Research And Implementation Of A Deep Learning-Based Image Segmentation And Classification System For Thyroid Nodules

As the incidence of thyroid nodules continues to rise,the number of patients affected by them increases each year,Analyzing the local geometric structure,intensity changes,and background feature information of nodules in ultrasound images is key to distinguishing between benign and malignant thyroid nodules,and it is also an important part of clinical diagnosis and treatment.Senior physicians can use medical ultrasound examinations to identify nodules through in-depth case studies and rich clinical experience,but there is a certain subjectivity,which makes the task of interpreting ultrasound images challenging.Therefore,in order to alleviate the diagnostic pressure on doctors,the application of objective,reliable,and automated methods to evaluate ultrasound image is of significant importance.With the revolutionary development in the field of deep learning,exploring new intelligent prediction methods is an effective means to improve the efficiency of diagnosis and provides a new way to identify the nature of thyroid nodules.Therefore,this thesis focuses on the application of artificial neural networks to the automatic segmentation and classification of thyroid nodules and develops a software system that implements the corresponding functions.The main functions of this system are supported by the basic theories and techniques in the field of deep learning.Firstly,a segmentation network based on UNet3+ and improved E＿UNet3+ is proposed for the task of thyroid nodule segmentation.While keeping the full-scale skip connection unchanged,structure modules dominated by depth separable convolution and EfficientNet＿B0 are used to modify the encoder and decoder parts of the network,making the network tend to be lightweight.After comparative verification,the improved network model achieved better results in nodule segmentation.Then this thesis selects EfficientNet as the baseline network for thyroid nodule classification task,and introduces the idea of integrated learning to achieve multi-branch feature fusion of nodule global image,texture image,and shape image,and finally obtains the classification results with higher accuracy than single feature and single model.This thesis develops a deep learning based thyroid nodule image segmentation and classification system based on requirements,which is an application supported by Python programming language,PyTorch framework,MySQL database and other related development technologies.The system is based on the general framework design and database design,and includes image processing,information management and other related functions,and is also well integrated with the key core technology of thyroid nodule segmentation and classification research to achieve the prediction function of nodule area segmentation and determination of benign and malignant properties,and the system has been tested to prove that the system meets the expected standard in terms of function and performance and satisfies the user requirements.