Magnetic resonance image reconstruction method and device and electronic equipment

Abstract

The invention provides a magnetic resonance image reconstruction method and device and electronic equipment, and the method comprises the steps: obtaining the measurement data of an undersampled preset space, and generating a zero-filling initial image according to the measurement data; performing forward estimation on the zero-filling initial image in an image domain to obtain an initial reconstructed image; performing iterative processing of adaptive data consistency correction and single-step diffusion refinement on the initial reconstructed image to obtain a target reconstructed image, and realizing efficient, stable and controllable magnetic resonance image reconstruction by taking the single-step diffusion refinement of'structure alignment 'as a core and combining the adaptive data consistency correction; high fidelity, strong semantics and low delay can be kept at an extremely low sampling rate, the reconstruction process of the magnetic resonance image is remarkably accelerated, and the image quality is improved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and specifically to a magnetic resonance image reconstruction method, apparatus, and electronic device. Background Technology

[0002] Magnetic resonance imaging (MRI) is widely used in medical fields such as neuroimaging, cardiovascular examination, and tumor monitoring due to its high-contrast imaging capabilities for soft tissues and its radiation-free nature. However, traditional MRI acquisition is relatively slow, with scan times typically lasting several minutes. This limits the turnaround efficiency of clinical equipment and is prone to image artifacts caused by patient movement, affecting the quality of subsequent diagnoses. In high-throughput scenarios such as pediatrics and emergency medicine, the trade-off between imaging efficiency and quality is becoming increasingly prominent.

[0003] In current MRI reconstruction, the mainstream diffusion reconstruction process requires starting from pure noise and going through dozens to hundreds of iterations. High-dimensional score estimation needs to be performed on each frame of the image, resulting in high inference costs and low reconstruction efficiency, which limits clinical applicability. The diffusion gradient in the image domain and the direction of the frequency domain measurement consistency are often inconsistent, which causes a conflict between consistency correction and generation path, and easily leads to oscillations and edge blurring in the generation process. Especially under high aliasing or low signal-to-noise conditions, the recovery of image details is unstable. Summary of the Invention

[0004] In view of this, the embodiments of the present invention aim to provide a magnetic resonance image reconstruction method, apparatus and electronic device, which takes "structure alignment" single-step diffusion refinement as the core and combines adaptive data consistency correction to achieve efficient, stable and controllable magnetic resonance image reconstruction; it can maintain high fidelity, strong semantics and low latency even at extremely low sampling rates, thereby significantly accelerating the magnetic resonance image reconstruction process and improving image quality.

[0005] According to one aspect of the present invention, an embodiment of the present invention provides a magnetic resonance image reconstruction method, comprising: acquiring measurement data of an undersampled preset space, and generating a zero-filled initial image based on the measurement data; performing forward estimation on the zero-filled initial image in the image domain to obtain an initial reconstructed image; and performing adaptive data consistency correction and single-step diffusion thinning processing on the initial reconstructed image to obtain a target reconstructed image.

[0006] In one embodiment, the iterative processing of adaptive data consistency correction and single-step diffusion refinement on the initial reconstructed image to obtain the target reconstructed image includes: using the initial reconstructed image as a first reconstructed image, performing adaptive data consistency correction based on measurement residuals according to the first reconstructed image and the zero-filled initial image to obtain a correction result; using the correction result as input, performing detail enhancement processing through a single-step diffusion refiner to obtain a second reconstructed image; updating the first reconstructed image to the second reconstructed image, iteratively performing adaptive data consistency correction and single-step diffusion refinement processing until a preset condition is met, wherein the preset condition is that the single-step diffusion refiner converges or reaches a preset number of iterations; and using the second reconstructed image that meets the preset condition as the target reconstructed image.

[0007] In one embodiment, the step of performing adaptive data consistency correction based on measurement residuals to obtain a correction result based on the first reconstructed image and the zero-filled initial image includes: obtaining a back projection of the measurement residual based on the first reconstructed image and the zero-filled initial image, wherein the back projection represents the residual between the zero-filled initial image and the result of processing the first reconstructed image using a preset acquisition operator; determining the step size for consistency correction; and obtaining the correction result based on the adaptive step size, the first reconstructed image, and the back projection.

[0008] In one embodiment, determining the step size for consistency correction includes: normalizing the back projection; applying a residual sensing network to process the normalized back projection, and combining it with interval mapping to obtain an adaptive step size.

[0009] In one embodiment, the step of using the correction result as input and performing detail enhancement processing through a single-step diffusion refiner to obtain a second reconstructed image includes: grouping the correction result into contextual groups based on structural similarity to obtain three-channel grouped data; processing the three-channel grouped data using a medical structural encoder and combining global average pooling and linear projection to obtain a guiding token, wherein the guiding token is a conditional signal input to the single-step diffusion refiner; and applying the single-step diffusion refiner to perform structural alignment single-step diffusion refinement on the three-channel grouped data according to the guiding token to obtain the second reconstructed image.

[0010] In one embodiment, the step of applying the single-step diffusion refiner to perform structure-aligned single-step diffusion refinement of the three-channel grouped data according to the guiding token to obtain a second reconstructed image includes: applying the variational autoencoder in the single-step diffusion refiner to map the three-channel grouped data to the latent space to obtain latent space image data; applying the backbone network in the single-step diffusion refiner in combination with the guiding token to perform single-step distillation on the latent space image data to obtain latent space prediction data; and restoring image domain details from the latent space prediction data using a decoder to obtain the second reconstructed image.

[0011] In one embodiment, before performing forward estimation on the zero-padded initial image in the image domain to obtain the initial reconstructed image, the process includes: training an initial structure predictor based on historical zero-padded initial images to obtain an initially trained initial structure predictor and a corresponding historical initial reconstructed image, wherein the initial structure predictor is used to perform forward estimation on the zero-padded initial image; training a single-step diffusion refiner based on the historical initial reconstructed image to obtain an initially trained single-step diffusion refiner; and jointly fine-tuning the initially trained initial structure predictor and the initially trained single-step diffusion refiner based on historical zero-padded initial images to obtain a trained initial structure predictor and a single-step diffusion refiner.

[0012] According to another aspect of the present invention, an embodiment of the present invention provides a magnetic resonance image reconstruction apparatus, comprising: an initial image generation unit, configured to acquire measurement data of an undersampled preset space and generate a zero-filled initial image based on the measurement data; an initial reconstruction unit, configured to perform forward estimation on the zero-filled initial image in the image domain to obtain an initial reconstructed image; and a target reconstruction unit, configured to perform iterative processing of adaptive data consistency correction and single-step diffusion refinement on the initial reconstructed image to obtain a target reconstructed image.

[0013] According to another aspect of the present invention, an electronic device is provided in one embodiment of the present invention, comprising: a memory for storing an executable computer program; and a processor for calling and running the executable computer program from the memory, such that the processor executes the above-described magnetic resonance image reconstruction apparatus.

[0014] This invention provides a magnetic resonance image reconstruction method, apparatus, and electronic device. The method acquires measurement data of an undersampled preset space and generates a zero-filled initial image based on the measurement data. Forward estimation is performed on the zero-filled initial image in the image domain to obtain an initial reconstructed image. Iterative processing of adaptive data consistency correction and single-step diffusion refinement is applied to the initial reconstructed image to obtain a target reconstructed image. Single-step diffusion refinement with "structure alignment" as the core, combined with adaptive data consistency correction, achieves efficient, stable, and controllable magnetic resonance image reconstruction. By replacing noise sampling with structure initialization, the iterative steps are significantly reduced, computational overhead is lowered, and the conflict between diffusion and physical constraints is alleviated, improving texture restoration quality and three-dimensional spatial consistency. This allows high fidelity, strong semantics, and low latency to be maintained even at extremely low sampling rates, providing a novel solution for the clinical deployability of diffusion models in medical image reconstruction. It can maintain high fidelity, strong semantics, and low latency at extremely low sampling rates, significantly accelerating the magnetic resonance image reconstruction process and improving image quality. Attached Figure Description

[0015] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0016] Figure 1 The diagram shown is a flowchart of a magnetic resonance image reconstruction method provided in an embodiment of this application. Figure 2 The diagram shown is a schematic representation of the initial structure predictor provided in an embodiment of this application. Figure 3 The diagram shown is a schematic representation of a magnetic resonance image reconstruction method provided in an embodiment of this application. Figure 4 The diagram shown is a schematic diagram of a one-time calibration provided in an embodiment of this application; Figure 5 The diagram shown is a structural schematic of a magnetic resonance image reconstruction device provided in an embodiment of this application; Figure 6 The diagram shown is a structural schematic of an electronic device provided in an embodiment of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Furthermore, in exemplary embodiments, since the same reference numerals denote the same components having the same structure or the same steps of the same method, if one embodiment has been described by way of example, then in other exemplary embodiments only structures or methods different from those described in the embodiment will be described.

[0019] Throughout the specification and claims, when a component is described as being “connected” to another component, that component may be “directly connected” to the other component or “electrically connected” to the other component via a third component. Furthermore, unless explicitly stated otherwise, the term “comprising” and its corresponding terms should be understood only to include the stated component and not to exclude any other component.

[0020] Among related technologies, magnetic resonance imaging (MRI) is widely used in medical fields such as neuroimaging, cardiovascular examination, and tumor monitoring due to its high-contrast imaging capabilities for soft tissues and its radiation-free nature. However, the traditional MRI acquisition process is relatively slow, with scan times typically lasting several minutes. This limits the turnaround efficiency of clinical equipment and is prone to image artifacts caused by patient movement, affecting the quality of subsequent diagnoses. In high-throughput scenarios such as pediatrics and emergency medicine, the conflict between imaging efficiency and quality is becoming increasingly prominent.

[0021] To improve imaging efficiency, the current MRI field mainly employs three types of acceleration techniques: parallel imaging, compressed sensing (CS), and deep learning methods. Parallel imaging relies on the spatial redundancy of multi-channel receiving coils, but the reconstruction quality degrades significantly at high acceleration rates. CS technology recovers high-quality images by introducing the assumption of transform domain sparsity, which can improve fidelity to some extent, but it is sensitive to sampling strategies, and its nonlinear optimization process is computationally expensive. Deep learning methods have developed rapidly in recent years, especially deconvolutional structures, which embed physical models into trainable networks, achieving a balance between performance and speed. However, these methods often rely on a fixed number of iterations, resulting in structural overfitting and long inference overhead, and still struggle to accurately recover details under certain undersampling conditions.

[0022] Diffusion models, as an emerging generative approach, model the probability distribution of images based on progressive denoising and have demonstrated leading performance in multiple image reconstruction tasks. Diffusion methods can adaptively learn data distributions, progressively recovering structural information while preserving physical constraints, theoretically possessing stronger generalization and detail modeling capabilities. Particularly in CS-MRI tasks, diffusion models combined with scoring functions have proven effective in mitigating structural artifacts and texture loss. However, the practical application of current diffusion models in MRI still faces three key challenges: First, mainstream diffusion reconstruction workflows require starting from pure noise and undergoing dozens to hundreds of iterations, performing high-dimensional score estimation in each frame, resulting in high inference costs and low reconstruction efficiency, limiting clinical applicability. Second, the diffusion gradient in the image domain and the direction of consistency in the frequency domain measurements are often inconsistent, causing conflicts between consistency correction and the generation path, easily leading to oscillations and edge blurring in the generation process, especially under high aliasing or low signal-to-noise conditions, resulting in unstable image detail recovery. Third, in the absence of structural guidance, the sampling path can easily introduce anatomical "illusions," potentially generating textures in unobserved areas that do not conform to physiological structures, affecting clinical reliability.

[0023] To alleviate these problems, existing research has attempted to accelerate optimization and stabilize modeling through conditions-guided approaches, latent space diffusion, or test-time adaptation. These methods reduce the number of inference steps to some extent and introduce physical priors to improve consistency performance, but they still generally rely on multi-step diffusion chains, resulting in limited overall efficiency improvements, insufficient structural control, and a lack of interpretability and local accuracy guarantees.

[0024] Therefore, the key issue in current MRI reconstruction is how to significantly reduce diffusion sampling costs without sacrificing image quality, and how to introduce structural priors to improve the anatomical consistency and clinical reliability of the generated images.

[0025] Specifically, the following three technical issues urgently need to be addressed: i) avoiding starting from pure noise to reduce unnecessary multi-step denoising processes; ii) establishing a unified structural prediction and adaptive consistency mechanism to achieve reliable reconstruction with fewer sampling steps; iii) achieving joint modeling of frequency domain consistency and spatial structure perception in single-step diffusion structures, thereby realizing an efficient, stable and structurally accurate MRI image restoration process.

[0026] To address the technical problems of existing diffusion models in magnetic resonance image reconstruction, such as reliance on multi-step sampling, high inference costs, and weak structural control, this application provides a magnetic resonance image reconstruction method. Figure 1 The diagram shown is a schematic flowchart of a magnetic resonance image reconstruction method provided in an embodiment of this application. Figure 1 As shown, the magnetic resonance image reconstruction method includes:

[0027] Step S11: Obtain measurement data of the undersampled preset space, and generate a zero-filled initial image based on the measurement data;

[0028] The preset space is preferably k-space. Undersampled k-space measurement data are acquired through scanning systems in medical fields such as neuroimaging, cardiovascular examination, and tumor monitoring. First, define the acquisition operator. The single-coil case can be represented as The case with multiple coils can be represented as follows: ,in, For two-dimensional Fourier operators, For the selection operator corresponding to the sampling mask, for Sensitivity diagrams of individual receiving coils and multiple receiving coils. It can be obtained through common algorithms such as ESPIRiT / JSENSE or pre-calibrated automatic methods. The measurement data meets the requirements. And noise . Measurement data After filling the full-frequency matrix grid and setting the unsampled positions to zero, an inverse Fourier transform is performed to obtain the zero-filled initial image. Among them, a single coil directly yields a zero-fill initial image. In one embodiment, multiple coils can be obtained by channel-by-channel reconstruction. Then, by combining coils based on the sensitivity map, a zero-fill initial image is obtained. The zero-padding initial image expands the complex numbers channel by channel with their real and imaginary parts to form the network input tensor. , These are the pixel height and width, respectively. Corresponding to real / virtual channels, This represents the number of slices.

[0029] Step S12: Perform forward estimation on the zero-filled initial image in the image domain to obtain the initial reconstructed image.

[0030] The zero-padding initial image is used as the sole input to the initial structure predictor. A forward estimation is performed directly on the zero-padding inverse Fourier result in the image domain to eliminate artifacts and perform preliminary image reconstruction. Initial Structure Predictor A lightweight encoder-decoder architecture is employed to perform a single forward prediction for large-scale anatomical structures to obtain the initial reconstructed image. Encoding and decoding can employ any or a combination of depthwise separable convolution, 3D or 2D convolution, state space units and / or attention units (including but not limited to Mamba-type units and Transformer-type units); multi-scale downsampling and upsampling paths can be set and multi-scale feature fusion can be performed through skip connections; the number of channels, depth and downsampling ratio can be adaptively set according to computational resources and clinical latency constraints.

[0031] Step S13: Perform iterative processing of adaptive data consistency correction and single-step diffusion refinement on the initial reconstructed image to obtain the target reconstructed image.

[0032] First, consistency correction is performed on the initial reconstructed image, and then single-step diffusion thinning is applied to the correction result to obtain a new reconstructed image. The next round of consistency correction and single-step diffusion thinning is then performed based on the zero-filled initial image and the new reconstructed image. After the iterations are complete, the final reconstructed image is used as the target reconstructed image.

[0033] This invention provides a magnetic resonance imaging (MRI) image reconstruction method. It acquires measurement data in an undersampled preset space and generates a zero-filled initial image based on the measurement data. Forward estimation of the zero-filled initial image in the image domain yields an initial reconstructed image. Iterative processing of the initial reconstructed image, including adaptive data consistency correction and single-step diffusion refinement, yields a target reconstructed image. This single-step diffusion framework, starting with a structural prediction map and combined with adaptive data consistency correction and anatomical prior guidance strategies, achieves efficient, stable, and controllable MRI image reconstruction. By replacing noisy sampling with structural initialization, it significantly reduces iteration steps and computational overhead, while mitigating the conflict between diffusion and physical constraints. This improves texture restoration quality and 3D spatial consistency, maintaining high fidelity, strong semantics, and low latency even at extremely low sampling rates. This provides a novel solution for the clinical deployability of diffusion models in medical image reconstruction, significantly accelerating the MRI image reconstruction process and improving image quality.

[0034] In a preferred embodiment, the step of forward estimation of the zero-padding initial image in the image domain to obtain an initial reconstructed image includes: calculating shallow features based on the zero-padding initial image; performing multi-level downsampling and encoding on the shallow features to obtain bottleneck features; performing multi-level upsampling and decoding on the bottleneck features, and fusing them with encoded features at each scale to obtain full-resolution features; and superimposing the full-resolution features and the shallow features and performing image reconstruction to obtain the initial reconstructed image.

[0035] This embodiment addresses the structural initialization and reconstruction stage of MRI undersampled reconstruction. It directly performs a forward estimation of the zero-filled inverse Fourier transform result in the image domain to eliminate artifacts and initially reconstruct the image. The input variable is the zero-filled initial image. Fill the initial image with zeros. Input Initial Structure Predictor An initial reconstructed image with preserved anatomical structure is obtained after one forward pass. This initial reconstructed image serves as the starting point for subsequent data consistency correction and diffusion refinement.

[0036] See Figure 2 Initial structure predictor A residual Mamba U-Net encoder-decoder architecture is adopted, consisting of shallow feature extraction, an encoder, multi-scale skip fusion and decoder, and an image reconstruction layer arranged in a "U" shape. The encoder and decoder both include downsampling and R3DMamba encoding. An initial structure predictor is also included. Only in the spatial dimension Perform upsampling and downsampling, slice axis To avoid losing cross-slice information, the tensor is kept unchanged; the last layout of the slice is used to reduce memory reordering overhead.

[0037] The shallow feature extraction module uses depthwise separable convolution (DWConv), which is... Depth convolution and Concatenated point convolutions map the input to baseline channel features. ,in, As the baseline number of channels, the preferred number in the embodiments of this application is... To balance throughput and expressive power, DWConv significantly reduces multiply-accumulate operations while preserving edge responses, making it a low-cost entry point for subsequent multi-scale coding.

[0038] encoder total Level, the preferred embodiment of this application is... The i-th level encoding consists of "downsampling + R3DMamba blocks": first, spatial convolution with stride 2 or nearest neighbor upsampling / downsampling are combined. Convolution yields Then it is fed into the R3DMamba block to get Tensor shape and Consistent. (The sentence is incomplete and lacks context.) Bottleneck characteristics are obtained after level. Bottleneck features can be cascaded with multiple R3DMambas to enhance the receptive field without changing the tensor size.

[0039] The decoder and encoder are set symmetrically, totaling Level. (Number) Level decoding first upsamples the result of the previous level to obtain... ,in Subsequently, it is encoded with features of the same scale. The jump adapter, defined as "DWConv→ReLU→UpSample", is used for fusion via lightweight jump connections and is responsible for combining the jumps. Aligned to the spatial resolution and number of channels of the current decoding scale; after fusion, it is fed into the R3DMamba block for refinement. .when Full-resolution features are obtained in time This skip path supplements shallow localization and texture information with minimal parameter cost, ensuring that large-scale anatomical contours are not overly smoothed during the decoding stage.

[0040] The image reconstruction layer uses a lightweight ReConv to incorporate full-resolution features. Projecting back to the number of input channels yields .

[0041] The R3DMamba block is a three-branch fusion module of spatial bidirectional and slice-oriented state space models (SSMs) used to couple long-range dependencies between the spatial and slice axes at each scale. Given input features... First, two tensors are obtained through linear projection. , For the number of channel extensions, take... ,in A non-negative modulation map is formed using ReLU. Then, three token streams are constructed: a spatially bidirectional token stream and a slice-oriented token stream. In the spatially bidirectional token stream, tokens are ordered according to their spatial indices. Flattened and rearranged in spatial dimension Perform one-dimensional state space recursion along the forward and backward directions of the space respectively (denoted as...). and And restored to its original state through 1D convolution. get and In the forward token stream of the slice, Rearranged along the slice axis and along Axis execution get After the three token streams are modulated and accumulated element by element along the same dimension, they are linearly mapped back to the channel number. get , and then input The block output is formed by summing the residuals at a learnable scale. Since the three recursions proceed along the forward spatial direction, the backward spatial direction, and the forward slice direction, respectively, and the sequence length along the slice axis is... Therefore, the overall complexity increases with the number of slices. It grows approximately linearly, thus it can naturally generalize between two-dimensional slices of different shapes and long-axis volume data, and stably output consistent structural representations across slices while maintaining throughput.

[0042] To facilitate engineering implementation and reproduction of the experimental procedure, the following sequence can be followed: S101, receive undersampled data. Space observation data With sampling mask The zero-filled initial image is obtained by performing zero-filled inverse Fourier transform and necessary density compensation. S102, Calculate shallow features S103, the bottleneck features are obtained by sequentially applying three layers of "downsampling + R3DMamba" encoding. S104, through three layers of "upsampling + R3DMamba" decoding and after being adapted with the encoded features of each scale via "DWConv→ReLU→UpSample", is added to achieve multi-scale jump fusion and obtain full-resolution features. S105, the initial reconstructed image is obtained by mapping to the input channel via ReConv. And can be pressed Apply the identity residual. At this point, the structural initialization has recovered a clear and robust large-scale anatomical structure in a single forward inward pass, providing a shape-consistent and cross-slice-stable starting solution for subsequent data consistency correction and single-step diffusion refinement.

[0043] In a preferred embodiment, the iterative processing of adaptive data consistency correction and single-step diffusion refinement on the initial reconstructed image to obtain the target reconstructed image includes: using the initial reconstructed image as a first reconstructed image, performing adaptive data consistency correction based on measurement residuals according to the first reconstructed image and the zero-filled initial image to obtain a correction result; using the correction result as input, performing detail enhancement processing through a single-step diffusion refiner to obtain a second reconstructed image; updating the first reconstructed image to the second reconstructed image, iteratively performing adaptive data consistency correction and single-step diffusion refinement processing until a preset condition is met, wherein the preset condition is that the single-step diffusion refiner converges or reaches a preset number of iterations; and using the second reconstructed image that meets the preset condition as the target reconstructed image.

[0044] See Figure 3 Consistency correction and single-step diffusion refinement are performed alternately and iteratively. That is, in the first iteration, the initial reconstructed image obtained in step S12 is used. As the first reconstructed image, adaptive data consistency correction based on measurement residuals is performed on the first reconstructed image and the zero-filled initial image to obtain the correction result. The correction result is then subjected to context grouping processing to obtain... Simultaneously, medical pre-treatment is performed, according to... The medical prior injection processing results are then enhanced with details using a single-step diffusion refiner to obtain the second reconstructed image, thus completing the first round of processing. In subsequent rounds, the second reconstructed image from the previous round is used as the first reconstructed image for the next round. Based on this first reconstructed image and the zero-filled initial image, consistency correction and single-step diffusion refinement are performed to obtain a new second reconstructed image. This process is repeated iteratively until a preset condition is met. The preset condition is that the single-step diffusion refiner converges or reaches a preset number of iterations. Convergence of the single-step diffusion refiner means that the loss function value of the single-step diffusion refiner is sufficiently small.

[0045] In a preferred embodiment, the step of performing adaptive data consistency correction based on measurement residuals to obtain a correction result based on the first reconstructed image and the zero-filled initial image includes: obtaining a back projection of the measurement residual based on the first reconstructed image and the zero-filled initial image, wherein the back projection characterizes the residual between the zero-filled initial image and the result of processing the first reconstructed image using a preset acquisition operator; determining the step size for consistency correction; and obtaining the correction result based on the step size, the first reconstructed image, and the back projection.

[0046] The one-time correction is an adaptive projection based on the measurement residual, and its update can be expressed as: ,in, For acquisition operators (single coil or multiple coils). Its accompanying operator; step size The back projection can be adaptively predicted by a lightweight network based on the back projection residuals, and is limited to a stable interval. Its form can be a scalar or pixel-level step-size map to achieve spatially variable updates. This one-time correction can be iteratively performed no more than a preset number of rounds to balance convergence and time delay, and can also be performed alternately or in parallel with diffusion refinement. The back projection of the measured residuals is then analyzed. Its shape is consistent with the input image. To avoid the image domain update deviating from the actual measurement constraints, the step size is set to [value missing]. The preferred range is [0,1], and an adaptive data consistency correction is inserted before diffusion refinement to bring the estimate back to the vicinity of the physical manifold. The k-th round correction is defined as... ,in For the collection operator The adjoint operator, For measured values, Based on current estimates, This is the step size for this round. Preferably, consistency correction and subsequent single-step diffusion refinement are performed alternately, and the total number of rounds does not exceed [a certain number]. To achieve stable convergence within a near real-time delay.

[0047] In one embodiment, determining the step size for consistency correction includes: normalizing the back projection; applying a residual sensing network to process the normalized back projection, and combining it with interval mapping to obtain an adaptive step size.

[0048] See Figure 4 The adaptive step size in the kth round Predicted by a residual sensing network. This network is a simplified convolutional neural network. The input is the back projection of the measured residual. The output is a step size map consistent with the spatial dimensions. The preferred structure is a stack of three convolutional layers: the first layer... Convolution (16 channels) + ReLU, second layer Convolution (16 channels) + ReLU, third layer Convolution (1 channel) + Sigmoid; for numerical stability, Before being fed into the network, the amplitude is normalized according to its range. The final step size is obtained through interval mapping. ,ensure Its preferred range is [0, 1, 2]. During training and inference, this residual perceptual network enables... In the first few rounds after initialization, larger values ​​are taken to enhance physical consistency, and the system naturally converges to a stable range near step 3, thereby suppressing residual aliasing caused by underprojection while avoiding over-smoothing of hard projection.

[0049] In a preferred embodiment, the step of using the correction result as input and performing detail enhancement processing through a single-step diffusion refiner to obtain a second reconstructed image includes: grouping the correction result into contextual groups based on structural similarity to obtain three-channel grouped data; processing the three-channel grouped data using a medical structural encoder and combining global average pooling and linear projection to obtain a guiding token, wherein the guiding token is a conditional signal input to the single-step diffusion refiner; and applying the single-step diffusion refiner to perform structural alignment single-step diffusion refinement on the three-channel grouped data according to the guiding token to obtain the second reconstructed image.

[0050] To improve the interlayer consistency of volumetric data, volumetric context modeling is performed without introducing heavy 3D diffusion: Centered on the target slice, a "three-slice group" is preferably formed by adaptively selecting one slice above and one slice below it as the input condition; the selection of adjacent slices can be filtered based on a structural similarity threshold (e.g., structural similarity index) with the central slice, reverting to only the central slice if the threshold is not met; the context of the slice axis can also be achieved through pyramid features, deformable sampling, or cross-slice gating units. The above strategy can also be replaced by any equivalent scheme capable of propagating interlayer structural consistency.

[0051] To suppress non-physiological textures in unobserved areas and enhance the credibility of pathology-related structures, anatomical priors can be introduced through conditional encoding. These priors can be derived from features generated by a pre-trained medical segmentation / detection / registration encoder, after dimensionality reduction, and used as guiding tokens / key-value pairs, injected into the cross-attention path or conditional normalization path of the diffusion network. The source, injection location, and form of the priors are not limited; equivalent alternatives such as radiomics priors, shape priors, or anatomical atlas priors can also be used.

[0052] In the embodiments of this application, see Figure 3 and Figure 4 After each round of consistency correction, with As input, perform structure-aligned single-step diffusion refinement on target slice index i to obtain First, group the body context: from Extract Three slices, ranked according to structural similarity thresholds Neighboring valid slices are selected based on the structural similarity index measurement (SSIM). Preferably, If the SSIM of adjacent slices and the center slice is lower than the similarity threshold... To preserve the input dimension, a central slice is used instead, forming a three-channel group. The three-channel group The type is real, and can be normalized to [−1, 1]. Then, a priori dissection injection is performed: [The text abruptly ends here, likely due to an incomplete or corrupted source]. High-resolution semantic features are obtained by feeding the frozen medical structural encoder (e.g., a multi-organ segmentation backbone network). The bootstrap token is obtained through global average pooling and linear projection, resulting in a compactly dissected token. (For example, when using SDv2.0 for the diffusion model) The guiding token is injected as a conditional signal into the diffusion refiner across the attention key / value or conditional normalization (FiLM) path to constrain the generated physiological credibility.

[0053] After obtaining the three-channel grouped data and the guide token, the single-step diffusion refiner is applied to perform structural alignment single-step diffusion refinement on the three-channel grouped data according to the guide token to obtain the second reconstructed image.

[0054] In a preferred embodiment, the step of applying the single-step diffusion refiner to perform structure-aligned single-step diffusion refinement of the three-channel grouped data according to the guiding token to obtain the second reconstructed image includes: applying the variational autoencoder in the single-step diffusion refiner to map the three-channel grouped data to the latent space to obtain latent space image data; applying the backbone network in the single-step diffusion refiner in combination with the guiding token to perform single-step distillation on the latent space image data to obtain latent space prediction data; and restoring image domain details from the latent space prediction data using a decoder to obtain the second reconstructed image.

[0055] The diffusion refiner uses an existing large-scale RGB pre-trained diffusion model as its base (the U-Net generator and VAE encoder / decoder maintain their original weights). A lightweight medical domain adaptation (LoRA rank optimization) is performed on the U-Net backbone and VAE encoder / decoder using a low-rank adapter. Fixed diffusion time step To achieve single-step distillation. Specifically, to through Mapped to latent space The latent space prediction of the detailed residuals is obtained by feeding the conditional U-Net. After Restored to image domain details If the MRI is in complex form, it can be set as follows: These correspond to the real and imaginary parts, respectively; if it is an amplitude diagram, then... Single-step diffusion refiner with or Instead of starting with Gaussian noise, only one generation denoising jump is performed to compensate for mid-to-high frequency details. To avoid noise amplification and emphasize diagnostically relevant bands, adversarial or divergence metrics for spectral focusing can be introduced during the training phase, such as matching constraints on the bandpass-filtered content. ,in It is a bandpass filter (the frequency band can be learned or preset). This serves as the discriminator. The above form can be equivalently replaced by other statistical divergence or perceived consistency measures.

[0056] In a preferred embodiment, before performing forward estimation on the zero-padded initial image in the image domain to obtain the initial reconstructed image, the process includes: training an initial structure predictor based on historical zero-padded initial images to obtain an initially trained initial structure predictor and corresponding historical initial reconstructed images, wherein the initial structure predictor is used for forward estimation on the zero-padded initial images; training a single-step diffusion refiner based on the historical initial reconstructed images to obtain an initially trained single-step diffusion refiner; and jointly fine-tuning the initially trained initial structure predictor and the initially trained single-step diffusion refiner based on historical zero-padded initial images to obtain trained initial structure predictors and single-step diffusion refiners.

[0057] In this embodiment of the application, each model needs to be trained first, including an initial structure predictor. Residual sensing network This includes training the diffusion refiner and other components. Training can employ a progressive strategy: first train the initial structure predictor, then train a single-step diffusion refiner, and finally perform end-to-end joint fine-tuning; alternatively, distillation, low-rank adaptation, pruning, and quantization can be used to adapt to deployment. A phased training strategy ensures stable convergence and low-computational-power adaptation. First, only the initial structure predictor is trained independently. To converge, to make In terms of structure and energy distribution, and Alignment This is a real magnetic resonance image; it is then fixed. Joint training for adaptive step size Predictive residual sensing network With single-step diffusion refiner Among them, single-step diffusion refiner Based on an existing large-scale RGB pre-trained diffusion model, the VAE is fully frozen, with only minor LoRA fine-tuning performed on the U-Net backbone. After these two stages, end-to-end fine-tuning is performed to eliminate amplitude and phase residuals while maintaining the one-step inference topology unchanged. To improve cross-device and cross-protocol robustness, the training phase performs adjustments to the image domain and... The space also employs data augmentation, including phase jitter, random undersampling mask perturbation, amplitude normalization, random field-of-view cropping, voxel flipping, and slight rotation, and is gradually increased through on-demand multi-scale cropping training courses. The optimizer employs the Adam family of algorithms, combined with cosine annealing or piecewise linear learning rates. Weight decay and gradient pruning are used to stabilize training of high-frequency adversarial components. To further reduce memory usage and latency while maintaining real-time performance, the fine-tuning stage can incorporate distillation (using multi-step diffusion as the teacher), structured pruning, and fixed-point quantization. Mixed-precision training and checkpoint recalculation are also used, without altering the inference flow of single-step diffusion refinement and adaptive consistency.

[0058] Single-step diffusion refiner To highlight the diagnostic relevance of high-frequency reconstruction and suppress noise amplification during the training phase, adversarial constraints on the high-frequency components are introduced. For the bandpass filter operator (preferably with a bandwidth of [0.15, 0.4] cycles / pixel), the discriminator... Only Domain to determine authenticity, but This is to narrow the difference between the reconstructed frequency band and the actual diagnostic frequency band distribution and to avoid amplifying ultra-high frequency noise.

[0059] The training objective simultaneously considers physical consistency, spectral detail, and perceptual quality, forming a stable one-step refinement learning framework. For example, the total loss can be defined as , in, Based on validation set grid search or clinical preferences, weights and metrics can be adaptively set according to the task or selected by the validation set. Learned Perceptual Image Patch Similarity (LPIPS) and Structural Similarity Index (SSIM) are used to measure perceptual and structural similarity, respectively.

[0060] The diffusion refiner employs a more efficient model distillation strategy. It first has a powerful, but multi-step (e.g., 1000-step) traditional RGB domain diffusion model as the teacher model. Then, a student model with the same or similar structure is built, but designed to complete reconstruction in one or very few steps (e.g., 1-4 steps). Through specific training strategies (such as the aforementioned training objective), the knowledge gained by the "teacher model" through multiple iterations is "distilled" and transferred to the "student model," enabling it to learn how to directly predict the final denoising result in one step. Various loss functions (such as...) The design (such as spectrum adversarial loss) is precisely to ensure the accuracy of this single-step prediction.

[0061] To meet clinical real-time requirements during the inference phase, the adaptive consistency and single-step diffusion refinement can be combined to perform a small number of iterations. The number of iterations is set according to the sampling mode, acceleration factor, and latency requirements of the target department, achieving near real-time reconstruction without sacrificing image quality. An iteration limit is preferably set during the inference phase. Step 0 obtains the initial reconstructed image from step S12. Then, the "adaptive consistency correction" is executed repeatedly. Body context grouping Anatomy token injection "Single-step diffusion refinement" to measure residuals or detail gain As an early stopping criterion (preferred) ), reaching any condition or iterating to That is, output the target reconstructed image .

[0062] Quantitative and qualitative evaluations of the magnetic resonance image reconstruction method of this application embodiment were conducted under various undersampling modes and acceleration ratios (preferably covering ×4 to ×32). The results show that quantitative and qualitative evaluations were performed on a conventional graphics processing unit (GPU) platform with a configuration of no more than 4 iterations. Due to the use of a single diffusion time step and lightweight conditional injection, the magnetic resonance image reconstruction method of this application embodiment generally improves peak signal-to-noise ratio, structural similarity, and perceptual metrics, and achieves high voxel-level consistency in downstream segmentation tasks. The single-chip inference time is approximately 0.81 seconds, significantly reducing latency compared to multi-step diffusion, meeting the high-throughput and near-real-time clinical needs of outpatient clinics. To meet different deployment scenarios, GPU acceleration or CPU edge acceleration can be selected; the former enables mixed precision and tensor cores, while the latter enables operator fusion, low-bit quantization, and automatic batch size scheduling.

[0063] The magnetic resonance imaging reconstruction method of this application does not depend on specific sampling modes or equipment manufacturer parameters, can be adapted to single-coil or multi-coil systems, and can be uniformly applied to two-dimensional slices and three-dimensional volume data. It replaces noise-based inference with structural prediction, and combines adaptive consistency and anatomical prior single-step diffusion refinement to maintain high fidelity, robustness, and low latency even under low sampling rates and high acceleration conditions, making it clinically feasible. Equivalent substitutions or adjustments can also be made to module partitioning, network topology, loss form, iteration rounds, and hardware carriers. For example, shape maps or registration fields can replace segmentation features as anatomical priors, bandpass weighted integral probability metrics can replace bandpass adversarial constraints, or changes can be made while ensuring single-step inference. The value of is related to the setting of the LoRA rank.

[0064] The magnetic resonance imaging reconstruction method in this application replaces noise initiation with structural prediction, reducing the sampling chain length by more than 100 times and lowering the inference time to approximately 1%–2% of traditional diffusion-based methods (0.81 s / image vs. 43.89–81.31 s / image), achieving second-level latency and approaching the deployment threshold of end-to-end (E2E) methods. Through adaptive data consistency, single-step refinement is performed near the physical manifold. Under multiple sampling modes and acceleration rates (×4–×32), compared to traditional diffusion-based methods, the peak signal-to-noise ratio (PSNR) is improved by approximately 2–17 dB, SSIM is increased by up to +0.49, and LPIPS is decreased by up to 0.43, resulting in clearer and more stable edges and textures. By introducing volume consistency and anatomical priors as conditional guidance, it remains robust across anatomical regions and at extreme accelerations (e.g., ×16–×32), with a high PSNR. In extreme scenarios, it can outperform by 6–17 dB, effectively suppressing structural illusions and improving 3D continuity. The magnetic resonance image reconstruction method of this application has a lightweight and easy-to-deploy overall framework, achieving stable convergence in just 4 iterations. It is compatible with single-coil / multi-coil systems and improves Dice to 0.9544 in downstream segmentation (approximately 1.7% improvement compared to traditional diffusion methods and approximately 6.2% improvement compared to E2E methods), demonstrating comprehensive clinical usability. The magnetic resonance image reconstruction method of this application focuses on single-step diffusion refinement with "structure alignment," enhancing details in diagnostically sensitive frequency bands while maintaining physical consistency. It achieves a balance between speed, quality, and stability (reducing latency to approximately 1%–2% compared to traditional diffusion methods, with significantly improved quality indicators), making it suitable for clinical application and large-scale deployment.

[0065] This invention provides a magnetic resonance image reconstruction method that acquires measurement data of an undersampled preset space and generates a zero-filled initial image based on the measurement data. Forward estimation of the zero-filled initial image is performed in the image domain to obtain an initial reconstructed image. Iterative processing of adaptive data consistency correction and single-step diffusion refinement is then applied to the initial reconstructed image to obtain a target reconstructed image. With "structure alignment" single-step diffusion refinement as the core, combined with adaptive data consistency correction, efficient, stable, and controllable magnetic resonance image reconstruction is achieved. By replacing noise sampling with structure initialization, the iterative steps are significantly reduced, computational overhead is lowered, and the conflict between diffusion and physical constraints is alleviated, improving texture restoration quality and three-dimensional spatial consistency. This allows high fidelity, strong semantics, and low latency to be maintained even at extremely low sampling rates, providing a novel solution for the clinical deployability of diffusion models in the field of medical image reconstruction. It can maintain high fidelity, strong semantics, and low latency at extremely low sampling rates, significantly accelerating the magnetic resonance image reconstruction process and improving image quality.

[0066] Figure 5 The diagram shown is a structural schematic of a magnetic resonance image reconstruction device provided in an embodiment of this application. Figure 5 As shown, the magnetic resonance image reconstruction device includes:

[0067] The initial image generation unit 501 is used to acquire measurement data of an undersampled preset space and generate a zero-filled initial image based on the measurement data;

[0068] The initial reconstruction unit 502 is used to perform forward estimation on the zero-filled initial image in the image domain to obtain the initial reconstructed image;

[0069] The target reconstruction unit 503 is used to perform iterative processing of adaptive data consistency correction and single-step diffusion refinement on the initial reconstructed image to obtain the target reconstructed image.

[0070] In one embodiment, the target reconstruction unit 503 is configured to: use the initial reconstructed image as a first reconstructed image, perform adaptive data consistency correction based on measurement residuals according to the first reconstructed image and the zero-filled initial image to obtain a correction result; use the correction result as input, perform detail enhancement processing through a single-step diffusion refiner to obtain a second reconstructed image; update the first reconstructed image to the second reconstructed image, and iteratively perform adaptive data consistency correction and single-step diffusion refinement processing until a preset condition is met, wherein the preset condition is that the single-step diffusion refiner converges or reaches a preset number of iterations.

[0071] In one embodiment, the target reconstruction unit 503 is further configured to: obtain a back projection of the measurement residual based on the first reconstructed image and the zero-filled initial image, wherein the back projection characterizes the residual between the zero-filled initial image and the result of processing the first reconstructed image using a preset acquisition operator; determine the step size for consistency correction; and obtain a correction result based on the step size, the first reconstructed image, and the back projection.

[0072] In one embodiment, the target reconstruction unit 503 is further configured to: normalize the back projection; apply a residual sensing network to process the normalized back projection, and combine it with interval mapping to obtain an adaptive step size.

[0073] In one embodiment, the target reconstruction unit 503 is further configured to: group the correction results based on structural similarity to obtain three-channel grouped data; process the three-channel grouped data using a medical structural encoder, and combine global average pooling and linear projection to obtain a guiding token, wherein the guiding token is a conditional signal input to the single-step diffusion refiner; and apply the single-step diffusion refiner to perform single-step diffusion refinement of the three-channel grouped data based on the guiding token to obtain a second reconstructed image.

[0074] In one embodiment, the target reconstruction unit 503 is configured to: apply the variational autoencoder in the single-step diffusion refiner to map the three-channel grouped data to the latent space to obtain latent space image data; apply the backbone network in the single-step diffusion refiner in combination with the guiding token to perform single-step distillation on the latent space image data to obtain latent space prediction data; and restore image domain details from the latent space prediction data using the decoder to obtain a second reconstructed image.

[0075] In one embodiment, the apparatus further includes a training unit, which is configured to: train an initial structure predictor based on a historical zero-filled initial image to obtain an initially trained initial structure predictor and a corresponding historical initial reconstructed image, wherein the initial structure predictor is used to perform forward estimation on the zero-filled initial image; train a single-step diffusion refiner based on the historical initial reconstructed image to obtain an initially trained single-step diffusion refiner; and jointly fine-tune the initially trained initial structure predictor and the initially trained single-step diffusion refiner based on the historical zero-filled initial image to obtain a trained initial structure predictor and a single-step diffusion refiner.

[0076] This invention provides a magnetic resonance image reconstruction device that acquires measurement data of an undersampled preset space and generates a zero-filled initial image based on the measurement data. Forward estimation is performed on the zero-filled initial image in the image domain to obtain an initial reconstructed image. Iterative processing of adaptive data consistency correction and single-step diffusion refinement is then applied to the initial reconstructed image to obtain a target reconstructed image. With single-step diffusion refinement based on "structure alignment" as the core, combined with adaptive data consistency correction, efficient, stable, and controllable magnetic resonance image reconstruction is achieved. By replacing noise sampling with structure initialization, the iterative steps are significantly reduced, computational overhead is lowered, and the conflict between diffusion and physical constraints is alleviated, improving texture restoration quality and three-dimensional spatial consistency. This allows high fidelity, strong semantics, and low latency to be maintained even at extremely low sampling rates, providing a novel solution for the clinical deployability of diffusion models in the field of medical image reconstruction. It can maintain high fidelity, strong semantics, and low latency at extremely low sampling rates, significantly accelerating the magnetic resonance image reconstruction process and improving image quality.

[0077] According to another aspect of the present invention, one embodiment of the present invention provides an electronic device, Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0078] For example, such as Figure 6As shown, the electronic device includes a memory 601 and a processor 602, wherein the memory 601 stores an executable computer program 6011, and the processor 602 is used to call and execute the executable computer program 6011 from the memory, so that the processor 602 performs a magnetic resonance image reconstruction method.

[0079] This embodiment can divide the electronic device into functional modules according to the above method embodiment. For example, each module can correspond to a separate functional module, or two or more functions can be integrated into one processing module. The integrated module can be implemented in hardware. It should be noted that the module division in this embodiment is illustrative and is only a logical functional division. In actual implementation, there may be other division methods.

[0080] When each functional module is divided according to its corresponding function, the electronic device may include: an initial image generation unit, an initial reconstruction unit, and a target reconstruction unit, etc.

[0081] It should be noted that all relevant content of each step involved in the above method embodiments can be referenced from the functional description of the corresponding functional module, and will not be repeated here.

[0082] The electronic device provided in this embodiment is used to execute the above-described co-magnetic resonance image reconstruction method, and thus can achieve the same effect as the above-described implementation method.

[0083] When using integrated units, the electronic device may include a processing module and a storage module. The processing module is used to control and manage the operation of the electronic device. The storage module is used to support the electronic device in executing computer programs and processing data.

[0084] The processing module may be a processor or a controller, which can implement or execute various exemplary logic blocks, modules, and circuits as disclosed in this application. The processor may also be a combination of computing functions, such as a combination of one or more microprocessors, a combination of digital signal processing (DSP) and microprocessors, etc., and the storage module may be a memory.

[0085] This embodiment also provides a computer-readable storage medium (including but not limited to disk storage, CD-ROM, optical storage, etc.) storing computer program code. When the computer program code is run on a computer, the computer executes the above-mentioned related method steps to implement the magnetic resonance image reconstruction method provided in the above embodiment. The computer-readable storage medium may include, but is not limited to, any type of disk, including floppy disks, optical disks, Digital Video Discs (DVDs), Compact Disc Read-Only Memory (CD-ROM), microdrives, and magneto-optical disks, read-only memory (ROM), random access memory (RAM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), dynamic random access memory (DRAM), video random access memory (VRAM), flash memory devices, magnetic cards or optical cards, nanosystems (including molecular memory ICs), or any type of medium or device suitable for storing instructions and / or data.

[0086] This embodiment also provides a computer program product that, when run on a computer, causes the computer to perform the aforementioned related steps to implement a magnetic resonance image reconstruction method provided in the above embodiment.

[0087] The beneficial effects of the above embodiments can be referred to the beneficial effects of the corresponding methods provided above, and will not be repeated here.

[0088] The basic principles of this application have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this application are merely examples and not limitations, and should not be considered as essential features of each embodiment of this application. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the application to the necessity of employing the aforementioned specific details for implementation.

[0089] The block diagrams of devices, apparatuses, devices, and systems involved in this application are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as “comprising,” “including,” “having,” etc., are open-ended terms meaning “including but not limited to,” and are used interchangeably with them. The terms “or” and “and” as used herein refer to the terms “and / or,” and are used interchangeably with them unless the context clearly indicates otherwise. The term “such as” as used herein refers to the phrase “such as but not limited to,” and is used interchangeably with it.

[0090] It should also be noted that in the apparatus or equipment of this application, the components or steps can be disassembled and / or recombined. These disassemblies and / or recombinations should be considered as equivalent solutions of this application.

[0091] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein can be applied to other aspects without departing from the scope of this application. Therefore, this application is not intended to be limited to the aspects shown herein, but rather to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0092] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this application to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

Claims

1. A magnetic resonance image reconstruction method, characterized in that, The method includes: Acquire measurement data of the undersampled preset space, and generate a zero-filled initial image based on the measurement data; The initial reconstructed image is obtained by forward estimation of the zero-padding initial image in the image domain. The initial reconstructed image is subjected to iterative processing of adaptive data consistency correction and single-step diffusion refinement to obtain the target reconstructed image.

2. The method according to claim 1, characterized in that, The iterative processing of adaptive data consistency correction and single-step diffusion refinement on the initial reconstructed image to obtain the target reconstructed image includes: Using the initial reconstructed image as the first reconstructed image, adaptive data consistency correction based on measurement residuals is performed on the first reconstructed image and the zero-filled initial image to obtain the correction result; Using the correction result as input, a detail enhancement process is performed through a single-step diffusion refiner to obtain a second reconstructed image; The first reconstructed image is updated to the second reconstructed image, and adaptive data consistency correction and single-step diffusion refinement are iteratively performed until a preset condition is met, wherein the preset condition is that the single-step diffusion refiner converges or reaches a preset number of iterations. The second reconstructed image that meets the preset conditions is used as the target reconstructed image.

3. The method according to claim 2, characterized in that, The adaptive data consistency correction based on measurement residuals, performed according to the first reconstructed image and the zero-filled initial image, to obtain the correction result, includes: The back projection of the measurement residual is obtained based on the first reconstructed image and the zero-filled initial image. The back projection represents the residual between the zero-filled initial image and the result of processing the first reconstructed image with a preset acquisition operator. Determine the step size for consistency correction; The correction result is obtained based on the step size, the first reconstructed image, and the back projection.

4. The method according to claim 3, characterized in that, The step size for determining consistency correction includes: The back projection is normalized. The normalized back projection is processed using a residual sensing network, and combined with interval mapping, an adaptive step size is obtained.

5. The method according to claim 2, characterized in that, The process of using the correction result as input and performing detail enhancement processing through a single-step diffusion refiner to obtain a second reconstructed image includes: The correction results are grouped in context based on structural similarity to obtain three-channel grouped data; The three-channel grouped data is processed by a medical structural encoder and combined with global average pooling and linear projection to obtain a guiding token, which is a conditional signal input to the single-step diffusion refiner. The single-step diffusion refiner is applied to perform structural alignment and single-step diffusion refinement on the three-channel grouped data according to the guiding token to obtain the second reconstructed image.

6. The method according to claim 5, characterized in that, The application of the single-step diffusion refiner to perform structural alignment and single-step diffusion refinement on the three-channel grouped data based on the guiding token, to obtain the second reconstructed image, includes: The variational autoencoder in the single-step diffusion refiner is used to map the three-channel grouped data to the latent space to obtain latent space image data. The backbone network in the single-step diffusion refiner is combined with the guiding token to perform single-step distillation on the latent space image data to obtain latent space prediction data. The decoder restores the image domain details from the latent space prediction data to obtain the second reconstructed image.

7. The method according to claim 1, characterized in that, Before performing forward estimation on the zero-padding initial image in the image domain to obtain the initial reconstructed image, the process includes: The initial structure predictor is trained based on the historical zero-padded initial image to obtain the initially trained initial structure predictor and the corresponding historical initial reconstructed image. The initial structure predictor is used to perform forward estimation on the zero-padded initial image. The single-step diffusion refiner is trained based on the historical initial reconstructed image to obtain the initially trained single-step diffusion refiner. The initial structure predictor and the initial single-step diffusion refiner trained based on the historical zero-filled initial image are jointly fine-tuned to obtain the trained initial structure predictor and single-step diffusion refiner.

8. A magnetic resonance image reconstruction device, characterized in that, The device includes: An initial image generation unit is used to acquire measurement data of an undersampled preset space and generate a zero-filled initial image based on the measurement data, wherein the measurement data includes multiple magnetic resonance images; An initial reconstruction unit is used to perform forward estimation on the zero-filled initial image in the image domain to obtain an initial reconstructed image; The target reconstruction unit is used to perform iterative processing of adaptive data consistency correction and single-step diffusion refinement on the initial reconstructed image to obtain the target reconstructed image.

9. An electronic device, characterized in that, The electronic device includes: Memory, used to store executable computer programs; A processor for calling and running the executable computer program from the memory, such that the processor performs the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed, implements the method as described in any one of claims 1 to 7.

Images