Research On Algorithm For Distinguishing Benign And Malignant Thyroid Nodules Based On Ultrasound Images

With the rapid development of medical technology and the continuous improvement of people’s awareness of scientific medical treatment,the screening rate of thyroid nodules continues to increase.Malignant thyroid nodules have adverse effects on the health of patients,and even endanger their lives.Ultrasound imaging is an effective method for diagnosing benign and malignant thyroid nodules,but the interpretation of ultrasound images has certain requirements on the qualifications and experience of doctors,and the diagnosis process is affected by subjective factors.Therefore,how to achieve objective,accurate and efficient ultrasound discrimination of benign and malignant thyroid nodules has become an urgent need for current clinical diagnosis.The method of distinguishing benign and malignant thyroid ultrasound based on deep learning mainly needs to solve two problems.The first is to segment the ultrasound image to extract the thyroid nodule area;then,to establish a classification model for the thyroid nodule area to distinguish between benign and malignant.This paper conducts algorithm research around these two issues,and the specific research content is as follows:(1)In this paper,we improve the feature extraction ability and fusion method of shallow and deep features in the thyroid ultrasound image segmentation model.The specific method is: we use a feature extraction network with better performance as encoder to capture richer semantic information;we use the channel attention module and the spatial attention module to model the dependency between the features,which can enhance the feature expression of fusion features and provide rich contextual information to the network.The experimental results show that the proposed method proposed improves the Dice coefficient by 1.95% and Io U by 2.25%.(2)Aiming at the problem of insufficient thyroid ultrasound data during classification network training,this paper proposes to build a multi-task learning network to improve the performance of the network.The specific method is to use the classification task as the main task and the segmentation task as the auxiliary task.Through joint learning,the feature expressions of different task models are shared to help the classification model learn more robust features during training.On the clinical thyroid ultrasound images collected by the Zhongnan Hospital and Zhongxin Hospital of Wuhan University,we use transfer learning to transfer the knowledge obtained in the multi-task learning of the network to the target data to improve the benign and malignant discrimination performance of the model.The experimental results show that the proposed method can achieve accuracy 88.39%、sensitivity 87.93%、specificity89.47% on the Zhongnan Hospital dataset and accuracy 92.86%、sensitivity 94.12 %、specificity 90.91% on the Zhongxin hospital dataset.