Computer Aided Diagnosis Of Thyroid Nodules Based On Ultrasound Images

Objectives:1.To design and train artificial neural network and support vector machine based computer-aided diagnosis systems in differentiating malignant from benign thyroid nodes with gray-scale ultrasound images2.Evaluate the clinical diagnosis value of the two trained models.Methods:1.We collect 543 patients' thyroid nodules ultrasound images from the First Affiliated Hospital of Nanchang University PACS system,all lesions were confirmed by surgery or biopsy pathology,according to the pathology,all images were divide into benign and malignant groups.2.Use multi-scale geometric analysis based shearlet algorithm for image noise reduction.Than extract the edge of the nodule by using the local gauss distribution fitting energy algorithm.3.According to the second step results,we calculate 2 morphological features: a.the boundary compactness;b.the absolute value of the angle between the x-axis and the major axis of the ellipse that has the same second-moments as the region in degrees.65 texture features which consisted of boundary features,first order statistics,spatial gray level dependence matrices,gray level difference statistics,Laws texture energy measures,fractal dimension texture,Fourier power spectrum and gray level run length matrix were extract.We use max-min algorithm to normalized the minimum and maximum values of each feature to [-1,1].4.We use matlab software to designed and trained two group classification model which based on artificial neural networks and support vector machine,Calculate the sensitivity,specificity,positive predictive value,negative predictive value,youden index and the diagnostic accuracy to evaluate the individual classifier.5.Use the best model of each method we trained from the fourth step to predict 50 cases of thyroid nodules.The sensitivity,specificity,positive predictive value,negative predictive value,youden index and diagnostic accuracy were also evaluated.Results:1.A total of 610 thyroid nodules(benign:403,malignant:207)were used in the training step.the sensitivity,specificity,positive predictive value,negative predictive value,youden index and the diagnostic accuracy of ANN model were 98.55%,99.26%,98.55%,99.26%,97.81% and 99.02%,respectively.And the result of SVM were 100%.2.The malignant rate was 34%(17/50),of which 3 were male and 14 were female,16 were papillary thyroid adenoma and 1 was medullary thyroid carcinoma,with an age of 38.71±13.02.In the benign group,6 were male and 27 were female,with an age of 53.36±12.15,of which 31 were goiter and 2 were thyroid adenoma.3.the sensitivity,specificity,positive predictive value,negative predictive value,youden index and diagnostic accuracy were 88.24%,90.91%,83.33%,93.75%,79.14% and 90%,respectively,based on the ANN model.The prediction results based on SVM model are 76.47%,90.91%,81.25%,88.24%,67.38% and 86% respectively.When two models were combined,the results were 100%,87.88%,80.95%,100%,87.88%,92%.Conclusion:1.The two diagnosis models based on artificial neural networks and support vector machine have some clinical value in classification benign and malignant thyroid nodules.2.The combination of two models can reduce the misdiagnosis rate of malignant nodules,and provide a more objective diagnostic reference for clinical diagnosis.