Research On Localization And Registration Technologies Of Hip Surgery Planning With Combined 2D And 3D Information

Total hip arthroplasty(THA)is an effective treatment for endstage hip lesions,and its surgical volume is increasing year by year.However,intraoperative and postoperative complications such as prosthesis loosening,lower limb inequality,and nerve injury still exist.Optimizing preoperative planning,improving prosthetic fit rates and reducing surgical complications play an increasingly important role in surgical quality.Surgeons typically use traditional C-arm X-ray images for intraoperative monitoring,but with only two-dimensional information,they need to adjust the shot several times based on experience,increasing the amount of X-ray radiation to the surgeon and patient.Although there are new three-dimensional C-arms and O-arms that can acquire threedimensional images,the equipment is expensive,does not reflect real-time conditions,and is difficult to popularize.2D/3D registration,which combines the advantages of these two types of data,has become a hot topic in surgical planning and design.In this thesis,we perform in-depth modeling of preoperative CT(Computed Tomography)and intraoperative X-ray images to investigate the localization and registration technologies of hip surgery planning with combined 2D and 3D information,and our main research work and contributions are as follows:(1)An adaptive hip joint center(HJC)localization method is proposed.The method is based on fast deep stacked network(FDSN)and dynamic registration graph,and the main steps are as follows: First,due to the different image information in single-pose and multi-pose situations,this thesis extracts the comprehensive prominent anatomical landmarks and proposes the FDSN-based HJC prediction method in single-pose situations and the dynamic registration graph-based HJC localization method in multi-pose situations.Second,this thesis introduces grey relational analysis(GRA)to construct an attribute optimization module to guide the network to focus on meaningful anatomical landmarks in cases where some of the prominent anatomical landmarks are damaged or anomalous;Third,to address the problem that the existing HJC prediction methods lack adaptive modeling and their model generalization performance needs to be improved,this thesis adds the multiuniverse optimizer(MVO)module to the framework to implement model initialization,and uses the regularized extreme learning machine(RELM)to train the stacking components and shares random feature maps to speed up the model training.Fourth,to address the problem of localization error in the dynamic registration graphbased HJC localization module,this thesis integrates dynamic fitting and two-step registration methods into the model to further improve the accuracy of HJC localization.The experimental results show that the proposed method outperforms the existing HJC prediction methods.(2)An explainable method for predicting the “safe zone” of the acetabular cup is proposed.The main steps are as follows: First,to address the problem of blurred acetabular prosthesis boundary,this thesis introduces an unsupervised ellipse detection algorithm based on arcsupport line segments to automatically and efficiently obtain the region of interest(ROI)occupied by the acetabular cup components,and proposes a post-processing strategy to extract the final target region contour.Second,to address the problem that the generalization ability of the model needs to be improved,this thesis extracts multi-scale features,introduces m RMR(Minimum Redundancy Maximum Relevance)attribute optimization,and builds a deep stacked network(DSN)prediction model.Third,to address the problem that the existing methods are difficult to explain the model prediction,this thesis adds SHAP(Shapley Additive ex Planations)module to the framework to improve the explainability of models through three main steps: local explanation,global explanation,and feature interactive visualization.Fourth,to tackle the problem that existing methods lack clinical utility evaluation,this thesis considers decision curve analysis(DCA)to assess the risk and benefit of the model.The experimental results demonstrate the effectiveness of the proposed method.(3)A set of 2D/3D registration methods for 2D X-ray and 3D CT images are proposed,including: 2D/3D feature point matching,2D/3D offline registration and 2D/3D intraoperative registration.(a)2D/3D feature point matching: To address the problem that the traditional digital reconstructed radiograph(DRR)method cannot accurately simulate clinical X-ray,this thesis improves a Deep DRR imaging method and constructs multi-directional multi-view DRRs.To address the problem of poor generalization of traditional feature descriptor/feature point methods for heterogenous cross-modal images,this thesis proposes a 2D/3D feature point matching method based on Super Glue+COTR(COrrespondence TRansformer)+multi-view.(b)2D/3D offline registration: To address the problems of high computational complexity and lack of registration guidance for existing 2D/3D offline registration methods,this thesis introduces feature point matching scores under multiple views to guide2D/3D offline registration methods,and this thesis introduces weight boundary values and constructs weight assignment functions.(c)The main steps of 2D/3D intraoperative registration are as follows: First,to address the problem of low real-time prediction of existing 2D annotations,this thesis constructs a model loss function and implements both bone tissue segmentation and landmark detection with a unified deep learning algorithm;Second,to address the situation of redundant small segmentation regions and holes in segmentation results,this thesis introduces the connected-components-3d module and morphological reconstruction methods for connected-components analysis and hole filling,respectively;Third,to address the problem of inefficient registration optimization strategies in 2D/3D intraoperative registration,the predicted landmarks are used to solve the PNP(Perspective-n-Point)problem and to calculate the pelvis initialization pose,and then this thesis incorporates them into the projection regularization factor and chooses to use both landmarks and segmentations for intraoperative registration.The experimental results demonstrate the effectiveness of the proposed method.