A registration method for total hip arthroplasty navigation system based on point cloud alignment Authors: Zhenling Wang, Qiurui He, Xinwei Yue Published: 2026 ABSTRACT Hip replacement surgery requires high precision in aligning intraoperative anatomical structure localization with preoperative imaging spatial alignment. To address the registration difficulties caused by the limited number and sparse distribution of intraoperative probe acquisition points, our paper proposes a method based on point cloud registration between an intraoperative probe tracked by a passive binocular optical tracking system (NDI) and a CT surface, aimed at improving patient registration accuracy in intraoperative navigation scenarios. First, we define the patient-registration problem and describe the methods for acquiring pre-operative and intra-operative data. Subsequently, the centroids of the NDI optical tracking markers are employed as common features shared by the patient and image spaces, and an initial registration is performed using a three-point method. Finally, an improved ICP algorithm is used to perform accurate registration between the two point clouds, and the registration results are validated using a fixture platform. In the experiments conducted on 3D-printed acetabular and femoral models, the target registration errors (TRE) of the proposed method for registering the acetabular model and the femoral model are (0.55 ± 0.15) mm and (0.67 ± 0.18) mm, respectively. The experimental results demonstrate that the proposed method achieves good registration accuracy and stability under phantom experimental conditions. Using the TRE as the primary evaluation metric to measure the spatial localization error of target points, the method exhibits superior geometric registration performance compared to the traditional iterative closest point (ICP) algorithm. Additionally, it features a concise workflow and ease of implementation, providing a feasible technical solution for sparse point cloud registration in intraoperative navigation scenarios. Introduction The integration of computer technology, medical image processing, and surgical navigation techniques has driven the development of digital orthopedic surgery [1]. Surgical navigation systems can provide surgeons with spatial reference information regarding anatomical structures and instrument positions during surgery, thereby improving the consistency and visualization level of surgical operations. Patient registration is a key technique in surgical navigation systems, aiming to obtain the rigid transformation between the physical space of the patient and the image space where the three-dimensional virtual model resides, thereby achieving consistency between the preoperative imaging data and the intraoperative spatial position of the patient. The accuracy and efficiency of registration largely determine the overall performance of the navigation system. In recent years, surgical navigation technology has been extensively studied in Total Hip Arthroplasty (THA) [2–7]. Related studies have shown that navigation systems can effectively improve the consistency and reproducibility of acetabular cup positioning during surgery. With the advancement of navigation technology, various techniques such as CT-based navigation, imageless navigation, and augmented reality navigation have been gradually applied in THA surgery. For example, Hasegawa et al. developed a CT-based augmented reality navigation system that overlays imaging information onto the patient’s anatomy, providing surgeons with intuitive spatial guidance and thereby improving the accuracy of acetabular cup positioning during surgery [3]. Additionally, some studies have proposed surgical navigation systems based on inertial sensors or imageless navigation to achieve more convenient and cost-effective intraoperative navigation [5], which can also significantly improve surgical positioning accuracy. However, in these navigation systems, achieving high-precision spatial registration between the preoperative CT model and the intraoperative patient anatomy remains one of the core technical challenges. Patient registration is accomplished by identifying at least three identical fiducial points described in both the image space and the patient space [8], after which a registration algorithm computes the transformation. Traditional fiducial-based approaches typically involve implanting medical markers into the patient’s bone or affixing them to the skin surface, achieving a registration accuracy of approximately 0.5–1.5 mm [9, 10]. Although this method offers high accuracy, it may increase operational complexity in practical applications and impose additional burdens on the patient. Lin et al. developed a real-time, automatic patient-registration method based on an optical tracking system, achieving a mean registration error of about 0.7 mm [11]. However, when this method is used, if the markers become displaced or the reconstructed model is defective, the registration error may increase significantly. With the rapid advancement of three-dimensional scanning and reconstruction technologies, point-cloud data acquisition and processing have become increasingly mature. Owing to advantages such as small data size, rapid processing, and retained structural information, point-cloud processing has found broad application in healthcare and other fields [12]. Numerous studies on 3D point-cloud registration have already been conducted worldwide. Koide et al. [13] proposed the Voxelized Generalized Iterative Closest Point (VGICP) algorithm for fast and accurate three-dimensional point-cloud registration. The method extends the Generalized Iterative Closest Point (GICP) algorithm by approximating the point-cloud geometry with a uniformly sized 3D voxel grid, thereby avoiding costly nearest-neighbor searches while preserving accuracy. Le et al. developed a robust, correspondence-free stochastic algorithm for point-cloud registration that solves 3D alignment via graph matching and semidefinite relaxation [14]. Yao et al. [15] proposed an enhanced Iterative Closest Point algorithm that leverages curvature-feature similarity of point clouds to improve the accuracy, robustness, and stability of the classical ICP in unstructured environments. Shi et al. [16] introduced a KD-tree-enhanced ICP algorithm that combines point-cloud filtering with a bioinspired stochastic search algorithm [17] for coarse registration. The method augments the conventional KD-tree-enhanced ICP algorithm with an additional point-cloud filtering stage and an adaptive fireworks coarse-registration step. He Y et al. [18] proposed an ICP registration algorithm based on local point-cloud features. First, leveraging the point cloud’s intrinsic feature information, they design robust and efficient 3D local feature descriptors—density, curvature, and normal angle. Based on these feature descriptors, correspondences between the point clouds are established and an initial registration result is obtained. Finally, the obtained result is supplied as the initial pose to ICP, enabling fine-tuning of the registration. Liu et al. [19] developed an improved registration algorithm based on Simulated Annealing (SA) and Markov Chain Monte Carlo (MCMC) theory to attain the global minimum from any initial condition. In recent years, deep learning methods have also been gradually introduced into surgical navigation registration. Some studies have combined neural networks with point cloud registration algorithms to improve the accuracy and robustness of spatial registration [20–29]. For example, Liu et al. proposed a hybrid framework combining neural networks with point cloud registration for spatial registration in augmented reality neurosurgical navigation, thereby improving registration accuracy [26]. The authors systematically summarized the feature extraction mechanisms of different deep learning models in LDCT denoising [29]. The study points out that traditional methods rely on manually designed filters, whereas deep learning methods can automatically extract hierarchical features through convolutional neural networks (CNNs), enabling modeling from noise distribution to structural information. In [22], the authors proposed a dual-stage deep learning framework, the core of which lies in the feature extraction and fusion of multimodal medical images. This method leverages the complementary information among different MRI modalities (such as T1, T2, and FLAIR) and extracts multi-channel features through deep convolutional networks. The study emphasizes that effective feature extraction not only requires capturing global information but also enhancing the discriminative capacity of local lesion regions. In [20], the authors classified radiological images based on a CNN model, with the core focus on learning discriminative features to distinguish different lung lesions. In this study, the CNN progressively extracts features through multiple convolutional operations, and the model maps features to the class space via fully connected layers, thereby completing the pneumonia type classification task. In summary, deep learning methods offer significant advantages in feature extraction capabilities, but their ”data-driven” nature also brings challenges such as limited interpretability and poor engineering controllability. Particularly in tasks involving spatial geometric relationship modeling, relying solely on deep features often makes it difficult to ensure the stability and consistency of the results. In the context of sparse point cloud registration, traditional methods are based on strict geometric constraints and optimization theory, achieving the solution of registration parameters by explicitly modeling the spatial relationships between point clouds. Their feature extraction process—such as local geometric features, normal vectors, and correspondences—has clear physical meaning, endowing the entire registration process with good interpretability. Moreover, these methods are less dependent on data scale, feature a clear algorithmic structure, and better meet the requirements of engineering systems for stability and reliability. Therefore, traditional point cloud registration methods continue to hold significant value in practical applications. This paper uniformly adopts the TRE as the core evaluation metric, supplemented by the RMSE, STD, and maximum error for comprehensive assessment. Motivation and contributions Although many ICP variants (such as point-to-plane ICP, GICP, SparseICP, and robust ICP) have been proposed to improve registration accuracy and robustness, these methods are primarily designed for dense point cloud registration scenarios. However, in surgical navigation systems, data acquired by intraoperative probes typically contains only a small number of sparse surface points. Direct application of traditional ICP variants may lead to unstable correspondences, thereby reducing registration accuracy. Therefore, this study focuses on the registration problem between sparse intraoperative probe points and CT surface models. Combining the application requirements of surgical navigation systems and the characteristics of ICP algorithms, with the important goals of improving registration accuracy and operability, we propose a method based on intraoperative NDI probe and CT surface point cloud registration to enhance the precision of surgical navigation systems. Therefore, our study focuses on the registration problem between sparse intraoperative probe points and CT surface models. Combining the application requirements of surgical navigation systems and the characteristics of ICP algorithms, with the important goals of improving registration accuracy and operability, we propose a method based on intraoperative NDI probe and CT surface point cloud registration to enhance the precision of surgical navigation systems. Unlike deep learning-based registration methods, the proposed method specifically targets sparse point cloud data acquired by intraoperative probes and improves upon traditional registration methods. Compared with deep learning methods that rely on large amounts of training data and have limited interpretability, traditional methods based on explicit geometric constraints still hold significant research value. It should be noted that the research in this paper is conducted based on 3D-printed phantom experiments and does not involve direct comparison with clinical navigation systems or surgical robots. The main contributions of our study are as follows: To address the problem of insufficient registration accuracy between intraoperative anatomy and preoperative CT models in Total Hip Arthroplasty (THA) navigation systems, we propose a point cloud registration method based on the NDI optical tracking system to achieve high-precision alignment between preoperative images and intraoperative space.Aiming at the characteristics of limited quantity and sparse distribution of points collected by intraoperative probes, we improve the traditional SparseICP algorithm: extending it to point-to-plane registration to improve the convergence speed and accuracy of sparse point clouds, and introducing a coarse registration stage to provide initial poses for fine registration. Through two-stage registration (coarse registration + local sparse point-to-plane ICP), high-precision registration between intraoperative local sparse point clouds and CT surface point clouds is achieved. This method is clearly distinguished from traditional global dense point cloud registration methods such as ICP, GICP, and SparseICP, and the proposed method is more suitable for addressing the problem of intraoperative local sparse point cloud registration, thereby demonstrating its methodological specificity for this particular application scenario.An experimental validation platform based on pelvic and femoral models and an optical tracking system was constructed: using 3D printed models and CT scan data to extract point clouds, the effectiveness of the proposed method in terms of registration accuracy, robustness, and repeatability is validated through multiple sets of experiments. The experimental results show that the spatial positioning errors of the proposed method in femoral and acetabular registration are 0.67 ± 0.18 mm and 0.55 ± 0.15 mm, respectively, which are significantly better than those of the traditional ICP algorithm and its improved variants. To address the problem of insufficient registration accuracy between intraoperative anatomy and preoperative CT models in Total Hip Arthroplasty (THA) navigation systems, we propose a point cloud registration method based on the NDI optical tracking system to achieve high-precision alignment between preoperative images and intraoperative space. To address the problem of insufficient registration accuracy between intraoperative anatomy and preoperative CT models in Total Hip Arthroplasty (THA) navigation systems, we propose a point cloud registration method based on the NDI optical tracking system to achieve high-precision alignment between preoperative images and intraoperative space. Aiming at the characteristics of limited quantity and sparse distribution of points collected by intraoperative probes, we improve the traditional SparseICP algorithm: extending it to point-to-plane registration to improve the convergence speed and accuracy of sparse point clouds, and introducing a coarse registration stage to provide initial poses for fine registration. Through two-stage registration (coarse registration + local sparse point-to-plane ICP), high-precision registration between intraoperative local sparse point clouds and CT surface point clouds is achieved. This method is clearly distinguished from traditional global dense point cloud registration methods such as ICP, GICP, and SparseICP, and the proposed method is more suitable for addressing the problem of intraoperative local sparse point cloud registration, thereby demonstrating its methodological specificity for this particular application scenario. Aiming at the characteristics of limited quantity and sparse distribution of points collected by intraoperative probes, we improve the traditional SparseICP algorithm: extending it to point-to-plane registration to improve the convergence speed and accuracy of sparse point clouds, and introducing a coarse registration stage to provide initial poses for fine registration. Through two-stage registration (coarse registration + local sparse point-to-plane ICP), high-precision registration between intraoperative local sparse point clouds and CT surface point clouds is achieved. This method is clearly distinguished from traditional global dense point cloud registration methods such as ICP, GICP, and SparseICP, and the proposed method is more suitable for addressing the problem of intraoperative local sparse point cloud registration, thereby demonstrating its methodological specificity for this particular application scenario. An experimental validation platform based on pelvic and femoral models and an optical tracking system was constructed: using 3D printed models and CT scan data to extract point clouds, the effectiveness of the proposed method in terms of registration accuracy, robustness, and repeatability is validated through multiple sets of experiments. The experimental results show that the spatial positioning errors of the proposed method in femoral and acetabular registration are 0.67 ± 0.18 mm and 0.55 ± 0.15 mm, respectively, which are significantly better than those of the traditional ICP algorithm and its improved variants. An experimental validation platform based on pelvic and femoral models and an optical tracking system was constructed: using 3D printed models and CT scan data to extract point clouds, the effectiveness of the proposed method in terms of registration accuracy, robustness, and repeatability is validated through multiple sets of experiments. The experimental results show that the spatial positioning errors of the proposed method in femoral and acetabular registration are 0.67 ± 0.18 mm and 0.55 ± 0.15 mm, respectively, which are significantly better than those of the traditional ICP algorithm and its improved variants. Organization The rest of this paper is organized as follows. In Section 2, the problem formulation and data acquisition for patient registration is presented. In Section 3, the registration method is investigated. Furthermore, a case study in registration experiments is presented in detail in Section 4. Finally, Section 5 concludes the paper. Problem formulation and data acquisition for patient registration A surgical navigation system is typically divided into three subsystems: the optical tracking subsystem, the image subsystem, and the patient subsystem. The optical positioning subsystem adopts the Polaris Vega passive optical 3D localizer (NDI, Canada). Its working components include a position sensor and passive localization tools—a probe pointer and rigid-body trackers. Each tracker is fitted with four reflective markers that are coplanar yet non-collinear. The position sensor tracks and outputs either the centroid coordinates of every marker or the pose of the tracker’s overall reference point. In the patient subsystem, the acetabulum serves as the treatment target; a rigid-body tracker is rigidly fixed to the acetabulum to coordinate intra-operative navigation. The image subsystem processes and visualizes the patient’s CT data and 3D model; the coordinate transformations among all subsystems are shown in Figure 1. Surgical navigation system components and coordinate transformations. The transformation matrices among the subsystems’ coordinate frames are derived as follows: Tprobendi and Ttrackerndi represent the real-time spatial relationship between the probe and the tracker in the NDI coordinate frame, acquired live by the NDI system.Tpointprobe denotes the position of the probe tip in the probe’s own coordinate frame; the transformation matrix T can be determined either by three-coordinate calibration or via an NDI calibrator, and it represents a fixed, constant relationship.During surgery, the tracker is rigidly fixed relative to the patient’s bone, maintaining an unchanging positional relationship. To ensure experimental accuracy, the probe-tip position must be expressed in the tracker coordinate frame. The transformation is given as Equation 1:Tpointtracker=(Ttrackerndi)−1∗Tprobendi∗Tpointprobe.(1)The surgical navigation system aims to determine the transformation that maps point Pndi in the tracker coordinate system to point Pct in the CT coordinate system, which expressed as Equation 2:Pct=Tprecise∗Tcoarse∗Pndi,(2)where Tcoarse is the coarse-registration matrix, and Tprecise is the fine-registration matrix.Matrix Tprecise is computed using a point-cloud registration algorithm. In an orthopedic surgical navigation system, the necessary pre-operative and intra-operative point-cloud data are acquired as follows: The NDI tracker is rigidly fixed to the patient’s bone and CT-scanned together with it. Next, the CT data are imported into 3D Slicer to reconstruct a 3D model, thereby defining the image space I. A complete surface point cloud MI is then extracted from the bone model, typically comprising tens to hundreds of thousands of points.In the patient’s physical space P, the probe tip is used to touch the exposed bone surface several times, acquiring a point cloud Vp that typically contains only a few dozen samples. This is due to the limited surgically accessible area imposed by the bone’s anatomical structure. Consequently, patient registration is reduced to aligning a dense target point cloud MI (in image space) with a sparse source point cloud MP (in patient space). These two sets differ dramatically in sample count and exhibit pronounced disparities in geometric characteristics such as contour and shape. Tprobendi and Ttrackerndi represent the real-time spatial relationship between the probe and the tracker in the NDI coordinate frame, acquired live by the NDI system. Tprobendi and Ttrackerndi represent the real-time spatial relationship between the probe and the tracker in the NDI coordinate frame, acquired live by the NDI system. Tpointprobe denotes the position of the probe tip in the probe’s own coordinate frame; the transformation matrix T can be determined either by three-coordinate calibration or via an NDI calibrator, and it represents a fixed, constant relationship. Tpointprobe denotes the position of the probe tip in the probe’s own coordinate frame; the transformation matrix T can be determined either by three-coordinate calibration or via an NDI calibrator, and it represents a fixed, constant relationship. During surgery, the tracker is rigidly fixed relative to the patient’s bone, maintaining an unchanging positional relationship. To ensure experimental accuracy, the probe-tip position must be expressed in the tracker coordinate frame. The transformation is given as Equation 1:Tpointtracker=(Ttrackerndi)−1∗Tprobendi∗Tpointprobe.(1)The surgical navigation system aims to determine the transformation that maps point Pndi in the tracker coordinate system to point Pct in the CT coordinate system, which expressed as Equation 2:Pct=Tprecise∗Tcoarse∗Pndi,(2)where Tcoarse is the coarse-registration matrix, and Tprecise is the fine-registration matrix. During surgery, the tracker is rigidly fixed relative to the patient’s bone, maintaining an unchanging positional relationship. To ensure experimental accuracy, the probe-tip position must be expressed in the tracker coordinate frame. The transformation is given as Equation 1:Tpointtracker=(Ttrackerndi)−1∗Tprobendi∗Tpointprobe.(1)The surgical navigation system aims to determine the transformation that maps point Pndi in the tracker coordinate system to point Pct in the CT coordinate system, which expressed as Equation 2:Pct=Tprecise∗Tcoarse∗Pndi,(2)where Tcoarse is the coarse-registration matrix, and Tprecise is the fine-registration matrix. The NDI tracker is rigidly fixed to the patient’s bone and CT-scanned together with it. Next, the CT data are imported into 3D Slicer to reconstruct a 3D model, thereby defining the image space I. A complete surface point cloud MI is then extracted from the bone model, typically comprising tens to hundreds of thousands of points. The NDI tracker is rigidly fixed to the patient’s bone and CT-scanned together with it. Next, the CT data are imported into 3D Slicer to reconstruct a 3D model, thereby defining the image space I. A complete surface point cloud MI is then extracted from the bone model, typically comprising tens to hundreds of thousands of points. In the patient’s physical space P, the probe tip is used to touch the exposed bone surface several times, acquiring a point cloud Vp that typically contains only a few dozen samples. This is due to the limited surgically accessible area imposed by the bone’s anatomical structure. Consequently, patient registration is reduced to aligning a dense target point cloud MI (in image space) with a sparse source point cloud MP (in patient space). These two sets differ dramatically in sample count and exhibit pronounced disparities in geometric characteristics such as contour and shape. In the patient’s physical space P, the probe tip is used to touch the exposed bone surface several times, acquiring a point cloud Vp that typically contains only a few dozen samples. This is due to the limited surgically accessible area imposed by the bone’s anatomical structure. Consequently, patient registration is reduced to aligning a dense target point cloud MI (in image space) with a sparse source point cloud MP (in patient space). These two sets differ dramatically in sample count and exhibit pronounced disparities in geometric characteristics such as contour and shape. Registration method Initial registration based on the three-point method Pct and Pndi denote the coordinates of the points in the CT and NDI coordinate systems, respectively. R is the rotation matrix between the two coordinate systems. T is the translation vector from the NDI coordinate system to the CT coordinate system. The relationship between Pct and Pndi is given by Equation 3:Pct=R∗Pndi+T.(3)To obtain the initial registration matrices for R and T, three point pairs are manually selected. As shown in Figure 2, points PCTi=[xCTi,yCTi,zCTi]T are marked on the acetabulum in the CT image, while the corresponding points PNDIi=[xNDIi,yNDIi,zNDIi]T are located on the patient’s bone surface at the same anatomical positions using a tracked probe. The initial registration matrices for R and T are computed as follows Equations 4–9:H=∑i=1N(PNDIi−centroidNDI)(PCTi−centroidCT)T,(4)centroidNDI=1N∑i=1NPNDIi,(5)centroidCT=1N∑i=1NPCTi,(6)[U,S,V]=SVD(H),(7)R=VUT,(8)T=−R∗centroidNDI+centroidCT,(9)Where centroidNDI and centroidCT denote the mean centers in the CT and NDI coordinate systems, respectively; H is the covariance matrix, and TNDICT=[R,T] represents the coarse registration matrix. Three-point coarse registration. Accurate registration based on the ICP algorithm Following coarse registration, an initial matrix transforming from the NDI to the CT coordinate system is established. A tracked probe is then used to sample points evenly on the hip joint and femoral head, as shown in Figure 3. Owing to the restricted incision of minimally invasive surgery, only 30 NDI points are acquired for fine registration. The CT point cloud is obtained by surface-rendering the CT volume and extracting its vertex coordinates, as shown in Figure 4. Femur (left) and acetabulum (right). Femoral point cloud (left) and Acetabular point cloud (right). The precise-registration procedure is as follows. The NDI point cloud comprises only 30 points, whereas the CT point cloud contains tens of thousands. For fine registration, the CT point cloud is restricted to the ipsilateral hip joint and femoral head within the target region.During surgery, cartilage covering the femoral head and acetabulum, together with variations in each surgeon’s probing technique, prevent the probe tip from consistently contacting the underlying bone surface.To mitigate the resulting error, fine registration must ensure that the majority of its selected points coincide. Therefore, this paper adopts a point-to-plane based sparse registration method, using the L2 distance and setting the penalty weight coefficient to 0.5. By switching from point-to-point to point-to-plane registration, the classical ICP objective is replaced, the registration problem is transformed into the form represented by Equation 10:arg⁡minR,T=∑i=1NL(niT(Rxi+T,Y))+ISO(k)(R),(10)where |niT(Rxi+T−yi)|p is the loss function, R is a 3×3 rotation matrix, T is a 3×1 translation vector, xi is the i-th point in the source point cloud, yi is the i-th point in the target point cloud, ni is the normal vector corresponding to the target point, Y is the set of point-to-plane residual vectors, ISO(k)(R) is the rotation constraint indicator function, R∈SO(3); Iy(yi) corresponds to the residual constraint, and N is the number of point pairs. However, when intraoperative sampling points are limited and measurement errors exist, squared errors are sensitive to outliers. If erroneous matching occurs, the error becomes extremely large, causing the registration to be biased. Therefore, the Lp sparse-induced norm is introduced, which automatically ignores large errors, and the weights of erroneous matches automatically decrease. Where 0