Study On Intelligent Diagnosis Of Thyroid Nodules Based On Ultrasound Image

Thyroid nodules are one of the most common endocrine diseases in the adult population.The incidence of thyroid nodules has been growing in China in recent years.Early screening and diagnosis of thyroid nodules are of great significance for subsequent treatment.Ultrasonography has become the most important imaging diagnostic method for early screening of thyroid nodules due to the advantages of real-time dynamic observation,non-invasive,and non-radiation.The intelligent diagnosis of thyroid nodules based on ultrasonic imaging can not only improve the efficiency of clinical diagnosis and reduce the working pressure of medical staff but also make the results more objective and reliable.In this thesis,an experimental data set was constructed based on clinically acquired thyroid ultrasound images,and the intelligent diagnosis method of thyroid nodules was carried out based on the data set.Firstly,the thyroid nodules were segmented by an intelligent algorithm to obtain the target region,and then from the perspectives of clinical interpretability and objective feature extraction,machine learning and deep learning models were designed to achieve the classification of benign and malignant thyroid nodules.the specific works include the following three aspects:Focusing on the complex structure of thyroid ultrasound images and difficulty in extracting the boundary of nodules,a multi-level feature fusion model with Transformer as the coding structure was proposed to achieve the segmentation of thyroid nodules.The local-global strategy was used to learn the joint scale distribution features in thyroid ultrasound images.the residual axial attention mechanism was used to extract the directional texture information in the image.15312 thyroid ultrasound images were obtained after the data augmentation of 3828 samples from the thyroid ultrasound data set.After several rounds of iterative training,the results of verification showed that the Dice coefficient was 92.2%,and the intersection over union was 85.5%.According to the ACR TI-RADS classification standard proposed by the American College of Radiology,34 imaging omics indicators were designed for the five types of descriptive features of thyroid nodules,including composition,echo,shape,boundary and calcification.Then,the principal component analysis was used to reduce the feature dimension of 34-dimensional feature vectors,and five principal component vectors were obtained.support vector machines and gradient boosting decision tree machine learning models were built for benign and malignant thyroid nodules.The classification accuracy of the two methods on the test set were 92.45%and 96.36% respectively,and the AUC were 0.929 and 0.950 respectively.Aiming at the subjectivity of the above-mentioned machine learning model in extracting features,the EfficientNet B3 lightweight network was used to automatically learn the characteristics of the nodule region,and the Focal Loss function was used to solve the problem of uneven distribution of benign and malignant samples in the data set.In the training process,the transfer learning strategy was combined to further improve the speed of the model.The results showed that the accuracy of benign and malignant classification of the EfficientNet B3 model was97.27%,the AUC was 0.970,and the number of model parameters was reduced to 1/6of the mainstream model,which realized the objectification of the feature extraction process and reduced the Computational load.Preliminary research showed that the work in this thesis had certain clinical application value in the early screening of thyroid nodules,which would also contribute to the popularization of intelligent auxiliary diagnosis systems in clinical application.