Study On The Computer-aided Diagnostic Of Liver Space-occupying Lesions Based On Deep Features In Abdominal CT Image

In clinical diagnosis,biopsy is the most effective method for the diagnosis of liver space-occupying lesions.However,as an invasive test,biopsy will inevitably bring physical and mental damage to the patient as well as the complication.The rapid development of medical imaging technology provides a new approach for the identification of liver space-occupying lesions.Doctors can diagnose and then analyze the lesions by observing their characteristics from the images.However,most of the existing medical imaging techniques for liver space-occupying lesions rely heavily on the experience and the technology of the operators.As a result,the diagnosis result will suffer from strong subjectivity,low reproducibility,high labor intensity and low efficiency.Therefore,the automated assisted-detection for liver space-occupying lesions is of great importance in clinical applications.Based on the research results on CT image of the liver at home and abroad,we propose a convolution neural network model to identify the type of liver lesion by detecting and segmenting on low-quality CT image.To explore the nature of the CT image for the liver further,we describe the mechanism and mathematical model in multi-layer network of deep learning method,so the precision of the automated assisted-detection for liver space-occupying lesions can be improved considerably.(1)CT scan has been widely applied in clinical examination and disease diagnosis because it can obtain the internal anatomical structure of patients.However,the reconstruction quality of CT image is heavily dependent on the dose of X-ray radiation.Thus,how to reduce the radiation dose without affecting the imaging quality of CT images has been a research hotspot in the field of image enhancement.Existing deep network-based image enhancement algorithms directly learn the mapping function between the degraded images and the clear ones,ignoring the constraints of the fidelity term of the observation model.We propose an improved enhancement algorithm based on neural network image for low-quality CT images in this Section,where the image enhancement tasks are embedded into a deep network.The data consistency can be guaranteed through alternative optimization of multiple enhancement network modules and back projection modules.In the enhancement network module,low-level features and high-level features are combined to form new features,and dual parameter loss function is used to optimize the training process and improve the generalization ability of deep model.The proposed image enhancement algorithm can not only use deep network to learn high-level features,but also use the fidelity prior of observation model to realize image enhancement.The experimental results on simulation test-data and real-data show that the algorithm proposed in this thesis can not only achieve a perfect enhancement effect,but also retain the details of low-quality CT image.Consequently,it would enhance the subsequent localization and segmentation accuracy of liver,the recognition for the focus,and the final disease diagnosis.(2)The abdominal CT images contain many organs with similar gray level.It will bring a large number of false alarms by directly segmenting the image and then increase the complexity of subsequent processing.The region of interest will be narrowed and the accuracy of the subsequent segmentation and recognition will be enhanced when directly detecting the liver region from the image sequence.However,the existing object detection algorithms are difficult to adapt to abdominal CT images that are characteristic of low contrast,complex background and multi-view changes.Therefore,it is still a challenge to accurately detect and locate the liver region.In order to improve the accuracy of liver detection and location,this thesis proposes an improved deep network by combining with edge perception.The network effectively retains the clear boundary of liver through the module of edge perception fusion,and then uses the multi-scale dense pyramid supervision module to capture the rich global context information of abdominal image.A large number of qualitative and quantitative experimental results show that the proposed model can effectively improve the accuracy of the existing network in liver detection and location,narrow the area of interest,and enhance the accuracy of subsequent segmentation and recognition.(3)Since the medical image boundaries in the liver region are not obvious and the internal texture is also quite different,the traditional image segmentation method represented by graph cut,energy functional,etc.,or the machine learning model represented by deep learning cannot be realized.That is because the traditional model based on artificial features has strong interpretability and robust representation effect,but it has poor universality in the application of complex background.On the other hand,the deep-learning-based methods can extract the features of the object through the learning from big-data,but it has poor interpretability to quantify the extracted features.Therefore,according to the detection results of the deep network liver detection algorithm,a two-stage liver region segmentation algorithm based on multi-level deep feature fusion is proposed.In this model,the most similar reference template is first searched from the standard abdominal image data set by using the improved SCNN measurement network,and then the SIFT-flow transform is used for dense matching to obtain the coarse segmentation result of the segmented liver.Finally,multi-level feature fusion is applied to realize the fine segmentation of liver region.Qualitative and quantitative experimental results show that the accuracy of liver segmentation as well as the accuracy of subsequent semantic segmentation and diagnosis for lesions can be enhanced simultaneously.(4)Most of the tumors are located in the interior of the liver.The diagnosis of liver diseases can be achieved only after segmenting the focus area.The existing semantic segmentation models are not able to segment small and weak focus,and their segmentation results are also not consistent in spatial.Therefore,in order to achieve accurate detection and segmentation for occupying lesions,this thesis proposes a supervised generation adversarial semantic segmentation model based on mutual learning.Firstly,the semantic mapping and semantic segmentation of the generator are trained separately in forward training.Afterward,more accurate focus area can be obtained by using semantic result with edge constraint in backward training.The weighted loss function is constructed by using the loss functions of generation network,segmentation network and adversarial network,so as to improve the coupling degree of each sub-module and enhance the generalization ability and training accuracy of the model.Finally,LSSVM classifier is used to detect and classify space occupying lesions.A large number of qualitative and quantitative experimental results show that the deep network architecture proposed in this thesis can steadily improve the performance of semantic segmentation model and the accuracy of tumor type diagnosis.