Research On Cirrhosis Recognition Based On Deep Learning

Cirrhosis recognition using B-model ultrasound images is important to diagnosis of liver diseases in clinics.But some influential factors like noise,different scales,edge blurring,and nonuniform of echo are inevitable.Traditional cirrhosis recognition methods have lower recognition rate,mainly due to conventional feature extraction methods have less capability in describing cirrhosis characteristics.Focusing on the above problems,an improved LBP(Local Binary Pattern)algorithm is proposed to solve the problem of edge blurring;then we apply deep learning theory and propose two kinds of Convolutional Neural Networks(CNNs)models and a Deep Belief Nets(DBNs)model.Finally,we obtained features with stronger describing capability.The main works include following aspects:(1)We improved LBP algorithm and introduce information entropy to measure the quantity of texture of samples transformed by improved LBP.Samples with largest quantity of texture are used as the input of SVM and ELM,experimental results show that the improved LBP transform is benefit for cirrhosis recognition.(2)We utilize deep learning thought and propose two kinds of CNNs models :Liver-CNN1 and Liver-CNN2.We use SVM and ELM to take the place of full-connected neural networks at the bottom normal CNNs structures.The combination of proposed CNNs models and SVM as well as ELM not only obtained higher recognition rate,but also less time consumption.The concept of receptive field and weights sharing of CNNs have seriously decreased the number of parameters,then the speed of learning has been enhanced.Comparison had shown that features learned by CNNs can result in higher recognition rate compared with traditional features,that is to say,it is able to describe the characteristics of cirrhosis precisely.(3)A method of combining DBNs and SVM as well as ELM is proposed,the layer of logistic regression is substituted by SVM or ELM,so the process of measurement between logistic values and class values can be avoided.Through pre-training and layer-wise learning we got the solution which is approximate to global optimum solution eventually,rate of convergence is faster.