Automatic Segmentation And 3D Reconstruction Of Thyroid And Nodules Based On Ultrasound Images

Ultrasound imaging is the most commonly used method in the diagnosis of thyroid disorders because it is real-time imaging,inexpensive,non-invasive and non-radioactive.Thyroid nodules can be diagnosed as benign or malignant,and malignant thyroid nodules need to be treated by surgical removal.Since 2D ultrasound imaging techniques present ultrasound transverse and ultrasound longitudinal videos that do not visualize the actual size and location of the thyroid nodule in the thyroid gland on a 3D level,small thyroid nodules are missed during surgery or the surgeon removes too much tissue,causing additional harm to the patient.Also in computer-aided diagnosis(CAD)systems for thyroid nodules,automatic detection and segmentation of thyroid nodules based on neural network is a key step in CAD and can effectively reduce the workload of clinicians in manually annotating data.Therefore,this thesis designs BC-UNet,a thyroid segmentation network based on the Spatio-Temporal Encoder,and MTN-Net,a thyroid nodule detection and segmentation network based on multi-scale feature fusion,and implements 3D reconstruction of the thyroid gland and thyroid nodules using Marching Cubes algorithm based on the thyroid segmentation results and thyroid nodule segmentation results.Finally,this thesis develops a 3D reconstruction system based on ultrasound video of thyroid diagnosis and treatment,which can provide valuable auxiliary information for surgeons’ surgery.The details of the study are as follows.(1)In order to accurately segment the cephalic and caudal cross-sections of the thyroid gland in ultrasound video,this thesis designs the BC-UNet,a spatial-temporal encoder-based thyroid segmentation network.The ultrasound transverse video is made up of a sequence of images of the thyroid gland in transverse sections,which change from small to large and then to small again because the cephalic and caudal sides of the thyroid gland are smaller.At the same time,it is difficult for the neural network to accurately segment the thyroid gland when the thyroid transverse section is small.Therefore,in order to solve this problem,this thesis designs the BC-UNet,which consists of a temporal encoder and a spatial encoder.The spatial encoder is used to segment the thyroid transverse plane in each frame,which consists of the UNet and the ASPP module,where the ASPP module enables the network to acquire multi-scale features of the image,thus enabling the network to identify small thyroid transverse planes.The temporal encoder consists of the BiConvLSTM,which combines contextual information from the video to enable the network to achieve an accurate segmentation of the thyroid gland.(2)In order to accurately detect and segment thyroid nodules in ultrasound images,this thesis designs the MTN-Net,a thyroid nodule detection and segmentation network based on multi-scale feature fusion.Thyroid nodules in ultrasound images vary greatly in size and shape and are very similar in appearance to surrounding tissues or organs,which makes it difficult for the neural network to identify thyroid nodules with extreme sizes and to accurately segment thyroid nodules with internal textures,and thyroid nodules with blurred boundaries.To address these problems,this thesis designs the MTN-Net,a thyroid nodule detection and segmentation network based on multi-scale feature fusion,in which the Trident Block is used to fully capture the multi-scale features of thyroid nodules,thus enabling the network to recognise thyroid nodules of different sizes.A TN-NMS algorithm is also proposed to merge the detection results of multiple branches,which can effectively suppress the false detection of internal thyroid nodules.Finally,a semantic segmentation branch is embedded in the detection network,which enables the segmentation of thyroid nodules based on the detection results.(3)Based on the above research,this thesis implements a 3D reconstruction algorithm for the thyroid gland and thyroid nodules and develops a 3D reconstruction system.Based on the results of thyroid and thyroid nodule segmentation,this thesis adopts the Marching Cubes algorithm to achieve 3D reconstruction of the thyroid and thyroid nodules.In addition,this thesis implements a 3D reconstruction system based on ultrasound video of thyroid treatment.The system has the functions of ultrasound video reading,thyroid segmentation,thyroid nodule segmentation,3D reconstruction,etc.The doctor can obtain more information by rotating the 3D model of thyroid and thyroid nodule,and can save the rendered 3D model locally so that the 3D model can be viewed again later.The thyroid segmentation network BC-UNet proposed in this thesis was trained,validated and tested on 31 transverse thyroid ultrasound videos.Experiments showed that the method in this thesis achieved Dice coefficient of 0.823 and IoU of 0.689 in the thyroid video segmentation task,outperforming other models,which proved the effectiveness of BC-UNet in the thyroid video segmentation task.The proposed MTN-Net thyroid nodule detection and segmentation network was also trained,validated and tested on the publicly available dataset TN3K with 3493 ultrasound images and the publicly available dataset DDTI with 347 ultrasound images,and the experiments showed that MTN-Net achieved an AP of 55.2%in the thyroid nodule detection task and 56.8%in the thyroid segmentation task.Compared with other models MTN-Net was the best in AP,AP50,APM and APL,which proved the effectiveness of MTN-Net in the thyroid nodule detection and segmentation tasks in ultrasound images.