| The liver is the most important metabolic organ in the human body,which is closely related to people’s life and health.Liver diseases can be mainly divided into two types,diffuse and focal lesions.Liver fibrosis is the most common liver disease in diffuse lesions,and its development may lead to the appearance of hepatocellular carcinoma(HCC)in the later stage,while HCC is the most harmful liver disease in focal lesions with high mortality.Improving the early detection rate of these diseases is beneficial to provide effective guidance for the treatment of patients and improve the cure rate of patients.Ultrasound imaging has been widely used in the clinical diagnosis of liver diseases due to its non-invasive,repeatable,and no radialization advantages.Therefore,the study of liver fibrosis and focal liver cancer based on ultrasound images is of great significance and can provide effective assistance for clinicians.However,the ultrasonic image has its inherent disadvantages,such as poor contrast,reduced resolution,and blurred tissue boundary due to the existence of speckle noise.This paper mainly focuses on the speckle noise suppression of ultrasonic images and the classification of the two typical liver lesions with ultrasonic images.The main contributions are as follows:(1)The key to speckle noise suppression in the ultrasonic image is to protect tissue details while smoothing the uniform area.In this paper,a hybrid deep neural network based on distribution prior and tissue structural prior is proposed to suppress ultrasonic speckle noise.The distribution parameters of ultrasonic speckle noise are estimated in the logarithmic domain according to the model of ultrasonic speckle noise.Meanwhile,a deep network is used to extract tissue structural information to fine-tune the speckle noise suppression network and improve the performance of tissue structural detail protection.Experimental results show that this method can protect the details of tissue structure well while suppressing speckle noise.(2)The differentiation between different grades of liver fibrosis based on ultrasound images is fine-grained,which affects the fine classification of liver fibrosis grades.In this paper,we propose a deep network based on local consistent attention integration to achieve a fine classification of liver fibrosis.The network uses an image pyramid structure to construct a multi-resolution representation of images and then extracts the detailed features that are distinguishing between different liver fibrosis grades.Then,a local consistent attention mechanism is embedded in each resolution to suppress interference information and enhance feature representation.The experimental results show that this method can effectively detect the distinguishing features and achieve the fine classification of liver fibrosis grades.(3)How to extract the time correlation information among different contrast-enhanced ultrasound perfusion stages and how to effectively fuse the characteristics of different perfusion stages are the key to achieving an accurate classification of HCC.In this paper,a deep neural network based on two-step orthogonal projection fusion is proposed for focal liver cancer classification.The network uses a views-related learning module to learn the temporal correlation among different perfusion stages and then uses the correlation to enhance the characteristics of a single perfusion stage.At the same time,the network uses two-step orthogonal projection to fuse the features of different perfusion stages and remove the redundant information among them.The experimental results show that this method can effectively explore the time correlation between different perfusion stages can effectively integrate the characteristics of different perfusion stages,and improve the diagnosis performance of HCC. |
