Research Of Vertebra Instance Segmentation Based On Deep Learning

Medical image segmentation aims at identifying the regions of organ,tissue or le-sion from medical images,and providing the key information such as shape,volume and location of these regions of interest.Medical image segmentation is an important part of clinical disease diagnosis and treatment,which is generally implemented manually by experienced doctors or experts.With the development of medical imaging technology,the scale of medical image is growing rapidly,which brings a lot of complicated work to medical experts.Therefore,researchers have proposed many automatic segmentation methods.Most of the previous medical image segmentation methods are traditional or machine learning based methods,which cannot meet the requirements of segmentation performance and efficiency in the clinical application.Deep learning can learn the most representative features from the original image by a data-driven approach,and get seg-mentation results significantly better than other methods.Accordingly,it has become the first choice for medical image segmentation tasks in recent years.Based on the deep learn-ing,this thesis intends to make an in-depth research on how to improve the performance and applicability of the segmentation algorithm by taking three-dimensional(3D)verte-brae instance segmentation as an analysis object.The main work and contribution of this thesis are as follows:(1)In order to solve the problem of limited application scenarios of most existing vertebrae detection algorithms,this thesis proposes a new vertebrae detection algorithm.Firstly,the likelihood of the vertebrae centroids of each point in the image is defined in terms of space and shape.Then a 2D CentroidNet is designed to predict the likelihood thermogram of the vertebrae centroids.Finally,LDBSCAN,an improved DBSCAN clus-tering algorithm,is developed and applied in the clustering of the likelihood thermogram of the vertebrae centroids to obtain the final vertebrae centroids.(2)Aiming at the problem of low segmentation accuracy caused by the high similarity of adjacent vertebras in the task of vertebrae instance segmentation,a semantic coarse-to-fine strategy is proposed which decomposes the task with complex semantics into two relatively simple ones.On this basis,the CascadeVertSegNet,a cascaded 3D vertebrae segmentation network,is presented.Concretely,the CascadeVertSegNet is composed of two 3D binary segmentation networks,which are CoarseVertSegNet and FineVertSeg-Net,respectively.The CoarseVertSegNet is utilized to identify the vertebrae from the background,and the FineVertSegNet is employed to further segment the current vertebrae from its adjacent vertebrae by using the segmentation result of the CoarseVertSegNet.(3)A new module is designed to improve the segmentation performance of the Cas-cadeVertSegNet.In this thesis,a new module,namely RISCSE,is designed by combing the residual learning and the scSE module,which enhances the attention on the channel and the space,and thus more easily for the network to identify the adjacent vertebrae.Moreover,the segmentation performance of the network is further improved by using the network-level skip connection.(4)The performance and robustness of the method proposed in this thesis are verified on three public datasets.Firstly,the detection rate and the Dice coefficient of the proposed method on MR Lower Spine and xVertSeg datasets both achieve 100% and 95% respec-tively,which is obviously better than most of the current advanced vertebrae instance segmentation methods.Secondly,in view of the lack of analysis on the segmentation per-formance of abnormal vertebrae in the most of the existing methods,the robustness of the proposed method on VerSe2019 dataset which contains a large number of abnormal verte-brae,metal implants,bone tumors and noise is analyzed.The experiment results show that the detection rate and Dice coefficient of the proposed method on the VerSe2019 dataset are 94.6% and 94.3% respectively,which indicates that the method is able to obtain satis-factory segmentation results in complex scenes.