Human tissue imaging method, system, medium and computer device

Abstract

The present invention discloses a method, system, medium and computer equipment for imaging human tissue, and relates to the technical field of imaging human tissue. The method includes a ray precision sampling module based on geometric prior, a geometric consistency constraint module and a two-stage reconstruction model. The CBCT image reconstruction task will be converted into a function optimization problem driven by an implicit neural field, and the geometric prior information of the CBCT imaging system is integrated to constrain the deep reconstruction method. The CBCT image reconstruction task is carried out in two stages. In the first stage, the geometric shape of the imaging system is used to discretely simulate the attenuation process, and the neural intensity field is optimized by comparing the real projection and the synthetic projection, and the CBCT shape is preliminarily learned to perform the initial reconstruction of the image; in the second stage, the preliminary reconstruction result of the first stage is first embedded, and then the neural intensity field is further optimized according to the real projection, and the initial reconstructed image is refined. The above method is helpful to achieve high-quality human tissue imaging.

Description

Technical Field

[0001] The present invention relates to the field of image imaging, in particular to the technical field of imaging of human tissue, and more particularly to an imaging method, system, medium and computer equipment for medical organs. Background Art

[0002] As a new type of CT imaging technology, CBCT plays an important role in human tissue imaging. In order to reduce radiation damage to patients, CBCT, which is commonly used in clinical practice, needs to improve its imaging quality while reducing X-ray radiation dose. Sparse angle CBCT imaging of human tissue can reduce the acquisition angle, effectively reduce the radiation dose during the scanning process, reduce radiation damage to patients, and also greatly shorten the scanning time, improving the inspection efficiency of CBCT.

[0003] However, compared with traditional CBCT imaging, sparse angle CBCT imaging faces more challenges. 1) Data incompleteness: In sparse angle CBCT reconstruction, since projection data can only be collected from limited angles, the data available for reconstruction is greatly reduced compared with full angle CBCT reconstruction, resulting in large errors in the reconstructed image, such as blur, distortion, and artifacts. 2) Computational complexity: Traditional CBCT reconstruction algorithms, such as filtered back projection (FBP), are not suitable for sparse angle reconstruction. Iterative reconstruction algorithms, such as algebraic reconstruction algorithm (ART), joint iterative reconstruction algorithm (SIRT), and compressed sensing (CS) algorithm are more effective when processing sparse sampling data, but these methods are computationally complex, especially when the number of iterations increases, there is a higher demand for computing resources. In addition, the parameter adjustment of the iterative reconstruction algorithm has a great impact on the final reconstruction quality, and often requires professional knowledge and a large number of experiments to optimize. 3) Recovering detail information from limited viewing angles: Under sparse conditions, the task of recovering object details from limited projections becomes more complicated, and it may be difficult to distinguish whether certain features are caused by actual anatomical structures or noise or artifacts in the reconstruction process. Therefore, the algorithm needs to be robust enough to distinguish these cases and recover realistic and accurate anatomical details.

[0004] In order to solve the problems of blurred reconstruction results and artifacts caused by incomplete projection data in sparse angle reconstruction, the current mainstream approach is to use image enhancement, using reconstructed images with artifacts from a few angles as input, improving image quality under the influence of big data, using the information extraction ability of deep learning models to repair the sparse angle sinusoidal information, and reconstructing CT images using the enhanced sinusoidal and original sinusoidal images. Although this type of method can achieve better imaging effects with a few angles, it still relies on projection data with more angles. In addition, the enhanced sinusoidal is only applicable to two-dimensional CT reconstruction tasks and is difficult to introduce into three-dimensional CBCT image reconstruction methods.

[0005] Large-scale training data can also be used to train convolutional neural networks to learn the mapping from original measurement data to reconstructed images. Although this method can achieve sparse angle or even single-angle three-dimensional CBCT imaging tasks, it has poor generalization and is difficult to apply in different imaging scenarios.

[0006] Based on the above existing technologies, it is not difficult to see that sparse-angle CBCT reconstruction of human tissue can only reconstruct a complete three-dimensional volume image from the projection data obtained from limited angles, resulting in a lack of integrity of spatial information in the reconstruction process, causing artifacts and blur in the reconstructed image; and although the reconstruction method based on deep learning has shown superiority in processing sparse data, the reconstruction results are unstable when there are subtle but significant structural changes. The generalization ability of its models and algorithms is insufficient for different human tissues (such as dentistry, orthopedics, breast imaging, etc.), and it is difficult to adapt to different imaging needs and conditions. Summary of the invention

[0007] The present invention provides a method, system, medium and computer equipment for imaging human tissue. The method is helpful to achieve high-quality human tissue imaging.

[0008] The present invention provides a method for imaging human tissue, comprising:

[0009] Sampling the attenuation field of human tissue penetrated by the ray to obtain the spatial coordinates of the human tissue attenuation field, and position encoding the spatial coordinates of the human tissue attenuation field to obtain the high-dimensional spatial coordinates of the encoded human tissue attenuation field;

[0010] The attenuation coefficient of the attenuation field of the human tissue is queried through the first multilayer perceptron neural network to synthesize the projection, the mean square error and similarity of the synthesized projection are compared with the actual projection to obtain a first loss function, and the first multilayer perceptron neural network is iteratively optimized using the first loss function; when the first multilayer perceptron neural network converges, the spatial coordinates of the encoded human tissue attenuation field are queried through the optimized first multilayer perceptron neural network to obtain a preliminary human tissue CBCT image;

[0011] embedding the preliminary human tissue CBCT image into the intensity field, sampling the spatial coordinates of the preliminary human tissue CBCT image, and position encoding the spatial coordinates of the sampled preliminary human tissue CBCT image to obtain the spatial coordinates of the encoded preliminary human tissue CBCT image;

[0012] The intensity value of the intensity field is queried through the second multi-layer perceptron neural network to obtain an intermediate human tissue CBCT image, the intermediate human tissue CBCT image is forward projected to obtain a target angle projection, the target angle projection is compared with the actual projection in terms of mean square error and structural similarity to obtain a second loss function, and the second multi-layer perceptual neural network is iteratively optimized using the second loss function; when the second multi-layer perceptual neural network converges, the spatial coordinates of the encoded preliminary human tissue CBCT image are queried through the optimized second multi-layer perceptual neural network to generate a final human tissue CBCT image.

[0013] Preferably, sampling the attenuation field of human tissue penetrated by the ray to obtain the spatial coordinates of the human tissue attenuation field specifically includes determining an X-ray path through the spatial coordinates p1 of the X-ray source and the spatial coordinates p2 of the detector pixel point, and the parameters of the ray path are expressed as follows:

[0014] X(α)=X1+α(X2-X1)

[0015] Y(α)=Y1+α(Y2-Y1)

[0016] Z(α)=Z1+α(Z2-Z1)

[0017] Among them, the value of parameter α at point P1 is 0, and the value at point P2 is 1, (X1, Y1, Z1) is the spatial coordinate of the X-ray source, and (X2, Y2, Z2) is the spatial coordinate of the detector pixel point;

[0018] The orthogonal set of equally spaced parallel planes of the CBCT array for the (Nx, Ny, Nz) spatial coordinates is expressed as follows:

[0019] X plane (i) = X plane (1)+(i-1)d x (i=1,...,N x )

[0020] Y plane (j) = Y plane (1)+(j-1)d y (j=1,...,N y )

[0021] Z plane (k) = Z plane (1)+(k-1)d z (k=1,…,N z )

[0022] Among them, dx, dy and dz are the distances between the x, y and z planes respectively, which are also the lengths of the spatial coordinates. The parameter values ​​amin and amax are obtained by the intersection of the ray with the boundary of the CBCT array. The parameters corresponding to each side of the spatial coordinate are expressed as follows:

[0023] α x (1) = [X plane (1)-X1] / (X2-X1)

[0024] α x (N x )=[X plane (N x )-X1] / (X2-X1)

[0025] X2–X1≠0

[0026] α y (1) α y (N y ), α z (1) and α z (N z) The calculation method of is the same as above;

[0027] For a given parameter value, αmin and αmax are calculated as follows:

[0028] α min =max{0,min[α x (1),α x (N x )],

[0029] min[α y (1),α y (N y )],min[α z (1),α z (N z )]}

[0030] α max =min{1,max[α x (1),α x (N x )],

[0031] max[α y (1),α y (N y )],max[α z (1),α z (N z )]}

[0032] Among them, the function min and the function max are the minimum and maximum items in their parameter lists respectively. If α min Greater than or equal to α max , then the ray does not intersect with the CBCT image;

[0033] The ray path is parameterized by α, and the entry and exit points of the ray passing through the CBCT image can be calculated by αmin and αmax:

[0034] X in =X1+αmin(X2-X1)

[0035] Y in =Y1+αmin(Y2-Y1)

[0036] Z in =Z1+αmin(Z2-Z1)

[0037] X out =X1+α max (X2-X1)

[0038] Y out =Y1+α max (Y2-Y1)

[0039] Z out =Z1+α max (Z2-Z1)

[0040] Among them, (X in , Y in , Z in ) is the spatial coordinate of the injection point, (X out , Y out , Z out ) is the spatial coordinate of the ejection point.

[0041] Preferably, the calculation formula of the synthetic projection is as follows:

[0042]

[0043] Among them, I is the intensity value of the ray after passing through the human tissue, I0 is the initial intensity, μ i is the attenuation coefficient of the sampling point, δ i is the ray sampling step size.

[0044] Preferably, the calculation formula of the first loss function is as follows:

[0045] Loss1=α·MSE(I pre ,I truth )+β·SSIM(projs_pre,projs_truth)

[0046] Among them, Loss1 is the first loss function, which is the projection value of the ray predicted by the attenuation field, I truth is the true projection value of the ray, projs_pre is the projection image generated by the CBCT image obtained by querying the corresponding voxel in the attenuation field, and projs_truth is the true projection image; in the first 500 iterations, α=1, β=0, and in the next 200 iterations, α=β=0.5.

[0047] Preferably, the calculation formula of the second loss function is as follows:

[0048] Loss2=MSE(pre_projs,train_projs)+SSIM(pre_projs,train_projs)

[0049] Among them, MSE is the mean square error, SSIM is the structural similarity, pre_projs represents the reconstructed image projection, and train_projs represents the target projection.

[0050] Preferably, the calculation formulas for mean square error and structural similarity are:

[0051]

[0052] Among them, MSE is the mean square error, SSIM is the structural similarity, n is the total number of pixels in the image, μ is the average brightness of the image, σ is the variance of the image, and C1 and C2 are constants added to avoid the denominator being zero.

[0053] Preferably, the basis for the convergence of the first multilayer perceptron neural network is that the value of the first loss function is zero; the basis for the convergence of the second multilayer perceptron neural network is that the value of the second loss function is zero.

[0054] Preferably, a hash encoder is used to position-encode the spatial coordinates.

[0055] The present invention also provides a human tissue imaging device, comprising:

[0056] The first data acquisition and processing module is used to sample the attenuation field of human tissue penetrated by the radiation to obtain the spatial coordinates of the human tissue attenuation field, and to position-code the spatial coordinates of the human tissue attenuation field to obtain the high-dimensional spatial coordinates of the encoded human tissue attenuation field;

[0057] A first image acquisition module is used to query the attenuation coefficient of the attenuation field of the human tissue through the first multilayer perceptron neural network to synthesize the projection, compare the synthesized projection with the actual projection by mean square error and similarity to obtain a first loss function, and iteratively optimize the first multilayer perceptron neural network using the first loss function; when the first multilayer perceptron neural network converges, query the spatial coordinates of the encoded human tissue attenuation field through the optimized first multilayer perceptron neural network to obtain a preliminary human tissue CBCT image;

[0058] The second data acquisition and processing module is used to embed the preliminary human tissue CBCT image into the intensity field, sample the spatial coordinates of the preliminary human tissue CBCT image, and position encode the spatial coordinates of the sampled preliminary human tissue CBCT image to obtain the spatial coordinates of the encoded preliminary human tissue CBCT image;

[0059] The second image acquisition module is used to query the intensity value of the intensity field through a second multi-layer perceptron neural network to obtain an intermediate human tissue CBCT image, perform forward projection of the intermediate human tissue CBCT image to obtain a target angle projection, compare the target angle projection with the actual projection in terms of mean square error and structural similarity to obtain a second loss function, and iteratively optimize the second multi-layer perceptual neural network using the second loss function; when the second multi-layer perceptual neural network converges, query the spatial coordinates of the encoded preliminary human tissue CBCT image through the optimized second multi-layer perceptual neural network to generate a final human tissue CBCT image.

[0060] The present invention also provides a computer-readable storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned human tissue imaging method is implemented.

[0061] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned human tissue imaging method when executing the program.

[0062] At least one of the above technical solutions adopted in this specification can achieve the following beneficial effects:

[0063] In a method for imaging human tissue provided in the present specification, the human tissue imaging process is divided into an initial reconstruction based on a neural attenuation field and a fine reconstruction based on continuous intensity. In the initial reconstruction process, more accurate spatial coordinates of the attenuation field are obtained through precise sampling, and the loss function obtained by comparing the synthetic projection of the neural network with the actual projection is optimized, and the image is initially acquired using the optimized neural network; in the fine reconstruction process, the initially acquired image is placed in an intensity field, and an intermediate image obtained by the neural network through the intensity field is orthographically projected to obtain a target angle projection, which is compared with the actual projection to obtain another loss function, and the neural network is optimized. The neural network is optimized through the loss function and the encoding is queried after optimization to initially acquire the final human tissue CBCT image.

[0064] The present invention constructs a two-stage reconstruction process, converts the image reconstruction task into an optimization problem of a continuous function in the preliminary reconstruction stage, and obtains discretized image space coordinates by uniformly grid sampling of the volume image after the neural network is optimized. The trained network is used to infer all image space coordinates in the intensity field to obtain the reconstructed image of the first stage. In the fine reconstruction stage, the volume image obtained in the previous stage is mapped to a neural intensity field, and the volume image is further refined using known projection data, so as to obtain a higher quality image imaging effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0066] Figure 1 A schematic diagram of a human tissue imaging method provided by the present invention;

[0067] Figure 2 A schematic diagram of an existing attenuation field sampling provided by the present invention;

[0068] Figure 3 A schematic diagram of attenuation field sampling provided by the present invention;

[0069] Figure 4 A schematic diagram of an optimization process of a neural attenuation field provided by the present invention;

[0070] Figure 5 A schematic diagram of the optimization process of a continuous intensity field provided by the present invention;

[0071] Figure 6 Some sample images in the medical imaging data set provided by the present invention;

[0072] Figure 7A comparison diagram of the imaging effects of images at different sparse angles provided by the present invention;

[0073] Figure 8 This is a comparison chart of the visual effects of the method of the present invention and the current mainstream method;

[0074] Fig. 9 A comparison chart of the imaging results of the region of interest (ROI) of the method of the present invention and other methods;

[0075] Fig.10 It is a visual display diagram of the reconstruction results of the mandible and foot by the method of the present invention;

[0076] Fig.11 This is a comparison diagram of imaging between the method of the present invention and other methods under high-resolution tasks; DETAILED DESCRIPTION

[0077] In order to make the purpose, technical solutions and advantages of this specification more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the specific embodiments of this specification and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments in the specification, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0078] The technical solutions provided by various embodiments of the present invention are described in detail below in conjunction with the accompanying drawings.

[0079] Method Example:

[0080] like Figure 1 The figure shows a schematic flow chart of a human tissue imaging method in this specification, which specifically includes the following steps:

[0081] Step S101, sampling the attenuation field of human tissue penetrated by the ray to obtain the spatial coordinates of the human tissue attenuation field, and position encoding the spatial coordinates of the human tissue attenuation field to obtain the high-dimensional spatial coordinates of the encoded human tissue attenuation field.

[0082] In the prior art, uniform sampling is performed directly on the intersection paths of each ray and the attenuation field uniformly specified according to the geometric parameters of the current CBCB imaging system. This will result in repeated sampling of free space that does not contribute to X-ray attenuation, which is inefficient and will cause reconstruction errors, such as Figure 2 Shown is a schematic diagram of existing attenuation field sampling.

[0083] like Figure 3The figure shows a schematic diagram of attenuation field sampling in the present invention. Based on the Siddon algorithm, the voxels in the CBCT volume image are regarded as the intersection volume of an orthogonal set of equidistant parallel planes. The accurate sampling of the rays passing through the attenuation field of human tissue to obtain the spatial coordinates of the attenuation field of human tissue specifically includes determining an X-ray path through the spatial coordinates p1 of the X-ray source and the spatial coordinates p2 of the detector pixel point. The parameters of the ray path are expressed as follows:

[0084] X(α)=X1+α(X2-X1)

[0085] Y(α)=Y1+α(Y2-Y1)

[0086] Z(α)=Z1+α(Z2-Z1)

[0087] The parameter α is 0 at point P1 and 1 at point P2. 1) is the spatial coordinate of the X-ray source, (X2, Y2, Z 2) is the spatial coordinate of the detector pixel;

[0088] The orthogonal set of equally spaced parallel planes of the CBCT array for the (Nx, Ny, Nz) spatial coordinates is expressed as follows:

[0089] X plane (i) = X plane (1)+(i-1)d x (i=1,…,N x )

[0090] Y plane (j) = Y plane (1)+(j-1)d y (j=1,…,N y )

[0091] Z plane (k) = Z plane (1)+(k-1)d z (k=1,…,N z )

[0092] Among them, d x d y and d z They are the distances between the x, y, and z planes, and are also the lengths of the spatial coordinates. The parameter value a min and a max Obtained by the intersection of the ray with the boundary of the CBCT array, the parameters corresponding to each edge of the spatial coordinate are expressed as follows:

[0093] α x (1) = [X plane(1)-X1] / (X2-X1)

[0094] α x (N x )=[X plane (N x )-X1] / (X2-X1)

[0095] X2–X1≠0

[0096] α y (1) α y (N y ), α z (1) and α z (N z) The calculation method of is the same as above;

[0097] For a given parameter value, α min and α max The calculation method is as follows:

[0098] α min =max{0,min[α x (1),α x (N x )],

[0099] min[α y (1),α y (N y )],min[α z (1),α z (N z )]}

[0100] α max =min{1,max[α x (1),α x (N x )],

[0101] max[α y (1),α y (N y )],max[α z (1),α z (N z )]}

[0102] Among them, the function min and the function max are the minimum and maximum items in their parameter lists respectively. If α min Greater than or equal to α max , then the ray does not intersect with the CBCT image;

[0103] The ray path is parameterized by α. min , α maxFind the entry and exit points of the ray passing through the CBCT image:

[0104] X in =X1+αmin(X2-X1)

[0105] Y in =Y1+αmin(Y2-Y1)

[0106] Z in =Z1+αmin(Z2-Z1)

[0107] X out =X1+α max (X2-X1)

[0108] Y out =Y1+α max (Y2-Y1)

[0109] Z out =Z1+α max (Z2-Z1)

[0110] Among them, (X in , Y in , Z in) is the spatial coordinate of the injection point, (X out , Y out , Z out) is the spatial coordinate of the emission point.

[0111] Based on the geometric prior information of the imaging system, the parameter information of each ray at each scanning angle is calculated, including the coordinates of the entry point (Xin, Yin, Zin) of the ray passing through the volume image and the coordinates of the exit point (Xout, Yout, Zout) of the ray exiting the volume image. The intersection of an X-ray and the volume image is accurately calculated through the coordinates of the two points.

[0112] The spatial coordinates of the attenuation field of human tissue are encoded by position encoding. Human tissue is usually composed of several homogeneous media, such as muscle and bone. The attenuation coefficient remains approximately uniform within a medium, but varies between different media. High-frequency features are not required unless the point is close to the edge. Natural objects like smoothness. Many organs have simple shapes, such as muscle or bone. Their smooth surfaces can be easily learned through low-dimensional features. In order to better utilize the above characteristics of human tissue, a hash encoder is used to encode the spatial coordinates.

[0113] The attenuation coefficient of the attenuation field of the ray passing through the human tissue is queried through the first multi-layer perceptron neural network to synthesize the projection, where the first multi-layer neural network is used to represent the continuous intensity field. It takes the encoded spatial coordinates as input and outputs the intensity value corresponding to the position. The first multi-layer perceptual network consists of 4 fully connected layers. The first three layers have 32 channels and use the ReLU activation function in the middle. The last layer has one neuron followed by a Sigmoid activation function. The first multi-layer neural perceptual network also includes a jump connection to connect the network input to the activation value of the second layer.

[0114] According to Beer's law, for each two-dimensional scan, the x-ray source emits a conical beam with an initial intensity. Each ray in the beam is attenuated by the different tissues it encounters, and is finally received by the flat-panel detector to form all pixel-by-pixel intensity values. The calculation formula for the synthetic projection is as follows:

[0115]

[0116] Among them, I is the intensity value of the ray after passing through the human tissue, I0 is the initial intensity, μ i is the attenuation coefficient of the sampling point, δ i is the ray sampling step size.

[0117] Step S102, querying the attenuation coefficient of the attenuation field of the human tissue through the first multilayer perceptron neural network to synthesize the projection, comparing the synthesized projection with the actual projection by mean square error and similarity to obtain a first loss function, and iteratively optimizing the first multilayer perceptron neural network using the first loss function; when the first multilayer perceptron neural network converges, querying the spatial coordinates of the encoded human tissue attenuation field through the optimized first multilayer perceptron neural network to obtain a preliminary human tissue CBCT image.

[0118] like Figure 4 It is a schematic diagram of the optimization process of the neural attenuation field in the present invention, where the mean square error and similarity of the synthetic projection and the actual projection are compared to obtain a loss function, and the loss function is used to iteratively optimize the multilayer perceptron neural network.

[0119] An accurate system geometry model is established based on the actual geometry parameters of the imaging system, including the precise position of the X-ray source relative to the detector, the size and resolution of the detector, and the rotation trajectory of the source and detector around the scanned object during the scanning process. Through the above geometry model, the ray direction relative to the X-ray source is calculated for each detector pixel. In human tissue imaging, these ray directions will replace the traditional observation direction input to ensure that the real X-ray propagation path is considered during the model learning process.

[0120] The imaging process follows specific physical laws, such as the attenuation and absorption of X-rays. Traditional models mainly focus on the color and volume density estimation at a given position and viewing direction. In imaging application scenarios, the geometric configuration has a direct impact on the X-ray projection image, and it is necessary to ensure that the model can accurately simulate the propagation and attenuation of X-rays in the object. In order to ensure the accuracy of reconstruction, the present invention designs a geometrically constrained loss to force the X-ray projection synthesized by the neural attenuation field to match the actual projection image, and constructs a differentiable orthographic projection module based on the geometric parameter information of the imaging system to simulate the propagation and attenuation of X-rays inside the attenuation field of human tissue.

[0121] In the process of optimizing the nerve attenuation field, the changes of the sampling points on the X-ray are simulated from the generated nerve attenuation field, and the projection value generated by the X-ray passing through the attenuation field of the human tissue is calculated through the calculation formula of the synthetic projection, and the projection value is compared with the actual projection value based on the mean square error (MSE) at the pixel level. In addition, after the internal anatomical structure information of the human tissue is initially learned from the attenuation field, the preliminary CBCT image is obtained by querying the corresponding voxels and sent to the forward projection module to obtain the projection image at the same angle as the real sparse angle projection data, and the error comparison with the real projection image is based on the more advanced structural similarity (SSIM).

[0122] The calculation formula of the first loss function of the neural attenuation field optimization process is as follows:

[0123] Loss1=α·MSE(I pre ,I truth )+β·SSIM(projs_pre,projs_truth)

[0124] Among them, Loss1 is the first loss function, which is the projection value of the ray predicted by the attenuation field, I truth is the true projection value of the ray, projs_pre is the projection image generated by the forward projection module of the CBCT image obtained by querying the corresponding voxel in the attenuation field, and projs_truth is the true projection image; in the first 500 iterations, α=1, β=0, and in the next 200 iterations, α=β=0.5.

[0125] MSE is used to measure pixel-level error, and SSIM is used to measure structural error. The calculation formula is as follows:

[0126]

[0127] When the multilayer perceptron neural network converges, the encoded spatial coordinates are queried by the optimized multilayer perceptron neural network to obtain a preliminary CBCT image of human tissue.

[0128] Step S103, embedding the preliminary human tissue CBCT image into the intensity field, sampling the spatial coordinates of the preliminary human tissue CBCT image, and position encoding the spatial coordinates of the sampled preliminary human tissue CBCT image to obtain the spatial coordinates of the encoded preliminary human tissue CBCT image.

[0129] In this process, the input of the network is the normalized coordinates in the preliminary CBCT image space field, and the output is the intensity value of the prior CBCT image at the coordinate learned by the network. The network is reversely optimized through the loss value between the network output image and the preliminary CBCT image. When the network converges, the process of embedding the prior image into the intensity field is completed.

[0130] Among them, a hash encoder is used to positionally encode the spatial coordinates of the preliminary human tissue CBCT image.

[0131] Step S104, querying the intensity value of the intensity field through the second multi-layer perceptron neural network to obtain an intermediate human tissue CBCT image, forward projecting the intermediate human tissue CBCT image to obtain a target angle projection, performing mean square error comparison and structural similarity comparison between the target angle projection and the actual projection to obtain a second loss function, and iteratively optimizing the second multi-layer perceptual neural network using the second loss function; when the second multi-layer perceptual neural network converges, querying the spatial coordinates of the encoded preliminary human tissue CBCT image through the optimized second multi-layer perceptual neural network to generate a final human tissue CBCT image.

[0132] The input of the intensity field is the spatial coordinate, and the output is the image intensity value corresponding to the coordinate. All voxel coordinates of the CBCT image are sent to the intensity field, and the intensity values ​​of all voxel points are queried to obtain the CBCT image.

[0133] During CBCT imaging, the X-ray source and the corresponding detector rotate around the scanned object. The penetrating rays emitted by the X-ray source pass through the scanned object and are projected onto the detection plane by integrating the intensities of intersecting voxels in the three-dimensional object. This process is called forward projection. Through forward projection, multiple projection images of the internal anatomical structure of the scanned object are obtained.

[0134] like Figure 5 It is a schematic diagram of the optimization process of the continuous intensity field in the present invention, and the target angle projection is compared with the actual projection by MSE mean square error and SSIM structural similarity to obtain the loss function, and the loss function is used to iteratively optimize the second multi-layer perception neural network.

[0135] The loss function of the continuous intensity field optimization process is as follows:

[0136] Loss2=MSE(pre_projs,train_projs)+SSIM(pre_projs,train_projs)

[0137] Among them, MSE is the mean square error, SSIM is the structural similarity, pre_projs represents the reconstructed image projection, and train_projs represents the target projection.

[0138] In the above technical scheme, the present invention constructs a two-stage reconstruction process, converts the image reconstruction task into an optimization problem of a continuous function in the preliminary reconstruction stage, and obtains discretized image space coordinates by uniformly grid sampling of the volume image after the neural network is optimized. The trained network is used to infer all image space coordinates in the intensity field to obtain the reconstructed image of the first stage. In the fine reconstruction stage, the volume image obtained in the previous stage is mapped to a neural intensity field, and the volume image is further refined using known projection data, so as to obtain a higher quality image imaging effect.

[0139] In order to further verify the image imaging quality, the present invention also compares the image imaging effects at different sparse angles, the imaging effects of the method of the present invention with other self-supervised methods and traditional methods, the imaging effects of the method of the present invention with other methods at 50 sparse angles, the imaging effects of the present invention on different human tissues, the imaging effects of the present invention in high-resolution scenes, and the imaging effects of ablation experiments.

[0140] This paper uses public datasets of human organ CT for method evaluation, including chest, mandible, and foot. The chest data comes from the LIDC-IDRI dataset, which is a public medical imaging dataset containing more than 1,000 chest CT scan cases from anonymous patients. The mandible and foot data come from Open Scientific Visualization Datasets. Figure 6 Shown are some examples from the medical imaging dataset.

[0141] The evaluation indexes of the present invention are mean absolute error (MAE), root mean square error (RMSE), structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR). MAE and RMSE are used to estimate the difference between the predicted image and the target image. The SSIM index is calculated using a window method in the image and is used to measure the overall similarity between the two images. The lower the MAE and RMSE values, the higher the SSIM value, indicating that the predicted image is closer to the real image. PSNR is the ratio of the maximum signal power to the noise power that affects the image quality, which is used to evaluate the quality of image imaging.

[0142] For experimental data, since only CT scan images are provided in these data sets, the present invention uses the tomography tool ODL to digitally generate sparse angle projection images in the range of 0° to 180°. These projections are used to optimize the model of the present invention, and its performance is evaluated using real CT scan image data. The specific CBCT scanning parameters are shown in the following table.

[0143] CBCT scanning parameters

[0144]

[0145] The method of the present invention is implemented in PyTorch, using the Adam optimizer, and the learning rate is gradually reduced from 1×10-3 to 1×10-4. All experiments are performed on a single NVIDIA Tesla V100 GPU. In the first stage of training, the batch size of each iteration is 2048 rays. The number of sampling points for each ray depends on the size of the CBCT data. When the CBCT slice image size is 128×128, the number of ray sampling points is 191, and when the CBCT slice image size is 256×256, the number of ray sampling points is 320. A total of 700 iterations are performed during the first stage of optimization. In the first 500 iterations, only the MSE loss value of the rendered (synthetic) projection value and the real projection value is used for optimization, and the geometric consistency loss constraint is added in the last 200 iterations. In the second stage of optimization, the total number of iterations of the coarse reconstruction image embedding process is 1000, and the total number of iterations of the intensity field parameter fine-tuning process is 1000, both of which use the geometric consistency loss constraint.

[0146] Comparative experiments under different sparse angles:

[0147] In order to explore the generalization of the method of the present invention to the degree of data sparsity, experiments were conducted at 20, 30, 40, and 50 projection angles. It should be noted that in the CBCT reconstruction task, 20 angles are already extremely sparse, usually 25 times less than the number of angles required clinically. The reconstruction results of the method of the present invention at different sparse angles are shown in the table below.

[0148] Reconstruction experimental results under different sparse angles

[0149]

[0150] In the table, PSNR and SSIM are the peak signal-to-noise ratio and structural similarity index between the reconstructed image and the real image, respectively. From the experimental results, it can be seen that good reconstruction results are achieved in 30-50 sparse angle scenes. Obviously, the more scanning angles there are, the better the quality of the reconstructed image. Figure 7 Shown is a comparison of image imaging effects at different sparse angles.

[0151] Comparison experiment of the imaging effect of the imaging method in the present invention with other self-supervision methods and traditional methods:

[0152] In order to verify the superiority of the two-stage sparse angle CBCT self-supervised reconstruction method based on NeRF, the current mainstream sparse angle CBCT self-supervised reconstruction method and traditional reconstruction methods were selected for comparative experiments, including FDK, SART, ASD-POCS, IntraTomo, and NAF. Among them, FDK is the representative of the traditional analytical method in CBCT reconstruction tasks, SART is a robust iterative method, and ASD-POCS is a widely used TV regularized iterative reconstruction method. IntraTomo and NAF are both self-supervised reconstruction methods based on neural radiation fields. In this embodiment, a CT scan image of a patient is randomly selected in the LIDC-IDRI dataset as a reference, the volume image size is preprocessed to 40×128×128, and the number of sparse angles is 50.

[0153] Comparative experimental results of the method of the present invention and the current mainstream method

[0154]

[0155] It can be seen from the results that the proposed method has achieved the best values ​​in all four numerical indicators and has a significant performance improvement compared with NAF, such as Figure 8The figure shows the visual effect comparison between the method of the present invention and the current mainstream method. The reconstruction result of FDK has significant stripe artifacts because the analytical reconstruction method requires a large number of projected images to achieve high-quality reconstruction. The iterative method SART suppresses noise at the expense of image detail features. The reconstruction result of ASD-POCS is severely blurred because the TV regularization method encourages the removal of high-frequency details, including unnecessary noise and subtle structures of the image. The reconstruction result of IntraTomo has no noise interference, but the edges are blurred, which indicates that the frequency encoder fails to guide the network to focus on edge features. The reconstruction result of NAF has missing pixels and the edges are also blurred. Compared with other methods, the method proposed in the present invention can restore the clearest edges and the least artifacts, and is closest to the real image.

[0156] Comparison of imaging effects between the method of the present invention and other methods at 50 sparse angles:

[0157] like Fig. 9 The figure shows the imaging result of the region of interest (ROI). It can be seen from the result that the method of the present invention recovers the most image information compared with other methods and well preserves the detailed features of the CBCT image.

[0158] Comparative experiment of imaging effects of different human tissues in the present invention:

[0159] In order to verify the generalization of the method proposed in this invention in different task scenarios, sparse reconstruction experiments were carried out at 50 angles on the mandible and foot respectively. The volumetric images of the mandible and foot have fewer detailed features and obvious bone areas, which are quite different from chest images. The experimental results are shown in the following table.

[0160] Reconstructive results in the jaw and foot

[0161]

[0162]

[0163] The experimental results show that the method introduced in this invention is suitable for reconstruction tasks of different parts. Compared with other comparative methods, this method has achieved the best reconstruction effect. In particular, in terms of the peak signal-to-noise ratio (PSNR), this method has shown a significant improvement, which further proves the superiority of the method of this invention. Fig.10 The imaging results of the mandible and foot of the present invention are shown. It can be clearly seen that the reconstruction results of the method of the present invention are extremely close to the real images, showing its excellent visual effect.

[0164] Comparative experiment of imaging effect of the present invention in high-resolution scene:

[0165] In order to verify the reconstruction performance of the proposed method under high-resolution tasks, FDK and NAF were selected as comparison methods. The experimental data were chest CT scan images in the LIDC-IDRI dataset. After preprocessing such as cropping and resizing, the size of the volume image was 80×256×256, and the number of sparse angles was 50. The experimental results are shown in the following table.

[0166] Reconstruction results of the method of the present invention in high-resolution tasks

[0167]

[0168] like Fig.11 The figure shows the comparison of the imaging of the method of the present invention and other methods under high-resolution tasks. It can be seen that when dealing with high-resolution reconstruction tasks, the method proposed by the present invention continues to maintain its superior performance, significantly better than NAF and other traditional reconstruction methods. In terms of structural similarity index (SSIM), the method of the present invention is 4.8 percentage points higher than NAF, and the peak signal-to-noise ratio (PSNR) is also significantly improved by 4.39, which is a significant improvement in performance. These data are consistent with Figure 3-11 This is consistent with the visualization results in , which further confirms the effectiveness and reliability of this method.

[0169] By analyzing the result graph in detail, it can be observed that, compared with the NAF and FDK methods, the image reconstructed by the method of the present invention is almost free from interference from artifacts and noise, showing the high accuracy of the image processing process. The detailed features in the image are extremely clear, and the contrast of the image is also maintained at an ideal level without a significant decrease. These characteristics illustrate the excellent performance of the method of the present invention in processing high-resolution images, and also highlight its advantages in maintaining image quality, especially in detail preservation and noise control. In general, these results fully verify the excellent performance and application potential of the method of the present invention in high-resolution reconstruction scenarios.

[0170] Ablation experiment of the imaging effect of the present invention:

[0171] This paper obtains better imaging effects through precise ray sampling, neural network optimization and two-stage reconstruction. In order to study the contribution of different methods to the imaging task, ablation experiments were conducted on the above parts using the same experimental data, sparse angles and CBCT imaging parameters. The experimental data is the chest CT scan image in the LIDC-IDRI dataset, the image size is preprocessed to 40×128×128, the sparse angle is 50, and NAF is selected as the baseline method.

[0172] Ablation experiment results of the proposed method on the LIDC-IDRI dataset

[0173]

[0174] The above table lists the results of the ablation experiment on the chest test data in LIDC-IDRI. The reconstructed image of the baseline method achieved the worst results in both PSNR and SSIM. Compared with the baseline, the addition of ray precise sampling significantly improved the peak signal-to-noise ratio and structural similarity index of the reconstructed image. Through the ray precise sampling module, the network can learn key visual features and capture effective scene details and depth information. In addition, the neural network optimization further improves the performance of the network. Finally, under the dual-stage reconstruction, the numerical indicators of the reconstruction results are further improved. The experimental results show that the method proposed in the present invention greatly improves the reconstruction quality of sparse angle CBCT images, showing the superiority of the method of the present invention.

[0175] A second embodiment of the present invention provides a human tissue imaging system, comprising:

[0176] The first data acquisition and processing module is used to sample the attenuation field of human tissue penetrated by the radiation to obtain the spatial coordinates of the human tissue attenuation field, and to position-code the spatial coordinates of the human tissue attenuation field to obtain the high-dimensional spatial coordinates of the encoded human tissue attenuation field;

[0177] A first image acquisition module is used to query the attenuation coefficient of the attenuation field of the human tissue through the first multilayer perceptron neural network to synthesize the projection, compare the synthesized projection with the actual projection by mean square error and similarity to obtain a first loss function, and iteratively optimize the first multilayer perceptron neural network using the first loss function; when the first multilayer perceptron neural network converges, query the spatial coordinates of the encoded human tissue attenuation field through the optimized first multilayer perceptron neural network to obtain a preliminary human tissue CBCT image;

[0178] The second data acquisition and processing module is used to embed the preliminary human tissue CBCT image into the intensity field, sample the spatial coordinates of the preliminary human tissue CBCT image, and position encode the spatial coordinates of the sampled preliminary human tissue CBCT image to obtain the spatial coordinates of the encoded preliminary human tissue CBCT image;

[0179] The second image acquisition module is used to query the intensity value of the intensity field through a second multi-layer perceptron neural network to obtain an intermediate human tissue CBCT image, perform forward projection of the intermediate human tissue CBCT image to obtain a target angle projection, compare the target angle projection with the actual projection in terms of mean square error and structural similarity to obtain a second loss function, and iteratively optimize the second multi-layer perceptual neural network using the second loss function; when the second multi-layer perceptual neural network converges, query the spatial coordinates of the encoded preliminary human tissue CBCT image through the optimized second multi-layer perceptual neural network to generate a final human tissue CBCT image.

[0180] A third embodiment of the present invention provides a computer-readable storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned human tissue imaging method is implemented.

[0181] A third embodiment of the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned human tissue imaging method when executing the program.

[0182] It should be noted that the specific embodiments described in the present invention are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the attached claims.

Claims

1. A method for imaging human tissue, characterized in that: include: Sampling the attenuation field of human tissue penetrated by the ray to obtain the spatial coordinates of the human tissue attenuation field, and position encoding the spatial coordinates of the human tissue attenuation field to obtain the high-dimensional spatial coordinates of the encoded human tissue attenuation field; The attenuation coefficient of the attenuation field of the human tissue is queried through the first multilayer perceptron neural network to synthesize the projection, the mean square error and similarity of the synthesized projection are compared with the actual projection to obtain a first loss function, and the first multilayer perceptron neural network is iteratively optimized using the first loss function; when the first multilayer perceptron neural network converges, the spatial coordinates of the encoded human tissue attenuation field are queried through the optimized first multilayer perceptron neural network to obtain a preliminary human tissue CBCT image; embedding the preliminary human tissue CBCT image into the intensity field, sampling the spatial coordinates of the preliminary human tissue CBCT image, and position encoding the spatial coordinates of the sampled preliminary human tissue CBCT image to obtain the spatial coordinates of the encoded preliminary human tissue CBCT image; The intensity value of the intensity field is queried through the second multi-layer perceptron neural network to obtain an intermediate human tissue CBCT image, the intermediate human tissue CBCT image is forward projected to obtain a target angle projection, the target angle projection is compared with the actual projection in terms of mean square error and structural similarity to obtain a second loss function, and the second multi-layer perceptual neural network is iteratively optimized using the second loss function; when the second multi-layer perceptual neural network converges, the spatial coordinates of the encoded preliminary human tissue CBCT image are queried through the optimized second multi-layer perceptual neural network to generate a final human tissue CBCT image.

2. The method for imaging human tissue according to claim 1, characterized in that: The sampling of the attenuation field of human tissue penetrated by the ray to obtain the spatial coordinates of the human tissue attenuation field specifically includes determining an X-ray path through the spatial coordinates p1 of the X-ray source and the spatial coordinates p2 of the detector pixel point. The parameters of the ray path are expressed as follows: X(α)=X1+α(X2-X1) Y(α)=Y1+α(Y2-Y1) Z(α)=Z1+α(Z2-Z1) The parameter α is 0 at point P1 and 1 at point P2. 1) is the spatial coordinate of the X-ray source, (X2, Y2, Z 2) is the spatial coordinate of the detector pixel; The orthogonal set of equally spaced parallel planes of the CBCT array for the (Nx, Ny, Nz) spatial coordinates is expressed as follows: X plane (i)=X plane (1)+(i-1)d x (i=1,...,N x ) Y plane (j)=Y plane (1)+(j-1)d y (j=1,...,N y ) Z plane (k)=Z plane (1)+(k-1)d z (k=1,…,N z ) Among them, d x ,d y and d z They are the distances between the x, y, and z planes, and are also the lengths of the spatial coordinates. The parameter value a min and a max Obtained by the intersection of the ray with the boundary of the CBCT array, the parameters corresponding to each edge of the spatial coordinate are expressed as follows: α x (1)=[X plane (1)-X1] / (X2-X1) α x (N x )=[X plane (N x )-X1] / (X2-X1) X2–X1≠0 α y (1) α y (N y ), α z (1) and α z (N z) The calculation method of is the same as above; For a given parameter value, α min and α max The calculation method is as follows: a min =max{0,min[α x (1),a x (N x )], min[a y (1),a y (N y )],min[a z (1),a z (N z )]} a max =min{1,max[α x (1),a x (N x )], max[a y (1),a y (N y )],max[a z (1),a z (N z )]} Among them, the function min and the function max are the minimum and maximum items in their parameter lists respectively. If α min Greater than or equal to α max , then the ray does not intersect with the CBCT image; The ray path is parameterized by α. min , α max Find the entry and exit points of the ray passing through the CBCT image: X in =X1+αmin(X2-X1) Y in =Y1+αmin(Y2-Y1) Z in =Z1+αmin(Z2-Z1) X out =X1+α max (X2-X1) Y out =Y1+α max (Y2-Y1) WITH out =Z1+α max (Z2-Z1) Among them, (X in , Y in , Z in ) is the spatial coordinate of the injection point, (X out , Y out , Z out ) is the spatial coordinate of the ejection point.

3. The method for imaging human tissue according to claim 1, characterized in that: The calculation formula of the synthetic projection is as follows: Among them, I is the intensity value of the ray after passing through the human tissue, I0 is the initial intensity, μ i is the attenuation coefficient of the sampling point, δ i is the ray sampling step size.

4. The method for imaging human tissue according to claim 1, wherein: The calculation formula of the first loss function is as follows: Loss1=α·MSE(I pre ,I truth )+β·SSIM(projs_pre,projs_truth) Among them, Loss1 is the first loss function, which is the projection value of the ray predicted by the attenuation field, I truth is the true projection value of the ray, projs_pre is the projection image generated by the forward projection module of the CBCT image obtained by querying the corresponding voxel in the attenuation field, and projs_truth is the true projection image; in the first 500 iterations, α=1, β=0, and in the next 200 iterations, α=β=0.

5.

5. The method for imaging human tissue according to claim 1, characterized in that: The calculation formula of the second loss function is as follows: Loss2=MSE(pre_projs,train_projs)+SSIM(pre_projs,train_projs) Among them, MSE is the mean square error, SSIM is the structural similarity, pre_projs represents the reconstructed image projection, and train_projs represents the target projection.

6. The method for imaging human tissue according to claim 4 or 5, characterized in that: The calculation formulas for the mean square error and structural similarity are: Among them, MSE is the mean square error, SSIM is the structural similarity, n is the total number of pixels in the image, μ is the average brightness of the image, σ is the variance of the image, and C1 and C2 are constants added to avoid the denominator being zero.

7. The method for imaging human tissue according to claim 1, wherein: The basis for the convergence of the first multilayer perceptron neural network is that the value of the first loss function is zero; the basis for the convergence of the second multilayer perceptron neural network is that the value of the second loss function is zero; Use a hash encoder to positionally encode spatial coordinates.

8. A human tissue imaging system, characterized in that: include: The first data acquisition and processing module is used to sample the attenuation field of human tissue penetrated by the radiation to obtain the spatial coordinates of the human tissue attenuation field, and to position-code the spatial coordinates of the human tissue attenuation field to obtain the high-dimensional spatial coordinates of the encoded human tissue attenuation field; A first image acquisition module is used to query the attenuation coefficient of the attenuation field of the human tissue through the first multilayer perceptron neural network to synthesize the projection, compare the synthesized projection with the actual projection by mean square error and similarity to obtain a first loss function, and iteratively optimize the first multilayer perceptron neural network using the first loss function; when the first multilayer perceptron neural network converges, query the spatial coordinates of the encoded human tissue attenuation field through the optimized first multilayer perceptron neural network to obtain a preliminary human tissue CBCT image; The second data acquisition and processing module is used to embed the preliminary human tissue CBCT image into the intensity field, sample the spatial coordinates of the preliminary human tissue CBCT image, and position encode the spatial coordinates of the sampled preliminary human tissue CBCT image to obtain the spatial coordinates of the encoded preliminary human tissue CBCT image; The second image acquisition module is used to query the intensity value of the intensity field through a second multi-layer perceptron neural network to obtain an intermediate human tissue CBCT image, perform forward projection of the intermediate human tissue CBCT image to obtain a target angle projection, compare the target angle projection with the actual projection in terms of mean square error and structural similarity to obtain a second loss function, and iteratively optimize the second multi-layer perceptual neural network using the second loss function; when the second multi-layer perceptual neural network converges, query the spatial coordinates of the encoded preliminary human tissue CBCT image through the optimized second multi-layer perceptual neural network to generate a final human tissue CBCT image.

9. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

10. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method described in any one of claims 1 to 7 is implemented.

Images