Remote reconstruction method of organ imaging model

Through the combination of deep learning engine and bidirectional predictor, super-resolution tiles are generated and moderately reconstructed using volume factor and position attenuation factor, the problems of image distortion and storage space bandwidth in medical images are solved, and efficient three-dimensional Image reconstruction.

CN113888404BActive Publication Date: 2024-12-03CHENGDU ZHONGCHUANG WULIAN TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202010623742.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2020-07-01
Publication Date
2024-12-03
Estimated Expiration
2040-07-01

AI Technical Summary

Technical Problem

The prior art has problems with image distortion, low signal-to-noise ratio and limited storage space and bandwidth in the three-dimensional reconstruction of medical images, making it difficult to take into account the needs of high resolution, three-dimensional image quality and cloud storage space.

Method used

The deep learning engine is used to perform multiple convolution operations, combine the bidirectional predictor and pixel summing unit to generate super-resolution tiles, improve the accuracy of image details through iterative training, and moderate reconstruction using volume factor and position attenuation factor , optimize image quality.

Benefits of technology

It effectively improves the quality of three-dimensional image and computing efficiency, solves the limitations of image distortion and bandwidth storage space, and realizes efficient remote image reconstruction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113888404B_ABST
    Figure CN113888404B_ABST
Patent Text Reader

Abstract

The present invention provides a remote reconstruction method for an organ imaging model, the method comprising: collecting high-resolution tiles and low-resolution tiles by a medical imaging scanning device; calculating multiple candidate high-resolution pixel sets and reference high-resolution pixel sets by a deep learning engine; the deep learning engine generates a post-processing residual value; a pixel summing unit performs a pixel summing operation on the post-processing residual value and the low-resolution tile to obtain a forward and reverse predicted high-resolution tile; a bidirectional predictor generates a super-resolution tile for the forward predicted high-resolution tile and the reverse predicted high-resolution tile. The present invention proposes a remote reconstruction method for an organ imaging model, which effectively improves the three-dimensional image quality and operation efficiency at a relatively low software and hardware cost, and solves the image distortion problem of the existing reconstruction method and the limited bandwidth and limited storage space problem of the existing remote reconstruction method.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to three-dimensional image reconstruction, and in particular to a remote reconstruction method of an organ image model. Background Art

[0002] High-resolution three-dimensional images have higher quality and clear details, which are very important for application fields such as medical imaging. The medical image scanning system needs to record the model file and transmit it to the cloud after image compression. After the cloud decompresses the model file, it can be transmitted to the medical terminal for real-time monitoring by the user; at the same time, the cloud can also be used to store multiple model files collected by the image scanning device for users to retrieve at any time. However, better picture smoothness and more image details will inevitably require improving the resolution and frame rate of the model file, but the model file capacity will also increase, which will put a greater burden on the storage space and cache of the cloud. If super-resolution reconstruction and other technologies are used, the image edges often have distortion problems such as jaggedness. In addition, the reconstruction method of the existing scheme still has the problem of low signal-to-noise ratio, and it is impossible to propose a suitable image enhancement mechanism for the medical image scanning system. Therefore, it is necessary to provide a method for high-resolution three-dimensional image reconstruction that takes into account the storage space, cache size, network speed and image quality of the cloud. Summary of the invention

[0003] In order to solve the problems existing in the above-mentioned prior art, the present invention proposes a remote reconstruction method of an organ image model, comprising:

[0004] Collecting multiple frames of high-resolution image tiles and multiple frames of low-resolution image tiles through a medical imaging scanning device;

[0005] The deep learning engine performs multiple convolution operations on the high-resolution image block to calculate multiple candidate high-resolution pixel sets, and performs at least one convolution operation on each candidate high-resolution pixel set to generate a reference high-resolution pixel set;

[0006] The deep learning engine generates a post-processing residual value after performing the convolution operation on the reference high-resolution pixel set and the low-resolution image block at least once;

[0007] A pixel summing unit performs a pixel summing operation on the post-processing residual value and the low-resolution image block to calculate a forward predicted high-resolution image block and a backward predicted high-resolution image block of the same time stamp; and

[0008] After receiving the forward predicted high-resolution image block and the backward predicted high-resolution image block by the bidirectional predictor, the convolution operation is performed at least once on the forward predicted high-resolution image block and the backward predicted high-resolution image block to generate a super-resolution image block;

[0009] Among them, when the high-resolution block of a frame and the low-resolution block of a frame of the previous timestamp sequentially execute the steps of generating candidate pixel sets and reference pixel sets, generating post-processing residual values ​​and generating predicted high-resolution blocks, the reverse predicted high-resolution blocks are generated.

[0010] Preferably, the deep learning engine has an input layer, an embedding layer and an output layer, the input layer is used to input the image blocks or residual values ​​to be subjected to the convolution operation, the embedding layer is used to store multiple parameters for determining the convolution kernel used in the convolution operation, and the output layer is used to output the operation result of the convolution operation.

[0011] Preferably, the multiple parameters are that the input layer can use different convolution kernels to perform the convolution operation on the high-resolution image blocks multiple times to calculate the multiple candidate high-resolution pixel sets, and the input layer can reuse the convolution kernel of the minimum pixel to perform the convolution operation on each candidate high-resolution pixel set at the same time to calculate the reference high-resolution pixel set.

[0012] Preferably, the multiple parameters are that the input layer can utilize the convolution kernel to simultaneously perform convolution operations on the reference high-resolution pixel set and the low-resolution image block to calculate the image superposition residual value.

[0013] Compared with the prior art, the present invention has the following advantages:

[0014] The present invention proposes a remote reconstruction method for an organ imaging model, which effectively improves the three-dimensional image quality and computing efficiency at a relatively low software and hardware cost, and solves the image distortion problem of the existing reconstruction method and the problems of limited bandwidth and limited storage space of the existing remote reconstruction method. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 4 is a flow chart of a remote reconstruction method of an organ imaging model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0016] The following and the accompanying drawings illustrate the principles of the present invention. Figure 1 The present invention is described in detail in conjunction with one or more embodiments of the present invention. The present invention is described in conjunction with such embodiments, but the present invention is not limited to any embodiment. The scope of the present invention is limited only by the claims, and the present invention encompasses numerous alternatives, modifications, and equivalents. In the following description, numerous specific details are set forth to provide a thorough understanding of the present invention. These details are provided for illustrative purposes, and the present invention may be implemented according to the claims without some or all of these specific details.

[0017] Aspects of the present invention provide a remote reconstruction method for an organ image model. The method is implemented by the medical image scanning and reconstruction system of the present invention. The system includes: a multi-layer cache module, a deep learning engine, a rendering and reconstruction engine, and a bidirectional predictor; the multi-layer cache module receives multiple frames of high-resolution tiles with a lower frame rate and multiple frames of low-resolution tiles with a higher frame rate collected by the image scanning device, and the deep learning engine, the rendering and reconstruction engine, and the bidirectional predictor perform super-resolution reconstruction model rendering to calculate the super-resolution tiles, and transmit them to the medical terminal in sequence for the user to view the three-dimensional image. The present invention uses the deep learning engine as the selection mechanism and model rendering mechanism of the image candidate samples in the super-resolution reconstruction method. Due to the iterative training and learning of the deep learning engine, more accurate image details are collected.

[0018] Figure 1 It is a flow chart of a remote reconstruction method of an organ imaging model according to an embodiment of the present invention. The medical imaging scanning and reconstruction system can establish communication connections with an imaging scanning device and a medical terminal respectively through a network. The imaging scanning device can collect high-resolution model files and low-resolution model files, and transmit them to the medical imaging scanning and reconstruction system for caching, decoding and super-resolution reconstruction. The frame rate of the high-resolution model files collected by the imaging scanning device can be less than or equal to the low-resolution model files. When the medical imaging scanning and reconstruction system completes the rendering of the super-resolution reconstruction model, super-resolution tiles with high frame rate and high resolution are generated. After being cached and image encoded, they can be transmitted to the medical terminal in sequence for users to view three-dimensional images with super-resolution image quality.

[0019] After the medical image scanning and reconstruction system receives multiple frames of high-resolution tiles with a lower frame rate and multiple frames of low-resolution tiles with a higher frame rate collected by the image scanning device, the number of tile frames collected by the image scanning device is determined by the sampling frame number T. If T=2, the image scanning device collects 1 frame of high-resolution tiles and 1 frame of low-resolution tiles; if T=5, the image scanning device collects 1 frame of high-resolution tiles and 4 frames of low-resolution tiles. The sampling frame number T value determines the number of tile frames continuously collected by the image scanning device, and the tiles collected by the image scanning device include at least 1 frame of high-resolution tiles. In a preferred embodiment, the remaining tiles are all low-resolution tiles, thereby solving the problem of limited storage space and transmission bandwidth of the image scanning device. When the image scanning device has collected the above-mentioned tiles, it is submitted to the image compression module for image compression, and the compression reduces the model file size by reducing the bit rate, and then submitted to the first transmission module for transmission. The first transmission module pre-divides the above-mentioned tiles compressed by the image compression module into multiple messages, and then transmits them to the medical image scanning and reconstruction system.

[0020] When the medical image scanning and reconstruction system receives multiple messages transmitted by the first transmission module, it is first temporarily stored in the model cache, and the model cache is then connected to the image decoder to decompress the aforementioned multiple messages, and then decode the above-mentioned blocks. Then, after the image decoder temporarily stores the above-mentioned blocks in the multi-layer cache module, the multi-layer cache module is then connected to the storage module to store the decoded three-dimensional images that do not need to be stored for a long time. In a preferred embodiment, the model cache can be directly connected to the storage module to effectively save storage space. When the multi-layer cache module receives the decoded blocks, the decoded blocks are transmitted to the deep learning engine, and then the deep learning engine can perform super-resolution reconstruction model rendering on the blocks collected by the image scanning device. Before performing super-resolution reconstruction model rendering, for the high-resolution blocks and low-resolution blocks received by the deep learning engine, the rendering and reconstruction engine can enlarge the low-resolution blocks to the same size as the high-resolution blocks through interpolation methods such as nearest neighbor interpolation and bilinear interpolation, and then calculate the super-resolution blocks and temporarily store them in the multi-layer cache module. Then, the image encoder encodes the super-resolution image tiles and temporarily stores them in the model cache. Finally, the model cache transmits the super-resolution image tiles to the medical terminal.

[0021] The deep learning engine includes an input layer, an embedding layer and an output layer, wherein the input layer is used to input a block or a residual value to be subjected to a convolution operation; the embedding layer is used to store multiple parameters to determine a pixel filtering unit for the deep learning engine to perform a convolution operation; the output layer is used to output a convolution operation result; the offset calculation unit of the rendering and reconstruction engine performs a pixel offset calculation to reduce the offset difference generated by each continuous block, thereby screening out a pixel set with a minimum pixel offset with a low-resolution block; the pixel summation unit of the rendering and reconstruction engine is used to perform a pixel summation operation to respectively generate blocks that can be received by a bidirectional predictor; the bidirectional predictor is used to receive a forward predicted high-resolution block and a backward predicted high-resolution block of the same timestamp, so that the deep learning engine can calculate a super-resolution block after performing at least one convolution operation.

[0022] The multiple parameters stored in the embedding layer of the deep learning engine may include: if it is a forward prediction mode and the number of sampling frames T=3, the input layer may perform multiple convolution operations on the high-resolution blocks using different convolution kernels, so that the input layer can calculate multiple candidate high-resolution pixel sets from the high-resolution blocks of the tth frame as the t+1th frame, and then, in order to reduce the number of candidate blocks for each candidate high-resolution pixel set, the input layer may perform at least one convolution operation on each candidate high-resolution pixel set of the t+1th frame with a smaller convolution kernel to generate a reference high-resolution pixel set. In a preferred embodiment, for the pixel set after the input layer performs a convolution operation on each candidate high-resolution pixel set, the pixel set with the minimum pixel offset from the low-resolution tile is first screened out as the reference high-resolution pixel set; if it is a reverse prediction mode, the input layer can perform multiple convolution operations on the high-resolution tile with different convolution kernels, so that the input layer can calculate multiple candidate high-resolution pixel sets from the high-resolution tile, and then perform at least one convolution operation on the multiple candidate high-resolution pixel sets with a smaller convolution kernel to generate a reference high-resolution pixel set, and for the pixel set after the input layer performs a convolution operation on each candidate high-resolution pixel set, the pixel set with the minimum pixel offset from the low-resolution tile is first screened out as the reference high-resolution pixel set.

[0023] In a preferred embodiment, for the pixel set generated after the deep learning engine performs a convolution operation on each candidate high-resolution pixel set, after screening out the pixel set with the minimum pixel offset with the low-resolution tile, the deep learning engine simultaneously performs a convolution operation on the reference high-resolution pixel set and the low-resolution tile to generate an image stacking residual value, and then to adjust the image quality of the image stacking residual value, the deep learning engine continues to perform a convolution operation on the image stacking residual value with the same or different convolution kernel to generate a first post-processing residual value, and then performs a convolution operation on the first post-processing residual value to generate a second post-processing residual value, and so on. More unnecessary image information can be filtered out from each post-processing residual value, and the image acquisition details can be increased through the learning mechanism of deep learning. Then the deep learning engine uses the finally generated post-processing residual value as the forward post-processing residual value and the reverse post-processing residual value, and transmits them to the rendering and reconstruction engine.

[0024] Optionally, the deep learning engine can also directly use the image superposition residual value as the final post-processing residual value and transmit it to the rendering and reconstruction engine. When the rendering and reconstruction engine receives the forward post-processing residual value and the reverse post-processing residual value, if it is in the forward prediction mode, the rendering and reconstruction engine performs a pixel summation operation on the forward post-processing residual value and the low-resolution image block to generate a forward predicted high-resolution image block; if it is in the reverse prediction mode, the rendering and reconstruction engine performs a pixel summation operation on the reverse post-processing residual value and the low-resolution image block to generate a reverse predicted high-resolution image block. Then, the forward predicted high-resolution image block and the reverse predicted high-resolution image block of the same timestamp are transmitted to the bidirectional predictor to continue to execute the step of generating a super-resolution image block, that is, submitting them to the deep learning engine for at least one convolution operation to generate a super-resolution image block.

[0025] When generating post-processing residual values, if it is a forward prediction mode, after the deep learning engine obtains the first post-processing residual value and the second post-processing residual value, the output layer of the deep learning engine determines that the final post-processing residual value is the forward post-processing residual value, and then submits it to the pixel summing unit to perform a pixel summing operation on the forward post-processing residual value and the low-resolution block to generate a forward predicted high-resolution block; conversely, if it is a reverse prediction mode, in order to adjust the image details of the image overlay residual value, the input layer can perform a convolution operation on the image overlay residual value with the same or different pixel filtering units, and calculate the first and second post-processing residual values ​​with the same process, and so on, and determine that the final post-processing residual value is the reverse post-processing residual value, and then perform a pixel summing operation on the reverse post-processing residual value and the low-resolution block to generate a reverse predicted high-resolution block.

[0026] Regardless of the forward or reverse prediction mode, the input layer can perform any number of convolution operations on the second post-processing residual value to generate a third, ... or Nth post-processing residual value, and use the Nth post-processing residual value as the forward post-processing residual value or the reverse post-processing residual value. If the input layer performs convolution with the same convolution kernel each time, more unnecessary image noise can be filtered out, thereby increasing the peak signal-to-noise ratio.

[0027] Finally, the pixel summation unit transmits the forward predicted high-resolution block and the backward predicted high-resolution block of the same time stamp to the bidirectional predictor to continue generating super-resolution blocks.

[0028] If the medical terminal wants to view non-real-time three-dimensional images, the multi-layer cache module first collects the decoded multi-frame high-resolution tiles and multi-frame low-resolution tiles from the storage module and transmits them to the deep learning engine. The storage module can also collect multiple undecoded messages, and the above messages are decoded by the image decoder, and then transmitted to the deep learning engine, rendering reconstruction engine and bidirectional predictor to calculate the super-resolution tiles, and then temporarily stored in the multi-layer cache module for caching, and transmitted to the image encoder for encoding, and then transmitted to the model cache after the encoding is completed. When the image caching and encoding steps are completed, the second transmission module transmits each super-resolution tile to the medical terminal so that the medical terminal can view the non-real-time three-dimensional image function.

[0029] In the process of generating candidate high-resolution pixel sets, generating post-processing residual values ​​and finally generating super-resolution blocks, the deep learning engine of the present invention performs convolution operations, which can capture more accurate image details after continuous training and learning.

[0030] In order to further improve the accuracy of image quantification, the present invention uses the product of the volume factor tb related to the size of the organ and tissue, and the position attenuation factor ta related to the position as the model variable t required for reconstruction, and interpolates the model variable t to perform moderate reconstruction.

[0031] The model variable t is used to describe the difference in attenuation factors at different positions, and a relationship can be established between the reconstruction process and the attenuation factors at different positions for rapid reconstruction. A preferred embodiment includes: determining an attenuation factor h(x, y) for each scanning position (x, y) in a scanning plane; determining a position model variable ta corresponding to an attenuation distance of each scanning position (x, y) in the scanning plane relative to a reference position in the scanning plane according to the attenuation factor h(x, y); generating a projection image g(x, y) corresponding to at least one light source on the scanning plane, wherein each pixel of the projection image corresponds to a scanning position (x, y) in the scanning plane, and a pixel in an original image corresponds to the reference position; determining a corresponding attenuation distance according to a distance between each pixel and a pixel corresponding to the reference position to determine a position model variable ta corresponding to each pixel; performing a reconstruction operation on each pixel of the projection image g(x, y) according to the projection image g(x, y) and the attenuation factor h(x, y) to generate an initial reconstructed image v(x, y); and performing an iterative operation on the scanning position (x, y) corresponding to each pixel of the projection image g(x, y), and after k iterative operations, an iterative reconstructed image v is formed. k+1 (x, y), that is:

[0032]

[0033] Where t is the product of the model variable ta at that position and the volume factor tb.

[0034] The attenuation factor may be the half-maximum full width of the resolution of other scanning positions relative to the reference position with the reference position as the center. The reference position is the center position on the scanning plane. The attenuation distance refers to the radial distance of the scanning position (x, y) on the scanning plane relative to the reference position. Each pixel v (k+1) (x, y) corresponds to different attenuation factors during the reconstruction process.

[0035] For the convenience of calculation, the present invention makes each pixel of the reconstructed image use the attenuation factor of the reference position as the attenuation factor corresponding to the current pixel. k+1 (x, y), the model variable t is added for appropriate compensation. Since the model variable t is the product of ta and the volume factor tb, the attenuation factor h(x, y) needs to be obtained first to determine the volume factor tb, which is the relationship between the distance of each scanning position (x, y) in the scanning plane relative to the center of the scanning plane and the half-height full width of the resolution of the medical imaging scanning device. The attenuation factor can give the volume factor tb a change relationship from 2 to 1 through a specific multiple range of the half-height full width of the system resolution, and when it exceeds the specific multiple range of the half-height full width, the volume factor tb is 1.

[0036] Continuing from the previous example, we iteratively reconstruct the image v k+1 (x, y) is the optimized reconstructed image obtained by the model of the present invention. The number of iterative operations k can be determined according to the needs and the image reconstruction status. The product of the number of iterative operations k and the model variable t is constant. If the image is over-reconstructed after iteration, the excessive impact caused by the number of iterations can be suppressed by lowering the value of the model variable t. According to the above method, the number of iterative operations k and the model variable t can be adjusted to control the over-reconstruction of the image caused by the iterative operation.

[0037] In another embodiment, the volume of the light source can be considered, that is, the value of the volume factor tb can be determined. Then, the result of multiplying the position model variable ta and the volume factor tb is used as the adjustment model variable t for image reconstruction. In this way, a light source with a small volume can also produce a clear image in image reconstruction. Even when there are multiple light sources with different volume sizes, a reconstructed image with better resolution effect can be obtained by compensating the volume factor tb.

[0038] If the multiplication result makes the model variable t greater than 2, the product of the model variable t and the number of iterations k is used to be a constant relationship to adjust and increase the number of iterations to reduce the value of the model variable t, thereby avoiding iterative divergence.

[0039] In addition, since medical images such as MRI are initially a large amount of two-dimensional grayscale data, a three-dimensional image model is formed by model rendering and reconstruction of the two-dimensional image. However, the three-dimensional image reconstructed from the two-dimensional image cannot fully present the geometric shape and structure of the original organ under test, and because the amount of input two-dimensional image data is too large, inputting each frame of two-dimensional image information one by one to reconstruct the original three-dimensional image will cause excessive computation. Therefore, in a further aspect of the present invention, under the premise of ensuring the accurate quantitative symmetry value, the quantifiable symmetry value is used to establish the optimal axial plane, which is used to assist medical personnel in identifying and correcting the deviation of the three-dimensional image in clinical diagnosis.

[0040] Because the grayscale value range of different tissues or organs of the human body after scanning will be different, first of all, after obtaining the medical image data and related parameters, a threshold range is set, that is, the grayscale value range for specific parts of the image; the error model containing the grayscale value is executed to solve the optimal parameters, which are used to judge the similarity of the binary medical image after mirroring, and the optimal parameters are obtained by continuously comparing the input parameters. The optimal parameters are substituted into the central axis model or the central axis surface model to obtain the best central axis surface. By entering a specific threshold in the set parameters, the amount of input data is effectively reduced, and the input image data is compared through the set grayscale value error model, so that a three-dimensional image can be quickly reconstructed, or the obtained best central axis surface or best central axis can be used to correct the deviation image caused by the scanning of the medical instrument after recalculation by interpolation.

[0041] In summary, the present invention proposes a remote reconstruction method for an organ imaging model, which effectively improves the three-dimensional image quality and computing efficiency at a lower software and hardware cost, while solving the image distortion problem of the existing reconstruction method and the problems of limited bandwidth and limited storage space of the existing remote reconstruction method.

[0042] Obviously, those skilled in the art should understand that the above modules or steps of the present invention can be implemented by a general computing system, they can be concentrated on a single computing system, or distributed on a network composed of multiple computing systems, and optionally, they can be implemented by program codes executable by the computing system, so that they can be stored in a storage system and executed by the computing system. Thus, the present invention is not limited to any specific combination of hardware and software.

[0043] It should be understood that the above specific embodiments of the present invention are only used to illustrate or explain the principles of the present invention, and do not constitute a limitation of the present invention. Therefore, any modifications, equivalent substitutions, improvements, etc. made without departing from the spirit and scope of the present invention should be included in the protection scope of the present invention. In addition, the appended claims of the present invention are intended to cover all changes and modifications that fall within the scope and boundaries of the appended claims or the equivalent forms of such scope and boundaries.

Claims

1. A remote reconstruction method of an organ image model, characterized in that: include: Collecting multiple frames of high-resolution image tiles and multiple frames of low-resolution image tiles through a medical imaging scanning device; The deep learning engine performs multiple convolution operations on the high-resolution image block to calculate multiple candidate high-resolution pixel sets, and performs at least one convolution operation on each candidate high-resolution pixel set to generate a reference high-resolution pixel set; The deep learning engine generates a post-processing residual value after performing the convolution operation on the reference high-resolution pixel set and the low-resolution image block at least once; A pixel summing unit performs a pixel summing operation on the post-processing residual value and the low-resolution image block to calculate a forward predicted high-resolution image block and a backward predicted high-resolution image block of the same time stamp; as well as After receiving the forward predicted high-resolution image block and the backward predicted high-resolution image block by the bidirectional predictor, the convolution operation is performed at least once on the forward predicted high-resolution image block and the backward predicted high-resolution image block to generate a super-resolution image block; When the high-resolution image block of a frame and the low-resolution image block of a frame at the previous timestamp sequentially execute the steps of generating a candidate pixel set and a reference pixel set, generating a post-processing residual value and generating a predicted high-resolution image block, the reverse predicted high-resolution image block is generated; The product of the volume factor tb related to the volume of the organ and tissue and the position attenuation factor ta related to the position is used as the model variable t required for reconstruction, and the model variable t is interpolated to perform moderate reconstruction; The model variable t is used to describe the difference in attenuation factors at different positions, and a relationship can be established between the reconstruction process and the attenuation factors at different positions to perform fast reconstruction, including: determining an attenuation factor h(x, y) for each scanning position (x, y) in the scanning plane; determining a position model variable ta corresponding to an attenuation distance of each scanning position (x, y) in the scanning plane relative to a reference position in the scanning plane according to the attenuation factor h(x, y); generating a projection image g(x, y) corresponding to at least one light source on the scanning plane, wherein each pixel of the projection image corresponds to a scanning position (x, y) in the scanning plane, and a pixel in the original image corresponds to the reference position; determining a corresponding attenuation distance according to a distance between each pixel and a pixel corresponding to the reference position, so as to determine a position model variable ta corresponding to each pixel; performing a reconstruction operation on each pixel of the projection image g(x, y) according to the projection image g(x, y) and the attenuation factor h(x, y) to generate an initial reconstructed image v(x, y); and performing an iterative operation on the scanning position (x, y) corresponding to each pixel of the projection image g(x, y), and forming an iterative reconstructed image v after k iterative operations. k+1 (x, y), which is: , Where t is the product of the model variable ta at that position and the volume factor tb; The attenuation factor may be the half-maximum full width value of the resolution of other scanning positions relative to the reference position with the reference position as the center; the reference position is the center position on the scanning plane; the attenuation distance refers to the radial distance of the scanning position (x, y) on the scanning plane relative to the reference position; each pixel v (k+1) (x, y) corresponds to different attenuation factors during the reconstruction process.

2. The method according to claim 1, characterized in that The deep learning engine has an input layer, an embedding layer and an output layer. The input layer is used to input a block or residual value to be subjected to a convolution operation. The embedding layer is used to store multiple parameters for determining a convolution kernel used in the convolution operation. The output layer is used to output a result of the convolution operation.