Research On Computer Aided Diagnosis Of Liver Cirrhosis Based On Deep Learning

Liver cirrhosis is a kind of liver damage caused by long-term or repeated action of one or more causes,which has diffuse characteristics and seriously endangers human health.Liver cirrhosis,a chronic progressive liver disease,is more common in clinic.Serological examination,CT,MRI and ultrasonography are the main diagnostic methods.Serological examination is reliable,but it is a kind of invasive examination,there are some limitations in its wide use in clinic.CT,MRI and other techniques are suitable for the diagnosis of space-occupying lesions,the auxiliary diagnosis of diffuse diseases such as cirrhosis needs to be further studied,CT and MRI are not suitable for screening large areas of people with cirrhosis.Ultrasound is a widely used examination method,but clinicians can judge the severity of liver cirrhosis subjectively by observing the texture of ultrasound images based on their own experience,which may lead to misjudgment.In this paper,a computer-aided diagnosis method of liver cirrhosis based on in-depth learning is proposed.By extracting patch features from high-frequency ultrasound images and combining voting method,the final stage of liver state can be achieved,and the efficiency and accuracy of clinical diagnosis can be improved.Quantitative staging method of liver cirrhosis mainly consists of four core modules: liver capsule location module,patch block acquisition module,network training module and two-level classification module.Firstly,the location of hepatic envelope is localized preliminarily by means of local block pixel mean search method,the envelope area and substance area are separated by the position of envelope line.Different sliding window strategies are adopted to obtain the envelope and substance patch blocks,the patch blocks are further filtered by using the statistical characteristics of gray level co-occurrence matrix to remove black background blocks and poor image quality blocks;Secondly,according to the characterization of patch patches in capsule and parenchyma at different stages of cirrhosis,the size of patch patches in two regions and the number of patch patches with acceptable quality etc,different deep neural network models are trained,and the network structure is improved before training,the appropriate convolution layer parameters,network layer types and number of patches are selected,the number of layers and the types of activation functions;Finally,two-level classification method is used to test the validated images,the first network combined with block voting method is used to discriminate severe liver cirrhosis,the second network combined with block voting method is used to discriminate the remaining stages,local features are used to aided diagnosis liver cirrhosis of the whole ultrasound image.Patch blocks are acquired from the high frequency ultrasound images provided by Shanghai Changzheng Hospital and the normal control group images(the liver status of these images was confirmed by various examinations).Patch blocks are trained and tested,and a set of patch blocks are obtained by changing the sliding window strategy of patch blocks.The results showed that the recognition rates of centralized normal control group,mild cirrhosis,moderate cirrhosis and severe cirrhosis are 95.00%,88.90%,94.10% and 92.30% respectively,which are superior to other advanced auxiliary diagnostic methods.This method can provide certain diagnostic basis for clinical application and assist doctors in the diagnosis of liver cirrhosis staging.It can also provide some ideological reference for researchers in the field of computer-aided diagnosis.It can migrate this method to other kinds of disease diagnosis,and further expand the application field of this method.