Ultrasound Diagnosis Of Knee Joint Disease Based On Multi-channel Stacking And Graph Embedding

Knee joint,as the most complex structure of human joints,is often infected or injured.Ultrasound diagnosis is used for common knee diseases,such as synovitis,synovial thickening,cysts and so on.In the ultrasound diagnosis,the symptoms of these knee diseases are black effusion area,which is the lesion area and the main basis for doctors to judge.So the accuracy of the delineation of this area affects the diagnosis of the doctor.At present,the diagnosis of knee joint diseases depends entirely on the naked eye of doctors,which wastes manpower,material resources and leads to subjective,artificial errors.Therefore,in order to assist doctors to diagnose or even replace doctors’ diagnosis,this paper studies two aspects of automatic segmentation of knee joint effusion area and automatic recognition of common knee joint diseases.For the task of automatic image segmentation,deep learning algorithm is usedforsemantic segmentation,and three commonly used deep networks are compared to achieve better segmentation effect.The reasons for the erroneous segmentation and edge segmentation deviation when the ultrasound image is directly input into the deep neural network are analyzed.The improvement is made from three aspects for the segmentation task.The first one is to use the dilated convolution to expand the receptive field,and the second one is shrinking the outer contour to remove the easily mis-segmented area using the Snakes model algorithm.The last one is to improve the edge region segmentation result using the multichannel learning method.For the task of automatic recognition of knee joint diseases,the classification based on this data set is different from that based on ordinary visual image.Firstly,this paper will use the residual network classify the data set,visualize the feature vectors used for classification in the network and analyze the reasons for the low accuracy of the classification network.Extracting better feature vectors is the key to improve the accuracy.The image is expressed as a vector and mapped to low-dimensional space by using the graph embeddingalgorithm,which makes the feature vectors of the same kind of image more similar in space and the distance between different kinds of feature vectors farther,so as to get more separable feature vectors,and then through the classifier.Classification accuracy is improved in training set and test set.By the above algorithm optimization,the average intersection in the segmentation index in the segmentation task is increased by about 10% in the training set and the test set.In the process of disease identification,the optimized classification accuracy of the secondary training increased by nearly 11% in both the training set and the test set.