Research And Application Of Transfer-learning For Diagnosis Of Thyroid Nodules Traits

If we can accurately recognize and classify thyroid nodules traits in the medical diagnosis of thyroid nodular disorders,it will play a significant role in the entire process of medical diagnosis.Thyroid nodule traits recognition based on convolutional neural networks has substantial clinical application value with the advancement of deep learning technology.However,several issues remain,such as a lack of data samples and a model with low recognition accuracy.In order to solve the problem that there are few thyroid ultrasound image samples,it is difficult to train an ideal recognition model,this work suggests a data augmentation strategy based on SinGAN and using transfer learning to increase the classification model’s performance.What an improved ResNet-152 model was built to improve the accuracy of thyroid nodule traits recognition due to the ResNet-152 transfer learning model’s poor performance in recognizing thyroid nodules traits because its failure to effectively extract the best characteristics.The following are the research topics and achievements for this thesis:(1)The research on thyroid ultrasound image data augmentation is carried out,and the method of data augmentation based on SinGAN is proposed,in order to address the problem that few samples of thyroid ultrasound images are not conducive to building an ideal recognition model.Multi-scale image generation and super-resolution image generation are the two data augmentation technologies used.Simultaneously,SinGAN is improved to address the issue of this method’s poor performance in superresolution image generation,namely,the activation function is changed to increase the performance of the depth image prior network during image generation.The results of a comparative experiment using real dataset reveal that the improved data augmentation method is both feasible and effective.(2)The research on thyroid nodule traits recognition model based on four kinds of transfer learning is being carried out in order to solve the problem of training an excellent deep learning classification model on a low amount of data.A suitable recognition model,namely ResNet-152 transfer learning model,is selected after a comparative experiment of four kinds of transfer models under three types of datasets.At the same time,the chosen model ResNet-152 has a poor accuracy problem due to the inability to entirely extract the best discriminant features during the training process,and the model is improved research.The ResNet-152 architecture is improved by adding an additional three-layer ’Conv’,’Batch＿Normaliz’,and ’Activation＿Relu’ layer,as well as a new fully connection layer.The improved model is called ResNet-155.Comparative experiments of the transfer learning model before and after optimization were performed using three types of datasets.Then the results of the experiments reveal that the improved model is effective.(3)Develop an auxiliary diagnosis system for thyroid nodule traits recognition.The improved ResNet-155 model is utilized as the foundation for designing and developing a thyroid nodule traits recognition auxiliary diagnosis system in this thesis,and the feasibility analysis,demand analysis,and architecture design are all done using the software engineering technique.Finally,the system completed the detailed design and function realization.