Research On Preoperative Auxiliary Diagnosis Method Of Thyroidectomy Based On Deep Learning

Recently,as the incidence of thyroid nodules increases,thyroid nodules have become one of the most prevalent diseases in the world.Thus,many patients require surgical intervention.Ultrasound examination is an essential step in the diagnosis of thyroid disease.The diagnostic results of ultrasonography determine the following treatment method and the formulation of the surgical plan for the patient.The characteristics of malignant thyroid nodules on the ultrasound image are not obvious,and it is prone to be confused with the surrounding normal tissues.Meanwhile,due to the differences in individual experience of clinicians,the misdiagnosis and missed diagnosis occurs frequently.In addition,patients with malignant thyroid nodules are usually treated with thyroidectomy.However,due to the presence of small and unfixed parathyroid glands(especially inferior parathyroid glands)around the thyroid gland,parathyroid glands are prone to damage or accidental remove during thyroidectomy,resulting in serious medical accidents.Preoperative localization and segmentation of malignant thyroid nodules and parathyroid glands in ultrasound can help doctors identify nodules and protect parathyroid glands during surgery.In summary,the study of computer-aided diagnosis methods before thyroidectomy has important clinical significance.Aiming at the malignant thyroid nodules and parathyroid glands that are not obvious in neck ultrasound images,this thesis studies the preoperative auxiliary diagnosis methods based on deep learning for thyroidectomy,so as to achieve accurate segmentation for malignant thyroid nodules and parathyroid glands before thyroidectomy.The main contributions and results of this thesis are as follows:(1)The construction of the relevant dataset is completed.In this thesis,a malignant thyroid nodule segmentation dataset and a lower parathyroid gland segmentation dataset are constructed.Meanwhile,the annotation work and data preprocessing are completed under the supervision of professional doctors.(2)The convolutional neural network model for malignant thyroid nodule segmentation in the preoperative ultrasound image is studied.The model proposed in this thesis uses a dual-route structure,namely the U-shaped subnet and the inversed U-shaped subnet.The U-shaped subnet is used to extract the context features of the ultrasound image to locate the malignant thyroid nodules in the ultrasound image.The inversed U-shaped subnet is used to capture more detailed features of malignant thyroid nodules to finely segment malignant thyroid nodules in the confused background.All the experimental results demonstrate the superiority of the proposed model on the task of thyroid nodule segmentation in ultrasound images.(3)The neural network model for parathyroid gland segmentation in preoperative ultrasound image is studied.The model proposed in this thesis uses an early-late multi-stage architecture.In the early stage,the Transformer model is used as the backbone,and its powerful semantic feature extraction ability can extract rich semantic information from the confused ultrasound image to segment parathyroid glands.At the same time,the cropping operation is introduced,the coordinate information of the target region is extracted from the segmentation results of the early stage,and the input ultrasound image,corresponding labels and intermediate feature maps are cropped and transferred to the late stage.In the late stage,the FR-Subnet is used as the backbone network in this stage to accurately segment the parathyroid gland from the cropped ultrasound image.In addition,to transfer the inter-pixel structured information extracted by the Transformer backbone in the early stage to the late stage,we also propose a "teacher-mate" training strategy.The proposed method achieves excellent segmentation results on the self-built parathyroid dataset.