A method for automatic registration of a SERF atomic magnetometer with head nuclear magnetic resonance images

Abstract

The application provides a SERF atomic magnetometer and a head nuclear magnetic resonance image automatic registration method, comprising the following steps: obtaining three-dimensional information of a subject's head based on an optical scanner, and dividing the scanning point cloud into two parts of a helmet and a face; performing coarse registration on the helmet model and the helmet point cloud based on an FPFH descriptor, performing fine registration by using an ICP algorithm, and determining the position and direction of the sensor on the three-dimensional scanning data; performing coarse registration on the face point cloud and the head nuclear magnetic resonance image based on face symmetry information, performing fine registration by using an ICP algorithm, and converting the sensor position and direction information to the nuclear magnetic resonance image coordinate system. The registration algorithm realized by the application has the characteristics of automation, high precision and high efficiency, greatly simplifies the registration process, is convenient for medical staff to use, and promotes the medical application of brain magnetism.

Description

Technical Field

[0001] This invention relates to the field of biomedical engineering, specifically to an automatic registration method for a SERF atomic magnetometer and head MRI images. Background Technology

[0002] Magnetoencephalography (MEG) is a non-contact functional imaging technique that measures neural signals in the scalp to infer information about endogenous brain activity, offering high spatial and temporal resolution. Because MEG signals are extremely weak, superconducting quantum interference devices (SQUIDs) are currently used for measurement. However, SQUIDs require liquid helium cooling, resulting in high maintenance costs and limiting their widespread adoption.

[0003] The emergence of SERF (Spin-Exchange Relaxation-Free) atomic magnetometers in recent years has made low-cost, wearable, and highly sensitive MEGs possible. It can be used at room temperature and, due to its small size, offers flexible configuration, allowing it to be placed close to the scalp for a higher signal-to-noise ratio. Subjects have identified it as a potential second-generation magnetoencephalography (MEG) device that will replace SQUIDs.

[0004] The wearable nature of the SERF atomic magnetometer is a prominent feature, but it also presents challenges for its registration with MRI. Current registration methods are mostly based on manual registration: manually selecting target points for coarse registration, and then manually cropping the corresponding region for fine registration. The entire process is time-consuming, inconvenient for processing large amounts of data, and unfriendly to medical personnel without a mathematical background, thus making it unsuitable for the SERF magnetometer's application in magnetoencephalography (MEG). Summary of the Invention

[0005] Current methods for registering SERF atomic magnetometers with MRI images require excessive manual steps, which is time-consuming and labor-intensive. To overcome the shortcomings of existing technologies, this invention provides an automated registration method for SERF atomic magnetometers and head MRI images, which can quickly and accurately complete the entire registration process.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] An automatic registration method for a SERF atomic magnetometer and head MRI images includes the following steps:

[0008] Step 1: Perform region growing segmentation on the original point cloud obtained from the scan, remove isolated data points, and divide the original point cloud into two parts: the face and the helmet. Also, remove the point cloud of the helmet surface slot according to the different surface curvatures to ensure that the helmet point cloud is basically consistent with the helmet model.

[0009] Step 2: Helmet model and helmet point cloud registration. First, downsample the helmet model and helmet point cloud, calculate the FPFH descriptor respectively, use the SAC-IA algorithm for coarse registration, and then use the ICP algorithm for fine matching to complete the first registration and obtain the transformation matrix T1.

[0010] Step 3: Registration of face point cloud with MRI. First, the face is cropped to make it a symmetrical point cloud. Then, the tip of the nose is located, and the nose region point cloud is extracted based on the tip of the nose. Similarly, the tip of the nose on the MRI is located, and the nose region is extracted. Based on the facial symmetry information, the face point cloud and MRI are coarsely registered. The ICP algorithm is used to perform fine registration between the two, and the transformation matrix T2 is obtained after the second registration. The transformation matrix of the whole process is T = T1 * T2. The MRI is a head MRI image.

[0011] Step 4: Since the helmet is a custom helmet, the position and orientation of the sensor relative to the helmet are known during the helmet design process. Transform the position and orientation of the sensor by T to obtain the position and orientation of the sensor relative to the MRI.

[0012] Further, step 1 includes:

[0013] The subject wearing a helmet was scanned using an optical scanner to obtain scanned images. Since the helmet point cloud and face point cloud in the scanned images are not connected, the region growing method was used to segment out regions with discontinuous curvature but continuous point clouds. Therefore, the region growing segmentation algorithm was used to separate them. After the region growing segmentation was completed, a series of point clusters were obtained. The two point clusters with the largest number were selected as candidate points for the helmet point cloud and the face point cloud. The point cloud with the larger volume was regarded as the helmet point cloud, and the point cloud with the smaller volume was regarded as the face point cloud.

[0014] Further, step 2 includes:

[0015] To reduce algorithm complexity and computation time, the helmet model and the segmented helmet point cloud are first downsampled, and the surface normals and FPFH descriptors of the downsampled point cloud are calculated. Based on the SAC-IA algorithm, a coarse matching of the helmet model and the helmet point cloud is performed. First, the fast point feature histogram features of the point cloud are extracted. Then, the random sampling consensus algorithm is used to match the FPFH descriptors, thus completing the coarse matching of the point cloud and giving the two point clouds a good initial position. The initial value corresponding to this initial position is used as the initial value for fine registration of the point cloud for rapid iteration. The ICP algorithm is then used to further match the coarsely registered point cloud to obtain the transformation matrix T1.

[0016] Furthermore, after transformation by transformation matrix T1, the helmet model coordinate system O1-X1Y1Z1 is transformed into coordinate system O2-X2Y2Z2, where plane O1-Y1Z1 is approximately the symmetry plane of the head, and O1-X2Z2 is basically parallel to the coronal plane of the head.

[0017] Furthermore, in the face point cloud, the farthest point along the O2Y2 axis is considered to be the approximate nose tip point p. nt A plane has its normal vector along the O2Y2 axis, at a distance p. nt The distance between the points is d, and the plane intersects with the face point cloud. The face point cloud is clipped using this plane to remove the point clouds in the shoulder, neck and ear areas, so that the remaining face point cloud is symmetrical.

[0018] Further, firstly, the symmetry plane of the face point cloud is determined. The intersection of this symmetry plane with the face point cloud yields the face contour C. The nose tip point is then searched on the contour C. The nose tip point should meet the following conditions: (1) The nose tip point P1 is located on the face contour C and is close to the centroid of the contour; (2) Among the points that meet condition (1), the nose tip point P1 has the largest value along the O2Y2 axis. Since the MRI pose has been corrected, the point with the largest Y-axis coordinate is considered to be the nose tip point P2.

[0019] Furthermore, for the face point cloud, a coordinate system P1-U1V1W1 is established with the nose tip point P1 as the origin. Principal component analysis is performed on the face contour C to obtain three feature vectors. The corresponding eigenvalues ​​are λ1 < λ2 < λ3. Corresponding to the P1U1 direction, Corresponding to the P1V1 direction, Corresponding to the P1W1 direction; for MRI, establish a coordinate system P2-U2V2W2 with the nasal tip point P2 as the origin. The X-axis direction of MRI is consistent with P2U2, the Y-axis direction is consistent with P2V2, and the Z-axis direction is consistent with P2W2; register the two coordinate systems as the coarse registration result between the face and MRI.

[0020] Furthermore, the nose region is extracted with the tip of the nose as the center and r as the radius. Since coarse registration has been completed and the two nose regions are roughly aligned, the ICP algorithm is used to perform fine registration on the two to obtain the transformation matrix T2.

[0021] Beneficial effects:

[0022] This invention implements an automatic registration algorithm between a SERF atomic magnetometer and an MRI scanner. Since no manual operation is required, it can greatly improve the registration efficiency and is very user-friendly for medical personnel without an engineering background. The average running time is 30 seconds, the sensor position error is 0.25±0.03mm, and the orientation error is 0.27±0.04°. Its accuracy is higher than that of manual registration. Therefore, this algorithm has the outstanding characteristics of high efficiency, high accuracy, and automation. Attached Figure Description

[0023] Figure 1 This is a flowchart of the automatic registration method between the SERF atomic magnetometer and head MRI images in this invention;

[0024] Figure 2 This is a schematic diagram of face point cloud cropping in this invention; wherein, (a) is a schematic diagram of helmet model coordinate system O1-X1Y1Z1, (b) is a schematic diagram of head coordinate system O2-X2Y2Z2, and (c) is a schematic diagram of planar face cropping.

[0025] Figure 3 This is a schematic diagram of obtaining the symmetry plane of the face point cloud in this invention; wherein, (a) is the initial symmetry plane Σ initial Schematic diagram, F about Σ initial After symmetry, F' is obtained. (b) shows that after the ICP algorithm registers F and F', F' is converted into F”. (c) is a schematic diagram of the face symmetry plane Σ.

[0026] Figure 4 This is a schematic diagram of the nose tip point positioning and nose region registration in this invention; wherein, (a) is a schematic diagram of the face outline and nose tip point, (b) is a schematic diagram of coordinate system P1-U1V1W1, and (c) is a schematic diagram of coordinate system P2-U2V2W2.

[0027] Figure 5 This is a schematic diagram of the registration results in this invention. Detailed Implementation

[0028] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention.

[0029] like Figure 1 As shown, the automatic registration method of a SERF atomic magnetometer and head MRI image of the present invention is implemented based on the PCL library in Windows and Linux operating systems. The specific steps are as follows:

[0030] Step 1: Scan the subject wearing the helmet using an optical scanner to obtain scanned images. Since the helmet point cloud and face point cloud in the scanned images are not connected, the region growing method can segment regions with discontinuous curvature but continuous point clouds. Therefore, these regions can be separated using a region growing segmentation algorithm, as specifically implemented below:

[0031] (1) There are unmarked points in the point cloud. Sort the points according to their curvature values, find the point with the smallest curvature value, and add it to the seed point set.

[0032] (2) For each seed point, the algorithm will find all its nearest neighbors:

[0033] a. Calculate the difference in normal angle between each nearest neighbor point and the current seed point. If the difference is less than the set threshold, the nearest neighbor point is given priority consideration and step b is performed.

[0034] b. The nearest neighbor point passes the normal angle difference test. If its curvature is less than the set threshold, the point is added to the seed point set, that is, it belongs to the current connected component.

[0035] (3) Remove points that pass both tests from the original point cloud;

[0036] (4) Set the minimum number of points in the point cluster to min, and the maximum number of points in the point cluster to max;

[0037] (5) Repeat steps (1)-(3). The algorithm will generate all planes with points in min and max and distinguish different planes by marking them with different colors.

[0038] (6) The algorithm stops working when the remaining number of points cannot satisfy min.

[0039] After the region growing and segmentation is completed, a series of point clusters will be obtained. The two point clusters with the largest number will be selected as candidate points for helmet point cloud and face point cloud. The point cloud with the larger volume will be regarded as helmet point cloud and the point cloud with the smaller volume will be regarded as face point cloud.

[0040] Step 2: To reduce algorithm complexity and computation time, the helmet model and the segmented helmet point cloud are first downsampled, and the surface normals and FPFH descriptors of the downsampled point cloud are calculated. Based on the SAC-IA algorithm, a coarse matching of the helmet models is performed. First, the fast point feature histogram features of the point clouds need to be extracted. Then, the random sampling consensus algorithm is used to calculate the relationship between these features, thereby completing the coarse matching of the point clouds and giving the two point clouds a good initial position. The algorithm principle is as follows:

[0041] (1) Randomly select n points from the reference point cloud R. In order to ensure that the selected points have certain differences in FPFH features, the distance between each pair of selected points should be greater than the preset threshold d.

[0042] (2) Find points in the target point cloud T that have similar FPFH features to the points in the reference point cloud R, and randomly select one of these similar points as the corresponding point in the target point cloud T.

[0043] (3) Calculate the rigid body transformation for these corresponding point pairs, and then determine the registration effect by solving the distance error and function after the transformation of the corresponding points. The distance error and function here are often represented by the Huber function, denoted as [equation missing]. in:

[0044]

[0045] Where: m l For a preset value, l i Let ||·|| represent the distance difference between corresponding points in the i-th group after transformation, and ||·|| refer to the 2-norm. The above transformation aims to find the optimal solution after transformation, minimizing the error function. The result of this optimization process is the registration matrix for coarse registration.

[0046] The transformation matrix obtained by the SAC-IA algorithm is not accurate, but this initial value can be used as the initial value for fine registration of point clouds for rapid iteration; the ICP algorithm is used to further match the coarsely registered point clouds to obtain the transformation matrix T1;

[0047] Step 3: Registration of face point cloud with MRI. First, the face is cropped to create a symmetrical point cloud. Then, the tip of the nose is located, and the nose region point cloud is extracted based on this point. Similarly, the tip of the nose on the MRI is located, and the nose region is extracted. Coarse registration is performed between the face point cloud and the MRI based on facial symmetry information. Fine registration is then performed using the ICP algorithm, completing the second registration and obtaining the transformation matrix T2. The transformation matrix for the entire process is T = T1 * T2. Specifically, this includes:

[0048] Step 3.1: Face region extraction. After transformation by transformation matrix T1, the helmet model coordinate system O1-X1Y1Z1 ( Figure 2 (a) is transformed into coordinate system O2-X2Y2Z2( Figure 2 In (b)), plane O1-Y1Z1 is approximately the plane of symmetry of the head, and O1-X2Z2 is basically parallel to the coronal plane of the head. In the face point cloud, the farthest point along the O2Y2 axis is considered to be the approximate nose tip point p. nt A plane has its normal vector along the O2Y2 axis, at a distance p. nt The distance between the points is d, and the plane intersects the face point cloud. The face point cloud is clipped using this plane, removing the point clouds in the shoulder, neck, and ear areas, so that the remaining face point cloud has a symmetrical shape. Figure 2 (c)

[0049] Step 3.2: Determine the plane of symmetry for the face point cloud; the cropped face region is F = {p i |p i =(x i ,y i ,z i ) T ,1≤i≤N},p i Points on the face region F, (x i ,y i ,z i ) is p i The three-dimensional coordinates of the point, where T represents the transpose; first, find the approximate symmetry plane Σ of the face. initial In this invention, a plane with normal vector O2X2 and passing through the centroid F is considered to be Σ. initial O2-Y2Z2 can also be considered as Σ initial ( Figure 3 (a)); F about Σ initial After symmetry, we get F'={p mi |p mi =(x mi ,y mi ,z mi ) T ,1≤i≤N},p mi Point F', (x mi ,y mi ,z mi ) is p mi The three-dimensional coordinates of the point; F and F' are registered using the ICP algorithm, and F' is converted to F” = {p ri |p ri =(x ri ,y ri ,z ri ) T ,1≤i≤N},p ri Point F, (x ri ,y ri ,z ri ) is p ri 3D coordinates Figure 3 (b)); the average F and F” are fitted to obtain a plane, which is the face symmetry plane Σ( Figure 3 (c)

[0050] Step 3.3: Nasal tip point localization; the plane of symmetry intersects with the face point cloud to obtain the face contour C, and the nasal tip point is found on the contour C; the nasal tip point should meet the following conditions: (1) the nasal tip point P1 should be located on the contour C and close to the centroid of the contour; (2) among the points that meet condition (1), point P1 should have the largest value along the O2Y2 axis; since the MRI pose has been corrected, the point with the largest Y-axis coordinate is considered to be the nasal tip point P2, such as Figure 4 As shown in (a).

[0051] Step 3.4: Nose region registration. For the face point cloud, establish a coordinate system P1-U1V1W1 with the nose tip point P1 as the origin. Perform principal component analysis (PCA) on the contour C to obtain three feature vectors. The corresponding eigenvalues ​​are λ1 < λ2 < λ3. Corresponding to the P1U1 direction, Corresponding to the P1V1 direction, Corresponding to the P1W1 direction; for MRI, establish a coordinate system P2-U2V2W2 with the nasal tip point P2 as the origin. Figure 4 In (b), the X-axis of the MRI is aligned with P2U2, the Y-axis with P2V2, and the Z-axis with P2W2. Figure 4 (c) The two coordinate systems are registered to obtain the coarse registration result between the face and the MRI. The nose region is cut off with the tip of the nose as the center and r as the radius. Since the coarse registration has been completed, the two nose regions are roughly aligned. The ICP algorithm is used to perform fine registration between the two to obtain the transformation matrix T2.

[0052] Figure 5 The final registration result is shown. The arrows indicate the direction of the sensor, and the spheres indicate the coordinates of the sensor. Both have been transformed into the MRI coordinate system.

[0053] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

[0054] The above embodiments are provided for the purpose of describing the present invention only, and are not intended to limit the scope of the present invention. The scope of the present invention is defined by the appended claims. Various equivalent substitutions and modifications made without departing from the spirit and principles of the present invention should be covered within the scope of the present invention.

Claims

1. An automatic registration method for a SERF atomic magnetometer and head MRI images, characterized in that, Includes the following steps: Step 1: Scan the subject wearing the helmet using an optical scanner to obtain scanned images; perform region growing segmentation on the original point cloud obtained from the scan, remove isolated data points, and divide the original point cloud into two parts: the face and the helmet. Also, remove the point cloud of the helmet surface slots according to the different surface curvatures to ensure that the helmet point cloud is basically consistent with the helmet model. Step 2: Register the helmet model and helmet point cloud. First, downsample the helmet model and helmet point cloud, calculate the FPFH descriptors for each, and use the SAC-IA algorithm for coarse registration. Then, use the ICP algorithm for fine matching to complete the first registration and obtain the transformation matrix. T 1 ; Step 3: Register the face point cloud with the MRI image. First, the face is cropped to create a symmetrical point cloud. Then, the tip of the nose is located, and the nose region point cloud is extracted based on this point cloud. Similarly, the tip of the nose on the MRI is located, and the nose region is extracted. Based on facial symmetry information, coarse registration is performed between the face point cloud and the MRI image. Then, the ICP algorithm is used for fine registration, completing the second registration to obtain the transformation matrix. T 2 The transformation matrix for the entire process is as follows: T=T 1 T 2 MRI is a head magnetic resonance imaging image; Step 4: Since the helmet is a custom-made helmet, the position and orientation of the sensors relative to the helmet are known during the helmet design process. The position and orientation of the sensors are then transformed. T The position and orientation of the sensor relative to the MRI were obtained.

2. The automatic registration method according to claim 1, characterized in that, Step 1 includes: Since the helmet point cloud and face point cloud in the scanned image are not connected, the region growing method is used to segment out regions with discontinuous curvature but continuous point clouds. Therefore, the region growing segmentation algorithm is used to separate them. After the region growing segmentation is completed, a series of point clusters are obtained. The two point clusters with the largest number are selected as candidate points for the helmet point cloud and face point cloud. The point cloud with the larger volume is regarded as the helmet point cloud, and the point cloud with the smaller volume is regarded as the face point cloud.

3. The automatic registration method according to claim 1, characterized in that, Step 2 includes: To reduce algorithm complexity and computation time, the helmet model and the segmented helmet point cloud are first downsampled, and the surface normals and FPFH descriptors of the downsampled point cloud are calculated. Based on the SAC-IA algorithm, a coarse matching of the helmet model and helmet point cloud is performed. First, the fast point feature histogram features of the point cloud are extracted. Then, the random sampling consensus algorithm is used to match the FPFH descriptors, thus completing the coarse matching of the point cloud and giving the two point clouds a good initial position. The initial value corresponding to this initial position is used as the initial value for fine registration of the point cloud for rapid iteration. The ICP algorithm is then used to further match the coarsely registered point cloud to obtain the transformation matrix. T 1 .

4. The automatic registration method according to claim 3, characterized in that: In step 3, after transformation by transformation matrix T1, the helmet model coordinate system... O 1 -X 1 Y 1 Z 1 Transformed into a coordinate system O 2 -X 2 Y 2 Z 2 , where the plane O 1 -Y 1 Z 1 Approximately the plane of symmetry of the head. O 1 -X 2 Z 2 It is basically parallel to the coronal plane of the head.

5. The automatic registration method according to claim 4, characterized in that: In step 3, in the face point cloud, along O 2 Y 2 The farthest point of the axis is considered to be the approximate nose tip. p nt A certain plane with O 2 Y 2 The axis direction is the normal vector, and the distance is... p nt The distance between the points is d Furthermore, the plane intersects with the face point cloud. The face point cloud is cropped using this plane, removing the point clouds in the shoulder, neck, and ear areas, so that the remaining face point cloud is symmetrical.

6. The automatic registration method according to claim 5, characterized in that: In step 3, the symmetry plane of the face point cloud is first determined, and the intersection of this symmetry plane with the face point cloud yields the face contour. C In the outline C Find the tip of the nose; the tip of the nose should meet the following conditions: (1) Tip of the nose P 1 Located in the outline of the face C It is located above and close to the center of gravity of the outline; (2) Among the points that satisfy condition (1), P 1 Point edge O 2 Y 2 The axis is at its maximum; since the MRI posture has been corrected, Y The point with the largest axial coordinate is considered the MRI nasal tip point. P 2 .

7. The automatic registration method according to claim 6, characterized in that: In step 3, for the face dot cloud, the tip of the nose is used as the dot. P 1 Establish a coordinate system with the origin. P 1 -U 1 V 1 W 1 For facial contours C Perform principal component analysis to obtain three eigenvectors. The corresponding eigenvalue is , Corresponding to P 1 U 1 direction, Corresponding to P 1 V 1 direction, Corresponding to P 1 W 1 Direction; for MRI, with the tip of the nose as the reference point. P 2 Establish a coordinate system with the origin. P 2 -U 2 V 2 W 2 MRI X Axial direction and P 2 U 2 Consistent Y Axial direction and P 2 V 2 Consistent Z Axial direction and P 2 W 2 Consistent; the two coordinate systems are registered as a coarse registration result between the face and the MRI.

8. The automatic registration method according to claim 7, characterized in that: In step 3, the tip of the nose is taken as the center of the ball. r The nose region is cropped with a radius; since coarse registration has been completed and the two nose regions are roughly aligned, the ICP algorithm is used to perform fine registration on the two to obtain the transformation matrix. T 2 .

Images