CN-121982078-A - Spatial registration method, device, equipment and medium without auxiliary marks

CN121982078ACN 121982078 ACN121982078 ACN 121982078ACN-121982078-A

Abstract

The application discloses a spatial registration method, device, equipment and medium without auxiliary marks, and belongs to the technical field of computer vision. The method comprises the steps of inputting three-dimensional image data into a pre-trained segmentation prediction joint model to obtain segmented image data and a first mark point set, extracting a first point cloud set from the segmented image data, selecting a hot point cloud set, obtaining a second mark point set based on the hot point cloud set, performing rough registration operation based on the first mark point set and the second mark point set to obtain a rough registration conversion matrix, obtaining a second point cloud set, determining a third point cloud set, and performing fine registration operation based on the second point cloud set, the third point cloud set and the first mark point set to obtain a fine registration conversion matrix. According to the technical scheme, on the premise that auxiliary markers are not needed, geometric characteristics and clinical importance of different anatomical regions are fully considered in registration optimization, and high-precision and high-robustness automatic spatial registration is realized.

Inventors

JIANG SHIZHONG
YANG WEI
LIN QINYONG
PENG KEHAI

Assignees

广州艾目易科技有限公司

Dates

Publication Date: 20260505
Application Date: 20251229

Claims (10)

1. A method of spatial registration without auxiliary marks, the method comprising: acquiring three-dimensional image data of a part to be operated, and inputting the three-dimensional image data into a pre-trained segmentation prediction joint model to obtain segmentation image data and a first mark point set output by the segmentation prediction joint model, wherein the segmentation prediction joint model performs joint optimization training based on segmentation loss and mark point prediction loss; extracting a first point cloud set from the segmented image data, and selecting a hot point cloud set from the first point cloud set; Acquiring a second mark point set based on the hot spot cloud set, and performing coarse registration operation based on the first mark point set and the second mark point set to obtain a coarse registration conversion matrix; acquiring a second point cloud set, and determining a third point cloud set in the first point cloud set based on the second point cloud set; and performing fine registration operation based on the second point cloud set, the third point cloud set and the first mark point set to obtain a fine registration conversion matrix, and combining the coarse registration conversion matrix and the fine registration conversion matrix to obtain a spatial registration result.
2. The method of co-label free spatial registration according to claim 1, wherein the process of pre-training the segmentation prediction joint model comprises: Acquiring training three-dimensional image data, and real segmentation image data and a real first mark point set corresponding to the training three-dimensional image data; Inputting the training three-dimensional image data into a segmentation prediction joint model to obtain prediction segmentation image data output by the segmentation prediction joint model and a prediction first mark point set; Determining a landmark prediction penalty based on the true first landmark set and the predicted first landmark set; Determining a segmentation loss based on the true segmented image data and the predicted segmented image data; and determining a joint optimization loss based on the marker point prediction loss and the segmentation loss, and optimizing the segmentation prediction joint model based on the joint optimization loss.
3. The spatial registration method without auxiliary marks according to claim 1, wherein the performing a fine registration operation based on the second point cloud set, the third point cloud set, and the first marker point set to obtain a fine registration transformation matrix includes: Determining a fine registration transformation matrix based on the second set of point clouds and the third set of point clouds; Determining an original error value set based on the second point cloud set, the third point cloud set, and the fine registration transformation matrix; Performing weighted summation calculation on each original error value in the original error value set based on the first mark point set to obtain a fine registration error; optimizing the fine registration transformation matrix based on the fine registration error; and repeatedly executing the steps until the fine registration error meets the optimal stop condition.
4. The spatial registration method without auxiliary marks according to claim 3, wherein said performing a weighted summation calculation on each original error value in the original error value set based on the first set of marker points to obtain a fine registration error comprises: calculating the shortest distance between a third point cloud corresponding to each original error value in the original error value set in the third point cloud set and the first mark point set; determining self-adaptive weights corresponding to the original error values in the original error value set based on the shortest distance; and carrying out weighted summation calculation on each original error value in the original error value set based on the self-adaptive weight to obtain a fine registration error.
5. The method of spatial registration without auxiliary marks according to claim 4, wherein said calculating a shortest distance between a corresponding third point cloud in the third point cloud set and the first set of landmark points for each original error value in the set of original error values comprises: Constructing a curved surface model based on the first point cloud set; And calculating the shortest path length of the curved surface between the third point cloud corresponding to each original error value in the third point cloud set and the first mark point set in the original error value set as the shortest distance based on the curved surface model.
6. The method of spatial registration without auxiliary marks according to claim 4, wherein said determining an adaptive weight corresponding to each original error value in the original error value set based on the shortest distance comprises: Under the condition that the shortest distance does not exceed a preset distance threshold, determining a preset maximum self-adaptive weight as the self-adaptive weight corresponding to the original error value corresponding to the shortest distance; And substituting the shortest distance into a preset distance attenuation function under the condition that the shortest distance exceeds a preset distance threshold value to obtain the self-adaptive weight corresponding to the original error value corresponding to the shortest distance.
7. The method of spatial registration without auxiliary marks according to claim 1, wherein selecting a set of hot point clouds from the first set of point clouds comprises: calculating the maximum curvature corresponding to each first point cloud in the first point cloud set; And adding the first point cloud with the maximum curvature exceeding a preset curvature threshold value to a hot point cloud set.
8. A spatial registration apparatus without auxiliary marks, the apparatus comprising: The system comprises a segmentation prediction module, a segmentation prediction model and a prediction model, wherein the segmentation prediction module is used for acquiring three-dimensional image data of a part to be operated and inputting the three-dimensional image data into a pre-trained segmentation prediction joint model to obtain segmentation image data and a first mark point set output by the segmentation prediction joint model; the point cloud acquisition module is used for extracting a first point cloud set from the segmented image data and selecting a hot point cloud set from the first point cloud set; The coarse registration module is used for acquiring a second mark point set based on the hot spot point cloud set, and performing coarse registration operation based on the first mark point set and the second mark point set to obtain a coarse registration conversion matrix; the fine registration module is used for acquiring a second point cloud set and determining a third point cloud set in the first point cloud set based on the second point cloud set; The result determining module is used for performing fine registration operation based on the second point cloud set, the third point cloud set and the first mark point set to obtain a fine registration conversion matrix, and combining the coarse registration conversion matrix and the fine registration conversion matrix to obtain a spatial registration result.
9. An electronic device comprising a processor, a memory and a program or instruction stored on the memory and executable on the processor, which when executed by the processor, implements the steps of the method of spatial registration without auxiliary marks as claimed in any one of claims 1-7.
10. A readable storage medium, characterized in that it has stored thereon a program or instructions which, when executed by a processor, implement the steps of the spatial registration method without auxiliary marks as claimed in any of claims 1-7.

Description

Spatial registration method, device, equipment and medium without auxiliary marks Technical Field The application belongs to the technical field of computer vision, and particularly relates to a spatial registration method, device, equipment and medium without auxiliary marks. Background Spatial registration is a central step in precisely aligning preoperative three-dimensional medical images (e.g., CT, MRI) with the actual anatomical location of the patient during surgery. Conventional registration methods often rely on pre-affixed physical markers (e.g., bone nails, registration balls) on the patient's body surface or bony structures, which not only add to the surgical time and complexity, but can also introduce additional risk of infection and patient discomfort. In recent years, registration techniques without auxiliary markers have become a research hotspot, which match intraoperatively acquired point cloud data by directly processing medical images. However, the existing method still faces challenges that firstly, the robustness of the anatomical features extracted by relying on manual interaction or a general algorithm is insufficient, registration failure is easy to cause when the anatomical structure is complex or the individual difference is large, and secondly, the difference of clinical precision requirements of different anatomical regions is not fully considered in the fine registration stage, so that the local registration precision in the critical region of the operation cannot meet the clinical requirements Disclosure of Invention The embodiment of the application provides a spatial registration method, device, equipment and medium without auxiliary marks, which aim to fully consider the geometric characteristics and clinical importance of different anatomical regions in registration optimization on the premise of not needing auxiliary marks and realize high-precision and high-robustness automatic spatial registration. In a first aspect, an embodiment of the present application provides a spatial registration method without auxiliary marks, where the method includes: acquiring three-dimensional image data of a part to be operated, and inputting the three-dimensional image data into a pre-trained segmentation prediction joint model to obtain segmentation image data and a first mark point set output by the segmentation prediction joint model, wherein the segmentation prediction joint model performs joint optimization training based on segmentation loss and mark point prediction loss; extracting a first point cloud set from the segmented image data, and selecting a hot point cloud set from the first point cloud set; Acquiring a second mark point set based on the hot spot cloud set, and performing coarse registration operation based on the first mark point set and the second mark point set to obtain a coarse registration conversion matrix; acquiring a second point cloud set, and determining a third point cloud set in the first point cloud set based on the second point cloud set; and performing fine registration operation based on the second point cloud set, the third point cloud set and the first mark point set to obtain a fine registration conversion matrix, and combining the coarse registration conversion matrix and the fine registration conversion matrix to obtain a spatial registration result. Optionally, the process of pre-training the segmentation prediction joint model includes: Acquiring training three-dimensional image data, and real segmentation image data and a real first mark point set corresponding to the training three-dimensional image data; Inputting the training three-dimensional image data into a segmentation prediction joint model to obtain prediction segmentation image data output by the segmentation prediction joint model and a prediction first mark point set; Determining a landmark prediction penalty based on the true first landmark set and the predicted first landmark set; Determining a segmentation loss based on the true segmented image data and the predicted segmented image data; and determining a joint optimization loss based on the marker point prediction loss and the segmentation loss, and optimizing the segmentation prediction joint model based on the joint optimization loss. Optionally, the performing a fine registration operation based on the second point cloud set, the third point cloud set, and the first marker point set to obtain a fine registration transformation matrix includes: Determining a fine registration transformation matrix based on the second set of point clouds and the third set of point clouds; Determining an original error value set based on the second point cloud set, the third point cloud set, and the fine registration transformation matrix; Performing weighted summation calculation on each original error value in the original error value set based on the first mark point set to obtain a fine registration error; optimizing the fine registration