Design And Implementation Of An Auxiliary Diagnosis System For Thyroid Nodules Based On The Improved MobileNetV3 Network

Thyroid nodules are one of the most common thyroid disorders.Most nodules are benign and do not require treatment.However,if the nodules become inflamed or show abnormal morphology,the risk of malignancy increases,so it is necessary to seek medical treatment promptly.Ultrasound examination is a commonly used method for thyroid nodule examination,which has the advantages of high safety and no ionizing radiation.Diagnosing thyroid nodules by physicians through manual ultrasound image observation is time-consuming and subject to inevitable subjectivity.Compared to manual observation by physicians,computer-aided diagnosis can provide objective and accurate results,reduce the workload of doctors,and lower the risk of misdiagnosis.Using computer technology to classify medical images can achieve an automated and accurate medical diagnosis.Thyroid nodule ultrasound images contain complex and diverse information,with different organs and tissues interconnecting and interweaving around them,making it difficult to differentiate between different thyroid nodules accurately.Traditional image classification methods perform well in handling simple classification tasks.However,they often need to improve when faced with classification tasks involving a lot of interference and intricate image details.With the continuous development of deep learning technology,deep learning-based image processing methods have been widely used in medical image classification and have achieved good results.The main work of this paper is as follows:1、This paper proposes an improved MobileNetV3 network model to differentiate between different categories of thyroid nodules in ultrasound images.The improvement includes two aspects: firstly,replacing the channel attention module in the original MobileNetV3 network with a convolutional block attention module.The convolutional block attention module comprises channel and spatial attention modules combined according to specific rules.The channel attention mechanism can enhance the importance of different features and improve the network’s perception of important image features.In contrast,the spatial attention mechanism can enhance the importance of different regions and improve the network’s perception of critical positional regions in the image.Secondly,a multiscale feature information extraction and fusion module is introduced.Firstly,different sizes of convolution kernels were used in multiple branches to extract features of different scales from the input image.Then,the features of different scales were fused to improve the model’s classification accuracy.Experimental results show that compared with Res Net34,Res Net50,Res Net101,Res Next50,Res Next101,VGG16,and Efficient Net B0,the improved MobileNetV3 network model in this paper has better classification performance for thyroid nodules,and its AUC index has increased by 5.56 percentage points compared with the original MobileNetV3 network model.2、A thyroid nodule assistant diagnosis system was designed and implemented based on an improved MobileNetV3 network.The system aims to quickly and accurately distinguish between benign and malignant thyroid nodules and improve nodule diagnosis efficiency.Feasibility analysis and functional and non-functional requirement analysis were performed to clarify each functional module’s design goals and details.The system supports three types of users:ordinary patient users,physician users,and administrator users,each with different system functional modules.After logging in,ordinary patient users can upload thyroid nodule ultrasound images that require diagnosis and submit a diagnosis application online.The administrator user reviews the diagnosis application.After approval,the system calls the improved MobileNetV3 network model to predict the benign or malignant category of the thyroid nodule in the ultrasound image.Physician users provide specific diagnostic opinions based on the system’s prediction results,and the system automatically generates corresponding diagnostic reports in PDF format.In addition,the system also provides an image annotation function,allowing physician users to label newly acquired thyroid nodule ultrasound images to expand the experimental dataset for subsequent model improvement and optimization.Finally,standardized testing was conducted on each functional module of the system.The test results show that all functional modules of the system can operate normally and meet the expected design requirements,assisting in diagnosing thyroid nodules and improving nodule diagnosis efficiency.