X-ray space calibration method, spatial self-calibration method based on residual network

Abstract

The application belongs to a kind of space position calibration method, to solve the technical problems that the space position calibration method used in current CT three-dimensional reconstruction is limited by precision or complex use condition, provide an X-ray space calibration method, a space self-calibration method based on residual network, without specific calibration object, through the idea of transfer learning, with similar reconstruction object data as training set, based on the principle of rotation invariance, with a pair of projection images with rotation angle difference of 180 degrees as data input, space offset error as output, through modified ResNet50 network, space self-calibration can be realized, the precision and effect of DR three-dimensional reconstruction are improved, and the error is smaller and more efficient.

Description

TECHNICAL FIELD

[0001] The application belongs to a spatial position calibration method, in particular to an X-ray spatial calibration method and a spatial self-calibration method based on a residual network. BACKGROUND

[0002] At present, non-destructive testing of industrial equipment and human diagnosis in clinical radiology have become a hot issue. Among them, the digital projection of the portable DR device is limited to two-dimensional plane. In a given projection, due to perspective distortion and overlapping structure, the observer often has difficulty in obtaining accurate three-dimensional shape and posture. From the early CT reconstruction technology of direct back projection method and filter back projection method to the later artificial intelligence algorithm, the CT reconstruction technology is mainly divided into three directions: analysis method, iterative method and artificial intelligence. The model of reconstruction is gradually developed from two-dimensional to three-dimensional.

[0003] However, no matter which CT three-dimensional reconstruction method is used, the spatial position needs to be calibrated. The commonly used calibration methods include self-calibration and calibration with calibration objects. The self-calibration algorithm is convenient to use, but the precision is poor, so the application is limited. The algorithm based on calibration objects has high precision, but the use condition is complex. SUMMARY

[0004] The application provides an X-ray spatial calibration method and a spatial self-calibration method based on a residual network to solve the technical problems that the spatial position calibration methods used in the current CT three-dimensional reconstruction are limited by precision or complex use conditions.

[0005] To achieve the above purpose, the application adopts the following technical solutions:

[0006] An X-ray spatial calibration method, which is characterized in that it comprises the following steps:

[0007] S1, taking the center axis of the measured object as the rotation axis, rotating the measured object around the rotation axis according to the step angle of the angle β, obtaining a group of mirror projections with the center axis of the measured object as the center after the measured object rotates by the angle β each time, and calculating the similarity coefficients of the group of mirror projections;

[0008] S2, determining the angle error ΔR of the rotation axis on the Z-axis of the fixed coordinate system according to the maximum value of the similarity coefficients of each group of mirror projections under all rotation angles in step S1 z And the position error of the detector in the X-axis component of the fixed coordinate system ΔTd x ; wherein the fixed coordinate system is a three-dimensional rectangular coordinate system with the center point of the detector as the origin and the line connecting the detector and the source position as the Z-axis, and the Y-axis is parallel to the rotation axis of the measured object;

[0009] S3, determining the middle plane on the detector according to the line where the maximum of the projection CT values of the line pixels of each line of all the mirror images corresponds;

[0010] S4, obtaining the component of the position error of the detector in the Y axis of the fixed coordinate system ΔTd y according to the intercept of the middle plane on the detector in the Y axis of the fixed coordinate system;

[0011] S5, calibrating the CT reconstruction space position according to ΔTd x , ΔR z and ΔTd y .

[0012] Further, in step S1, the similarity coefficient is calculated by the following formula:

[0013]

[0014] wherein S represents the similarity coefficient, LA i and LB i respectively represent the projection CT values of each group of mirror i-th line pixels after the measured object is rotated by θ degrees, N num is the number of detector pixels with the maximum length, and respectively represent the average of the projection CT values of each group of mirror N num -th line pixels after the measured object is rotated by θ degrees, and θ is greater than or equal to 0° and less than 180°;

[0015] In step S2, the angle error ΔR z of the rotation axis on the Z axis of the fixed coordinate system is determined according to the maximum of the similarity coefficients of each group of mirror images in all the rotation angles in step S1. z .

[0016] Further, in step S2, the component of the position error of the detector in the X axis of the fixed coordinate system ΔTd x is obtained by the following manner:

[0017] The line where the maximum of the similarity coefficients of each group of mirror images in all the rotation angles in step S1 corresponds is denoted as y / , and the intercept of y / on the X axis of the fixed coordinate system is the X axis component ΔTd x of the position error parameter of the detector.

[0018] Further, step S3 is specifically:

[0019] S3.1, y / For each mirror projection in a set of mirror projections, obtain multiple lines related to y. / Vertical pixel lines;

[0020] S3.2, calculate the root mean square error of each pixel connection line. The pixel connection line corresponding to the minimum root mean square error in the two mirror projections is denoted as x. / x / Located and perpendicular to y / The plane is the position of the upper middle plane of the detector.

[0021] Further, in step S3.2, the root mean square error is calculated using the following formula:

[0022]

[0023] Among them, M num For the detector and y / Total number of vertical pixel connections, P center+j (R y ) and P center-j (R y +π) are respectively y / In a set of mirror projections, y in the two mirror projections / The CT value at the intersection of the line connecting the upper pixel and the j-th pixel.

[0024] This invention also provides a spatial self-calibration method based on residual networks, which is characterized by including the following steps:

[0025] S1, Construct a ResNet50 residual network; The ResNet50 residual network includes an X-ray spatial calibration method for the residual network;

[0026] The X-ray spatial calibration method used for the residual network includes the following steps:

[0027] S1-1, calculate the similarity coefficients of multiple sets of mirror projections;

[0028] S1-2, Based on the maximum similarity coefficient of multiple sets of mirror projections, determine the angular error ΔR of the rotation axis on the Z-axis of the fixed coordinate system. z The position error of the detector in the X-axis component of the fixed coordinate system, ΔTd x The fixed coordinate system is a three-dimensional rectangular coordinate system with the center point of the detector as the origin and the line connecting the detector and the source position as the Z-axis. The Y-axis is parallel to the rotation axis of the object being measured.

[0029] S1-3, Determine the midplane on the detector based on the line pixel corresponding to the maximum value of the projected CT value of each line pixel in all mirror projections;

[0030] S1-4, obtaining a component of the position error of the detector in the Y axis of the fixed coordinate system according to the intercept of the median plane on the Y axis of the fixed coordinate system on the detector y ;

[0031] S2, obtaining a plurality of error-free standard CT data, performing forward projection on the standard CT data according to a set error parameter to obtain a plurality of projection images, and forming a plurality of mirror projection groups with the plurality of standard CT data and the plurality of projection images;

[0032] S3, inputting the plurality of projection images and the corresponding set error parameter into the ResNet50 residual network, and taking the corresponding set error parameter as the control label of each projection image;

[0033] S4, performing an X-ray space calibration method for residual network in the ResNet50 residual network for each projection image to obtain a calibration error parameter, and comparing the calibration error parameter with the control label;

[0034] S5, repeatedly performing steps S2 to S4 to train the ResNet50 residual network until the comparison result of the calibration error parameter and the control label meets a preset requirement, complete the training, and obtain the trained ResNet50 residual network;

[0035] S6, inputting a to-be-calibrated projection image into the trained ResNet50 residual network to obtain ΔR z , ΔTd x , ΔTd y , and performing space calibration according to ΔR z , ΔTd x , ΔTd y .

[0036] Further, in step S1-1, the similarity coefficient is calculated by the following formula:

[0037]

[0038] wherein S represents the similarity coefficient, LA i and LB i respectively represent the projection CT value of each group of mirror image i-th line pixels after the measured object is rotated by θ degrees, N num is the maximum length of the detector pixel number, and respectively represent the average value of the projection CT value of each group of mirror image N num line pixels after the measured object is rotated by θ degrees, and θ is greater than or equal to 0° and less than 180°;

[0039] In step S1-2, the angle error ΔR of the rotation axis on the Z-axis of the fixed coordinate system is determined according to the maximum value of the similarity coefficients of the plurality of mirror projections z Specifically, the rotation angle corresponding to the maximum value of the similarity coefficients of the mirror projections is taken as ΔR z ;

[0040] In step S1-2, the position error of the detector on the X-axis of the fixed coordinate system is ΔTd x Specifically, ΔTd is obtained by the following method:

[0041] The line on which the line pixel corresponding to the maximum value of the mirror projection similarity coefficients is located is denoted as y / , and the intercept on the X-axis of the fixed coordinate system is ΔTd / , that is, the position error parameter of the detector on the X-axis. x .

[0042] Further, step S1-3 is specifically:

[0043] S1-3.1, a plurality of pixel connecting lines perpendicular to y / are obtained on each mirror projection in the corresponding group of mirror projections. /

[0044] S1-3.2, the root mean square error of each pixel connecting line is calculated, and the pixel connecting line corresponding to the minimum value of the root mean square errors in the two mirror projections is denoted as x / , and the plane in which x / is located and perpendicular to y / is the position of the middle plane of the detector.

[0045] Further, in step S1-3.2, the root mean square error is calculated by the following formula:

[0046]

[0047] Wherein, M num is the total number of pixel connecting lines perpendicular to y / , P center+j (R y ) are the CT values at the intersection of y / and the jth pixel connecting line in the two mirror projections corresponding to the group of mirror projections. /

[0048] Further, in step S2, the standard CT data is derived from the official CT data provided by TCIA;

[0049] In step S2, the forward projection of the standard CT data is specifically performed by the open-source X-ray simulation code RTK.​​

[0050] Compared with the prior art, the present application has the following beneficial effects:

[0051] 1. The present application provides an X-ray space calibration method, based on the invariance of the rotation axis, by calculating the similarity coefficient of each group of mirror projection, the angle error of the rotation axis on the Z axis of the fixed coordinate system and the position error of the detector on the X axis of the fixed coordinate system are determined, and then combined with the position of the midplane on the detector, the position error of the detector on the Y axis of the fixed coordinate system can be determined, and then the CT reconstruction space position is calibrated, the calibration method is accurate and simple.

[0052] 2. The present application also provides a space self-calibration method based on residual network, compared with the calibration method based on the invariance of the rotation axis, no specific calibration object is needed, through the idea of transfer learning, taking the data of similar reconstruction objects as the training set, based on the principle of rotational invariance, taking a pair of projection images with a rotation angle difference of 180° as data input and spatial offset error as output, through the modified ResNet50 network, the space self-calibration can be realized, the accuracy and effect of DR three-dimensional reconstruction are improved, the error is smaller and more efficient.

[0053] 3. In the present application, the standard CT data is derived from the official CT data provided by TCIA, and the standard CT data is forward projected by the open source X-ray simulation code RTK, so that the training of the residual network is more accurate. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 is a three-dimensional schematic diagram of the DR shooting space. DETAILED DESCRIPTION

[0055] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations.

[0056] The present application provides an X-ray space calibration method, the shooting space of DR (direct digital X-ray photography) is based on a three-dimensional coordinate system, as shown in Figure 1 is a three-dimensional schematic diagram of the DR shooting space, including source position (X source Ts), model (measured object) and detector Td, taking the central axis of the model itself as the rotation axis, when shooting by the DR equipment, the model rotates around the rotation axis. The description of the geometric shape is based on the international standard IEC 61217, which is designed for cone beam imaging instruments on isocentric radiotherapy systems, and the rotation axis is perpendicular to the detector. Figure 1the Y axis in the model is parallel. Each projection uses 9 parameters to define the source position and the detector relative position in the fixed coordinate system, which takes the center point of the detector as the origin, takes the line connecting the detector and the source position as the Z axis, and takes the direction in which the detector points to the source position as the positive direction of the Z axis, takes the center axis of the model itself as the Y axis, and takes the X axis, the Y axis and the Z axis to form a three-dimensional orthogonal coordinate system. The 9 parameters respectively represent certain meanings, which are the source position Ts = {Ts x Ts y Ts z} T , the detector position Td = {Td x Td y Td z} T , and the detector rotation angle Rd = {R x R y R z} T , wherein Ts x , Ts y , Ts z respectively represent the X axis coordinate, the Y axis coordinate and the Z axis coordinate of the source position in the three-dimensional coordinate system, Td x , Td y , Td z respectively represent the X axis coordinate, the Y axis coordinate and the Z axis coordinate of the detector position in the three-dimensional coordinate system, and R x , R y , R z respectively represent the rotation component of the detector rotation angle relative to the X axis, the rotation component relative to the Y axis and the rotation component relative to the Z axis. Assuming that the source position is fixed relative to the detector, the geometric parameters become the position of the detector with 6 degrees of freedom and the rotation angle of the detector, and the corresponding error parameters are: ΔTd = {ΔTd x ΔTd y ΔTd z} T , ΔRd = {ΔR x ΔR y ΔR z} T , wherein ΔTd x , ΔTd y , ΔTd z respectively represent the X axis component, the Y axis component and the Z axis component of the detector position error parameter, and ΔR x , ΔR y , ΔR z respectively represent the X axis component, the Y axis component and the Z axis component of the detector rotation angle error parameter. The error parameters here specifically refer to the error between the actual situation and the projected result.

[0057] ΔTd z The magnification scale mainly affects the reconstructed model, but has little effect on the geometric artifacts of the model, so ΔR is generally considered z ΔTd x and ΔTd y After the CT reconstruction model, the size of the reconstructed model is compared with the actual size, and the actual magnification coefficient Mag real is:

[0058]

[0059] ΔTd z =(Mag real -1)(Ts z +Td z ).

[0060] Where Size reconstructed calibration represents the size of the simple calibration model, and Size actual calibration represents the actual size.

[0061] In the DR scanning process, the rotation axis keeps rotating motion without any translation motion, so the projection of the rotation axis on the detector is always on the same straight line, and the rotation axis has invariance. Based on this characteristic of the rotation axis, the calibration algorithm can calculate ΔRz through multiple sets of mirror projections, each of which is two projection images taken on the rotation axis with an interval of 180°. In the CT scanning process, the measured object is on a circular trajectory centered on the rotation axis, and the projection at each angle can be collected by the detector. In particular, when the projection interval on the rotation axis is 180°, they show the mirror characteristics in the projection image due to their symmetry through the rotation axis. In each set of mirror projection images, the CT value on the rotation axis is the same, while the others are different. Therefore, the rotation axis can be determined by the similarity of the mirror projection, which is represented by a similarity coefficient S, and the line corresponding to the maximum similarity coefficient is the rotation axis. The similarity coefficient S is defined as follows:

[0062]

[0063] Where LA i and LB i represent the projection CT values of the i-th line pixel in each set of mirror images after the measured object is rotated by θ degrees, N num is the number of detector pixels with the maximum length. and represent the average projection CT values of N num line pixels in each set of mirror images after the measured object is rotated by θ degrees. θ is greater than or equal to 0° and less than 180°.

[0064] After the object under test rotates by a certain angle β, which is generally equal to 1°, each set of mirror images centered on the axis of rotation is N. num For each pixel in a line, calculate the similarity coefficient, i.e., calculate N for each angle θ. num The similarity coefficients of the line pixels, based on the maximum value of all similarity coefficients, can determine the projection of the rotation axis onto the detector plane. The θ corresponding to the maximum value of the similarity coefficients of all line pixels at all angles is taken as ΔR. z .

[0065] Let y be the line containing the pixel with the maximum similarity coefficient among all line pixels at all angles. / y / The intercept on the X-axis of the fixed coordinate system is the X-axis component ΔTd of the detector's position error parameter. x .

[0066] In y / For each mirror projection in a set of mirror projections, obtain multiple lines related to y. / For vertical pixel connections, calculate the root mean square error (RMSE) of each connection. The pixel connection corresponding to the minimum RMS error in both mirror projections is denoted as x. / x / Located and perpendicular to y / The plane is the position of the upper middle plane of the detector, and ΔTd can be calculated based on the intersection of this middle plane and the Y-axis. y Specifically, the intercept of the mid-plane on the Y-axis is the Y-axis component of the detector's position error parameter ΔTd. y .

[0067] The root mean square error (RMSE) of each pixel connection is calculated using the following formula:

[0068]

[0069] Among them, M num For the detector and y / Total number of vertical pixel connections, P center+j (R y ) are respectively y / In a set of mirror projections, y in the two mirror projections / The CT value at the intersection of the line connecting the upper pixel and the j-th pixel.

[0070] In x / The intercept on the Y-axis of the fixed coordinate system is the Y-axis component ΔTd of the detector's position error parameter. y .

[0071] In this invention, for practical application considerations, the spatial offset error considers three variables: ΔR z ΔTdx , ATd y . Wherein, AR z The range of AR x , ATd y The range of ATd

[0072] In addition, the open source deep learning framework Pytorch can be used to build a ResNet50 residual network, and the above method is used for calibration based on the residual network. A pair of projection images (2*224*224) with a difference of 180° in rotation angle are input into the trained ResNet50 residual network, and the output of the ResNet50 residual network is the spatial offset error, that is, ATd x , AR z and ATd y .

[0073] When calibrating with a residual network, the official CT data provided by the cancer imaging archive (TCIA) is used as the data set for network training. The real CT data is simulated by the open source X-ray simulation code Reconstruction Toolkit (RTK), the error parameters are set, and the forward projection is performed. The simulated projection data is calculated by the foregoing calibration method, the error parameters are compared with the set point, and finally the difference between the reconstructed model and the real model is compared. The specific steps are as follows:

[0074] 1. Construct a ResNet50 residual network;

[0075] 2. Based on the official CT data provided by the cancer imaging archive (TCIA), prepare CT data of similar shooting objects, and the prepared CT data is error-free standard CT data;

[0076] 3. Set different error parameters, use RTK to perform forward projection on the CT model corresponding to the CT data, obtain the projection image after forward projection as the data, and the training data with error parameters as labels; wherein, the projection image after projection is obtained according to the set error parameters, and the projection with error;

[0077] 4. The ResNet50 residual network constructed in step 1 can calculate AR z , ATd x , ATd y, the error parameter as a label in the training data is a training control in the ResNet50 residual network, the input data is calculated in the ResNet50 residual network through the calibration method of the application, and the obtained ΔR z , ΔTd x , ΔTd y respectively compared with the error parameter as a label, until the cross entropy meets the minimum threshold value, the test requirement is met, the training of the ResNet50 residual network is completed, and the trained ResNet50 residual network is obtained; wherein the minimum threshold value is a set value, which can be set according to the calibration requirement;

[0078] 5. Input the projection diagram of the measured object into the trained ResNet50 residual network, so that the deviation ΔR z , ΔTd x , ΔTd y of the spatial position of the measured object can be obtained, and the spatial calibration is realized;

[0079] 6. According to the spatial calibration structure, three-dimensional reconstruction work can be carried out.

[0080] The above is only a preferred embodiment of the application and is not used to limit the application. For those skilled in the art, the application can have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the application shall be included in the protection scope of the application.

Claims

1. An X-ray spatial calibration method, characterized in that, Includes the following steps: S1. Using the central axis of the object under test as the rotation axis, the object under test is rotated around the rotation axis in increments of β angles. After each rotation of the object under test by β angles, a set of mirror projections centered on the central axis of the object under test is obtained, and the similarity coefficient of the set of mirror projections is calculated. S2, Based on the maximum similarity coefficient of each group of mirror projections under all rotation angles in step S1, determine the angular error ΔR of the rotation axis on the Z-axis of the fixed coordinate system. z The position error of the detector in the X-axis component of the fixed coordinate system, ΔTd x The fixed coordinate system is a three-dimensional rectangular coordinate system with the center point of the detector as the origin and the line connecting the detector and the source position as the Z-axis. The Y-axis is parallel to the rotation axis of the object being measured. The position error of the detector in the fixed coordinate system is expressed as ΔTd along the X-axis. x Specifically, it is obtained through the following methods: In step S1, the line containing the line pixel corresponding to the maximum value of the similarity coefficient of each group of mirror projections at all rotation angles is denoted as y. / y / The intercept on the X-axis of the fixed coordinate system is the X-axis component ΔTd of the detector's position error parameter. x ; S3, based on the line containing the line pixel with the maximum projected CT value among all line pixels in the mirror projections, determine the midplane on the detector, specifically: S3.1, in y / For each mirror projection in a set of mirror projections, obtain multiple lines related to y. / Vertical pixel lines; S3.2, calculate the root mean square error of each pixel connection line. The pixel connection line corresponding to the minimum root mean square error in the two mirror projections is denoted as x. / x / Located and perpendicular to y / The plane is the position of the upper and middle plane of the detector; the root mean square error is calculated by the following formula: Among them, M num For the detector and y / Total number of vertical pixel connections, P center+j (R y ) and P center-j (R y +π) are respectively y / In a set of mirror projections, y in the two mirror projections / The CT value at the intersection of the line connecting the upper pixel and the j-th pixel; S4. Based on the intercept of the mid-plane on the detector on the Y-axis of the fixed coordinate system, obtain the component ΔTd of the detector's position error on the Y-axis of the fixed coordinate system. y ; S5, according to ΔTd x ΔR z and ΔTd y The spatial location of the CT reconstruction is calibrated.

2. The X-ray spatial calibration method according to claim 1, characterized in that: In step S1, the similarity coefficient is calculated using the following formula: Where S represents the similarity coefficient, LA i and LB i N represents the projected CT value of the i-th line pixel in each mirror image after the measured object is rotated by an angle θ. num The maximum number of pixels for the detector. and Each set of mirror images (N) represents the number of mirror images after the object under test has been rotated by θ degrees. num The average projected CT value of the line pixels, θ is greater than or equal to 0° and less than 180°; In step S2, the angular error ΔR of the rotation axis on the Z-axis of the fixed coordinate system is determined based on the maximum similarity coefficient of each group of mirror projections under all rotation angles in step S1. z Specifically, the rotation angle corresponding to the maximum value of the similarity coefficient of each group of mirror projections under all rotation angles in step S1 is taken as ΔR. z .

3. A spatial self-calibration method based on residual networks, characterized in that, Includes the following steps: S1, Construct a ResNet50 residual network; The ResNet50 residual network includes an X-ray spatial calibration method for the residual network; The X-ray spatial calibration method used for the residual network includes the following steps: S1-1, calculate the similarity coefficients of multiple sets of mirror projections; S1-2, Based on the maximum similarity coefficient of multiple sets of mirror projections, determine the angular error ΔR of the rotation axis on the Z-axis of the fixed coordinate system. z The position error of the detector in the X-axis component of the fixed coordinate system, ΔTd x The fixed coordinate system is a three-dimensional rectangular coordinate system with the center point of the detector as the origin and the line connecting the detector and the source position as the Z-axis. The Y-axis is parallel to the rotation axis of the object being measured. The position error of the detector in the fixed coordinate system is expressed as ΔTd along the X-axis. x Specifically, it is obtained through the following methods: The line containing the pixel corresponding to the maximum value of the mirror projection similarity coefficient is denoted as y. / y / The intercept on the X-axis of the fixed coordinate system is the X-axis component ΔTd of the detector's position error parameter. x ; S1-3, Based on the line containing the pixel with the maximum projected CT value among all line pixels in the mirror projections, determine the midplane on the detector, specifically: S1-3.1, in y / For each mirror projection in a set of mirror projections, obtain multiple lines related to y. / Vertical pixel lines; S1-3.2, calculate the root mean square error of each pixel connection line. The pixel connection line corresponding to the minimum root mean square error in the two mirror projections is denoted as x. / x / Located and perpendicular to y / The plane is the position of the upper and middle plane of the detector; the root mean square error is calculated by the following formula: Among them, M num For the detector and y / Total number of vertical pixel connections, P center+j (R y ) are respectively y / In a set of mirror projections, y in the two mirror projections / The CT value at the intersection of the line connecting the upper pixel and the j-th pixel; S1-4, based on the intercept of the mid-plane on the detector on the Y-axis of the fixed coordinate system, obtain the component ΔTd of the detector's position error on the Y-axis of the fixed coordinate system. y ; S2, acquire multiple error-free standard CT data, and perform forward projection on the standard CT data according to the set error parameters to obtain multiple projection images, so that the multiple standard CT data and multiple projection images form multiple sets of mirror projections. S3, input multiple projection images and corresponding set error parameters into the ResNet50 residual network, and use the corresponding set error parameters as reference labels for each projection image; S4. In the ResNet50 residual network, the X-ray spatial calibration method used by the residual network is performed on each projection map to obtain the calibration error parameters and compare them with the control label. S5. Repeat steps S2 to S4 to train the ResNet50 residual network until the comparison results of the calibration error parameters and the control labels meet the preset requirements. The training is then completed, and the trained ResNet50 residual network is obtained. S6, input the projection image to be calibrated into the trained ResNet50 residual network to obtain the ΔR output by the ResNet50 residual network. z ΔTd x ,ΔTd y And according to ΔR z ΔTd x ΔTd y Perform spatial calibration.

4. The spatial self-calibration method based on residual networks according to claim 3, characterized in that: In step S1-1, the similarity coefficient is calculated using the following formula: Where S represents the similarity coefficient, LA i and LB i N represents the projected CT value of the i-th line pixel in each mirror image after the measured object is rotated by an angle θ. num The maximum number of pixels for the detector. and Each set of mirror images (N) represents the number of mirror images after the object under test has been rotated by θ degrees. num The average projected CT value of the line pixels, θ is greater than or equal to 0° and less than 180°; In steps S1-2, the angular error ΔR of the rotation axis on the Z-axis of the fixed coordinate system is determined based on the maximum value of the similarity coefficients of multiple sets of mirror projections. z Specifically, the rotation angle corresponding to the maximum value of the similarity coefficient of each group of mirror projections is taken as ΔR. z .

5. The spatial self-calibration method based on residual networks according to claim 3 or 4, characterized in that: In step S2, the standard CT data is derived from the official CT data provided by TCIA; In step S2, the forward projection of the standard CT data specifically involves forward projection of the standard CT data using the open-source X-ray simulation code RTK.

Images