Research On Detection And Segmentation Of Neck Tubular Organs Based On Ultrasound Images

Ultrasound examination is one of the most important medical diagnostic methods due to the advantages of real-time,non-invasive and economical.It is widely used in the medical examination of muscles,soft tissues and organs.The results show that the detection and segmentation of tubular organs in thyroid ultrasound images is of great reference value for the localization and diagnosis of thyroid and its nodules.With the rapid development of deep learning and computer vision technology in recent years,the performance and efficiency of organ detection and segmentation in medical ultrasound images have been greatly improved.Medical ultrasound image detection and segmentation tasks often require models with strong feature expression ability.However,existing deep learning models have poor ability to represent tubular organs in thyroid ultrasound images,resulting in unsatisfactory detection and segmentation effects.As a data-driven method,deep learning often relies on a large amount of annotated data,while medical ultrasound image data often has the problem of small amount of data and expensive annotation.In this paper,we studied the detection and segmentation methods of tubular organs in the neck region based on thyroid medical ultrasound images.Aiming at the problems of limited feature expression ability of existing models and less annotated data of medical ultrasound images,we proposed a neck tubular organ detection model based on Faster R-CNN called SSL-DH-Faster R-CNN.To solve the problem of network degradation caused by deep convolutional neural network,an improved neck tubular organ segmentation network model RDPA-U-NET based on U-NET was proposed.After completing the experiments based on the above model,a three-dimensional modeling system of neck tubular organs using medical ultrasound images was developed to assist clinical doctors in the diagnosis of thyroid and its nodules.The main research contents of this paper are as follows:(1)Due to the lack of public data sets of neck tubular organs in thyroid regional medical ultrasound images,and the relatively expensive manual annotation,deep learning model training is prone to overfitting results.Therefore,this paper proposes SSL-DHFaster R-CNN object detection model based on semi-supervised learning method,which makes full use of a large number of unlabeled medical image data.In the semi-supervised learning framework,unlabeled medical image data can be pseudolabeled and re-input to the network to further improve the network performance.(2)The existing target detection models usually use feature pyramid network(FPN)as the Neck part of the detection framework.FPN can fuse the feature maps with strong semantic information in low resolution and the feature maps with weak semantic information but rich spatial information in high resolution without increasing the amount of computation.However,due to the long propagation path of FPN’s shallow feature to the top layer,the shallow feature information is seriously lost.Therefore,based on the Faster R-CNN model,this paper uses PAFPN to enhance the bottom-up path in the Neck part,so as to better preserve the shallow feature information.In addition,in order to solve the balance between localization and classification tasks,the single detection head structure of Faster R-CNN is changed into the dual head structure of convolutional head and fully connected joint,so as to achieve accurate localization of neck tubular organs.(3)Due to the network degradation problem caused by the use of convolutional neural network in the segmentation of tubular organs in U-NET network,in order to optimize the segmentation performance of U-NET network,this paper proposes RDPA-U-NET network by introducing cyclic residual structure and dual path attention module in the segmentation network based on U-Net.Among them,the cyclic residual structure is used to alleviate the problem of gradient disappearance and gradient explosion caused by the degradation of the network,and the dual path attention module integrates spatial features and channel features to teach the network to look ”where” and ”what”.Experimental results show that these two improved methods can effectively improve the segmentation performance of neck tubular organs on ultrasound images.(4)Based on the above the neck of tubular organs detection and segmentation model research,this paper designs and implements the neck tubular organs medical auxiliary system,implements for medical ultrasonic image read,complete the trachea,carotid artery and jugular vein detection and segmentation,and use the image segmentation,generate the three-dimensional model of neck tubular organ,as the doctor provides effective medical assistance.In this paper,deep learning model training and experiments were conducted on the data sets made of 877 medical ultrasound images of cervical tubular organs with correct labels and 3736 medical ultrasound images of thyroid regions without labels.In the detection and segmentation of carotid artery,jugular vein and trachea,the results are better than the conventional target detection and image segmentation methods at present.Taking U-net as baseline,the Dice coefficient of carotid artery segmentation increased from 0.862 to 0.931,the Dice coefficient of jugular vein segmentation increased from 0.875 to 0.912,and the Dice coefficient of tracheal segmentation increased from 0.824 to 0.891.The segmentation model proposed in this paper can be used to complete the three-dimensional modeling of tubular organs in the neck region with good segmentation images,which proves the effectiveness of the research method in this paper.