Research On Automatic Segmentation Of Thyroid Ultrasound Image Based On Deep Learning

Segmentation of regions of interest in medical images is a key step for the quantitative analysis and computer-aided diagnosis of lesions.Automatic segmentation of medical images has become a challenging topic in the field of computer vision due to the low contrast,high noise and less annotated images.In recent years,the breakthrouths of deep learning in various fields have brought new opportunities for comprehensively solving the problem.Based on this background,this paper takes the thyroid ultrasound images as the research object,deep learning algorithm as the research method,and uses the deep learning algorithm to achieve a certain breakthrough on the automatic segmentation of medical images.Based on intensive analysis of the U-shape network represented by U-Net,we proposed Re Ag U-Net model,which embedded the improved residual units into the skip connection among the encoding path and decoding path of U-Net in order to reduce the semantic gap between the feature maps from shallow layers and deep layers.Afterwards,batch-normalization operation is introduced to let the input fall into the range where the nonlinear function is sensitive,which is conducive to enlarge the gradient of the back propagation and speed up the convergence of the model training process.Besides,introduce the attention gate(AG)mechanism to multiply the weight feature maps obtained from shallow layers and the abstract features obtained from deep layers,which beneficial to correct the target position while deepening the network horizontally.In order to exploit the respective advantages of the existing loss functions,a hyperparameter is introduced to combine Focal-Tversky Loss,Dice Loss and Cross-entropy Loss to jointly guide the model training process.On the basis of in-depth analysis of the Generative Adversarial Network,we proposed U-Seg AN model.We use Re Ag U-Net to replace the generator of the Seg AN model,which greatly reduce the parameter numbers and improve the segmentation accuracy.On the basis of considering L 1 and L 2 loss,LMSE(Log Mean Square Error)is proposed as the loss function of Seg AN model.The convergence speed of Seg AN is accelerated,the space occupation,generalization ability and accuracy are further improved.The rationality and effectiveness of the proposed model are verified from several aspects by setting performance comparison experiments between different models under the same dataset,model performance under different datasets,and space occupancy of each model.Experiment results show that the proposed algorithm is effective and has good segmentation accuracy and generalization ability.On this basis,the paper implements the segmentation prototype system,analyzes and designs the system,and implements the data management,results display and performance evaluation modules.