Application Research Of Convolution Neural Networks In Medical Image Processing

Convolution neural networks are often superior to other algorithms in image classification,which convolution and sub-sampling layers have the ability to extract sample features.In addition,the characteristics of shared weights greatly reduce the parameters of network training.With the continuous progress of science and technology,medical technology has also been rapid development.At the same time,this phenomenon has led to a great many medical images.Doctors need to get rid of all kinds of heavy medical image screening work,and how to find out the similar characteristics of some diseases from the numerous cases.Therefore,the study of medical image has become a hot topic because of many difficult problems.This thesis studies the application of CNNs in two kinds of medical images.One of them is to reflect the physical disease of the eye bloodshot pictures,and the other is a series of various types of brain gliomas magnetic resonance imaging.All the work can be summarized as follows:(1)This thesis introduces the development process of CNNs,including the domestic and foreign research results.In addition,the thesis describes the structure,algorithm and derivation of CNNs in detail,and the superiority of CNNs for complex image classification is discussed.(2)The classical LeNet-5 structure has been improved,which has different convolution kernels,different sub-sampling methods and different classifiers.And this structure is used to solve the problem of identifying eye bloodshot pictures.And that,the sample size of the input layer and the number of iterations of the network are studied in the experiment.The difference between the improved structure and the LeNet-5 structure is compared in the case of same data set.In the experiment,the improved structure can classify the disease well.(3)According to the characteristics of multiple images of brain gliomas and base on eyeball blood network model,a multi-column CNNs is designed.Each sample is used as the input of each column,and the convolution and the number of sub-sampling layers are increased.In particular,the Maxout function is used instead of the Sigmoid function that is often used in traditional neural networks.In the experiment,the multi-column structure is compared with the single column structure and the manual extraction method.The results show that the multi-column structure is better.In addition,the samples are optimized to improve the classification accuracy.For the sake of learning more about how the network is extracted and learned,the thesis also gives the visual processing of the multi-column CNNs.