Image reconstruction method and system based on joint optimization of intensity and residual error dual domains
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-13
- Publication Date
- 2026-08-11
AI Technical Summary
[0010]针对现有深度学习图像重建/恢复方法仅使用图像强度域 MSE 容易导致回归均值与过平滑，以及直接在图像强度域应用散度约束可能导致高灰度背景主导目标函数、掩蔽微小细节并引入解剖结构失真等技术缺陷，本发明提供了一种基于强度与残差双域联合优化的图像重建方法及系统
Smart Images
Figure CN122550754A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the interdisciplinary fields of computer vision, medical image processing, and deep learning. Specifically, this invention relates to an image reconstruction method and system based on joint optimization of intensity and residual domains. Background Technology
[0002] In the field of modern medical imaging, technologies such as computed tomography (CT), positron emission tomography (PET), and single-photon emission computed tomography (SPECT) provide crucial three-dimensional anatomical and functional information for the screening, diagnosis, staging, and treatment planning of clinical diseases. However, due to limitations imposed by radiation dose safety principles (such as the ALARA principle), scanning time windows, or hardware limitations, raw medical images acquired clinically often suffer from severe degradation.
[0003] Taking low-dose CT (LDCT) as an example, the reduction in tube current leads to a significant decrease in the number of photons received at the detector, introducing complex Poisson-Gaussian mixture noise in the projection domain. After analytical reconstruction algorithms (such as filtered back projection FBP), this noise is transformed into non-stationary, heteroscedastic, or highly signal-correlated stripes and speckle artifacts in the image space. This degradation reduces the signal-to-noise ratio and contrast of the image, adversely affecting the detection of minute lesions and quantitative texture analysis in computer-aided diagnostic (CADx) systems.
[0004] In recent years, with the improvement of computing power, deep learning technologies, represented by convolutional neural networks (CNNs), Transformer architectures, and various diffusion models, have been widely used in medical image reconstruction and restoration tasks. Current mainstream deep learning image reconstruction training paradigms typically employ supervised learning frameworks, which involve constructing data pairs between low-quality images and corresponding high-quality reference images, and training model parameters to minimize a certain target loss function.
[0005] However, analysis of existing deep learning image reconstruction and restoration methods reveals the following shortcomings in terms of loss construction and optimization strategies:
[0006] First, there is the problem of mean regression and oversmoothing. Existing methods often use mean squared error (MSE) or L2 norm as a single loss function in the image intensity domain. The optimization results tend to converge to the conditional mean, resulting in blurred details and loss of texture information in the reconstructed image.
[0007] Second, the masking effect of divergence constraints in the image intensity domain. Some studies have attempted to apply statistical divergence directly to the image intensity domain, but high grayscale regions may dominate the loss function calculation, resulting in poor recovery of low-contrast details and potentially introducing artifacts of unrealistic anatomical structures.
[0008] Third, it fails to decouple structure preservation and texture restoration at the objective function level. Existing loss designs typically couple macroscopic structure preservation with local texture restoration, making it difficult to clearly distinguish between the two optimization objectives at the level of the supervision signal.
[0009] In summary, overcoming the oversmoothing problem caused by a single image intensity domain MSE, while reducing the high-frequency masking risk caused by pure divergence constraints, and restoring low-contrast clinical texture details while ensuring the fidelity of global anatomical structures, remains a technical problem that needs to be solved in the field of deep learning medical image reconstruction. Summary of the Invention
[0010] To address the shortcomings of existing deep learning image reconstruction / restoration methods, such as the tendency to oversmooth regression mean when using only the image intensity domain MSE, and the potential for high-grayscale backgrounds to dominate the objective function, masking minute details, and introducing anatomical distortion when directly applying divergence constraints in the image intensity domain, this invention provides an image reconstruction method and system based on joint optimization of the intensity and residual domains.
[0011] In a first aspect, the present invention provides an image reconstruction method based on joint optimization of intensity and residual domains, comprising:
[0012] Acquire low-quality images and their corresponding high-quality reference images;
[0013] The low-quality image is input into the deep learning image reconstruction model to be trained for processing to obtain the reconstructed image;
[0014] In the image intensity domain, the mean square error between the reconstructed image and the reference high-quality image is calculated;
[0015] Construct a reference residual and a prediction residual, wherein the reference residual is the difference between the low-quality image and the reference high-quality image, and the prediction residual is the difference between the low-quality image and the reconstructed image;
[0016] The reference residual and the predicted residual are positively valued, and a stability constant greater than zero is introduced.
[0017] In the residual domain, the Itakura-Saito divergence induced by the negative logarithmic function is calculated, which includes at least a logarithmic penalty term, a reference baseline term, and a linearization term;
[0018] A loss function is constructed based on the mean squared error and the Itakura-Saito divergence. The parameters of the deep learning image reconstruction model are iteratively optimized using the loss function to obtain the trained image reconstruction model.
[0019] In a second aspect, the present invention provides an image reconstruction system, comprising:
[0020] The data acquisition module is used to acquire low-quality images and corresponding high-quality reference images;
[0021] The model training module is used to execute the above image reconstruction method to obtain the trained image reconstruction model.
[0022] The inference reconstruction module is used to input the low-quality image to be reconstructed into the image reconstruction model and output the inference reconstructed image.
[0023] Thirdly, the present invention provides an electronic device, including a memory and a processor; the memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the above-described image reconstruction method.
[0024] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described image reconstruction method.
[0025] Compared with existing single-domain reconstruction or traditional loss function network optimization techniques, the present invention has at least the following technical advantages:
[0026] First, a separation constraint is achieved between preserving global anatomical structure and restoring local texture. This is accomplished by introducing L... MSE As a structural anchoring term, it can reduce the risk of structural deviation caused by statistical constraints.
[0027] Second, it helps to alleviate the oversmoothing problem caused by meanization. The residual construction mechanism can separate the high-amplitude background in the image intensity domain from the fine-grained residual information and introduce IS divergence constraints in the residual domain.
[0028] It should be noted that directly transferring IS divergence to the image residual domain faces significant technical obstacles:
[0029] First, the Itakura-Saito divergence, mathematically defined, requires that the input variables be strictly positive real numbers, while the image residuals ( In terms of physical representation, it must contain positive and negative two-way grayscale differences (such as light texture and dark texture).
[0030] Secondly, low-contrast details in medical images often correspond to extremely small residual amplitudes, and direct calculation can lead to numerical singularities (log(0)) or be submerged by background noise.
[0031] Therefore, this invention is not a simple transfer of methods, but rather proposes an adaptive positive mapping and stability constant coupling mechanism. It maps the difference magnitude of the residuals to the positive domain through absolute value extraction or a smoothing function, while simultaneously introducing a stability constant. i This approach balances numerical stability with sensitivity to weak signals. This reconstruction of the loss function's domain allows the inverse intensity-weighted property of the IS divergence to precisely apply to the residual amplitude, rather than being dominated by the absolute grayscale values of the original image, thus enhancing texture recovery capabilities.
[0032] Third, inverse intensity weighting is used to improve the recovery of low-contrast details.
[0033] Linearization term It can provide higher constraint sensitivity for low-amplitude residuals, thereby enhancing the ability to recover weak textures and fine structures. This is because in the traditional mean squared error (MSE) loss, the errors of all pixels are treated with equal weight, which makes the optimization process prone to being dominated by high gray-scale regions or large noise, while weak texture differences (small residuals) are easily ignored.
[0034] This invention utilizes the denominator term of the IS divergence. A nonlinear sensing bias was constructed. When the residual amplitude... When the value approaches 0 (representing that image details tend to be smoother or weaker), the value of the denominator is mainly determined by the stability constant. i The decision is made at this point, where the sensitivity of the loss function to unit error approaches [value missing]. (A relatively large constant); however, the sensitivity decays rapidly when the residual magnitude is large. This mechanism is equivalent to opening a green channel for low-contrast details in the loss function, forcing the model to focus on those minute structural differences that are difficult to detect with the naked eye but have diagnostic value.
[0035] Fourth, it exhibits good network compatibility and cross-modal applicability. The dual-domain joint optimization strategy can be integrated into different network architectures as a general training module and can be used in scenarios such as low-dose CT, as well as CBCT, digital X-ray images, and mammograms with strong scattering interference. Attached Figure Description
[0036] The above features, structural mechanisms, and other technical effects of the present invention will become clearer and more apparent by referring to the following accompanying drawings and the description of relevant specific embodiments:
[0037] Figure 1 is an overall flowchart of dual-domain joint training of a deep learning image reconstruction model provided in an embodiment of the present invention;
[0038] Figure 2 is an analytical flowchart of the residual construction, positive value processing and LIS divergence component calculation process provided by an embodiment of the present invention;
[0039] Figure 3 is a schematic diagram comparing the dual-domain joint optimization effect and mechanism of image intensity domain MSE and residual domain IS divergence provided in an embodiment of the present invention;
[0040] Figure 4 shows the reconstruction results provided by the embodiment of the present invention;
[0041] Figure 5 is a block diagram of the functional modules of an electronic device and hardware system for performing image reconstruction and model optimization according to an embodiment of the present invention. Detailed Implementation
[0042] To enable those skilled in the art to more fully understand the inventive spirit, design principles, and specific technical solutions of this invention, the invention will be further described below in conjunction with the accompanying drawings and several detailed embodiments. It should be understood that the specific embodiments described herein are merely for explaining the technical solutions claimed by this invention.
[0043] It should be stated beforehand that the modular and step-by-step descriptions such as "acquisition," "calculation," "positiveing," and "anchoring" appearing in the specification and figures not only represent the mathematical calculation logic used, but also correspond to the information processing process executed collaboratively by the underlying processor and memory. Furthermore, the term "dual domain" used in this specification refers to "the image intensity domain plus the residual domain" throughout the entire text.
[0044] To ensure the accuracy of subsequent implementation methods, the core inputs, outputs, intermediate parameters, and measurement symbols involved in the technical solution of this invention are first defined in the following table:
[0045]
[0046] Example 1: A General Dual-Domain Joint Optimization Training Method
[0047] This embodiment describes the working mechanism of the core model training / optimization method of this invention. This mechanism is network-independent; regardless of whether the backbone network is CNN, Transformer, or U-Net, it can be connected according to the following steps. See [link to relevant documentation]. Figure 1 .
[0048] Step S1: Acquisition of paired data and forward inference of the intensity domain
[0049] The system extracts normalized low-quality input images from the database. and corresponding high-quality reference images After feeding the data into the deep learning image reconstruction model, the network outputs a reconstructed image at the end. This process establishes a mapping relationship from low-quality images to high-quality images.
[0050] Step S2: Construct the anatomical structure anchoring term in the image intensity domain
[0051] In the intensity domain branch, the reconstructed image is calculated. With reference high-quality image Mean square error between:
[0052]
[0053] in For spatial domain The total number of pixels / voxels within, As an anatomical anchoring term, it is used to constrain the consistency of overall organ morphology, large-scale tissue boundaries, and grayscale distribution, and maintain the stability of global topology and grayscale structure.
[0054] Step S3: Explicit construction of the residual domain and separation of high-frequency signals
[0055] To perform difference subtraction, see Figure 2 Construct the reference residual and the predicted residual:
[0056] ,
[0057] This transformation shifts the optimization focus from the original image intensity domain to the subtle differences in the residual domain. Low-contrast details in medical images (such as soft tissue boundaries, microvessels, and microcalcifications) appear as weak signals close to zero in the residual domain, facilitating the application of targeted statistical constraints.
[0058] Step S4: Normalize the mapping to fit the domain of the Bregman divergent family.
[0059] The Itakura-Saito (IS) divergence, mathematically defined, requires that the input variables be strictly positive real numbers, while the image residuals... Physically, residual data necessarily includes bidirectional grayscale differences (such as bright and dark textures). Therefore, it is necessary to perform positive value processing on the residual data.
[0060] This embodiment preferably uses the absolute value operation, that is, taking positive values for both the reference residual and the predicted residual: and
[0061] Simultaneously introduce the stability constant iIt is used to smooth numerical calculations near zero and avoid logarithmic singularities.
[0062] Step S5: Calculate the Itakura-Saito statistical divergence in the residual domain.
[0063] Within the residual domain, based on the strict negative logarithm function Induced IS divergence Calculate the total spatial divergence per pixel / voxel:
[0064]
[0065] Should It must contain at least three terms: a logarithmic penalty term ( ), reference baseline item ( ) and linearization term ( ). Among them, the denominator term An "inverse intensity weighting" mechanism was constructed: when the residual amplitude When the value approaches 0 (corresponding to low contrast details), the denominator mainly consists of... i The decision is that the sensitivity of the loss function to unit error approaches 1 / 2. This significantly amplifies the penalty for subtle texture differences, forcing the network to prioritize optimizing minute structural differences that are difficult to discern with the naked eye but have diagnostic value. Figure 3 .
[0066] Step S6: Dual-domain joint optimization and model convergence
[0067] Based on image intensity domain With the residual domain A joint loss function is constructed to iteratively optimize the internal parameters of the deep learning image reconstruction model. After the convergence condition is met, the trained model is output.
[0068] Example 2: Positive mapping and configuration variant of stability constant
[0069] Positive value processing method and stability constant i The configuration provides multiple equivalent implementation methods.
[0070] Regarding variations of positive value processing:
[0071] In addition to the absolute value operation described in Example 1, the positive value processing can also be performed using any of the following equivalent methods:
[0072] Positiveening after translation: Apply a global offset c to the residual data, such that... and Then take a positive value to ensure that all residual elements fall into the positive real domain of the IS divergence;
[0073] Extracting the square amplitude: direct calculation and This maps the bidirectional difference to positive energy.
[0074] Smoothing positive value mapping function: Using smoothing nonlinear functions such as the Softplus function or exponential mapping, the residuals are mapped to the positive value domain to maintain the smoothness of gradient propagation.
[0075] Regarding the stability constant i Configuration variants:
[0076] Stability constant i It is not merely a conventional numerical correction term to prevent logarithmic singularities, but is configured as an adaptive hyperparameter coupled with residual statistical properties. Its settings include:
[0077] Fixed global constants: These are preset to empirical values before training and are suitable for scenarios where the residual distribution is relatively stable.
[0078] Dynamic decay scalar: decays according to a preset function with the number of training iterations, so that the numerical stability is emphasized in the early stage of training and the sensitivity to weak signals is emphasized in the later stage.
[0079] Network learnable parameters: i As a trainable variable in the loss function, it is automatically optimized through backpropagation to adaptively match the residual distribution characteristics of the current batch of data.
[0080] Data-driven constants: based on the reference residual of the current batch The percentiles of the amplitude distribution are adaptively calculated, and the thresholds for "noise floor" and "effective detail" are dynamically defined.
[0081] Example 3: Derivative Configurations of Optimization Strategies and Network Topology
[0082] Regarding optimization strategies:
[0083] In optimizing non-convex loss surfaces in deep learning, directly applying full dual-domain joint constraints may lead to instability in the initial stage of the model. Therefore, this invention provides two complementary optimization mechanisms:
[0084] Simultaneous joint optimization: During the same training phase, based on the joint objective function... Joint optimization, among which and These are the weights for the mean squared error term and the divergence term, respectively. This method is suitable for scenarios where the network has already been well initialized or the data distribution is relatively simple.
[0085] Phased, course-based optimization: In the first training phase (e.g., the first 50% of iterations), Set to 0, or set This allows the network to prioritize learning the overall geometric structure and grayscale distribution, achieving large-scale convergence; in the second training phase, it gradually improves... By introducing texture statistical constraints in the residual domain, a coarse-to-fine course-based learning approach is achieved.
[0086] Regarding network output format:
[0087] To accommodate the design requirements of different network architectures, the output format of deep learning image reconstruction models can adopt any of the following:
[0088] Single-output reconstructed image format: The network only outputs the reconstructed image. Predicting residuals Through external calculation Implicit acquisition. This form is applicable to standard encoder-decoder architectures;
[0089] Explicit residual output format: The network directly outputs the predicted residual. Reconstructing images Through external calculation This form is applicable to residual learning architectures (such as ResNet).
[0090] Dual-head output configuration: The network has two parallel output branches, each outputting a reconstructed image. Compared with the predicted residual The richness of feature learning is enhanced through dual-task supervision.
[0091] Regarding the backbone network:
[0092] The aforementioned dual-domain joint optimization strategy, as a pluggable training module, can be integrated into any two-dimensional, 2.5-dimensional, or three-dimensional network, including but not limited to the following backbone networks: Convolutional Neural Network (CNN), Transformer architecture, State Space Model (Mamba), Residual Network (ResNet), or U-Net structure, without changing the main structure of the backbone network.
[0093] Example 4: Application of Typical Medical Imaging Modal
[0094] Regarding the adaptation of high-quality and low-quality reference images:
[0095] Reference high-quality images I N The appropriate reference image can be flexibly selected based on actual clinical conditions, including: reference high-quality image (full-dose CT), high signal-to-noise ratio image (long-term acquisition PET), reference image registered with low-quality image, pseudo-reference image (approximate high-quality image reconstructed based on traditional algorithm), or multi-frame fused reference image (obtained after fusion and noise reduction of multiple low-dose images).
[0096] low quality images I L This includes, but is not limited to, low-dose CT (LDCT) images, cone-beam computed tomography (CBCT) images, digital X-ray images, mammograms, PET / SPECT images, or OCT images. These modalities share a common underlying physical mechanism (all affected by photon statistical noise, electronic noise, or scattering interference), and are all suitable for the dual-domain joint optimization framework of this invention.
[0097] Specific applications, such as low-dose CT:
[0098] In low-dose CT, the reduced tube current causes a decrease in the number of photons received by the detector, introducing Poisson-Gaussian mixed noise in the projection domain. After filtered back projection (FBP) reconstruction, this noise is transformed into non-stationary stripes and speckle artifacts in the image space. Traditional denoising frameworks that rely solely on image intensity domain MSE tend to smooth out both non-stationary stripes and subtle anatomical structures simultaneously.
[0099] When using the method of this invention, full-dose CT is used as a reference high-quality image. I N , with low-dose images I L Subtraction forms the reference residual I R This preserves the statistical characteristics of noise and artifacts. During training:
[0100] Constraining the overall anatomical structure to prevent organ morphological distortion;
[0101] By constraining the statistical distribution of noise / artifacts in the residual domain, its inverse intensity weighting properties are utilized to enhance the recovery capability for low-contrast soft tissue and small lesions.
[0102] A horizontal comparison of the effects of single image intensity domain MSE, image intensity domain divergence, and the intensity-residual dual-domain strategy of this invention on global structure preservation and fine-grained texture restoration is provided in [link to original text]. Figure 4 The figure shows eight sets of reconstruction results and their magnified local details (bottom right corner). (a) is the input low-dose CT image, showing obvious noise and artifacts; (b) to (f) are, respectively, pure MSE, pure IS divergence (image intensity domain), MSE+SSIM, MSE+residual domain KL divergence, and MSE+residual domain IS divergence (without positive value / without i (g) is the reconstruction result of the present invention (MSE+ residual domain IS divergence, including positiveization, iThe reconstruction result (adaptive and phased optimization) is shown in (h), which is the target reference image. The red box marks the low-contrast soft tissue area. The magnified view shows that the dual-domain strategy of this invention can more clearly restore texture and micro-structural details while suppressing noise, which is significantly better than other contrast methods.
[0103] Example 5: Deployment and System Restructuring in the Inference Phase
[0104] Since the core of the "dual-domain joint" strategy of this invention resides in the training / optimization phase, its network weights are fixed once the model training is complete. During the inference phase, there is no need to input a high-quality reference image again; only the following steps need to be performed:
[0105] Step R1: Acquire low-quality medical images to be reconstructed
[0106] Acquire low-quality test images to be processed from the hospital's PACS system, imaging equipment, or edge acquisition equipment.
[0107] Step R2: Model Inference
[0108] Low-quality medical images are input into the trained image reconstruction model. The model is based on optimized weights. Directly output the inference-reconstructed image.
[0109] Step R3: Output the results
[0110] The inference-reconstructed image is output to a display terminal, storage system, or subsequent computer-aided diagnostic (CADx) analysis process.
[0111] Underlying hardware implementation:
[0112] When deployed in hospital PACS servers or edge AI devices, the electronic computing device includes a system-level bus, persistent storage media, and a multi-core parallel execution engine, see [link to relevant documentation]. Figure 5 When the core processor schedules the computer program, it allocates tensor buffers in video memory for "image intensity domain feature extraction" and "residual domain statistical mapping," respectively, and performs tensor element-wise operations on the IS divergence. The above calculations are performed via backpropagation using an automatic differentiation module, thus achieving joint optimization across both domains. During the inference phase, only the forward propagation path needs to be invoked; there is no need to calculate the loss function or perform backpropagation.
[0113] The preferred embodiments, core component architecture, and operation flow of the present invention have now been explained. Those skilled in the art should understand that variations, modifications, and equivalent substitutions based on this invention all fall within the protection scope of the appended claims.
[0114] While the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Those skilled in the art can make various modifications and refinements without departing from the spirit and scope of the invention. Therefore, the scope of protection of the present invention shall be determined by the claims.
Claims
1. An image reconstruction method based on joint optimization of intensity and residual domains, characterized in that, include: Acquire low-quality images and their corresponding high-quality reference images; The low-quality image is input into the deep learning image reconstruction model to be trained for processing to obtain the reconstructed image; In the image intensity domain, the mean square error between the reconstructed image and the reference high-quality image is calculated; Construct a reference residual and a prediction residual, wherein the reference residual is the difference between the low-quality image and the reference high-quality image, and the prediction residual is the difference between the low-quality image and the reconstructed image; The reference residual and the predicted residual are positively valued, and a stability constant greater than zero is introduced. In the residual domain, the Itakura-Saito divergence induced by the negative logarithmic function is calculated, which includes at least a logarithmic penalty term, a reference baseline term, and a linearization term; A loss function is constructed based on the mean squared error and the Itakura-Saito divergence. The parameters of the deep learning image reconstruction model are iteratively optimized using the loss function to obtain the trained image reconstruction model.
2. The method according to claim 1, characterized in that, The positive value processing can be performed using any of the following methods: Take the absolute value of the reference residual and the predicted residual; The reference residual and the predicted residual are shifted and then positive values are taken. Extract the squared magnitude values of the reference residual and the predicted residual; A smooth positive value mapping function is used to map the reference residual to the predicted residual.
3. The method according to claim 1 or 2, characterized in that, The stability constant can be set in any of the following ways: Preset fixed global constants; A scalar that decays dynamically with the number of training iterations; The parameters that the network can learn; The data-driven constant is adaptively calculated based on the percentile of the reference residual amplitude distribution.
4. The method of claim 3, wherein, The Itakura-Saito divergence induced by the negative logarithmic function in the residual domain is calculated using the following formula: wherein, is a spatial index, is a spatial domain, is a magnitude of a prediction residual, is a magnitude of a reference residual, is a stable constant greater than 0.
5. The method of claim 3, wherein, The iterative optimization of the parameters of the deep learning image reconstruction model using the loss function can be achieved through one of the following methods: Simultaneous joint optimization: During the same training phase, based on the joint objective function... Joint optimization, among which and These are the weights of the mean square error term and the divergence term, respectively. Mean square error, Itakura-Saito divergence; Phased course-based optimization: In the first training phase, as described above... The main structure is initialized, and then introduced in the second training phase. or increase The weights are used for texture restoration in the residual domain.
6. The method of claim 3, wherein, The mean square error serves as an anatomical structure anchoring term to maintain the stability of the overall organ morphology, tissue boundary topology, and grayscale structure; the Itakura-Saito divergence serves as a texture statistical constraint term to constrain the statistical distribution of the predicted residual to be consistent with the reference residual distribution, thereby enhancing the recovery ability for low-contrast textures, soft tissue boundaries, and small blood vessels.
7. The image reconstruction method of claim 1, wherein, After obtaining the trained image reconstruction model, the method also includes an inference phase: Acquire low-quality medical images to be reconstructed; The low-quality medical image is input into the trained image reconstruction model to obtain an inferred reconstructed image; wherein, the inference stage does not require the input of a high-quality reference image.
8. An image reconstruction system characterized by, include: The data acquisition module is used to acquire low-quality images and corresponding high-quality reference images; The model training module is used to execute the image reconstruction method described in claim 1 to obtain a trained image reconstruction model. The inference reconstruction module is used to input the low-quality image to be reconstructed into the image reconstruction model and output the inference reconstructed image.
9. An electronic device, comprising: It includes a memory and a processor; the memory stores a computer program, which, when executed by the processor, causes the processor to implement the steps of the image reconstruction method of claim 1.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the image reconstruction method of claim 1.