Reasearch On Automatic Stratification For Thyroid Nodules In Ultrasound Images

In last years,the incidence rate of thyroid nodules increased year by year.Since the cause of the disease remains unknown,early detection and diagnosis is the key for thyroid nodules control and the success of treatment.Ultrasound is a widely used non-invasive imaging study used for diagnosis of thyroid gland.However,low quality of ultrasound images and the diverse characteristics of thyroid nodules make it very difficulty to diagnose thyroid nodule.Meanwhile,computer-aided diagnosis system can improve not only the efficiency of diagosis but also its accuracy.Therefore,the study of computer aided diagnosis of thyroid nodule has got more and more research attentions.In this thesis,the related technologies of computer aided diagnosis of thyroid nodule were studied.The research mainly includes four parts: image pre-processing,nodule segmentation,feature extraction,classification and stratification of malignant nodules.The main contributions of this thesis are as follows:(1)Considering the low quality of ultrasound image,we proposed a semi-automatic segmentation algorithm based on active contour model.(2)Based on the sonographic features used in thyroid diagnostic criteria,we designed several statistic features,and combined them with some traditional shape and texture features for classification.(3)Two algorithms were proposed for benign/malignant thyroid nodule classification,both of which adopt feature selection mechanism.One algorithm uses filter-based feature selection and support vector machine,another adopts the LASSO logic regression,in which L1 regularization is used to enforce the sparsity among features.(4)Based on a TI-RADS diagnostic standard,an ordinal regression based algorithm was proposed for stratification of malignant thyroid nodules.