An x-ray detector array imaging system

Abstract

The application discloses an X-ray detector array imaging system and relates to the technical field of array imaging, which is based on a three-dimensional tensor input constructed by continuous frames, combined with a generation mechanism of an acceleration channel, so that the system can sensitively capture sudden displacement changes in surgery, such as non-uniform motion caused by electric saw sudden stop or respiratory muscle tremor. The curvature change of the displacement curve is accurately quantified through Gaussian weighted second-order difference, and weak jitter interference is filtered by using an adaptive suppression coefficient, so as to provide high-robustness time-varying features for motion compensation; a spatial feature extraction branch of a double-branch processing unit retains anatomical details of bones and blood vessels, and a time motion modeling branch analyzes long-term dependence of a displacement field sequence through an LSTM network; after being spliced through a fusion layer, the two are input into an adversarial training module, under a game mechanism of a generator-discriminator, so that overall motion artifacts are corrected, and microstructure loss caused by excessive smoothing is avoided.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of array imaging technology, in particular to an X-ray detector array imaging system. BACKGROUND

[0002] In the minimally invasive surgery of orthopedics and thoracic surgery, the C-arm X-ray imaging system needs to take multiple-angle continuous shooting around the patient to guide the positioning of surgical instruments in real time. The target of shooting moves irregularly in a millisecond interval due to the unconscious muscle tremor of the patient under the anesthesia state, the fluctuation of breathing and the sudden displacement of the surgical instrument operated by the doctor. This compound motion contains low-frequency tremor and high-frequency stop, which is beyond the compensation range of the conventional uniform motion model, causing local trailing and organ edge virtualization in the reconstructed image, and affecting the identification accuracy of fine needles and micro bone seams.

[0003] To deal with motion artifacts, recent solutions mainly fall into two categories: one is the inter-frame displacement compensation algorithm based on optical flow tracking, which predicts the motion trajectory through the displacement vector field of adjacent two frames; the other is the hardware synchronization of the detector and the mechanical arm, which adjusts the exposure timing according to the preset trajectory. The former relies on the continuity assumption of motion and produces false compensation vectors for sudden acceleration scenarios such as instrument stop. The latter needs to modify the mechanical control system, increasing the complexity and failure risk of the device.

[0004] In current clinical practice, the above-mentioned solutions still have three limitations: the optical flow method is prone to misjudgment of blood vessel texture as motion displacement in soft tissue deformation areas; the hardware synchronization scheme cannot adapt to sudden body position adjustment during surgery; the existing deep learning deblurring network excessively smooths bone crack textures below 0.5mm in pursuit of overall clarity; especially in spinal minimally invasive surgery, such defects can lead to misjudgment of the nail track by the operator, requiring repeated shooting verification and prolonging the operation time. SUMMARY

[0005] In view of the above existing problems, the present application is proposed.

[0006] The present application provides an X-ray detector array imaging system to solve the problem of image trailing caused by non-uniform motion during C-arm surgery, false compensation of existing optical flow method and loss of microvessel texture by deep learning deblurring.

[0007] To solve the above technical problems, the present application provides the following technical solutions:

[0008] The present application provides an X-ray detector array imaging system, which comprises,

[0009] An image acquisition module for acquiring continuous multiple frames of X-ray projection images of a target object;

[0010] A motion compensation module connected to the image acquisition module for motion artifact correction of the continuous multiple frames of projection images.

[0011] an image reconstruction module connected to the motion compensation module and outputting a three-dimensional reconstructed image;

[0012] The motion compensation module comprises:

[0013] an input unit for stacking three consecutive projection images along the time dimension into a three-dimensional tensor;

[0014] an acceleration estimation unit for calculating a second-order derivative of the inter-frame displacement based on the three-dimensional tensor to generate an acceleration channel;

[0015] a double-branch processing unit comprising a spatial feature extraction branch and a temporal motion modeling branch connected in parallel.

[0016] As a preferred scheme of the X-ray detector array imaging system, the spatial feature extraction branch is composed of multiple residual convolution blocks for extracting anatomical structure features in the projection images.

[0017] The temporal motion modeling branch comprises a long short-term memory network (LSTM) with the input being an inter-frame displacement vector field calculated by an optical flow method.

[0018] As a preferred scheme of the X-ray detector array imaging system, the double-branch processing unit further comprises a fusion layer for channel splicing the output of the spatial feature extraction branch and the output of the temporal motion modeling branch.

[0019] The fusion layer is connected to an adversarial training module comprising a generator and a discriminator, wherein:

[0020] The generator outputs a corrected image.

[0021] The discriminator calculates an adversarial loss based on the difference between a real high-resolution blood vessel texture image and a generated image.

[0022] As a preferred scheme of the X-ray detector array imaging system, the generation of the acceleration channel comprises the following steps:

[0023] a) calculating a pixel displacement vector field of two adjacent images;

[0024] b) performing a time dimension differentiation operation on the displacement vector field to obtain an acceleration field;

[0025] c) superimposing the acceleration field as an independent channel to the three-dimensional tensor.

[0026] As a preferred scheme of the X-ray detector array imaging system, in the acceleration channel generation step, the second-order differentiation in the time dimension to generate the acceleration field comprises:

[0027] At time index i, take the neighboring three-frame displacement vector field d i-1 (x,y),d i (x,y),d i+1 (x,y) form a local time window, apply a Gaussian weighted second-order central difference to the window to obtain a pixel-level acceleration vector:

[0028]

[0029] where w +1 represents the time weight when shifting forward by one frame, d i+1 (x,y) is the pixel displacement vector when shifting forward by one frame, with unit px, w0 is the current frame time weight, d i (x,y) is the current frame pixel displacement vector, with unit px, w -1 is the time weight when shifting backward by one frame, d i-1 (x,y) is the pixel displacement vector when shifting backward by one frame, with unit px, Δt is the sampling interval of adjacent two frames, with unit s, a i (x,y) is the pixel-level acceleration vector, with unit px·s -2 ;

[0030] The Gaussian kernel gives three weights:

[0031]

[0032] where w r is the normalized weight at time offset r frames, r is the time offset, with unit frame, σ t is the time diffusion coefficient, with unit frame;

[0033] An adaptive suppression is introduced to the acceleration modulus:

[0034]

[0035] A i (x,y) =‖a i (x,y)‖2,

[0036]

[0037] where c i (x,y) is the pixel-level suppression coefficient, k a is the slope adjustment coefficient, px -1 , A i (x,y) is the acceleration modulus, with unit px·s -2 , γ a is the acceleration threshold, with unit px·s -2 , To complete the inhibited acceleration channel vector, unit: px·s -2 .

[0038] As a preferred scheme of the X-ray detector array imaging system, the system is integrated into a C-arm X-ray machine, and the target object is an intraoperative patient.

[0039] The acquisition angles of the continuous multiple frames of projection images cover an arc track of 0° to 180°.

[0040] As a preferred scheme of the X-ray detector array imaging system, the loss function includes a weighted sum of the following three terms:

[0041] a) The sum of the squares of the intensity differences between the corrected images output by the generator and the corresponding pixels of the true high-resolution images;

[0042] b) The binary cross-entropy loss of the discriminator on the distribution difference between the generated images and the true images;

[0043] c) The second-order difference operation along the time axis of the displacement vector field output by the time motion modeling branch, and the calculation of its Euclidean norm.

[0044] As a preferred scheme of the X-ray detector array imaging system, in the loss function, the time second-order difference of the displacement vector field and the Euclidean norm calculation specifically include:

[0045] The time motion modeling branch gives a pixel displacement vector field s t (x,y) at the t-th frame; the second-order central difference is performed on the sequence:

[0046]

[0047] wherein s t+1 (x,y) is the pixel displacement vector at time index t+1, unit: px, s t (x,y) is the pixel displacement vector at time index t, unit: px, s t-1 (x,y) is the pixel displacement vector at time index t-1, unit: px, Δt is the sampling interval of adjacent two frames, unit: s, g t (x,y) is the pixel-level displacement second-order time difference vector, unit: px·s -2 ;

[0048] The Euclidean norm of the difference vector is taken to obtain the local curvature intensity:

[0049] r t (x,y) =‖g t (x,y)‖2,

[0050] wherein r t (x,y) is the pixel-level curvature intensity, unit: px·s -2 ;

[0051] A weight mask based on the acceleration channel is introduced to balance the contribution of different spatial regions in the loss function:

[0052]

[0053] wherein ω(x,y) is the spatial weight, β is the weight amplification coefficient, α is the decay rate coefficient, A(x,y) is the pixel-level acceleration module, unit: px·s -2 , A max is the maximum value of the acceleration module in the current batch, unit: px·s -2 ;

[0054] The time smoothing regularization term is obtained comprehensively:

[0055]

[0056] wherein, is the time smoothing regularization term, λ tem is the regularization weight, N f is the input sequence frame number, N p is the total number of pixels in a single frame, unit: px;

[0057] The regularization weight is adaptively adjusted according to the overall motion intensity, and the adjustment formula is:

[0058]

[0059] wherein η is the reference proportion factor, is the average value of the acceleration module, unit: px·s -2 , A ref is the acceleration reference value, unit: px·s -2 .

[0060] As a preferred scheme of the X-ray detector array imaging system, wherein: when the image reconstruction module executes the filtered back-projection algorithm, the projection data input is the image sequence processed by the motion compensation module, and the image sequence is weighted filtered before reconstruction, and the filter kernel function is negatively related to the amplitude of the acceleration channel.

[0061] As a preferred scheme of the X-ray detector array imaging system, wherein: the spatial feature extraction branch contains at least three residual convolution blocks, and each residual convolution block is composed of a convolution layer, a batch normalization layer and a ReLU activation layer connected in turn.

[0062] The hidden layer dimension of the long short-term memory network (LSTM) of the time motion modeling branch is the same as the output channel number of the spatial feature extraction branch.

[0063] The X-ray detector array imaging system provided by the application has the advantages that: the multi-dimensional collaborative processing of the motion compensation module significantly optimizes the dynamic imaging quality in surgical operation; the three-dimensional tensor input constructed based on continuous frames, combined with the generation mechanism of the acceleration channel, enables the system to sensitively capture sudden displacement changes in the operation, such as non-uniform motion caused by the sudden stop of an electric saw or the tremor of respiratory muscles. The curvature change of the displacement curve is accurately quantified through Gaussian weighted second-order difference, and the weak jitter interference is filtered by using an adaptive suppression coefficient, thereby providing high-robustness time-varying features for motion compensation.

[0064] The spatial feature extraction branch of the double-branch processing unit of the application retains the anatomical details of bones and blood vessels, and the time motion modeling branch analyzes the long-term dependence of the displacement field sequence through an LSTM network; after being spliced by the fusion layer, the two are input into the adversarial training module, which, under the game mechanism of the generator-discriminator, not only corrects the overall motion artifacts, but also avoids the loss of microstructure caused by excessive smoothing; the curvature intensity penalty term in the loss function, combined with the acceleration-weighted spatial mask, effectively suppresses the timing jitter in the motion compensation process; the constraint is automatically relaxed in the smoothing limit during fast-paced surgical actions, avoiding the false suppression of real motion;

[0065] An acceleration amplitude negatively correlated filter kernel is introduced in the image reconstruction stage, which further weakens the influence of residual artifacts on three-dimensional reconstruction, so that the output image still maintains sharp organ edges in the fast-moving instrument scene;

[0066] The application realizes end-to-end optimization in the 0°-180° multi-angle continuous shooting of a C-arm machine, reduces the need for repeated shooting of the same part during the operation, and reduces the risk of radiation exposure of the patient. BRIEF DESCRIPTION OF DRAWINGS

[0067] In order to more clearly illustrate the technical solutions of the embodiments of the application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0068] Figure 1 It is a schematic diagram of the framework of an X-ray detector array imaging system in Example 1. DETAILED DESCRIPTION

[0069] In order to make the above objectives, characteristics and advantages of the present application more obvious and comprehensible, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0070] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. The present application, however, can be practiced in a variety of ways other than those specifically described herein without departing from the scope of the present application, and it is understood that it covers all technical and structural equivalents of the elements described and practiced in the same manner for a similar purpose. Therefore, the present application is not limited to the specific embodiments disclosed below.

[0071] Secondly, the term "one embodiment" or "an embodiment" as used herein means that a particular implementation can include a particular feature, structure, or characteristic, but every embodiment can not necessarily include the particular feature, structure, or characteristic. Moreover, such phrases do not necessarily refer to the same embodiment. Further, when a particular feature, structure, or characteristic is described in connection with an embodiment, it is submitted that it is within the purview of those skilled in the art to effect such feature, structure, or characteristic in connection with other embodiments whether or not explicitly described or illustrated.

[0072] Embodiment 1, with reference to Figure 1 The embodiment provides an X-ray detector array imaging system, comprising:

[0073] An image acquisition module is configured to acquire continuous multiple frames of X-ray projection images of a target object.

[0074] A motion compensation module is connected to the image acquisition module and configured to correct motion artifacts of the continuous multiple frames of projection images.

[0075] An image reconstruction module is connected to the motion compensation module and configured to output a three-dimensional reconstructed image.

[0076] The motion compensation module comprises:

[0077] An input unit is configured to stack three continuous frames of projection images along a time dimension into a three-dimensional tensor.

[0078] An acceleration estimation unit is configured to calculate a second derivative of inter-frame displacement based on the three-dimensional tensor to generate an acceleration channel.

[0079] A double-branch processing unit comprises a spatial feature extraction branch and a temporal motion modeling branch connected in parallel.

[0080] When the image reconstruction module executes a filtered back-projection algorithm, the projection data input is the image sequence processed by the motion compensation module, and the image sequence is weighted filtered before reconstruction, and the filter kernel function is negatively correlated with the amplitude of the acceleration channel.

[0081] The generation of the acceleration channel comprises the following steps:

[0082] a) calculating a pixel displacement vector field of two adjacent frames of images;

[0083] b) time-dimension differential operation on the displacement vector field to obtain an acceleration field;

[0084] c) superimposing the acceleration field as an independent channel to the three-dimensional tensor;

[0085] In the acceleration channel generation step, the second-order time-dimension differential operation is performed to generate the acceleration field, comprising:

[0086] At time index i, take the adjacent three frames of displacement vector fields d i-1 (x,y), d i (x,y), d i+1 (x,y) form a local time window, and a Gaussian weighted second-order central difference is applied to the window to obtain a pixel-level acceleration vector:

[0087]

[0088] where w +1 represents the time weight when shifting forward by one frame, d i+1 (x,y) is the pixel displacement vector when shifting forward by one frame, with the unit of px, w0 is the current frame time weight, d i (x,y) is the current frame pixel displacement vector, with the unit of px, w -1 is the time weight when shifting backward by one frame, d i-1 (x,y) is the pixel displacement vector when shifting backward by one frame, with the unit of px, Δt is the sampling interval of adjacent two frames, with the unit of s, a i (x,y) is the pixel-level acceleration vector, with the unit of px·s -2 ;

[0089] The Gaussian kernel gives three weights:

[0090]

[0091] where w r is the normalized weight at the time offset r frames, r is the time offset, with the unit of frame, σ t is the time diffusion coefficient, with the unit of frame;

[0092] To highlight fast motion and suppress weak jitter, an adaptive suppression is introduced to the acceleration module value:

[0093]

[0094] A i (x,y) = ‖a i (x,y)‖2,

[0095]

[0096] where c i(x, y) is a pixel-level suppression coefficient, k a is a slope adjustment coefficient, px -1 , A i (x, y) is an acceleration module, unit: px s -2 , γ a is an acceleration threshold, unit: px s -2 , is the acceleration channel vector after completing suppression, unit: px s -2 ;

[0097] Specifically, the flow constructs time-dependent weights through a Gaussian kernel, making the difference more sensitive to short-term continuous motion and less affected by random jitter; the second-order central difference can directly reflect the curvature change of the displacement curve, locate the acceleration peak position, and thus capture the real tissue rapid movement; the adaptive suppression link uses a logic function to smooth and clip the module value, and the weak motion area signal is low-pass filtered, while the strong motion area retains the original amplitude to avoid injecting background noise into the subsequent network; the overall forms a closed loop of weight distribution curvature extraction amplitude modulation, and the acceleration channel dynamically adjusts with the rapid change of the shooting angle, providing high-contrast and high-robustness time-varying prior features for motion compensation, significantly reducing the false image interference in angiographic sequences;

[0098] The spatial feature extraction branch is composed of multiple residual convolution blocks for extracting anatomical structure features in the projection image;

[0099] The temporal motion modeling branch includes a long short-term memory network (LSTM), and the input is the inter-frame displacement vector field calculated by the optical flow method;

[0100] The dual-branch processing unit further includes a fusion layer for channel splicing the output of the spatial feature extraction branch and the output of the temporal motion modeling branch;

[0101] The fusion layer is connected to an adversarial training module, which includes a generator and a discriminator, wherein:

[0102] The generator outputs the corrected image;

[0103] The discriminator calculates an adversarial loss based on the difference between the real high-resolution blood vessel texture image and the generated image;

[0104] The spatial feature extraction branch includes at least three residual convolution blocks, and each residual convolution block is composed of a convolution layer, a batch normalization layer, and a ReLU activation layer connected in sequence;

[0105] The hidden layer dimension of the long short-term memory network (LSTM) of the temporal motion modeling branch is the same as the output channel number of the spatial feature extraction branch;

[0106] The loss function includes a weighted sum of the following three terms:

[0107] a) The sum of the square of the intensity difference between the corrected image output by the generator and the corresponding pixels of the real high-resolution image;

[0108] b) The binary cross-entropy loss of the distribution difference between the generated image and the real image by the discriminator;

[0109] c) The Euclidean norm of the second-order difference operation along the time axis of the displacement vector field output by the time motion modeling branch;

[0110] In the loss function, the time second-order difference of the displacement vector field and the Euclidean norm calculation are performed, which specifically includes:

[0111] The time motion modeling branch gives the pixel displacement vector field s t (x,y) at the t-th frame; The second-order central difference is performed on the sequence:

[0112]

[0113] where s t+1 (x,y) is the pixel displacement vector at time index t+1, unit: px, s t (x,y) is the pixel displacement vector at time index t, unit: px, s t-1 (x,y) is the pixel displacement vector at time index t-1, unit: px, Δt is the sampling interval of adjacent two frames, unit: s, g t (x,y) is the pixel-level displacement second-order time difference vector, unit: px·s -2 ;

[0114] The Euclidean norm of the difference vector is taken to obtain the local curvature intensity:

[0115] r t (x,y) = ‖g t (x,y)‖2,

[0116] where r t (x,y) is the pixel-level curvature intensity, unit: px·s -2 ;

[0117] A weight mask based on the acceleration channel is introduced to balance the contributions of different spatial regions in the loss function:

[0118]

[0119] where ω(x,y) is the spatial weight, β is the weight increase coefficient, α is the decay rate coefficient, A(x,y) is the pixel-level acceleration module, unit: px·s -2 , A max is the maximum value of the acceleration module in the current batch, unit: px·s -2;

[0120] The time smoothing regularization term is obtained by integration:

[0121]

[0122] wherein, is the time smoothing regularization term, λ tem is the regularization weight, N f is the number of input sequence frames, N p is the total number of single-frame pixels, unit: px;

[0123] The regularization weight is adaptively adjusted according to the overall motion intensity, and the adjustment formula is:

[0124]

[0125] wherein, η is a reference scale factor, is the average value of acceleration module, unit: px·s -2 , A ref is the acceleration reference value, unit: px·s -2 ;

[0126] Specifically, the second central difference directly reflects the curvature of the displacement sequence, which can mark the position of the acceleration mutation; after the Euclidean norm is converted into a scalar, the curvature intensity of each pixel in the same frame is convenient for unified measurement; the weight mask makes the area with high acceleration amplitude occupy a larger proportion in the regularization term, and the low-amplitude static area has limited influence on training, thereby highlighting the real anatomical motion; the regularization weight depends on the global average acceleration and is adaptively changed, which automatically relaxes the smoothing constraint when the intraoperative fast-paced action occurs, avoids excessive inhibition, and improves the constraint when the action is gentle, thereby suppressing the time noise; the curvature extraction-space weighting-global adaptive chain improves the sensitivity of the motion modeling branch to the temporal consistency, while reducing the overfitting tendency of the network to the pseudo-random transition;

[0127] The system is integrated in a C-arm X-ray machine, and the target object is an intraoperative patient;

[0128] The acquisition angles of the continuous multiple frames of projection images cover an arc trajectory of 0° to 180°.

[0129] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application, and they should be covered in the scope of the claims of the present application.

Claims

1. An X-ray detector array imaging system, characterized in that, The system comprises: an image acquisition module configured to acquire continuous multiple frames of X-ray projection images of a target object; a motion compensation module connected to the image acquisition module and configured to perform motion artifact correction on the continuous multiple frames of projection images; an image reconstruction module connected to the motion compensation module and configured to output a three-dimensional reconstructed image; wherein the motion compensation module comprises: an input unit configured to stack three continuous frames of projection images along a time dimension into a three-dimensional tensor; an acceleration estimation unit configured to calculate a second-order derivative of inter-frame displacement based on the three-dimensional tensor to generate an acceleration channel; a dual-branch processing unit comprising a spatial feature extraction branch and a temporal motion modeling branch connected in parallel; the spatial feature extraction branch is composed of multiple residual convolution blocks and is configured to extract anatomical structure features in the projection images; the temporal motion modeling branch comprises a long short-term memory network (LSTM) and is configured to input an inter-frame displacement vector field calculated by an optical flow method; the dual-branch processing unit further comprises a fusion layer configured to channel splice outputs of the spatial feature extraction branch and the temporal motion modeling branch; the fusion layer is connected to an adversarial training module comprising a generator and a discriminator, wherein: the generator outputs a corrected image; the discriminator calculates an adversarial loss based on a difference between a real high-resolution blood vessel texture image and a generated image.

2. An X-ray detector array imaging system as in claim 1, characterized in that The generation of the acceleration channel comprises the following steps: a) calculating a pixel displacement vector field of adjacent two frames of images; b) performing a time dimension differential operation on the displacement vector field to obtain an acceleration field; c) superimposing the acceleration field as an independent channel to the three-dimensional tensor.

3. An X-ray detector array imaging system as in claim 2, wherein, In the acceleration channel generation step, the generation of the acceleration field by the time dimension second-order differential operation comprises: At time index The adjacent three-frame displacement vector fields are taken A local time window is formed, and a Gaussian weighted second-order central difference is applied to the window to obtain a pixel-level acceleration vector: ， wherein, represents the time weight when offsetting one frame forward, is a pixel displacement vector offsetting one frame forward, in px, is the current frame time weight, is a pixel displacement vector of the current frame, in px, is the time weight offsetting one frame backward, is a pixel displacement vector offsetting one frame backward, in px, is the sampling interval between two adjacent frames, in s, is a pixel-level acceleration vector, in ; a Gaussian kernel gives three weights: ， ， wherein, is a time offset a normalized weight at a frame, is a time offset in frames, is a time diffusion coefficient in frames; an adaptive suppression is introduced for the acceleration modulus value: ， ， ， wherein, is a pixel-level suppression coefficient, is a slope adjustment coefficient, , is an acceleration module in , is an acceleration threshold in , is a completed acceleration channel vector after suppression in .

4. An X-ray detector array imaging system as claimed in any one of claims 1 to 3, characterized in that The system is integrated into a C-arm X-ray machine, and the target object is an intraoperative patient; the acquisition angle of the continuous multiple frames of projection images covers an arc trajectory of 0° to 180°.

5. An X-ray detector array imaging system as in claim 4, wherein, The loss function comprises a weighted sum of the following three terms: a) a sum of squares of intensity difference values of corresponding pixels between the corrected image output by the generator and a real high-resolution image; b) a binary cross-entropy loss of the discriminator on the distribution difference between the generated image and the real image; c) a quadratic difference operation is performed on the displacement vector field output by the temporal motion modeling branch along the time axis, and a Euclidean norm thereof is calculated.

6. An X-ray detector array imaging system as in claim 5, wherein, In the loss function, the time quadratic difference of the displacement vector field and the Euclidean norm calculation comprise: The temporal motion modeling branch in the first frame gives a pixel displacement vector field ; performing a second central difference on the displacement vector field: ， wherein, is a time index is a pixel displacement vector at time t, unit px, is a time index is a pixel displacement vector at time t, unit px, is a time index is a pixel displacement vector at time t, unit px, is a sampling interval between two adjacent frames, unit s, is a pixel-level displacement second-order time difference vector, unit ; a local curvature intensity is obtained by taking the Euclidean norm of the difference vector: ， wherein, is the pixel-level curvature intensity in units of ; a weight mask based on the acceleration channel is introduced to balance the contributions of different spatial regions in the loss function: ， wherein, is a spatial weight, is a weight increase coefficient, is a decay rate coefficient, is a pixel-level acceleration modulus, in , is a current batch acceleration modulus maximum, in ; a time smoothing regularization term is obtained by integration: ， wherein, is a time smoothing regular term, is a regular weight, is the number of input sequence frames, is the total number of single frame pixels, unit px; the regularization weight is adaptively adjusted according to the overall motion intensity, and the adjustment formula is: ， ， wherein, is a reference scale factor, is an acceleration module average value in , is an acceleration reference value in .

7. An X-ray detector array imaging system as in claim 1, wherein, When the image reconstruction module executes a filtered back-projection algorithm, the projection data input is the image sequence processed by the motion compensation module, and the image sequence is weighted filtered before reconstruction, and the filter kernel function is negatively correlated with the amplitude of the acceleration channel.

8. An X-ray detector array imaging system as in claim 1, wherein, The spatial feature extraction branch comprises at least three residual convolution blocks, and each residual convolution block is composed of a convolution layer, a batch normalization layer and a ReLU activation layer connected in sequence. The hidden layer dimension of the long short-term memory network (LSTM) of the time motion modeling branch is the same as the output channel number of the spatial feature extraction branch.

Images