Ultrasound Image Segmentation Of Thyroid Nodules Based On Improved U-Net

In recent years,with the development of computer performance,the application of artificial intelligence technology in medical imaging has become a hot spot in medical research.Through deep learning and other methods,the machine can assist clinicians in lesions identification and auxiliary diagnosis with good accuracy,saving labor costs and enabling doctors to focus on more meaningful work.Among them,automatic segmentation of thyroid nodules based on ultrasound images is a both meaningful and challenging task.Due to the advantages of low cost and no radiation,ultrasonography is widely used in the diagnosis of thyroid diseases.However,the imaging quality of ultrasonography is poor,and the segmentation effect of the nodule area in the image is poor,and it is difficult to segment through the classic single image method.Thus,algorithmic solution based on the mining of big data is needed.The research group that this study relies on has cooperation with the clinic and accumulated a number of ultrasound data of thyroid nodules.Therefore,it is expected that the data-based deep learning method will be used to segment the ultrasound thyroid image nodules and apply it in real clinical scenarios.Commonly used deep learning image segmentation networks mainly include U-Net and Deeplab V3+,among which U-Net has achieved good results in the field of medical image processing with its unique jump layer connection.However,in the actual use process,we found that U-Net still has poor segmentation effect on nodule boundary and wrong segmentation shape.Therefore,in view of the above problems,combined with the local convexity of the shape of thyroid nodules in ultrasound images,this paper designs an improved U-Net based on level set loss function and shape prior regularization term to improve the network’s ability to detect nodule boundaries and properly segment shape.A large number of numerical experiments were conducted on the datasets using GPU-assisted computing.The results show that the method proposed in this paper is 5.9% and 7.4% higher than the original U-Net in terms of Dice coefficient and average intersection ratio.The proposed method has certain advantages.Some of these methods have been compiled into patents and papers.The main contents of this paper include:1.Combined with clinical problems,the shortcomings of the original U-Net in segmenting thyroid nodules in ultrasound images were found and analyzed,and an improved U-Net was constructed according to the characteristics of thyroid nodules in ultrasound images.The model was validated based on real clinical data and the results show that the proposed method is effective;2.A loss function based on the level set method is designed.By approximately replacing the level set function with the probability map output by the network,the probability of the constraint probability map within the target boundary is close to 1,while the probability outside the target boundary is close to 0,which can enhance the learning of the model’s ability to segment boundaries.At the same time,the image in the level set function is replaced with the labeled image of the clinician,so that the model can be trained more stably under supervision;3.Aiming at the problem of discontinuous segmentation results caused by poor imaging quality and high noise in ultrasound images,a softmax activation function containing smooth prior information is introduced.By establishing connections between adjacent pixels in the feature map,the position information of the pixels in the image is fully utilized to output smooth and continuous segmentation results.In summary,the research content of this thesis comes from real clinical problems.Based on the deep learning U-Net network framework,an improved U-Net network framework is proposed.Combining the traditional level set method with deep learning technology,adding interpretable feature priors to the network improves the interpretability of deep learning,the proposed method obtained superior experimental results by experimental application carried out on the real data set.The thesis has both theoretical and practical research significance.