CN-122023451-A - Edge feature point extraction and registration method for three-dimensional volume data rigid body registration

CN122023451ACN 122023451 ACN122023451 ACN 122023451ACN-122023451-A

Abstract

The invention relates to the technical field of medical image processing, and discloses an edge feature point extraction and registration method for three-dimensional volume data rigid body registration, which comprises the steps of acquiring first and second three-dimensional volume data to be registered; the method comprises the steps of synchronously outputting an edge characteristic point thermodynamic diagram and a tangential plane attitude parameter diagram by utilizing a full convolution neural network, analyzing sub-voxel positions of characteristic points based on the thermodynamic diagram, constructing tangential plane normal vectors in a physical space by utilizing attitude parameters, establishing a matching relation according to the category of the characteristic points, constructing an objective function based on a point-to-tangential plane distance, and solving rigid transformation parameters by adopting a Gaussian-Newton iterative algorithm based on Litsea parameterization. According to the invention, the geometric constraint mechanism from the point to the tangential plane is introduced, so that the characteristic points are allowed to slide in the target tangential plane, the sensitivity of the registration result to the extraction position error of the characteristic points is effectively reduced, and the accuracy and stability of the rigid registration of the medical image are obviously improved.

Inventors

LIANG YUEQIANG
WANG ZHIWU
ZHAO XUEQIANG

Assignees

张家港赛提菲克医疗器械有限公司

Dates

Publication Date: 20260512
Application Date: 20260203

Claims (11)

1. The edge feature point extraction and registration method for three-dimensional volume data rigid body registration is characterized by comprising the following steps of: Step S1, acquiring first three-dimensional volume data and second three-dimensional volume data to be registered, wherein the first three-dimensional volume data is defined as source data, and the second three-dimensional volume data is defined as target data; S2, respectively inputting the first three-dimensional data and the second three-dimensional data into a pre-trained UNet network, and outputting a corresponding edge characteristic point thermodynamic diagram and a tangential plane attitude parameter diagram; S3, analyzing geometric features and constructing a tangent plane aiming at each edge feature point category, determining edge feature point positions by utilizing the edge feature point thermodynamic diagram aiming at each edge feature point category, and constructing a tangent plane normal vector by utilizing the tangent plane attitude parameter diagram; Step S4, establishing a one-to-one correspondence matching relationship between the edge feature points of the first three-dimensional volume data and the second three-dimensional volume data based on the edge feature point types, wherein the edge feature points in the first three-dimensional volume data and the edge feature points of the same type in the second three-dimensional volume data directly form matching point pairs; and S5, constructing an objective function based on the distance from the point to the tangent plane, and solving rigid body transformation parameters from the first three-dimensional volume data to the second three-dimensional volume data.
2. The edge feature point extraction and registration method for three-dimensional volume data rigid body registration according to claim 1, wherein the UNet network is constructed as a full convolution neural network structure, and the specific process of outputting the corresponding edge feature point thermodynamic diagram and the tangential plane pose parameter diagram in the step S2 includes the following steps: Generating an edge characteristic point thermodynamic diagram through a first output branch, wherein the first output branch is configured with a three-dimensional convolution layer and a Sigmoid nonlinear activation function layer, and each channel of the edge characteristic point thermodynamic diagram corresponds to a class of anatomically fixed-semantic edge characteristic points; And generating a tangential plane attitude parameter map through a second output branch, wherein the second output branch is configured with a three-dimensional convolution layer, a hyperbolic tangent activation function layer and a scaling operation, and the numerical values of three channels of the tangential plane attitude parameter map respectively represent a pitch angle, a yaw angle and a roll angle in an Euler system.
3. The edge feature point extraction and registration method for three-dimensional volume data rigid body registration according to claim 1, wherein the specific process of determining the edge feature point position by using the edge feature point thermodynamic diagram in the step S3 comprises the following steps: performing full-map maximum value search in a corresponding channel of the edge feature point thermodynamic diagram, and selecting a voxel coordinate with the largest response value as an integer-level peak coordinate; And defining a local cube window by taking the integer-level peak coordinate as a center, and performing weighted average operation by taking the numerical value of the prediction thermodynamic diagram in the local cube window as a weight to obtain the sub-voxel-level coordinate.
4. The edge feature point extraction and registration method for three-dimensional volume data rigid body registration according to claim 3, wherein said step S3 further comprises a process of performing physical space coordinate mapping and restoration on said sub-voxel level coordinates, comprising the steps of: Reading an affine transformation matrix in metadata of the first three-dimensional volume data or the second three-dimensional volume data, the affine transformation matrix describing a transformation relationship from an image array coordinate system to a DICOM patient coordinate system; And carrying out coordinate system mapping on the sub-voxel level coordinates by utilizing the affine transformation matrix to obtain physical feature point coordinates defined under the DICOM patient coordinate system.
5. The edge feature point extraction and registration method for three-dimensional volume data rigid body registration according to claim 1, wherein the specific process of constructing a tangent plane normal vector by using the tangent plane posture parameter map in the step S3 includes the following steps: Reading corresponding pitch angle, yaw angle and roll angle at the coordinates corresponding to the edge characteristic point positions in the tangential plane attitude parameter diagram; Constructing a rotation matrix according to a preset rotation sequence of Z-Y-X, wherein the rotation matrix is obtained by multiplying the matrix of rotating the rolling angle around a Z axis, rotating the yaw angle around a Y axis and rotating the pitch angle around an X axis; and selecting a preset reference axis vector, transforming the reference axis vector by using the rotation matrix, and calculating to obtain the tangent plane normal vector.
6. The edge feature point extraction and registration method for three-dimensional volumetric data rigid body registration of claim 5, wherein said process of constructing a tangent plane normal vector using said tangent plane pose parameter map further comprises the step of transforming said tangent plane normal vector to a DICOM patient coordinate system: Reading an affine transformation matrix of the first three-dimensional volume data or the second three-dimensional volume data, extracting a rotation scaling matrix in the affine transformation matrix, and defining a spatial position under a DICOM patient coordinate system; calculating an inverse transformation matrix of the rotation scaling matrix; And transforming the normal vector of the tangential plane by using the inverse transformation matrix and executing normalization operation to obtain the normal vector of the physical tangential plane.
7. The edge feature point extraction and registration method for rigid body registration of three-dimensional volume data according to claim 1, wherein the edge feature point class has definite medical semantic properties and is defined as strictly unique in each three-dimensional volume data, the physical position corresponding to the edge feature point class is located on the surface boundary of the human organ or tissue, and the process of establishing the matching point pair in step S4 does not perform similarity calculation based on image gray scale or local descriptors.
8. The edge feature point extraction and registration method for three-dimensional volume data rigid body registration according to claim 1, wherein the specific process of constructing the objective function based on the point-to-tangent plane distance in step S5 includes: And calculating the distance between the edge characteristic points of the first three-dimensional data in the normal direction of the tangential plane of the edge characteristic points of the second three-dimensional data corresponding to the edge characteristic points of the first three-dimensional data after the edge characteristic points of the first three-dimensional data are transformed by the rigid transformation parameters, and carrying out square summation on the distances to construct an objective function based on the distance from the point to the tangential plane.
9. The edge feature point extraction and registration method for rigid body registration of three-dimensional volume data according to claim 1, wherein the specific process of solving the rigid body transformation parameters of the first three-dimensional volume data to the second three-dimensional volume data in step S5 includes: When the rigid body transformation needs to translate and rotate at the same time, a Gaussian-Newton iterative algorithm or a conjugate gradient method based on the lie algebra parameterization is adopted to parameterize the rotation matrix into a lie algebra vector or a residual vector; in each iteration step, a weighted linear equation system is constructed, a jacobian matrix and a residual vector are calculated, and a linearized normal equation is solved to obtain the optimal disturbance quantity; And updating the rotation matrix and the translation vector according to the optimal disturbance quantity by using an exponential mapping operation until a preset convergence condition is met.
10. The edge feature point extraction and registration method for three-dimensional volume data rigid body registration according to claim 9, wherein the specific process of determining the preset convergence condition in step S5 includes the steps of: calculating the overall numerical amplitude of the optimal disturbance quantity, and comparing the overall numerical amplitude with a preset minimum error threshold; Simultaneously monitoring the cycle times of iterative computation, and comparing the cycle times with a preset maximum iteration times limit value; And when the overall numerical amplitude is smaller than the minimum error threshold, or when the cycle number reaches the maximum iteration number limit, judging that the preset convergence condition is met, and ending iterative computation.
11. The edge feature point extraction and registration method for rigid body registration of three-dimensional volume data according to claim 9, wherein the specific process of solving the rigid body transformation parameters of the first three-dimensional volume data to the second three-dimensional volume data in step S5 includes: When the rigid body transformation only needs translation, the Cholesky method is adopted to directly calculate the translation parameters.

Description

Edge feature point extraction and registration method for three-dimensional volume data rigid body registration Technical Field The invention relates to the technical field of medical image processing, in particular to an edge feature point extraction and registration method for three-dimensional volume data rigid body registration. Background Medical image registration is a fundamental key technology in medical image analysis and clinical auxiliary diagnosis and treatment, and aims to convert two or more images acquired by the same patient at different times, different modalities or different viewing angles into a unified coordinate system. For three-dimensional volume data such as CT (computed tomography) and MRI (magnetic resonance imaging), rigid body registration mainly solves the problem of spatial alignment of anatomical structures with small deformations such as bones and brains, and the core task of the rigid body registration is to solve a rotation matrix and a translation vector which can describe the spatial position difference between a source image and a target image. The accurate registration result can provide multisource fusion spatial information for doctors, and has important clinical significance for making radiotherapy plans, executing operation navigation and monitoring focus development. In the existing three-dimensional rigid body registration technology based on characteristics, a processing flow of characteristic extraction-matching-transformation solving is generally adopted. Specifically, the algorithm first extracts a representative set of anatomical landmark points or surface contour points from the three-dimensional volume data using an edge detection operator or a deep learning model. After the corresponding relation between the source data characteristic points and the target data characteristic points is established, an optimized objective function is established, and the iterative nearest point algorithm or the variant thereof is utilized for solving. The method generally adopts Euclidean distance as a measurement standard, and estimates rigid transformation parameters by minimizing the coordinate distance between matching point pairs, so that the characteristic point set of the source image is overlapped with the characteristic point set of the target image as far as possible in space. However, the point-to-point distance constraint mechanism commonly adopted in the prior art has limitations in practical application, and the main problem is that the accuracy of the feature point extraction position is too sensitive. When medical images affected by partial volume effects or imaging noise are processed, feature points extracted by an algorithm are often difficult to accurately locate to completely consistent anatomical positions, and tiny position deviations are unavoidable between corresponding points of a source image and a target image. In this case, strict point-to-point constraints may force the source feature points to fully coincide with the target feature points in coordinates, and such forced constraints may cause positioning errors in the feature extraction stage to be directly introduced into the solution process of the transformation parameters, so that the finally calculated rigid transformation parameters deviate from the true values, thereby reducing the overall accuracy of registration. Disclosure of Invention Aiming at the defects of the prior art, the invention provides an edge feature point extraction and registration method for three-dimensional volume data rigid body registration, which solves the problems mentioned in the background art. In order to achieve the above purpose, the invention is realized by the following technical scheme: The first aspect of the invention provides an edge feature point extraction and registration method for three-dimensional volume data rigid body registration. The object processed by the method is three-dimensional medical volume data stored in the form of a voxel matrix and having spatially positioned metadata. According to the invention, the position information and the geometric posture information of the anatomic edge feature points are synchronously extracted through the deep learning network, and an optimization model is constructed by utilizing the geometric constraint from the points to the tangential planes, so that the high-precision rigid registration of the source data and the target data under a physical coordinate system is realized. The method comprises the following main processes: First, first three-dimensional volume data to be registered is acquired as source data, and second three-dimensional volume data is acquired as target data. Both need to read their metadata to determine the spatial location under a unified coordinate system. And secondly, performing feature extraction on the input three-dimensional data by using a full convolution neural network structure. The network is configured with a dual-