined with MR images was used. The training of the neural network can easily fall into the local optimal points.This thesis used a new and very effective predator-prey particle swarm optimization(PP-PSO) algorithm to optimize the neural network. It improved the classification performance of the SPBD for new samples, which improved the detection efficiency and accuracy.(2) Methods: This thesis used DA-160 dataset and used Hu moment invariant(HMI) to extract brain image features. HMI features are in translation, rotation, and scale invariance, it is widely used in target recognition, image matching, shape analysis and other fields. In this thesis, a single-hidden layer neural-network (SLN) was used as the classifier. Artificial neural network (ANN) constructs neural network by imitating human brain, and then the distributed information processing is implemented. It has good adaptability, self-organization and strong self-learning ability. It is a powerful tool for data classification and image recognition. A series of matrix information composed of seven characteristic moments extracted by HMI was fed into SLN. After SLN training, the output result obtain information of 0 or 1 (0 means healthy brain image, 1 means pathological brain image). The particle swarm optimization (PSO) algorithm above was easily falling into local optimal points. To improve its performance, this thesis adopted an improved algorithm — the predator-prey particle swarm optimization(PP-PSO) algorithm to train the weights of SLN. The five-fold stratified cross validation (FFSCV) will be used for statistical analysis. FFSCV was used to guarantee that the limited dataset was learned and trained as much as possible. Finally, the classification accuracy was used as an excellent criterion for the evaluation of the experiment.(3) Results: The experimental results was compared with the other six state-of-the-art SPBD algorithms. Based on the comparison of the training results, the predator-prey particle swarm algorithm and the single hidden layer neural network algorithm(HMI + SLN + PP-PSO) was found out the best. On average, the method achieved a sensitivity of 96.00±5.16%,a specificity of 98.57±0.75%,and an accuracy of 98.25±0.65% for the DA-160. Of course,we also compared the PSO and PP-PSO,respectively, in terms of accuracy. PSO achieved an accuracy of 96.44%.(4) Conclusion: By comparison, the method (HMI + SLN + PP-PSO) performed best. Moreover, the advantages and disadvantages of the method can be found by the comparative analysis of the experimental results, which further paved the way for further research and optimization of SPBD.