Study On Computer Aided Diagnosis Of Thyroid Nodules In Ultrasound Image

Thyroid cancer is one of the malignancies which growth fastest, its clinical manifestations are thyroid nodules, B-mode ultrasound technology is the first choice to examine thyroid nodules. The description of ultrasonographic signs and qualitative evaluation plan in Thyroid Imaging Reporting and Data System(TI-RADS) are important standard references for clinician. TI-RADS can help doctors improve the accuracy of clinical diagnosis, but the quantitative analysis methods for TI-RADS are lacked in clinical. This thesis studies quantitative description method for ultrasonographic signs in TI-RADS and nodule boundary segmentation method.The first part of the thesis implements a semi-automatic segmentation of thyroid nodule edges. Firstly, bilateral filtering is chosen to construct multi-scale image of nodules, canny, phase coherence and watershed algorithm are used to segment the boundary of nodules; suitable local cost function is constructed by different weighting way. Secondly, preliminary segmentation is achieved by seeking the shortest path between reference points which are provided by users. Finally, the optimized segmentation is completed through morphology operations. Experimental results show that the proposed segmentation algorithm can achieve better segmentation. The second part studies the feature extraction of thyroid tumor. The shape, echo, posterior acoustic, calcifications, margin feature are extracted through the analysis of TI-RADS and other features of clinical application. These algorithms include calcification and internal echo detection algorithm based MSER, shape feature detection algorithm by the least squares method and polar coordinate conversion, rear echo region and the edge region detection algorithm, instrument orientation detection based on Hough transform and so on.The third part studies the feature extracting of thyroid nodules in color Doppler image, including extracting blood strength characteristics and ranking scheme based on Alder standard, extracting distribution of blood flow and the establishment of regression model. The fourth part implements the benign and malignancy classification of thyroid tumor. Firstly, the shape, calcifications, echo, posterior echo attenuation, edges, aspect ratio and other features of nodules are extracted. Secondly, feature selection algorithm is used to select a better performance of the feature set. Finally, training classifiers and classification the test sets by cross-validation are done.The experiments results show that the proposed TI-RADS features are more suitable for classification of thyroid nodules in ultrasound image compared to the conventional image texture features.