The Research On Thyroid Nodules Ultrasound Image Segmentation And Feature Extraction Algorithm

Thyroid nodule is a kind of common disease,and timely detection and treatment canavoid it into malignant tumor. With the continuous improvement of ultrasound diagnostictechniques, it has become the preferred examination of thyroid nodular disease.The benignand malignant thyroid nodules can be distinguished on the ultrasound image features.So theresearch on how to get the accurate segmentation of thyroid nodules and extract the effectivefeatures of distinguishing benign and malignant nodules in order to distinguish the benign andmalignant has important theoretical significance and broad prospects clinical application.Clinical examination, doctors mark suspicious lesion area in ultrasound videos to judgeand draw conclusions according to the judgment nature of the nodule and clinical experience.This method has many problems: doctors need to manually mark the lesion area which needlarge workload of manually diagnosis.Also the diagnosis result is subjective because thejudgment on the image is mainly by a doctor’s qualitative evaluation and the characteristicboundaries of benign and malignant nodules are not distinguished.To solve the above problems, this article did exploratory research on the thyroid nodulesegmentation and better extracting the best features that distinguish the thyroid nodulesbenign or malignant,taking thyroid ultrasound images as the research objects.The followingresults were obtained:1．In the aspect of image segmentation:This paper introduced DRLSE model which iscommonly used in ultrasound image segmentation at present and analyze the shortcomings ofthe DRLSE model in a single direction evolution and the boundary leakage. In this article,Themodel had been improved in three areas: Firstly, it constructed the boundary indicatorfunction which combined phase and gradient information and enhanced the ability of themodel to identify the location of the weak edge. Secondly, it defined the adaptive weightingfactor and realized that the curve evolved inward or outward adaptively. So it solved theshortcomings of the DRLSE model in a single direction evolution and the problem thatposition sensitive of the initial curve. Thirdly, it combined the fidelity term of CV model, strengthened the model’s capacity of global segmentation and improved the model’ssegmentation rate in a certain extent.2．In the aspect of feature extraction: It proposed a feature extraction algorithm ofthyroid nodule ultrasound image combining textures, shapes and attenuation characteristicinformation. In the texture feature extraction, it treated the traditional Local Binary Patternalgorithm as a prototype and improved it’s neighborhood and distance coding on the basis ofthe algorithm. It’s more conducive to show the nodules using elliptical neighborhood andfuzzy distance coding,also it solved the LBP algorithm’s problems that it’s singletransformation mapping and sensitive to noise. In addition,it extracted the nodules’ shapefeatures and the attenuation coefficient.All of them together constituted the feature vectors ofthyroid nodules.We applied the above methods to the thyroid nodule image database which is created byourselves. Improved segmentation algorithm is better than traditional model of DRLSEverified by experiments. It can be better segmented nodule region; And on this basis, weextract the features of the nodules and then enter them into the SVM classifier to distinguishbenign and malignant nodules. The feature fusion algorithm described in this paper has highaccuracy and a higher classification performance compared with other feature extractionmethods.