Research On Segmentation Algorithm Of Thyroid And Nodules In Ultrasonography Based On Multi-stage U-Net

Thyroid is a butterfly-shaped endocrine organ located in the anterior part of the neck,which plays an important role in regulating metabolism and promoting growth and development.In recent years,as the incidence of thyroid disease has continued to increase,it has gradually become a high incidence disease.Early detection and diagnosis are the key to treatment.Ultrasound imaging has become the preferred method for thyroid disease examination because of its advantages of real-time,cheapness and non-invasiveness.The size of thyroid gland can be used to analyze the secretion of thyroid hormone,which is an important feature in the diagnosis of thyroid abnormalities.In addition,the size and shape of thyroid nodule is an important basis for its benign or malignant diagnosis in clinical practice.Therefore,the accurate segmentation of thyroid gland and thyroid nodule in ultrasonic images can more accurately describe the thyroid lesion area and surrounding tissues,which has important clinical significance for the effective diagnosis of thyroid lesion.However,accurate segmentation of thyroid and thyroid nodules is difficult due to the presence of speckle noise and low contrast,variable shape and size,fuzzy edge of thyroid and thyroid nodules in ultrasound images.To solve the problems above,this thesis proposes a deep convolutional network model based on multi-stage U-Net to realize automatic segmentation of thyroid glands and thyroid nodules.Firstly,limited contrast adaptive histogram equalization is used to improve the image contrast and bilateral filter is used to reduce the noise of ultrasonic image.Then,a deep convolutional network model based on multi-stage U-Net is designed.The model takes U-Net as the basic network framework,and further features are extracted in the decoding path.In this model,multi-stage U-Net with the same depth is formed to overcome the influence of image noise and realize the depth information extraction of image edge.At the same time,a multi-scale residual convolution module is used in the model to enhance the segmentation ability of objects with different scales and further improve the segmentation accuracy.Finally,dice-like loss function integrating binary cross entropy is used to ensure the model convergence speed and robustness of prediction.Multiple comparative experimental results show that the Dice coefficient of the proposed algorithm in thyroid and thyroid nodule ultrasound image segmentation is0.7712,0.8347 respectively,and the Io U is 0.6713 and 0.7450,respectively.Precision is 0.8289 0.8668 and Recall is 0.7817 0.8534,respectively.Compared with other common baseline algorithms,this method can achieve better segmentation results and has certain clinical application value.