Research On Ultrasound Thyroid Nodule Segmentation Based On Segmentation Adversarial Networks

As a common disease of the endocrine system,thyroid nodules have a prevalence rate of32.4%,and the probability of turning into malignant nodules is as high as 5%～10%,posing a serious threat to human health.In recent years,medical imaging technology has developed rapidly,and the clinical detection rate is about 20% to 70%.More and more small nodules are revealed under ultrasonic diagnostic equipment,which means that patients can be detected and treated in time.It is of great significance to improve the cure rate of patients’ diseases.In order to solve the problems of doctors’ manual segmentation of nodule regions,such as heavy workload,long time consuming,subjective ideas,missed diagnosis and misdiagnosis,and different standards,computer-aided diagnosis technology based on ultrasound images is rapidly promoted in clinical medicine.In recent years,artificial intelligence technology has made in-depth research progress in medical imaging and has gradually become a hot spot in the industry.Medical image segmentation is one of them.Therefore,the development of a robust ultrasound thyroid nodule segmentation system is of great clinical significance.This article is based on the segmentation confrontation network to achieve the segmentation of the ultrasound thyroid nodule area.This study deeply analyzed the current state of thyroid nodule segmentation technology,and carried out research with ultrasound images as the main research object.(1)Aiming at the problems of low resolution and low contrast of ultrasound thyroid images,low segmentation accuracy and lack of edge information caused by different shapes and sizes of nodules,an ultrasound thyroid nodule segmentation method based on joint upsampling is proposed.This method inputs the original ROI image,and realizes the segmentation of the nodule area by learning the intrinsic characteristics of healthy and unhealthy thyroid tissues such as grayscale,edge,shape,and texture.First,a multi-layer convolutional neural network is used to extract the features of thyroid nodules.Secondly,the design uses a multi-expansion rate convolution block to accurately locate the target area.Under the premise of the same calculation cost,the context information of the expansion convolution under different sampling coefficients is merged to capture a larger range of dependencies.Finally,jump connections are used to fuse the shallow and deep features of the network.The experimental results show that this research method obtains a pixel accuracy of93.19%,a dice similarity coefficient of 0.8558,and a jaccard distance value of 0.0824,which achieves a good segmentation result.(2)The existing thyroid nodule segmentation network model generally has the problem of inaccurate segmentation of the nodule area,which often over-refines or coarsens the edge information of the nodule.When it is necessary to quantitatively calculate the size of nodules,the phenomenon of over-segmentation or under-segmentation is very disadvantageous in the diagnosis process of doctors.In order to obtain a more robust model,while improving the accuracy of segmentation.Based on the research of ultrasonic thyroid nodule segmentation method based on joint upsampling,this study introduces an adversarial training,and proposes an ultrasonic thyroid nodule segmentation method based on conditional segmentation adversarial network to achieve more thyroid nodule area Accurate segmentation.This research model is composed of two parts: a segmenter network and a discriminator network.The segmenter network design uses a method based on joint upsampling for segmentation,and extracts the depth and shallow feature information of the nodule through network learning,and obtains the second area of the nodule.Value mask;the discriminator network compares the gap between the segmentation result and the gold standard to evaluate the segmentation result.Through multiple adversarial training,the experimental results show that the pixel accuracy of the segmentation network model in this study reaches 95.31%,the dice similarity coefficient reaches 0.8994,and the jaccard distance value is 0.0687.Compared with the segmentation network with joint upsampling,the performance is improved.More accurate segmentation of the ultrasound thyroid nodule area also provides a solid foundation for subsequent research.