Research On Texture Feature Extraction And Semi-supervised Classification Of Thyroid Nodules Ultrasound Image

Thyroid nodule is a common clinical disease,caused by a variety of factors.B-ultrasonic imaging technology has become the most commonly used method because of its low cost,non-invasive,fast imaging,real-time diagnosis,repeatability and other characteristics.In this thesis,ultrasound images are used as the study object to extract the texture features of the ultrasound subgraph and to quantify the ultrasound features of the region of interest(ROI)of thyroid nodules.The ultrasound images and clinical data are used to analyze the image texture feature and the ultrasonic signs of thyroid imaging reporting and data system(TI-RADS),providing the feature set for the identification model.First of all,we collect and analyze 449 cases of thyroid nodules ultrasound image.All the ultrasound images are taken from the ultrasound video system,and then the boundary of the nodule is marked.According to the standard of TI-RADS,the nodule signs,the performance of each and the final diagnosis of nodules are sorted out.Then the thesis introduces two common image segmentation techniques:Normalized Cut and Snakes,and applied them to the segmentation of thyroid.Next,based on the dual texture complex wavelets transform(DT-CWT)and Gabor,a texture feature extraction method of thyroid nodule based on multi-scale fusion is proposed.In the method,the ROI of thyroid ultrasound image is subjected to DT-CWT and Gabor transform to obtain the texture image,and then calculate the mean and variance of the texture image.The feature fusion is realized by the end-to-end method,to achieve the benign and malignant of thyroid nodules by classifier.Finally,a semi-supervised classification method for thyroid nodules ultrasound image is proposed.TI-RADS is the standard for diagnosing thyroid nodules.TI-RADS can normatively and accurately guide ultrasound workers to perform ultrasonic examination.As a computer-aided diagnostic system,we firstly quantitatively analyze the ultrasound features of TI-RADS,and then these signs are used as the eigenvectors to judge the thyroid nodules,and then the semi-supervised fuzzy C-means clustering model is used to cluster the results.The experimental results show that the method can discriminate the different thyroid nodules.In this thesis,we study the thyroid ultrasound image from two aspects:the texture feature extraction method and semi-supervised classification method,and achieve a better result.As a computer-aided diagnostic method,it is hoped that it will assist doctors in diagnosing thyroid nodule disease,reduce the subjective judgment of doctors,provide effective diagnosis suggestions,and further promote the application of machine learning method in medical.