A sparse view reconstruction system and method based on neural attenuation field

Abstract

The present invention relates to a sparse view reconstruction system and method based on a neural attenuation field. On the basis of setting an X-ray tube, a collector, a line point sampling module, a hash coding module and a construction module are also set. The line point sampling module can extract multiple X-ray paths in each imaging image respectively through the global-local ray sampling method of the Sobel operator, and adaptively extract equidistant point columns on each X-ray path respectively to obtain the three-dimensional position information of each point, which helps to ensure a more efficient and higher quality computer tomography view reconstruction process under a sparse view. Combined with the hash coding module, the position information of each point obtained by the line point sampler can be three-dimensionally hash coded and two-dimensionally hash coded point by point, and weighted summation is performed to obtain the hash coding results corresponding to each point, which can overcome the image quality problems and technical defects of insufficient data information in the prior art.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a sparse view reconstruction system and method based on a neural attenuation field. Background Art

[0002] The images obtained by cone beam computed tomography (CBCT) can be reconstructed in three dimensions, which helps to observe and analyze the features of the observed objects or scenes from different angles. This technology is extremely beneficial in medical fields such as surgical planning, implant position determination, and orthodontic treatment design. Since the quality of images obtained by cone beam computed tomography is closely related to the radiation dose of X-rays, hundreds of views need to be captured to ensure the imaging quality, and excessive ionizing radiation from X-rays is harmful to the human body, reducing the projection views to reduce the radiation dose is indispensable in this technology.

[0003] In the existing technology, cone-beam computed tomography technology based on deep network models (such as convolutional neural network models and adversarial network models) can improve image quality in terms of details, but such models require a large number of projections for model training and rely on the three-dimensional shape of the object, which makes the model lack of generalization and cannot be generalized from one scene to another.

[0004] At the same time, cone-beam computed tomography based on neural radiance fields (NeRF) is an effective new view synthesis and scene reconstruction technology, which is widely applicable to many natural scene stereo reconstruction tasks. This technology can parameterize the attenuation coefficient field in the acquired image into an attenuation field to complete image reconstruction, and then obtain high-frequency details in the scene through position encoding to improve the accuracy of reconstruction. However, this technology has technical defects such as low sampling efficiency and limited data information. Summary of the invention

[0005] The technical problem to be solved by the present invention is how to overcome the technical defects of low sampling efficiency of images obtained by cone beam computed tomography in the prior art and lack of generalization of the model based on it. In order to overcome the above defects of the prior art, the present invention provides a sparse view reconstruction system and method based on neural attenuation field, including a sparse view reconstruction system based on neural attenuation field and a sparse view reconstruction method based on neural attenuation field.

[0006] The present invention provides a sparse view reconstruction system based on neural attenuation field, comprising:

[0007] An X-ray tube, used to emit a cone beam of X-rays to the object being measured;

[0008] A collector is configured to capture a plurality of imaging images of the object to be measured from a plurality of different angles by means of computer tomography;

[0009] The line point sampling module is configured to extract multiple X-ray paths in each of the imaging images by a global-local ray sampling method of a Sobel operator, and adaptively extract equidistant point arrays on each X-ray path to obtain three-dimensional position information of each point;

[0010] A hash coding module is configured to perform three-dimensional hash coding and two-dimensional hash coding on the three-dimensional position information of each point obtained by the line point sampling module point by point, and perform weighted summation to obtain a hash coding result corresponding to each point;

[0011] A construction module is configured to obtain predicted attenuation values ​​corresponding to each point using hash coding results corresponding to each point obtained by the hash coding module through a multi-layer perception algorithm, and then obtain a three-dimensional scanning image showing X-ray intensity distribution through a numerical integration method;

[0012] The line point sampling module communicates with the collector, the hash coding module communicates with the line point sampling module, and the construction module is electrically connected to the hash coding module.

[0013] The sparse view reconstruction system based on neural attenuation field disclosed in the present invention is provided with a collector, a line point sampling module, a hash coding module and a construction module on the basis of providing an X-ray tube. The line point sampling module can extract multiple X-ray paths in each imaging image respectively through the global-local ray sampling method of the Sobel operator, and adaptively extract equidistant point columns on each X-ray path respectively to obtain the three-dimensional position information of each point. The global-local ray sampling method of the Sobel operator can not only perform structurally aware X-ray sampling of pixel points in the imaging image according to global and local information, but also dynamically adjust the number of points sampled on each X-ray to better focus on the effective sampling area, thereby achieving sampling flexibility and efficiency, which helps to ensure a more efficient and higher quality computed tomography view reconstruction process under sparse views and a precise simulation of X-ray attenuation. On this basis, combined with the hash coding module, the position information of each point obtained by the line point sampling module can be 3D hash coded and 2D hash coded one by one, and weighted summation is performed to obtain the hash coding results corresponding to each point. The fusion of 3D hash coding and 2D hash coding results actually constitutes an Enhanced Tri-hash Representation (ETR) coding mechanism, which shows that richer context and spatial information are embedded in the input features. It not only considers the information of the 3D space point itself, but also pays attention to the mapping of the point on three orthogonal 2D planes. It can predict higher quality attenuation values ​​and achieve more accurate output, so that the output results of the hash coding module can extract more context information and spatial information, thereby overcoming the technical defects of image quality problems and insufficient data information in the existing technology and the technical defects of lack of generalization of the model. Finally, the construction module set up uses the hash coding results corresponding to each point obtained by the hash coding module with a multi-layer perception algorithm to obtain the predicted attenuation values ​​corresponding to each point, and then obtains a 3D scanning image showing the X-ray intensity distribution through numerical integration, and finally reproduces the pixel values ​​in the projection image, and improves the image sampling efficiency.

[0014] In a possible implementation, the line point sampling module includes:

[0015] A ray sampling module is configured to receive all the imaging images acquired by the collector, and then extract multiple X-ray paths in each of the imaging images respectively through a global-local ray sampling method of a Sobel operator, and respectively determine the range of each X-ray path in the imaging image to which it belongs;

[0016] A point sampling module is configured to adaptively extract at least three points at equal intervals on each X-ray path according to the range of each X-ray path in the imaging image to which it belongs, to obtain the equal-distance point sequence of the X-ray path, and obtain three-dimensional position information of each point;

[0017] The ray sampling module communicates with the collector, the point sampling module communicates with the ray sampling module, and the hash coding module communicates with the point sampling module;

[0018] This solution can combine local-global ray sampling based on the Sobel operator with adaptive equal-step point sampling by setting up point sampling modules and ray sampling modules. The former can sense the information of edge contours and perform structure-aware X-ray sampling based on global and local pixel information, while the latter can cleverly adjust the number of sampling points on each X-ray to adapt the sampling process to different scenarios, effectively solving the technical problems of insufficient point sampling, low point utilization, and blur or aliasing caused by ignoring corresponding areas.

[0019] In a possible implementation, the ray sampling module includes:

[0020] A horizontal encoder is configured to perform a convolution operation on each of the imaged images in a horizontal direction convolution operation mode to obtain a horizontal encoding result corresponding to each of the imaged images;

[0021] A vertical encoder is configured to perform convolution operations on each of the imaged images in a vertical direction convolution operation manner to obtain a vertical encoding result corresponding to each of the imaged images;

[0022] an integrator, configured to perform Sobel operator operations on the horizontal coding results and the vertical coding results corresponding to each of the imaging images, respectively, to obtain Sobel operation result images corresponding to each of the imaging images;

[0023] A generator is configured to binarize the Sobel operation result images corresponding to the imaging images according to a threshold value to obtain edge areas and non-edge areas corresponding to the imaging images;

[0024] A line sampler is configured to perform global pixel-level X-ray sampling on the non-edge areas corresponding to the imaging images to obtain a plurality of X-ray paths corresponding to the respective images, and to use the intersection of each X-ray path with the boundary line of the edge area as its endpoint;

[0025] A block sampler, configured to obtain batches and quantities of X-ray paths corresponding to all of the imaging images by local block sampling;

[0026] The horizontal encoder and the vertical encoder both communicate with the collector, the integrator communicates with the horizontal encoder and the vertical encoder respectively, the generator communicates with the integrator, the line sampler communicates with the generator, the block sampler communicates with the line sampler, and the point sampling module communicates with the block sampler;

[0027] This scheme sets up an integrator to perform Sobel operator operations on the horizontal and vertical coding results corresponding to each imaging image, and uses a generator to extract the effective edge area of ​​the image based on the Sobel operator operation results, thereby sampling only the area containing useful information and avoiding invalid sampling. This not only ensures the comprehensive capture of information, but also significantly reduces the calculation and time cost of sampling.

[0028] In a possible implementation, the point sampling module is configured to perform the following steps:

[0029] A1: taking the first X-ray path as the current X-ray path according to the batch of X-ray paths obtained by the block sampler, and then executing the next step;

[0030] A2: determining the distance between the two ends of the current X-ray path according to the endpoints of the X-ray path obtained by the line sampler, and determining the number of points of the current X-ray path according to the distance between the two ends, and the number of points is not less than three;

[0031] A3: extracting the endpoints at both ends of the current X-ray path and at least one point between the endpoints at both ends on the current X-ray path, and making the distances between two adjacent points equal and the number of extracted points equal to the number of points determined in step A2, to obtain the equidistant point sequence of the current path;

[0032] A4: determining the three-dimensional position information of each point in the equidistant point array of the current X-ray path according to the shooting position of the imaging image corresponding to the current X-ray path;

[0033] A5: determining whether the number of X-ray paths currently traversed is the number of X-ray paths determined by the block sampler;

[0034] If yes, the obtained three-dimensional position information of all points is transmitted to the hash coding module;

[0035] If not, the next X-ray path is used as the current X-ray path, and then the process returns to step A2;

[0036] The adaptive equal-step point sampling represented by this solution achieves the goal of flexibly adjusting the number of X-ray sampling points according to the actual length of the object traversed while maintaining the consistency of the sampling point spacing, so as to improve rendering efficiency and quality. This can optimize the performance of the neural attenuation field represented by the X-ray intensity distribution law output by the building module in different complex scenarios and improve the generalization ability of the system.

[0037] In a possible implementation, the hash coding module includes:

[0038] A three-dimensional hash encoder is configured to perform three-dimensional hash coding on the three-dimensional position information of all points obtained by the point sampling module point by point to obtain a three-dimensional hash coding result corresponding to each point;

[0039] The first two-dimensional hash encoder is configured to perform orthogonal projection on the horizontal and vertical coordinate planes one by one on the three-dimensional position information of all points obtained by the point sampling module to obtain first orthogonal projection coordinates corresponding to each point, and then perform two-dimensional hash encoding on the first orthogonal projection coordinates corresponding to each point to obtain first two-dimensional hash encoding results corresponding to each point;

[0040] The second two-dimensional hash encoder is configured to perform orthogonal projection on the horizontal and vertical coordinate planes one by one on the three-dimensional position information of all points obtained by the point sampling module to obtain second orthogonal projection coordinates corresponding to each point, and then perform two-dimensional hash encoding on the second orthogonal projection coordinates corresponding to each point to obtain second two-dimensional hash encoding results corresponding to each point;

[0041] The third two-dimensional hash encoder is configured to perform orthogonal projection on the vertical and horizontal coordinate planes one by one on the three-dimensional position information of all points obtained by the point sampling module to obtain third orthogonal projection coordinates corresponding to each point, and then perform two-dimensional hash encoding on the third orthogonal projection coordinates corresponding to each point to obtain third two-dimensional hash encoding results corresponding to each point;

[0042] A fuser is configured to perform weighted summation on the three-dimensional hash coding result, the first two-dimensional hash coding result, the second two-dimensional hash coding result, and the third two-dimensional hash coding result corresponding to each point to obtain the hash coding result corresponding to each point;

[0043] The three-dimensional hash encoder, the first two-dimensional hash encoder, the second two-dimensional hash encoder, and the third two-dimensional hash encoder all communicate with the point sampling module, the fusion device communicates with the three-dimensional hash encoder, the first two-dimensional hash encoder, the second two-dimensional hash encoder, and the third two-dimensional hash encoder respectively, and the construction module communicates with the fusion device;

[0044] This scheme decomposes the three-dimensional space into three mutually orthogonal two-dimensional planes based on the three-dimensional position information of all points obtained by the point sampling module, so as to perform in-depth processing and expansion of the three-dimensional position information of these points, enrich the information of point features, and finally obtain a feature representation that can more comprehensively and accurately describe the scene.

[0045] In a possible implementation, the three-dimensional hash encoder is configured to use the following formula for the hash function used when performing hash encoding:

[0046] , (1)

[0047] The first two-dimensional hash encoder is set to use the following hash function when performing hash encoding:

[0048] , (2)

[0049] The second two-dimensional hash encoder is set to use the following hash function when performing hash encoding:

[0050] , (3)

[0051] The third two-dimensional hash encoder is set to use the following hash function when performing hash encoding:

[0052] , (4)

[0053] Where:

[0054] V represents XOR operation;

[0055] represents the remainder operation;

[0056] Represents the vertices of the voxel grid where the point to be hashed is located;

[0057] Represents the size of the feature lookup table;

[0058] ;

[0059] ;

[0060] ;

[0061] Using this type of hash function can not only reduce computational complexity and save energy, but also ensure efficient operation of hash coding, further enriching the feature representation content describing the scene.

[0062] In a possible implementation, the building blocks include:

[0063] A multi-layer perceptron is configured to obtain the predicted attenuation value corresponding to each point by using the hash coding result corresponding to each point obtained by the fusion device through a multi-layer perceptron algorithm;

[0064] an integrator, configured to obtain predicted attenuation value distribution information according to all predicted attenuation values ​​obtained by the multilayer perceptron, and to obtain the three-dimensional scanned image according to the predicted attenuation value distribution information by numerical integration;

[0065] The multilayer perceptron communicates with the fusion device, and the integrator communicates with the multilayer perceptron;

[0066] The three-dimensional scanning image obtained by this scheme focuses on the radiation density captured by the object under test when X-rays penetrate rather than the color information. It can reflect the degree of absorption or attenuation of X-rays or other forms of ionizing radiation by the object under test, that is, the "attenuation value", thereby achieving the purpose of three-dimensional reconstruction of sparse views by obtaining an image that describes the attenuation relationship of the intensity when light passes through an absorbing medium through a sparse image.

[0067] In a possible implementation, the multilayer perceptron includes an input end, a first fully connected layer unit, a second fully connected layer unit, a third fully connected layer unit, and an activation unit;

[0068] The first fully connected layer unit communicates with the input end, the second fully connected layer unit communicates with the first fully connected layer unit, the third fully connected layer unit communicates with the input end and the second fully connected layer unit respectively, and the activation unit communicates with the third fully connected layer unit;

[0069] The input end is configured to receive the hash coding result obtained by the fuser and output it point by point;

[0070] The activation unit is configured to perform an activation operation of a Sigmoid function on an output result of the third fully connected layer unit to obtain a predicted attenuation value;

[0071] This solution can achieve the prediction results reflecting the absorption or attenuation degree of the object under test to X-rays or other forms of ionizing radiation with less computing power, thereby ensuring the efficient operation of three-dimensional reconstruction of sparse views.

[0072] Another technical solution of the present invention is to provide a sparse view reconstruction method based on neural attenuation field, comprising the following steps:

[0073] S1: Optimizing parameters of the building module by using a square error loss function and an absolute error loss function to obtain the optimized building module;

[0074] S2: emits cone beam X-rays to the object to be measured through the X-ray tube;

[0075] S3: using a collector to capture a plurality of imaging images of the object to be measured from a plurality of different angles in a computer tomography manner;

[0076] S4: extracting a plurality of X-ray paths from each of the imaging images by a line point sampling module in a global-local ray sampling manner of a Sobel operator, and adaptively extracting equidistant point sequences on each X-ray path to obtain three-dimensional position information of each point;

[0077] S5: performing three-dimensional hash coding and two-dimensional hash coding on the position information of each point obtained by the line point sampling module point by point through a hash coding module, and performing weighted summation to obtain a hash coding result corresponding to each point;

[0078] S6: The predicted attenuation value corresponding to each point is obtained by using the hash coding result corresponding to each point obtained by the hash coding module through the optimized construction module with a multi-layer perception algorithm, and then a three-dimensional scanning image showing the X-ray intensity distribution is obtained through numerical integration.

[0079] The method disclosed in the present application first optimizes the parameters of the construction module through the square error loss function and the absolute error loss function to ensure that the error between the three-dimensional image constructed by the construction module and the reality can be reduced and controlled. Then, after collecting several imaging images of the object under test about X-rays, a line point sampling module is used to extract multiple X-ray paths in each imaging image in a global-local ray sampling method of the Sobel operator, and an equidistant point column is adaptively extracted on each X-ray path to obtain the three-dimensional position information of each point. This process can not only perform structure-aware X-ray sampling of pixel points in the imaging image according to global and local information, but also dynamically adjust the number of points sampled on each X-ray to better focus on the effective sampling area, thereby achieving sampling flexibility and efficiency, which helps to ensure a more efficient and higher-quality computer tomography view reconstruction process under sparse views and a precise simulation of X-ray attenuation. Then, the hash coding module performs three-dimensional hash coding and two-dimensional hash coding on the position information of each point obtained by the line point sampling module one by one, and performs weighted summation to obtain the hash coding results corresponding to each point. The hash coding results obtained by fusion of the three-dimensional hash coding and the two-dimensional hash coding results can extract more contextual information and spatial information, thereby overcoming the technical defects of image quality problems and insufficient data information in the prior art and the technical defects of lack of generalization of the model. Finally, the hash coding results corresponding to each point obtained by the hash coding module are used by the optimized building module with a multi-layer perception algorithm to obtain the predicted attenuation values ​​corresponding to each point, and then a three-dimensional scanning image showing the X-ray intensity distribution is obtained by numerical integration, and finally the pixel values ​​in the projection image are reproduced, and the image sampling efficiency is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 Schematic diagram of the structure of a sparse view reconstruction system based on neural attenuation field disclosed in an embodiment of the present application;

[0081] Figure 2 It is a schematic diagram of the structure of the ray sampling module disclosed in the embodiment of the present application;

[0082] Figure 3 It is an operation flow chart of the point sampling module disclosed in the embodiment of the present application;

[0083] Figure 4 This is a schematic diagram of the structure of the hash coding module disclosed in the embodiments of the present application;

[0084] Figure 5 This is a schematic diagram of the operating principle of the hash coding module disclosed in the embodiments of the present application;

[0085] Figure 6 A schematic diagram of the structure of a multilayer perceptron disclosed in an embodiment of the present application;

[0086] Figure 7 It is a flow chart of the method disclosed in the embodiments of the present application;

[0087] Figure 8 This is a schematic diagram of the process principle of the method disclosed in the embodiments of this application.

[0088] Description of reference numerals:

[0089] 1. X-ray tube, 2. Collector, 3. Line point sampling module, 31. Ray sampling module, 311. Horizontal encoder, 312. Vertical encoder, 313. Integrator, 314. Generator, 315. Line sampler, 316. Block sampler, 32. Point sampling module, 4. Hash coding module, 41. Three-dimensional hash encoder, 42. First two-dimensional hash encoder, 43. Second two-dimensional hash encoder, 44. Third two-dimensional hash encoder, 45. Fusion, 5. Building module, 51. Multilayer perceptron, 511. Input end, 512. First fully connected layer unit, 513. Second fully connected layer unit, 514. Third fully connected layer unit, 515. Activation unit, 52. Integrator. DETAILED DESCRIPTION

[0090] First, those skilled in the art should understand that these implementations are only used to explain the technical principles of the embodiments of the present application, and are not intended to limit the protection scope of the embodiments of the present application. Those skilled in the art can make adjustments to them as needed to adapt to specific application scenarios.

[0091] In the embodiments of the present application, unless otherwise clearly specified and limited, the communication or communication connection between the first feature and the second feature refers to the transmission of information between the first feature and the second feature. This information transmission can be either unidirectional or bidirectional, and the communication connection can be achieved by electrical connection of wires, radio connection, electrical connection of electromagnetic media (such as semiconductors), communication achieved by channels, etc.

[0092] In the embodiments of the present application, unless otherwise specified and limited, the distance between two points should be understood according to the mathematical meaning of distance, that is, the so-called point With point Distance It refers to the mapping of any two points in the set in question (such as any two points in the image in this embodiment or any two points on the X-ray path) to the set of real numbers. , and At the same time, the following three requirements must be met: (1) , if and only if When the equality sign holds; (2) ; (3) Specifically, it can be the Euclidean distance, or the distance induced by the inner product or the norm, and those skilled in the art can adjust it according to specific needs. In addition, all bold letters represent images.

[0093] In the embodiments of the present application, unless otherwise clearly specified and limited, a first feature being "above" or "below" a second feature may mean that the first and second features are in direct contact, or the first and second features are in indirect contact through an intermediate medium. Moreover, a first feature being "above", "above" or "above" a second feature may mean that the first feature is directly above or obliquely above the second feature, or simply means that the first feature is higher in level than the second feature. A first feature being "below", "below" or "below" a second feature may mean that the first feature is directly below or obliquely below the second feature, or simply means that the first feature is lower in level than the second feature.

[0094] The present application is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0095] See also Figure 1 to Figure 6 As shown, the embodiment of the present application discloses a sparse view reconstruction system based on neural attenuation field, and the structural schematic diagram of the system is shown in Figure 1 As shown, the system includes an X-ray tube 1, a collector 2, a line point sampling module 3, a hash coding module 4 and a construction module 5, wherein the line point sampling module 3 communicates with the collector 2, the hash coding module 4 communicates with the line point sampling module 3, and the construction module 5 is electrically connected to the hash coding module 4.

[0096] See also Figure 1 As shown, in the system, the X-ray tube 1 is used to emit a cone beam X-ray to the object to be measured; the collector 2 is configured to capture a plurality of imaging images of the object to be measured from a plurality of different angles by means of computer tomography; the line point sampling module 3 is configured to extract a plurality of X-ray paths in each imaging image by means of a global-local ray sampling method of a Sobel operator, and adaptively extract equidistant point sequences on each X-ray path to obtain the three-dimensional position information of each point; the hash coding module 4 is configured to perform three-dimensional hash coding and two-dimensional hash coding on the three-dimensional position information of each point obtained by the line point sampling module 3 point by point, and perform weighted summation to obtain the hash coding results corresponding to each point; the construction module 5 is configured to obtain the predicted attenuation values ​​corresponding to each point by means of a multi-layer perception algorithm using the hash coding results corresponding to each point obtained by the hash coding module 4, and then obtain the three-dimensional scanning image showing the X-ray intensity distribution by means of numerical integration.

[0097] Please continue to see Figure 1As shown, in this embodiment, the line point sampling module 3 includes a ray sampling module 31 and a point sampling module 32 , the ray sampling module 31 communicates with the collector 2 , the point sampling module 32 communicates with the ray sampling module 31 , and the hash coding module 4 communicates with the point sampling module 32 .

[0098] In the online point sampling module 3, the ray sampling module 31 is configured to receive all imaging images obtained by the collector 2, and then extract multiple X-ray paths in each imaging image respectively through the global-local ray sampling method of the Sobel operator, and respectively determine the range of each X-ray path in the imaging image to which it belongs; the point sampling module 32 is configured to equidistantly and adaptively extract no less than three points on this X-ray path according to the range of each X-ray path in the imaging image to which it belongs to obtain an equidistant point series of this X-ray path, and obtain the three-dimensional position information of each point.

[0099] According to the terminology used in this field, the so-called equidistant point series refers to an ordered set of no less than three points. , and the distance between two adjacent points is a constant ,Right now

[0100] ,

[0101] In the formula, .

[0102] See also Figure 2 As shown, in this embodiment, the ray sampling module 31 includes a horizontal encoder 311, a vertical encoder 312, an integrator 313, a generator 314, a line sampler 315 and a block sampler 316, wherein the horizontal encoder 311 and the vertical encoder 312 both communicate with the collector 2, the integrator 313 communicates with the horizontal encoder 311 and the vertical encoder 312 respectively, the generator 314 communicates with the integrator 313, the line sampler 315 communicates with the generator 314, the block sampler 316 communicates with the line sampler 315, and the point sampling module 32 communicates with the block sampler 316.

[0103] In the ray sampling module 31, the horizontal encoder 311 is configured to perform convolution operations on each imaged image in a horizontal direction convolution operation mode to obtain horizontal encoding results corresponding to each imaged image. The convolution kernel size of the convolution unit used by the horizontal encoder 311 in this embodiment is 3×3. The step size is 1. ,

[0104] ,

[0105] Among them represents the pixel intensity, , , let the convolution kernel used by the horizontal encoder 311 be ,

[0106] ,

[0107] In the specific design of this embodiment, the convolution kernel used here is:

[0108] .

[0109] Imaging images After a horizontal convolution operation, the feature image is obtained ,

[0110] ,

[0111] ,

[0112] In the formula, , After several convolution operations, the image can be obtained. The corresponding horizontal encoding result. Of course, those skilled in the art will In the process of corresponding horizontal encoding results, a pooling unit can be set based on the convolution unit to perform a pooling operation for sampling as needed.

[0113] In the ray sampling module 31, the vertical encoder 312 is configured to perform convolution operations on each imaged image in a vertical direction convolution operation mode to obtain vertical encoding results corresponding to each imaged image. The convolution kernel size of the convolution unit used by the vertical encoder 312 in this embodiment is 3×3 and the step size is 1. Let the convolution kernel of the convolution unit used by the vertical encoder 312 be ,

[0114] ,

[0115] In the specific design of this embodiment, the convolution kernel used here is:

[0116] .

[0117] For imaging images After a horizontal convolution operation, the feature image is obtained ,

[0118] ,

[0119] ,

[0120] After several convolution operations, the imaging image can be obtained. The corresponding vertical coding result. Of course, those skilled in the art will In the process of obtaining the vertical encoding result, a pooling unit can be set based on the convolution unit to perform a pooling operation for sampling as needed, but it should be noted that the result must have the same image size as the result obtained by the horizontal encoder 311.

[0121] In the ray sampling module 31, the integrator 313 is configured to perform Sobel operator operations on the horizontal encoding results and the vertical encoding results corresponding to each imaging image, respectively, to obtain the Sobel operation result images corresponding to each imaging image. The corresponding horizontal encoding result is the feature image , the corresponding vertical encoding result is the feature image , then the Sobel operation result image obtained by performing the Sobel operator operation is ,

[0122] .

[0123] In the ray sampling module 31, the generator 314 is configured to binarize the Sobel operation result images corresponding to each imaging image according to the threshold value to obtain the edge area and non-edge area corresponding to each imaging image. The line sampler 315 is configured to perform global pixel-level X-ray sampling on the non-edge area corresponding to each imaging image to obtain multiple X-ray paths corresponding to each image, and use the intersection of each X-ray path with the boundary line of the edge area as its endpoint. The block sampler 316 is configured to obtain the batches and quantities of X-ray paths corresponding to all imaging images by local block sampling.

[0124] See also Figure 3 As shown, in this embodiment, the point sampling module 32 is configured to perform the following steps:

[0125] A1: According to the batch of X-ray paths obtained by the block sampler 316, the first X-ray path is used as the current X-ray path, and then the next step is executed.

[0126] A2: Determine the distance between the two ends of the current X-ray path based on the endpoints of the X-ray path obtained by the line sampler 315, and determine the number of points of the current X-ray path based on the distance between the two ends, and the number of points is not less than three.

[0127] A3: Extract the endpoints of the two ends of the current X-ray path and at least one point between the endpoints at the two ends on the current X-ray path, and make the distance between two adjacent points equal and the number of extracted points is the number of points determined in step A2, so as to obtain an equidistant point sequence of the current path.

[0128] A4: Determine the three-dimensional position information of each point in the equidistant point sequence of the current X-ray path according to the shooting position of the imaging image corresponding to the current X-ray path.

[0129] A5: Determine whether the number of X-ray paths currently traversed is the number of X-ray paths determined by the block sampler 316;

[0130] If yes, the obtained three-dimensional position information of all points is transmitted to the hash coding module 4;

[0131] If not, the next X-ray path is used as the current X-ray path, and the process returns to step A2.

[0132] See also Figure 4 and Figure 5 As shown, in this embodiment, the hash coding module 4 includes a three-dimensional hash encoder 41, a first two-dimensional hash encoder 42, a second two-dimensional hash encoder 43, a third two-dimensional hash encoder 44 and a fusion device 45, wherein the three-dimensional hash encoder 41, the first two-dimensional hash encoder 42, the second two-dimensional hash encoder 43, and the third two-dimensional hash encoder 44 all communicate with the point sampling module 32, the fusion device 45 communicates with the three-dimensional hash encoder 41, the first two-dimensional hash encoder 42, the second two-dimensional hash encoder 43, and the third two-dimensional hash encoder 44 respectively, and the construction module 5 communicates with the fusion device 45.

[0133] See also Figure 5 As shown, in the hash coding module 4, the three-dimensional hash encoder 41 is configured to perform three-dimensional hash coding on the three-dimensional position information of all points obtained by the point sampling module 32 point by point, and obtain the three-dimensional hash coding result corresponding to each point. The three-dimensional hash encoder 41 is configured to use the following formula for the hash function used when performing hash coding:

[0134] , (1)

[0135] Where:

[0136] V represents XOR operation;

[0137] represents the remainder operation;

[0138] Represents the vertices of the voxel grid where the point to be hashed is located, represents the horizontal coordinate of the point, represents the vertical coordinate of the point, Represents the vertical coordinate of the point;

[0139] Represents the size of the feature lookup table;

[0140] ;

[0141] ;

[0142] .

[0143] In the specific implementation process, the three-dimensional hash encoder 41 provided in this embodiment is configured to perform the following steps:

[0144] In the first step, the three-dimensional position information of all points obtained by the point sampling module 32 is used to determine the common area of ​​these points, and then the area is divided into a plurality of square grids, which are also called voxel grids.

[0145] The second step is to determine the coordinates of the eight vertices of the voxel grid where these points are located, and use formula (1) to obtain the hash function values ​​of the eight vertices, thereby obtaining the first hash transformation results corresponding to each point.

[0146] The third step is to use the first hash transformation result corresponding to each point to find the corresponding hash feature in the established hash table to obtain the first hash feature corresponding to each point.

[0147] Step 4: Linearly interpolate the first hash features corresponding to each point to obtain the features at multiple levels corresponding to each point. The features are the three-dimensional hash coding results corresponding to the corresponding points. In this embodiment, each point corresponds to features at 16 levels, and the length of the feature vector at each level is 2. Therefore, the three-dimensional hash coding result corresponding to each point is a feature vector with a length of 32. For point F, the corresponding three-dimensional hash coding result is recorded as .

[0148] Please continue to see Figure 5 As shown, in the hash coding module 4, the first two-dimensional hash encoder 42 is configured to orthogonally project the three-dimensional position information of all points obtained by the point sampling module 32 on the horizontal and vertical coordinate planes (i.e., the XOY plane) point by point, obtain the first orthogonal projection coordinates corresponding to each point, and then perform two-dimensional hash coding on the first orthogonal projection coordinates corresponding to each point, respectively, to obtain the first two-dimensional hash coding results corresponding to each point. The first two-dimensional hash encoder 42 is configured to use the following formula for the hash function used when performing hash coding:

[0149] ; (2)

[0150] In the specific implementation process, the first two-dimensional hash encoder 42 provided in this embodiment is configured to perform the following steps:

[0151] In the first step, the three-dimensional position information of all points obtained by the point sampling module 32 is used to determine the common area of ​​these points, and then the area is divided into a plurality of voxel grids.

[0152] The second step is to determine the horizontal and vertical coordinates of the eight vertices of the voxel grid where these points are located, and use formula (2) to obtain the hash function values ​​of the eight vertices, thereby obtaining the second hash transformation results corresponding to each point.

[0153] In the specific implementation process, each of the eight vertices has another point with the same horizontal and vertical coordinates, so only four points need to be taken out, but these four points must satisfy the condition that any two points have different horizontal coordinates, different vertical coordinates, or different horizontal and vertical coordinates.

[0154] The third step is to use the second hash transformation result corresponding to each point to find the corresponding hash feature in the established hash table to obtain the second hash feature corresponding to each point.

[0155] Step 4: Linearly interpolate the second hash features corresponding to each point to obtain the features at multiple levels, which are the first two-dimensional hash coding results corresponding to the corresponding points. Matching the three-dimensional hash coding results, each point corresponds to features at 16 levels, and the length of the feature vector at each level is 2. Therefore, the first two-dimensional hash coding result corresponding to each point is a feature vector with a length of 32. For point F, its corresponding first two-dimensional hash coding result is recorded as .

[0156] Please continue to see Figure 5 As shown, in the hash coding module 4, the second two-dimensional hash encoder 43 is configured to orthogonally project the three-dimensional position information of all points obtained by the point sampling module 32 on the horizontal and vertical coordinate planes (i.e., the XOZ plane) point by point, obtain the second orthogonal projection coordinates corresponding to each point, and then perform two-dimensional hash coding on the second orthogonal projection coordinates corresponding to each point, respectively, to obtain the second two-dimensional hash coding results corresponding to each point. The second two-dimensional hash encoder 43 is configured to use the following formula for the hash function used when performing hash coding:

[0157] ; (3)

[0158] In the specific implementation process, the second two-dimensional hash encoder 43 provided in this embodiment is configured to perform the following steps:

[0159] In the first step, the three-dimensional position information of all points obtained by the point sampling module 32 is used to determine the common area of ​​these points, and then the area is divided into a plurality of voxel grids.

[0160] The second step is to determine the horizontal and vertical coordinates of the eight vertices of the voxel grid where these points are located, and use formula (3) to obtain the hash function values ​​of the eight vertices, thereby obtaining the third hash transformation results corresponding to each point.

[0161] In the specific implementation process, each of the eight vertices has another point with the same horizontal and vertical coordinates, so only four points need to be taken out, but these four points must satisfy the condition that any two points have different horizontal coordinates, different vertical coordinates, or different horizontal and vertical coordinates.

[0162] The third step is to use the third hash transformation result corresponding to each point to find the corresponding hash feature in the established hash table to obtain the third hash feature corresponding to each point.

[0163] Step 4: Linearly interpolate the third hash features corresponding to each point to obtain the features at multiple levels, which are the second two-dimensional hash coding results corresponding to the corresponding points. Similarly, matching the three-dimensional hash coding results, each point corresponds to features at 16 levels, and the feature vector length at each level is 2, so the second two-dimensional hash coding result corresponding to each point is a feature vector with a length of 32. For point F, its corresponding second two-dimensional hash coding result is recorded as .

[0164] Please continue to see Figure 5 As shown, in the hash coding module 4, the third two-dimensional hash encoder 44 is configured to orthogonally project the three-dimensional position information of all points obtained by the point sampling module 32 on the vertical and horizontal coordinate plane (i.e., the YOZ plane) point by point, obtain the third orthogonal projection coordinates corresponding to each point, and then perform two-dimensional hash coding on the third orthogonal projection coordinates corresponding to each point, respectively, to obtain the third two-dimensional hash coding result corresponding to each point. The third two-dimensional hash encoder 44 is configured to use the following formula for the hash function used when performing hash coding:

[0165] ; (4)

[0166] In the specific implementation process, the third two-dimensional hash encoder 44 provided in this embodiment is configured to perform the following steps:

[0167] In the first step, the three-dimensional position information of all points obtained by the point sampling module 32 is used to determine the common area of ​​these points, and then the area is divided into a plurality of voxel grids.

[0168] The second step is to determine the ordinates and vertical coordinates of the eight vertices of the voxel grid where these points are located, and use formula (4) to obtain the hash function values ​​of the eight vertices, thereby obtaining the fourth hash transformation results corresponding to each point.

[0169] In the specific implementation process, each of the eight vertices has another point with the same ordinate and ordinate, so only four points need to be taken out, but these four points must satisfy the condition that any two points have different ordinates, different ordinates, or different ordinates and ordinates.

[0170] The third step is to use the fourth hash transformation result corresponding to each point to find the corresponding hash feature in the established hash table to obtain the fourth hash feature corresponding to each point.

[0171] Step 4: Linearly interpolate the fourth hash features corresponding to each point to obtain the features at multiple levels, which are the third two-dimensional hash coding results corresponding to the corresponding points. Similarly, matching the three-dimensional hash coding results, each point corresponds to features at 16 levels, and the feature vector length at each level is 2, so the third two-dimensional hash coding result corresponding to each point is a feature vector with a length of 32. For point F, its corresponding second two-dimensional hash coding result is recorded as .

[0172] Please continue to see Figure 5 As shown, in the hash coding module 4, the fuser 45 is configured to perform weighted summation of the three-dimensional hash coding result, the first two-dimensional hash coding result, the second two-dimensional hash coding result, and the third two-dimensional hash coding result corresponding to each point to obtain the hash coding result corresponding to each point. Specifically, for point F, the corresponding hash coding result is:

[0173] ,

[0174] ,

[0175] In the formula,

[0176] The hash encoding result corresponding to the representative point F;

[0177] , , and are all non-negative real numbers.

[0178] See also Figure 1 and Figure 6As shown, in this embodiment, the building module 5 includes a multilayer perceptron 51 and an integrator 52, wherein the multilayer perceptron 51 communicates with the fusion unit 45, and the integrator 52 communicates with the multilayer perceptron 51. In the building module 5, the multilayer perceptron 51 is configured to obtain the predicted attenuation value corresponding to each point by using the hash coding result corresponding to each point obtained by the fusion unit 45 through the multilayer perception algorithm. Figure 6 As shown, in this embodiment, the multilayer perceptron 51 includes an input terminal 511, a first fully connected layer unit 512, a second fully connected layer unit 513, a third fully connected layer unit 514 and an activation unit 515; the first fully connected layer unit 512 communicates with the input terminal 511, the second fully connected layer unit 513 communicates with the first fully connected layer unit 512, the third fully connected layer unit 514 communicates with the input terminal 511 and the second fully connected layer unit 513 respectively, and the activation unit 515 communicates with the third fully connected layer unit 514; the input terminal 511 is configured to receive the hash coding result obtained by the fusion device 45, and output it point by point; the activation unit 515 is configured to perform a Sigmoid function activation operation on the output result of the third fully connected layer unit 514 to obtain a predicted attenuation value. In addition, the activation function used by the first fully connected layer unit 512, the second fully connected layer unit 513, the third fully connected layer unit 514 and the activation unit 515 is a Relu activation function.

[0179] In the building block 5, the integrator 52 is configured to obtain predicted attenuation value distribution information based on all predicted attenuation values ​​obtained by the multilayer perceptron 51, and obtain a three-dimensional scanning image based on the predicted attenuation value distribution information by numerical integration. The integral formula used in this embodiment is:

[0180] ,

[0181] In the formula,

[0182] Represents the X-ray intensity value corresponding to the point on a certain X-ray path;

[0183] represents the initial intensity value of the X-ray emitted by the X-ray tube 1;

[0184] represents the predicted attenuation value of the th point located on a certain X-ray path obtained by the multilayer perceptron 51;

[0185] Represents the first Point and The distance of points.

[0186] The following will further introduce the use method of the sparse view reconstruction system based on neural attenuation field disclosed by this embodiment. Figure 7 and Figure 8 As shown, the method comprises the following steps:

[0187] S1: Parameter optimization is performed on the building block 5 by using the square error loss function and the absolute error loss function to obtain the optimized building block 5. The specific optimization object may be only the multilayer perceptron 51, and the weight parameters and hyperparameters of the multilayer perceptron 51 are determined by parameter optimization.

[0188] S2: A cone beam X-ray is emitted toward the object to be measured through the X-ray tube 1.

[0189] S3: Using the collector 2 to acquire a plurality of imaging images of the object to be measured from a plurality of different angles in a computer tomography manner.

[0190] S4: extracting multiple X-ray paths in each imaging image by the line point sampling module 3 using the global-local ray sampling method of the Sobel operator, and adaptively extracting equidistant point sequences on each X-ray path to obtain the three-dimensional position information of each point.

[0191] S5: The hash coding module 4 performs three-dimensional hash coding and two-dimensional hash coding on the position information of each point obtained by the line point sampling module 3 point by point, and performs weighted summation to obtain the hash coding result corresponding to each point.

[0192] S6: The predicted attenuation value corresponding to each point is obtained by using the hash coding result corresponding to each point obtained by the hash coding module 4 through the optimized construction module 5 with a multi-layer perception algorithm, and then a three-dimensional scanning image showing the X-ray intensity distribution is obtained through numerical integration.

[0193] Subsequently, these three-dimensional scanned images can be orthogonally projected onto different surfaces to obtain multiple two-dimensional images. Using these two-dimensional images, combined with the square error loss function and the absolute error loss function, the parameters of the construction module 5 can be optimized again, so that the prediction results of the construction module 5 are further accurate.

[0194] In order to verify the effectiveness of the sparse view reconstruction system based on neural attenuation field disclosed in this embodiment (in order to reduce the occupied space in the table, THP-NAF is used to represent the sparse view reconstruction system based on neural attenuation field disclosed in this embodiment) in practical applications, this embodiment uses six data sets (including four public data sets and two private data sets) to compare THP-NAF with various popular systems. The analysis results are shown in Table 1. Among them, the bold content represents the optimal result, and the underlined content represents the suboptimal result.

[0195] Table 1 Peak signal-to-noise ratio / structural similarity measures of seven systems on six datasets

[0196]

[0197] The results in Table 1 show that THP-NAF has superiority on four public datasets and two private datasets. In addition, after verification, THP-NAF can complete the CT model reconstruction of 256×256 resolution projection images in about 20 minutes.

[0198] In summary, the sparse view reconstruction system based on the neural attenuation field disclosed in this embodiment, on the basis of setting the X-ray tube 1, also sets a collector 2, a line point sampling module 3, a hash coding module 4 and a construction module 5. The line point sampling module 3 can extract multiple X-ray paths in each imaging image respectively through the global-local ray sampling method of the Sobel operator, and adaptively extract equidistant point columns on each X-ray path respectively to obtain the three-dimensional position information of each point. The global-local ray sampling method of the Sobel operator can not only perform structurally aware X-ray sampling of pixel points in the imaging image according to global and local information, but also dynamically adjust the number of points sampled on each X-ray, and better pay attention to the effective sampling area, so as to achieve sampling flexibility and efficiency, which helps to ensure a more efficient and higher quality computed tomography view reconstruction process under sparse views and a precise simulation of X-ray attenuation. On this basis, combined with the hash coding module 4, the position information of each point obtained by the line point sampling module can be three-dimensionally hashed and two-dimensionally hashed one by one, and weighted summation is performed to obtain the hash coding results corresponding to each point. The fusion of the three-dimensional hash coding and the two-dimensional hash coding results can show that richer context and spatial information are embedded in the input features, not only considering the information of the three-dimensional space point itself, but also paying attention to the mapping of the point on three orthogonal two-dimensional planes, and can predict higher quality attenuation values ​​and achieve more accurate output, so that the output result of the hash coding module can extract more context information and spatial information, thereby overcoming the technical defects of image quality problems and insufficient data information in the prior art and the technical defects of lack of generalization of the model. Finally, the construction module 5 is set to use the hash coding results corresponding to each point obtained by the hash coding module with a multi-layer perception algorithm to obtain the predicted attenuation value corresponding to each point, and then obtain a three-dimensional scanning image showing the X-ray intensity distribution through numerical integration, and finally reproduce the pixel value in the projection image, and improve the image sampling efficiency.

[0199] In the description of the embodiments of the present application, it should be noted that in the description of the present application, terms such as "inside" and "outside" indicating directions or positional relationships are based on the directions or positional relationships shown in the drawings. This is only for the convenience of description, and does not indicate or imply that the device or component must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it cannot be understood as a limitation on the present application.

[0200] In the description of the present application, the description with reference to the terms "one embodiment", "some embodiments", "in the present embodiment", "specific example", or "some examples" etc. means that the specific features, mechanisms, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, mechanisms, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, unless they are contradictory.

[0201] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application shall be based on the protection scope of the claims.

Claims

1. A sparse view reconstruction system based on neural attenuation field, characterized in that: include: An X-ray tube (1) is used to emit a cone beam of X-rays toward the object being measured; A collector (2) is configured to capture a plurality of imaging images of the object to be measured from a plurality of different angles by means of computer tomography; Line point sampling module (3); A hash coding module (4) is configured to perform three-dimensional hash coding and two-dimensional hash coding on the three-dimensional position information of each point obtained by the line point sampling module (3) point by point, and perform weighted summation to obtain a hash coding result corresponding to each point; A construction module (5) is configured to obtain the predicted attenuation value corresponding to each point by using the hash coding result corresponding to each point obtained by the hash coding module (4) through a multi-layer perception algorithm, and then obtain a three-dimensional scanning image showing the X-ray intensity distribution through a numerical integration method; The line point sampling module (3) comprises: A ray sampling module (31) is configured to receive all the imaging images acquired by the collector (2), and then extract multiple X-ray paths in each of the imaging images using a global-local ray sampling method of a Sobel operator, and determine the range of each X-ray path in the imaging image to which it belongs; A point sampling module (32) is configured to adaptively extract at least three points at equal intervals on each X-ray path according to the range of each X-ray path in the imaging image to which it belongs, to obtain the equal-distance point sequence of the X-ray path, and to obtain three-dimensional position information of each point; The ray sampling module (31) comprises: A horizontal encoder (311) is configured to perform a convolution operation on each of the imaged images in a horizontal direction convolution operation manner to obtain a horizontal encoding result corresponding to each of the imaged images; A vertical encoder (312) is configured to perform a convolution operation on each of the imaged images in a vertical direction convolution operation manner to obtain a vertical encoding result corresponding to each of the imaged images; An integrator (313) is configured to perform Sobel operator operations on the horizontal coding results and the vertical coding results corresponding to each of the imaging images, respectively, to obtain a Sobel operation result image corresponding to each of the imaging images; A generator (314) is configured to binarize the Sobel operation result images corresponding to each of the imaging images according to a threshold value, so as to obtain edge areas and non-edge areas corresponding to each of the imaging images; A line sampler (315) is configured to perform global pixel-level X-ray sampling on the non-edge areas corresponding to each of the imaging images to obtain a plurality of X-ray paths corresponding to each of the imaging images, and to use the intersection of each X-ray path with the edge area boundary line as its endpoint; A block sampler (316) is configured to obtain batches and quantities of X-ray paths corresponding to all the imaging images by local block sampling; The line point sampling module (3) communicates with the collector (2), the hash coding module (4) communicates with the line point sampling module (3), and the construction module (5) is electrically connected to the hash coding module (4).

2. The sparse view reconstruction system based on neural attenuation field according to claim 1, characterized in that: The ray sampling module (31) communicates with the collector (2), the point sampling module (32) communicates with the ray sampling module (31), and the hash coding module (4) communicates with the point sampling module (32).

3. The sparse view reconstruction system based on neural attenuation field according to claim 2, characterized in that: The horizontal encoder (311) and the vertical encoder (312) both communicate with the collector (2); the integrator (313) communicates with the horizontal encoder (311) and the vertical encoder (312) respectively; the generator (314) communicates with the integrator (313); the line sampler (315) communicates with the generator (314); the block sampler (316) communicates with the line sampler (315); and the point sampling module (32) communicates with the block sampler (316).

4. The sparse view reconstruction system based on neural attenuation field according to claim 3, characterized in that: The point sampling module (32) is configured to perform the following steps: A1: according to the batch of X-ray paths obtained by the block sampler (316), the first X-ray path is taken as the current X-ray path, and then the next step is executed; A2: determining the distance between the two ends of the current X-ray path according to the endpoints of the X-ray path obtained by the line sampler (315), and determining the number of points of the current X-ray path according to the distance between the two ends, wherein the number of points is not less than three; A3: extracting the endpoints at both ends of the current X-ray path and at least one point between the endpoints at both ends on the current X-ray path, and making the distances between two adjacent points equal and the number of extracted points equal to the number of points determined in step A2, to obtain the equidistant point sequence of the current path; A4: determining the three-dimensional position information of each point in the equidistant point array of the current X-ray path according to the shooting position of the imaging image corresponding to the current X-ray path; A5: determining whether the number of X-ray paths currently traversed is the number of X-ray paths determined by the block sampler (316); If yes, the obtained three-dimensional position information of all points is transmitted to the hash coding module (4); If not, the next X-ray path is used as the current X-ray path, and then the process returns to execute step A2.

5. The sparse view reconstruction system based on neural attenuation field according to claim 4, characterized in that: The hash coding module (4) comprises: A three-dimensional hash encoder (41) is configured to perform three-dimensional hash encoding on the three-dimensional position information of all points obtained by the point sampling module (32) point by point, to obtain a three-dimensional hash encoding result corresponding to each point; The first two-dimensional hash encoder (42) is configured to orthogonally project the three-dimensional position information of all points obtained by the point sampling module (32) on the horizontal and vertical coordinate planes one by one to obtain first orthogonal projection coordinates corresponding to each point, and then perform two-dimensional hash encoding on the first orthogonal projection coordinates corresponding to each point to obtain first two-dimensional hash encoding results corresponding to each point; The second two-dimensional hash encoder (43) is configured to perform orthogonal projection on the horizontal and vertical coordinate planes one by one on the three-dimensional position information of all points obtained by the point sampling module (32) to obtain second orthogonal projection coordinates corresponding to each point, and then perform two-dimensional hash encoding on the second orthogonal projection coordinates corresponding to each point to obtain second two-dimensional hash encoding results corresponding to each point; The third two-dimensional hash encoder (44) is configured to orthogonally project the three-dimensional position information of all points obtained by the point sampling module (32) on the vertical and horizontal coordinate planes point by point to obtain third orthogonal projection coordinates corresponding to each point, and then perform two-dimensional hash encoding on the third orthogonal projection coordinates corresponding to each point to obtain third two-dimensional hash encoding results corresponding to each point; A fuser (45) is configured to perform weighted summation on the three-dimensional hash coding result, the first two-dimensional hash coding result, the second two-dimensional hash coding result, and the third two-dimensional hash coding result corresponding to each point to obtain a hash coding result corresponding to each point; The three-dimensional hash encoder (41), the first two-dimensional hash encoder (42), the second two-dimensional hash encoder (43), and the third two-dimensional hash encoder (44) all communicate with the point sampling module (32); the fusion device (45) communicates with the three-dimensional hash encoder (41), the first two-dimensional hash encoder (42), the second two-dimensional hash encoder (43), and the third two-dimensional hash encoder (44) respectively; and the construction module (5) communicates with the fusion device (45).

6. The sparse view reconstruction system based on neural attenuation field according to claim 5, characterized in that: The three-dimensional hash encoder (41) is set to use the following hash function formula when performing hash encoding: ， The first two-dimensional hash encoder (42) is set to use the following hash function when performing hash encoding: ， The second two-dimensional hash encoder (43) is set to use the following hash function when performing hash encoding: ， The third two-dimensional hash encoder (44) is configured to use the following hash function when performing hash encoding: ， Where: V represents XOR operation; represents the remainder operation; Represents the vertex of the voxel grid where the hashed point is located; Represents the size of the feature lookup table; ； ； 。 7. The sparse view reconstruction system based on neural attenuation field according to claim 6, characterized in that: The building block (5) comprises: A multi-layer perceptron (51) is configured to obtain the predicted attenuation value corresponding to each point by using the hash coding result corresponding to each point obtained by the fusion device (45) through a multi-layer perceptron algorithm; an integrator (52) configured to obtain predicted attenuation value distribution information based on all predicted attenuation values ​​obtained by the multilayer perceptron (51), and to obtain the three-dimensional scanning image based on the predicted attenuation value distribution information by numerical integration; The multilayer perceptron (51) communicates with the fusion device (45), and the integrator (52) communicates with the multilayer perceptron (51).

8. The sparse view reconstruction system based on neural attenuation field according to claim 7, characterized in that: The multilayer perceptron (51) comprises an input terminal (511), a first fully connected layer unit (512), a second fully connected layer unit (513), a third fully connected layer unit (514) and an activation unit (515); The first fully connected layer unit (512) communicates with the input end (511), the second fully connected layer unit (513) communicates with the first fully connected layer unit (512), the third fully connected layer unit (514) communicates with the input end (511) and the second fully connected layer unit (513) respectively, and the activation unit (515) communicates with the third fully connected layer unit (514); The input end (511) is configured to receive the hash coding result obtained by the fusion device (45) and output it point by point; The activation unit (515) is configured to perform an activation operation of a Sigmoid function on the output result of the third fully connected layer unit (514) to obtain a predicted attenuation value.

9. A sparse view reconstruction method based on neural attenuation field, characterized in that: The sparse view reconstruction system based on neural attenuation field according to any one of claims 1 to 8 comprises the following steps: S1: Optimizing the parameters of the building module (5) by using a square error loss function and an absolute error loss function to obtain the optimized building module (5); S2: emits a cone beam X-ray to the object to be measured through the X-ray tube (1); S3: using a collector (2) to capture a plurality of imaging images of the object to be measured from a plurality of different angles by means of computer tomography; S4: extracting a plurality of X-ray paths from each of the imaging images by using a line point sampling module (3) in a global-local ray sampling manner of a Sobel operator, and adaptively extracting equidistant point sequences on each of the X-ray paths to obtain three-dimensional position information of each point; S5: performing three-dimensional hash coding and two-dimensional hash coding on the position information of each point obtained by the line point sampling module (3) point by point through a hash coding module (4), and performing weighted summation to obtain a hash coding result corresponding to each point; S6: The predicted attenuation value corresponding to each point is obtained by using the hash coding result corresponding to each point obtained by the hash coding module (4) through the optimized construction module (5) with a multi-layer perception algorithm, and then a three-dimensional scanning image showing the X-ray intensity distribution is obtained by numerical integration.

Images