Identification Lung Nodules From CT Image Based On Improved BP Neural Network

Microscopic technical development makes the computer aided technology in lung CT images, the role of nodular detection is more and more obvious. Computer-aided technique can be applied to every aspect of image data analysis of the test results, and then shows, reliable test results must rely on the right to do image segmentation, feature extraction and rely on subsequent machine learning techniques. The results are better able to deal with the more secure help the doctor to finish the clinical diagnosis is difficult to human body, make the doctor internal lesion site observation more direct and clearer, diagnose rate is also higher.In CT images,there are large vessels and bronchi to be easy to create false positive, because blood vessels and bronchi in the CT images scan faults in general terms the shape of the lung nodules is very similar, it is difficult to distinguish. Therefore, in CT image segmentation and pulmonary nodules recognition, existence many difficulties.The BP neural network is a traditional machine learning techniques, in recognition of images in a wide range of applications, but BP neural network, there also exist many shortcomings, such as network instability (for initialized weights sensitive), convergence and local minimal issues, the serious influence recognition result. Thesis main job is that:1.The present medical image segmentation method was summarized, the analysis of the advantages and disadvantages, and gives some common segmentation method of experimental results, according to the mathematical morphology method in lung CT images edge noise sensitive faults, and carry on the simple improved lung CT images, the first after simple pretreatment (remove background, etc), and then after wavelet transformation of mathematical morphology after recycling of high frequency and low frequency conversion, forming repair operations after reconstruction multiscale image segmentation results, obtained.2. The BP neural network in the lung nodules recognition for the shortcomings of weights of the neural network are unstable, large amount of calculation, easy lead convergence and into the local extremum shortcomings, came up with a based on improved genetic algorithm to optimize the MTANN (massive training artificial neural network) algorithm, using the improved genetic algorithm to MTANN weights of the genetic algorithm and optimized adjustment in ways for floating-point coding method, with the sort of genetic algorithm to optimize the MTANN after adjusting network weights, networking as initialized weights. So after the adjustment MTANN network with genetic algorithm not only global stochastic search ability, and has MTANN network stability and stronger self learning ability.The genetic algorithm was applied in the MTANN lung CT images pulmonary nodules recognition, through the simulation result research shows that improved MTANN based on genetic algorithm, not only can improve network convergence speed, avoid MTANN network in training into the local extremum, and can be a very good overcome MTANN algorithm training for a long time, such as the insufficient training instability.