High-precision preoperative and intraoperative 3D point cloud registration during pedicle screw placement surgery is crucial for improving surgical safety and success rates. However, preoperative and intraoperative point clouds are obtained using different imaging devices and acquisition techniques, which give rise to challenges concerning noise, variations in densities, and initial poses of the two point clouds. In addition, the independent nature of keypoint features within the point cloud after encoding leads to a lack of global contextual correlation. The absence of feature interaction between keypoints of the preoperative and intraoperative point clouds further reduces the relevance of the features, resulting in suboptimal registration accuracy. To address these issues in preoperative and intraoperative point cloud registration for pedicle screw placement navigation systems and then improve the robustness and accuracy of the registration task, a cross-source point cloud registration network with enhanced attention mechanisms is proposed.

A convolutional neural network is presented for preoperative and intraoperative point cloud registration with enhanced attention mechanisms. First, a voxel-filtering algorithm is applied to adjust the density of the intraoperative point cloud based on the density of the preoperative point cloud. Next, farthest point sampling (FPS) is employed to construct local regions of the preoperative point cloud. For local feature extraction, a multilayer perceptron (MLP) is used to build the encoder. Three feature extraction (FE) and feature propagation (FP) layers are employed to encode the point cloud into keypoints and their corresponding high-dimensional feature representations. The feature aggregation module, consisting of graph self-attention and cross-attention mechanisms, is used to enhance the feature representation of the point clouds. In the graph self-attention mechanism, K-nearest neighbors (KNN) is employed to connect each keypoint in the preoperative point cloud to its neighboring points. By calculating the differences between the keypoint features and neighboring point features, the expression of local geometric features is enhanced. The cross-attention mechanism captures the similarity between preoperative and intraoperative point clouds and identifies deep-level correlations to strengthen the global relevance. Then, the features obtained from cross-attention are enhanced using the graph self-attention mechanism to further improve the local contextual relationships. A similarity function is used to compute point-cloud-matching probabilities to obtain a set of corresponding keypoint pairs. Finally, the random sample consensus (RANSAC) algorithm is applied to eliminate incorrectly matched keypoint pairs. The accuracy of the calculated transformation matrix is improved.

To verify the registration performance of the proposed cross-source point-cloud registration network with enhanced attention mechanisms in surgical navigation of pedicle screw placement, the following actions were conducted: algorithm comparisons, ablation experiments, and registration experiments on noise influenced by intraoperative data. The experiments were conducted on preoperative and intraoperative point cloud datasets, which comprise data from the Capital Medical University Affiliated Hospital and SpineWeb dataset, both of which exhibit substantial initial pose variation and angular changes. (Table 1). The proposed model successfully completes precise preoperative and intraoperative point cloud registration (Fig. 5). To evaluate the performance of the algorithm, the FPS+FPFH and FPS+FastReg methods are compared. The results (Table 2) demonstrate that the proposed method achieves the lowest error in coarse registration, with an average rotational error of 2.87° and a translational error of 3.22 mm, meeting clinical accuracy requirements. Additionally, to further analyze the impact of different attention mechanisms on the overall registration performance, ablation studies were designed to quantitatively assess the contributions of each module to the performance of the network. The results (Table 4) indicate that the combined use of graph self-attention and cross-attention mechanisms significantly improves the expression of point-cloud features and registration accuracy. Noise experiments were conducted to validate the robustness of the proposed model. The results (Table 5) show that although noise degrades performance, the proposed method still achieves good coarse registration accuracy under noisy conditions, demonstrating the robustness of the model to noise interference.

To address the challenges of significant initial poses and density differences during point-cloud registration in a pedicle screw placement navigation system, graph self-attention and cross-attention mechanisms are employed to aggregate and enhance features generated by the encoder. Graph self-attention refines the local feature representation, whereas cross-attention strengthens the global correlations between preoperative and intraoperative point clouds. Consequently, the integration of attention mechanisms allows the model to effectively capture the geometric structure of point clouds, improving both registration accuracy and robustness. The experimental results of the preoperative and intraoperative point cloud registration show that the proposed algorithm improves registration accuracy and efficiency in the navigation system for pedicle screw placement, even in cases with large initial pose differences and cross-source data. In addition, the proposed model demonstrates good coarse registration accuracy under noisy intraoperative point clouds, verifying its robustness against noise interference. Compared to the FPS+FPFH method and FPS+FastReg network model, the proposed model achieves better coarse registration accuracy with shorter executions. This algorithm improves the success rate of point-cloud registration and provides technical support for clinical applications.