The Study Of Thyroid Nodule Recognition Based On CT Images

Thyroid nodules(TN)are the local lesions caused by abnormal growth of thyroid cells in the glands.Computed tomography(CT)examination is an effective method for the diagnosis of thyroid nodules,and has obvious advantages for the diagnosis of retrosternal goiters and in detecting intra-thyroid calcifications.In clinical practice,the diagnosis of thyroid in CT images is easily influenced by the training knowledge and experience of the individual doctors.Radiologists visually inspect a large amount of thyroid CT images in rutine work,which is tedious and error-prone.Some subtle CT features,like micro-calcifications,could be missed in visual inspection.In order to overcome the above shortcomings and assist radiologists to improve efficiency and provide objective and quantitative diagnostic information,the imaging biomarkers based on CT image were considered to identify the thyroid nodule.The computer aided thyroid nodule recognition technique and differential diagnosis technology of benign and malignant nodules based on CT images were presented in this thesis.Firstly,the technique of thyroid nodule recognition based on CT image threshold was considered in this thesis.The grey intensity in CT images increased or decreased owing to the changing of the tissue density which was caused by the nodule pathology.According to this theory,we preposed a technique of thyroid nodule recognition based on the threshold of pixel grey intensity.Median filtering and mean filtering were used to reduce the noise.The thyroid contour was delineated manually,and the digital image processing steps were applied to extract region of interest(ROI).Then,the 2 x 2 gray detection operator was developed.High density and low density threshold were set up through the optimization process.The abnormal region where four pixel grey intensity were more or less than the high or low threshold,considered as the thyroid nodule,was identified using detection operator in the ROI.The experimental results showed that the thyroid nodule could be effectively detected and the accuracy was 85.1%.To improve the performance of recognition technique on thyroid nodule,another recognition method based on the CT image texture analysis was presented in this thesis.The first and second order texture features including statistic feature and special relationship of each pixel were extracted from the ROI.The first order texture features include entropy,uniformity,mean intensity,standard deviation,kurtosis,and skewness.The second order texture feature include 28 features of gray level co-occurrence matrix(GLCM)and gray level gradient co-occurrence matrix(GLGCM).The classifiers of back propagation artificial neural network(BP-ANN)and support vector machine(SVM)etc were applied in thyroid nodule detection.The experimental results showed that this method outperformed the technique based on pixel threshold,the accuracy was improved to 89%.Then,a computer aided diagnostic(CAD)system for differential diagnosis based on CT images is proposed to assist radiologists to identify cancer and benign nodules.The texture features including histogram,GLCM and GLGCM were extracted from the ROI.The original feature set was optimized by feature selection algorithm using T test,Relief,and particle swarm optimization(PSO)method.The classification performances were evaluated using classifiers including SVM,random forest(RF),bagging and linear discriminant analysis(LDA).The strategy of leave one cross validation method was considered in classification.A good performance was achieved(accuracy=89.4%).Finally,a comparative study between the CAD system and radiologists was carried out.The accuracy,sensitivity,specificity and area under receiver operating characteristic curve(AUC)were analyzed.The results showed that CAD system outperformed radiologist.The results showed that the computer aided diagnosis system proposed in this thesis can effectively detect nodules and differentiate benign and malignant nodules,the accuracy reached 89%and 89.4%respectively.The results also indicated that some featues of the first order texture feature and GLCM have important value in the recognition of thyroid nodules,such as entropy,uniformity and mean intensity.The CAD system can assist radiologists to improve the work efficiency and observe subtle signs in CT image.The diagnosis accuracy could be improved by presenting objective and quantitative information of radiomics in this system to differential diagnosis of benign and malignant thyroid nodules.