Classification Of Deep Convolution Neural Networks In Thyroid Ultrasound Images

With the improvement of medical image equipment and the successful application of machine learning algorithms in medical image field,computer aided diagnosis technology has developed rapidly.It is of great clinical value to use depth learning theory to analyze and extract important features of thyroid ultrasound images,and to predict the patient's development trend reasonably.Construction of classification model of thyroid ultrasound images based on depth convolution neural network:1.Classification of thyroid ultrasound images.The present situation of the classification of thyroid ultrasound images was analyzed,and the importance of classification of thyroid ultrasound images was expounded according to the specific clinical data of patients with thyroid nodules.According to the TI-RADS standard system,a reasonable classification performance index is designed.Compared based on ultrasonic image classification method based on ensemble learning and ultrasonic image classification method of artificial neural network,aiming at the shortage of the two methods,put forward a kind of thyroid ultrasound image classification measure of convolutional neural network and residual meshwork,provides theoretical support for the establishment of the model.2.Construction of thyroid echocardiography image classification model based on Convolution NN.The idea that neurons extend to neural networks is explained,and the concepts of convolution and pooling are introduced.On this basis,construct the convolutional neural network model,discusses the different activation application function in thyroid ultrasound image,and a detailed analysis of the training process of thyroid ultrasound image classification algorithm of convolution model based on neural network.In view of the particularity of ultrasonic image,we use the unsaturated nonlinear function to improve the convolutional neural network.Especially,in order to avoid serious overfitting,we introduce the method of stochastic gradient descent and regularization.3.The construction of the thyroid ultrasound image classification model based on the depth convolution neural network of residual structure.Through depth analysis of residual network advantages and disadvantages,in-depth analysis and deep influence on the parallel convolution convolution model,calculated on the residuals of each layer of the network output,introducing residual block structure of cross layer connection,realize the complementary advantages of improved convolutional neural network CNN6 and residual network Resnet.Based on the above research,set as training and testing samples with color ultrasound image data of the Second Affiliated Hospital of Ningxia Medical University in the design of experiments on the feasibility of this method in TensorFlow deep learning framework for verification.The experimental results suggest that this method can efficaciously detect and classify malignant thyroid nodules with higher accuracy.