A fast magnetic resonance imaging reconstruction and segmentation method, system, medium and product

Abstract

The application discloses a fast magnetic resonance imaging reconstruction and segmentation method, system, medium and product, and relates to the technical field of image processing. The method comprises the following steps: decomposing original three-dimensional magnetic resonance data into a plurality of independent two-dimensional magnetic resonance data slices in a slice order; performing image reconstruction on the two-dimensional magnetic resonance data slices by using a pre-trained forward generator in a frequency domain Schrodinger bridge diffusion model to generate two-dimensional magnetic resonance images, wherein the forward generator is integrated with a Fourier convolution module for performing frequency domain transformation and feature enhancement processing on feature mapping in the image reconstruction process; inputting each two-dimensional magnetic resonance image into a pre-trained segmentation model in sequence to perform segmentation on a target region and a preset dangerous organ and generate corresponding two-dimensional segmentation masks; and rearranging and stacking all the two-dimensional segmentation masks in the slice order, and generating a three-dimensional segmentation result through a three-dimensional fusion operation. The above technical scheme can quickly and accurately perform magnetic resonance imaging reconstruction and segmentation.

Description

Technical Field

[0001] This application relates to the field of image processing, and in particular to a fast magnetic resonance imaging reconstruction and segmentation method, system, medium, and product. Background Technology

[0002] In the field of medical imaging, magnetic resonance imaging (MRI) technology has developed rapidly, providing crucial information for disease diagnosis and treatment. MRI can clearly present information about the internal tissue structure and physiological functions of the human body, helping doctors to more accurately detect lesions and assess conditions. With the deepening of medical research and increasing clinical needs, higher demands are being placed on the speed and accuracy of MRI. On the one hand, rapid imaging can reduce patient examination time and improve hospital examination efficiency; on the other hand, accurate imaging and segmentation results help doctors make more precise diagnoses and treatment plans.

[0003] Traditional magnetic resonance imaging (MRI) takes about 3 minutes (e.g., a standard T2-weighted 3D turbine spin echo sequence scan lasts 3 minutes and 22 seconds), and manual delineation of the target area and organs at risk (OAR) takes an average of 15-20 minutes. This results in a total procedure of 40-50 minutes for a single online adaptive radiotherapy session, which is far beyond the patient's tolerance range (usually ≤30 minutes). This not only reduces patient compliance (especially in children, the elderly, patients with pain or claustrophobia, who are prone to positional shifts and examination failures), but also affects treatment accuracy due to anatomical variations within the fractional range, while increasing the burden on clinical resources. Summary of the Invention

[0004] This application provides a fast magnetic resonance imaging reconstruction and segmentation method, system, medium, and product, which can perform magnetic resonance imaging reconstruction and segmentation quickly and accurately, improving imaging processing efficiency and quality.

[0005] In a first aspect, this application provides a fast magnetic resonance imaging reconstruction and segmentation method, the method comprising:

[0006] Acquire raw three-dimensional magnetic resonance data collected by a magnetic resonance scanning device, and decompose the raw three-dimensional magnetic resonance data into multiple independent two-dimensional magnetic resonance data slices along the slice sequence;

[0007] The two-dimensional magnetic resonance data slices are sequentially input into a pre-trained frequency domain Schrödinger bridge diffusion model. Image reconstruction is performed through the pre-trained forward generator in the frequency domain Schrödinger bridge diffusion model to generate a two-dimensional magnetic resonance image. The forward generator integrates a Fourier convolution module, which is used to perform frequency domain transformation and feature enhancement processing on the feature mapping during the image reconstruction process.

[0008] Each of the two-dimensional magnetic resonance images is sequentially input into a pre-trained segmentation model to segment the target area and preset dangerous organs, generating a corresponding two-dimensional segmentation mask.

[0009] All the two-dimensional segmentation masks are rearranged and stacked according to the slice order, and a three-dimensional segmentation result is generated through a three-dimensional fusion operation.

[0010] By employing the above technical solution, 3D data is split into multiple 2D slices for parallel processing, and an optimized frequency-domain Schrödinger bridge diffusion model is used. This significantly shortens the total time from magnetic resonance imaging to segmentation results, meeting the stringent speed requirements of online adaptive radiotherapy. Combining the frequency-domain Schrödinger bridge diffusion model with a Fourier convolution module enables more accurate learning of data distribution and enhanced high-frequency detail capture, resulting in reconstructed images with fewer artifacts and higher resolution, laying the foundation for subsequent accurate segmentation. Integrating reconstruction and segmentation, two independent tasks, into an automated serial processing flow avoids information loss and time delays in traditional step-by-step operations, ensuring accuracy matching and smooth workflow between the two stages. The entire process has low computational load and controllable memory consumption, making it well-suited for existing clinical hardware (such as MR-Linac devices).

[0011] In some embodiments, performing image reconstruction using a pre-trained forward generator in the frequency domain Schrödinger bridge diffusion model to generate a two-dimensional magnetic resonance image specifically includes:

[0012] Preset shallow feature extraction is performed on the two-dimensional magnetic resonance data slices to obtain initial latent features. The initial latent features are then combined with preset noise scheduling to form noisy latent variables.

[0013] The latent variables are input into the pre-trained forward generator, where a denoising process is performed on the latent variables through multiple cascaded residual blocks that integrate the Fourier convolution module, outputting clean latent features.

[0014] The clean latent features are upsampled and spatially mapped to generate the two-dimensional magnetic resonance image.

[0015] The above technical solution transforms the complex image reconstruction problem into precise denoising and optimization of the latent feature space. Compared to end-to-end black-box models, this structured denoising path offers a more controllable and stable reconstruction process, enabling more effective recovery of high-quality image structures from undersampled or noisy data. A key innovation is the integration of a Fourier convolution module into the core denoising stage (residual block). This module significantly enhances the model's ability to capture and reconstruct high-frequency information such as image edges and textures through feature transformation and enhancement in the frequency domain. This effectively compensates for the shortcomings of traditional spatial convolution in modeling global information and high-frequency details, resulting in fewer artifacts and clearer anatomical structures in the generated MRI images. Using cascaded residual blocks as the backbone network, combined with Fourier convolution, a deep and powerful feature extraction and transformation architecture is constructed. This design not only alleviates the gradient vanishing problem during deep network training, ensuring effective model training, but also ensures that the final output of clean latent features and reconstructed images possesses high visual and diagnostic realism through progressive feature processing, providing reliable input for subsequent accurate segmentation.

[0016] In some embodiments, the denoising process performed on the latent variables to output clean latent features includes:

[0017] In the current residual block, spatial domain feature extraction and frequency domain feature enhancement are performed in parallel. The spatial domain feature extraction includes extracting spatial dimension features from the input features through a spatial domain convolutional layer to obtain a first output result. The frequency domain feature enhancement includes: performing a two-dimensional fast Fourier transform on the input features to convert them into frequency domain features; decomposing the frequency domain features into real components and imaginary components; performing convolutional modulation on the real components to generate a first modulation result; performing convolutional modulation on the imaginary components to generate a second modulation result; and reconstructing the first modulation result and the second modulation result into spatial domain features through an inverse Fourier transform to obtain a second output result.

[0018] The first output result and the second output result are fused to obtain intermediate features. The intermediate features are then added to the input features of the current residual block through a skip connection to form the output features of the current residual block, which are then used as the input features of the next residual block.

[0019] The output features of the last residual block are used as the clean latent features.

[0020] By employing the aforementioned technical solution, spatial convolution and frequency domain enhancement are performed in parallel, simultaneously capturing both local spatial contextual information (through spatial convolution) and global texture detail information (through frequency domain transformation) of the image. This dual-path parallel architecture overcomes the limitations of single-domain processing, achieving more comprehensive and richer feature representations and laying an information foundation for high-quality image reconstruction. By transforming features to the frequency domain and modulating the real and imaginary parts separately, high-frequency components representing image details and edges can be directly and accurately manipulated. This enables the effective recovery of subtle structures and clear boundaries that are easily lost in traditional spatial domain models, significantly reducing blur and artifacts in the reconstructed image. Within each residual block, the fused intermediate features are added to the original input features through skip connections. This design constitutes a densely connected network, effectively mitigating the gradient vanishing problem in deep networks and ensuring the stability of the training process. Simultaneously, it directly transmits low-level information to deeper layers, protecting the integrity of the original data and preventing the decay of important features.

[0021] In some embodiments, the method further includes modulating the second output result, specifically including:

[0022] Generate a time-varying noise embedding vector, which is associated with the current time step of the frequency domain Schrödinger bridge diffusion model;

[0023] The time-varying noise embedding vector is input into a multilayer perceptron to generate scale parameters and displacement parameters.

[0024] The second output result is subjected to adaptive scaling and displacement compensation using the scale and displacement parameters.

[0025] By employing the above technical solution, and introducing a noise embedding vector associated with the current time step, the model can perceive and adapt to different noise intensities during the diffusion process, achieving dynamic adjustment. Using the generated scale and displacement parameters, the features output by the frequency domain path are finely and adaptively scaled and translated, optimizing feature representation. Through precise calibration of the frequency domain features, artifacts or distortions that may appear in the reconstructed image are effectively reduced, thereby improving the realism and detail fidelity of the final generated image.

[0026] In some embodiments, the method further includes training the forward generator by means of the following steps:

[0027] A training dataset is constructed, which contains multiple pairs of preset low-quality magnetic resonance images and paired high-quality magnetic resonance images. An initial bidirectional diffusion generation network containing a forward generator and a backward generator is constructed.

[0028] The low-quality magnetic resonance image is input into the forward generator to obtain a predicted high-quality image. The predicted high-quality image is then input into the backward generator to obtain a reconstructed noisy variable. A Schrödinger bridge loss function is calculated, which includes a first difference and a second difference. The first difference is the difference between the predicted high-quality image and the high-quality magnetic resonance image. The second difference is the second difference between the reconstructed noisy variable and the noisy variable obtained from the low-quality magnetic resonance image after a preset noise schedule. Based on the Schrödinger bridge loss function, the network parameters of the forward generator are updated through gradient backpropagation.

[0029] Repeat the previous step until the network parameters of the forward generator converge, thus obtaining the pre-trained forward generator.

[0030] By employing the aforementioned technical solution, and introducing a backward generator to form a closed loop with the forward generator, the training process no longer merely learns a unidirectional mapping from low-quality to high-quality data, but rather a bidirectional, consistent transformation relationship between the two. This guides the model to find a smoother and more reasonable transformation path between the two distributions, avoiding shortcuts that might deviate from the true data distribution, as can occur with traditional methods. The core of the Schrödinger Bridge loss function lies in its dual difference constraints: the first difference directly constrains the output result, ensuring that the predicted high-quality image closely approximates the real target in content; the second difference constrains the intermediate process, ensuring the reversibility and consistency of the entire generation path. This dual constraint provides a stronger and more precise guiding signal for model training, effectively forcing the generator to output images that are not only realistic but also structurally accurate, significantly improving the fidelity of the reconstruction. The closed-loop training paradigm based on the Schrödinger Bridge adopted in this invention, through the consistency constraints of the forward and backward directions, plays a stabilizing role similar to an "autoencoder." This inherent cyclic consistency reduces oscillations during training, helping the model to converge more robustly to an optimal solution.

[0031] In some embodiments, the Schrödinger bridge loss function takes the following specific form:

[0032] ;

[0033] in, This represents the set of learnable parameters of the forward generator; The first step in representing the diffusion process i Each discrete time step takes values ​​in the range [0, 1]. This represents a noisy latent variable generated by combining initial latent features with noise scheduling. This represents the predicted high-quality image features output by the forward generator; Indicates the preset hyperparameters; The forward generator represents the noisy latent variables. Predicting high-quality image features under the given conditions The conditional probability distribution; Represents information entropy; This represents the KL divergence constraint formula.

[0034] By employing the aforementioned technical solution, the core objective of the Schrödinger Bridge loss function is to find the optimal stochastic path from a low-quality image distribution to a high-quality image distribution, rather than simply learning a one-way mapping function. This avoids the biases that may be introduced by traditional diffusion models relying on the "Gaussian prior" assumption, enabling the model to learn the essential transformation relationship between two complex distributions more directly and efficiently, thus improving the rationality and efficiency of the path. By simultaneously constraining both the forward generation (predicting high-quality images) and backward reconstruction (reconstructing noisy variables), a strong bidirectional consistency constraint is formed. This closed-loop training mechanism forces the transformation process learned by the model to be coherent and reversible, thereby greatly ensuring the high fidelity of the generated images in terms of structure and detail, effectively reducing artifacts and distortion. The information entropy term in the loss function plays a crucial regularization role. It not only requires minimizing transmission costs but also encourages the path to maintain a certain degree of randomness (i.e., higher entropy). This prevents the model from falling into a simplistic "pattern collapse" solution in order to simply reduce differences, ensuring the diversity of generated results. This makes the reconstructed images not only accurate but also possess the texture and variation that natural images should have, avoiding overly smooth or patterned outputs.

[0035] In some embodiments, the method further includes:

[0036] The joint probability distribution of fake images is calculated using the following formula:

[0037] ;

[0038] in, The forward generator represents the noisy latent variables. Predicting high-quality image features under the given conditions The conditional probability distribution; Represents the noisy latent variable The prior probability distribution.

[0039] By employing the above technical solution, the abstract generation process is concretized into a computable joint probability distribution, providing a solid and operable mathematical foundation for the implementation of the entire Schrödinger bridge theory in the model, enabling subsequent loss calculation and optimization. The complex joint distribution is decomposed into two parts with clear physical meaning: the capability of the forward generator, i.e., the conditional probability of generating high-quality output from noisy input; and the prior distribution of the input data. This decomposition allows model training to specifically optimize the forward generator without altering the inherent prior distribution of the data, making the optimization objective clearer.

[0040] In a second aspect, embodiments of this application provide a computer system including a memory, a processor, and a computer program stored in the memory; the processor executes the computer program to implement the steps of the method described in any possible implementation of the first aspect.

[0041] Thirdly, embodiments of this application provide a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the method described in any possible implementation of the first aspect.

[0042] Fourthly, embodiments of this application provide a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the method described in any possible implementation of the first aspect.

[0043] It is understood that the computer system provided in the second aspect, the storage medium provided in the third aspect, and the computer program product provided in the fourth aspect are all used to execute the method provided in this application. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods, and will not be repeated here.

[0044] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0045] 1. By decomposing the original 3D magnetic resonance data into multiple independent 2D slices along the slice direction, the problem of high computational load and high memory consumption when directly processing 3D data is fundamentally solved. This allows subsequent reconstruction and segmentation steps to be executed efficiently and in parallel at the 2D level, significantly reducing the algorithm's hardware computing power requirements. This enables the solution to be deployed on commonly used clinical computing devices, meeting the stringent speed requirements of the treatment process.

[0046] 2. In the core image reconstruction stage, a pre-trained frequency-domain Schrödinger bridge diffusion model is used for the reconstruction of 2D slices. Combined with its integrated Fourier convolution module, this achieves a dual improvement in reconstruction speed and quality. The frequency-domain Schrödinger bridge diffusion model itself ensures efficient and accurate mapping from low-quality data to high-quality images through its theoretical framework. The introduction of the Fourier convolution module, by transforming and enhancing features in the frequency domain, compensates for the shortcomings of traditional spatial domain convolution in capturing global image information and high-frequency details (such as tissue edges and textures). This combination results in reconstructed 2D magnetic resonance images that are not only fast but also have fewer artifacts and clearer details, providing a high-quality input foundation for subsequent accurate segmentation.

[0047] 3. The reconstructed 2D image is seamlessly input into the pre-trained segmentation model, achieving integration and automation of the "reconstruction-segmentation" process. This serial design avoids the time overhead of data transfer and format conversion caused by the independent operation of the two stages in traditional solutions, and also ensures that the segmentation model receives optimized and enhanced images, thereby directly improving the accuracy and reliability of segmentation. Segmentation is performed at the 2D level, thus inheriting the advantage of computational efficiency;

[0048] 4. By rearranging, stacking, and fusing all the two-dimensional segmentation masks in their original order into a three-dimensional segmentation result, the complete three-dimensional anatomical structure of the target region was restored. This step integrates the efficient two-dimensional processing results into the three-dimensional spatial information required by clinicians for radiotherapy planning adjustments, completing a full closed loop from rapid imaging to accurate structural depiction. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating a fast magnetic resonance imaging reconstruction and segmentation method according to an embodiment of this application;

[0050] Figure 2 This is a schematic diagram of the serial processing joint learning architecture in an embodiment of this application;

[0051] Figure 3 This is a schematic diagram of the operation flow of the serial processing joint learning model in the embodiments of this application;

[0052] Figure 4 This is a schematic diagram of the magnetic resonance imaging reconstruction process in an embodiment of this application;

[0053] Figure 5 This is a schematic diagram of training the frequency domain Schrödinger bridge diffusion model in an embodiment of this application;

[0054] Figure 6 This is a schematic diagram of the Fourier convolution module in an embodiment of this application;

[0055] Figure 7This is a schematic diagram of an exemplary hardware structure of a computer system in an embodiment of this application. Detailed Implementation

[0056] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification and appended claims of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to any or all possible combinations including one or more of the listed items.

[0057] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0058] The following is combined with Figure 1 The method of the embodiments of this application will be described below.

[0059] Please see Figure 1 This is a flowchart illustrating a fast magnetic resonance imaging reconstruction and segmentation method in an embodiment of this application.

[0060] S101. Acquire the original three-dimensional magnetic resonance data collected by the magnetic resonance scanning device, and decompose the original three-dimensional magnetic resonance data into multiple independent two-dimensional magnetic resonance data slices along the slice sequence.

[0061] S102. The two-dimensional magnetic resonance data slices are sequentially input into the pre-trained frequency domain Schrödinger bridge diffusion model. Image reconstruction is performed through the pre-trained forward generator in the frequency domain Schrödinger bridge diffusion model to generate a two-dimensional magnetic resonance image. The forward generator integrates a Fourier convolution module, which is used to perform frequency domain transformation and feature enhancement processing on the feature mapping during the image reconstruction process.

[0062] S103. Each of the two-dimensional magnetic resonance images is sequentially input into the pre-trained segmentation model to segment the target area and the preset dangerous organs, and generate the corresponding two-dimensional segmentation mask.

[0063] S104. All the two-dimensional segmentation masks are rearranged and stacked according to the slice order, and a three-dimensional segmentation result is generated through a three-dimensional fusion operation.

[0064] In clinical settings, existing MRI scanners are used for accelerated scanning to acquire raw 3D MRI data for online adaptive radiotherapy. This accelerated scanning sequence (such as a T2-weighted 3D turbine spin echo sequence) can be completed in approximately 22 seconds, but this comes at the cost of image signal-to-noise ratio and resolution, resulting in k-space undersampling, which cannot be directly used for precise delineation. To achieve rapid processing without sacrificing accuracy, this approach reduces the 3D spatial problem to multiple 2D planar problems. Specifically, the acquired 3D data (e.g., containing 80 consecutive slices) is decomposed along its inherent slice orientation to generate a series of independent 2D MRI data slices. Directly processing 3D data (e.g., a 200×268×80 matrix) places extremely high demands on GPU memory, resulting in slow computation. After decomposition into 80 2D slices, each slice (200×268) can be processed independently and in parallel on the GPU, significantly reducing memory usage and enabling rapid processing on conventional workstations (e.g., those configured with RTX A6000 GPUs). During the decomposition process, the pixel arrangement and grayscale information of each 2D slice are completely preserved, ensuring that its spatial resolution is consistent with the original data, laying the foundation for subsequent high-precision reconstruction and segmentation.

[0065] Each 2D undersampled data slice is sequentially input into a pre-trained frequency-domain Schrödinger bridge diffusion model, where its forward generator performs image reconstruction, outputting a high-quality, high-resolution 2D MRI image. Unlike traditional diffusion models that start with Gaussian noise for denoising, the Schrödinger bridge diffusion model directly learns the optimal random path from a "low-quality MRI data distribution" to a "high-quality MRI data distribution." It guides the model along the most probable path by minimizing the transmission cost under KL divergence constraints, thus achieving more efficient and accurate reconstruction with fewer time steps. The frequency-domain Schrödinger bridge diffusion model includes a bidirectional diffusion generation network, which comprises a forward generator and a backward generator. During training, the forward generator is responsible for denoising the noisy input to generate a clean image; the backward generator reconstructs the noisy input from the clean image. This bidirectional closed-loop training, through consistency constraints, ensures the semantic coherence of the generation process and the realism of the image. The forward generator integrates a Fourier convolution module, which operates in parallel with spatial domain convolution within the residual blocks. The process is as follows: A Fast Fourier Transform (FFT) is performed on the input feature map to transform it from the spatial domain to the frequency domain; the frequency domain features are decomposed into real and imaginary parts, which are modulated separately through convolutional layers to capture global context and detailed texture; the real and imaginary components are reconstructed into complex forms, and an Inverse Fourier Transform (IFFT) is performed to revert to the spatial domain. A time-varying noise embedding vector is introduced, and a small Multilayer Perceptron (MLP) is used to generate scale parameter β and displacement parameter γ to scale and compensate for the features in the reverted spatial domain. Through this step, the reconstruction time for a single 2D slice is extremely short (the total reconstruction time for all 80 slices is approximately 20 seconds), enabling the rapid recovery of high-quality images suitable for diagnosis and segmentation from 22 seconds of accelerated scanning data.

[0066] Automated structural segmentation is performed using reconstructed high-quality 2D magnetic resonance images. Each reconstructed 2D image is sequentially input into a pre-trained segmentation model (such as DpnUNet). This model automatically identifies and delineates the target region (such as liver or pancreatic tumors) and pre-defined dangerous organs (such as the spinal cord, intestines, etc.), and outputs the corresponding 2D segmentation mask. This transforms the traditional 15-20 minute manual delineation work by doctors into a second-level (approximately 1.6 seconds to process 80 slices) automated segmentation, significantly reducing the workload of clinicians. Because the input to the segmentation model is a high-quality reconstructed image using the frequency domain Schrödinger bridge diffusion model, the high image quality and few artifacts provide a solid foundation for segmentation accuracy, ensuring the accuracy of the delineation results. The accuracy matching between the frequency domain Schrödinger bridge diffusion model and the segmentation model, and the improvement in reconstruction quality directly contributes to the improvement in segmentation accuracy, achieves a synergistic effect of 1+1>2.

[0067] All output 2D segmentation masks are rearranged and stacked strictly according to the original slice order during decomposition. Through 3D fusion operations (typically including simple pixel-level logical operations and possible morphological post-processing), these 2D contours are reconstructed into a coherent and complete 3D segmentation result. This 3D segmentation result accurately depicts the shape, size, and location of tumors and hazardous organs in 3D space. The efficient 2D processing results are seamlessly converted back into 3D clinical space. The entire process (from scanning to generating the 3D segmentation result) is controlled to approximately 42 seconds, allowing the entire online adaptive radiotherapy workflow to be completed within 20-30 minutes, meeting urgent clinical needs.

[0068] Overall, the embodiments of this application adopt a three-stage processing architecture of "split-serial processing-merging" to achieve efficient and accurate magnetic resonance image reconstruction and structural segmentation. This architecture systematically solves several key bottlenecks in existing rapid MRI imaging and segmentation, and achieves significant overall technical effects:

[0069] In the splitting stage, by decomposing the 3D data into multiple independent 2D slices along the slicing direction, the basis for data dimensionality reduction and parallel processing is realized, which significantly reduces the computational load and memory usage. This overcomes the problems of excessive memory consumption and low computational efficiency caused by directly processing 3D data in existing technologies, and provides the possibility for subsequent efficient serial processing.

[0070] In the serial processing stage, each 2D slice is reconstructed and segmented independently, forming a slice-level pipeline. This design ensures both high processing efficiency and precise matching and coordination between the reconstruction and segmentation stages, solving the problem of fragmented reconstruction and segmentation processes and the lack of an integrated serial mechanism in existing technologies. The frequency-domain Schrödinger bridge diffusion model used in this stage abandons the dependence of traditional diffusion models on Gaussian prior assumptions and directly learns the optimal transmission path from low-quality to high-quality data distribution, improving the model's ability to model complex MRI data distributions and reconstruction speed. At the same time, an adaptive noise modulation mechanism is introduced, which dynamically optimizes feature processing through time-varying noise embedding vectors, reducing the dependence on empirical parameter tuning and further improving reconstruction quality and robustness.

[0071] During the merging phase, by rearranging and stacking all two-dimensional segmentation masks in the original order, the complete structural restoration from slice to whole and from two-dimensional to three-dimensional is achieved, forming a closed-loop output of the "splitting-serial processing-merging" architecture, providing clinicians with three-dimensional anatomical information that can be directly used for radiotherapy planning adjustments.

[0072] In summary, this architecture achieves computational optimization through "splitting," synergistic quality and efficiency through "serial processing," and structural integrity through "merging," forming a highly efficient, accurate, and complete processing workflow from data input to 3D result output. This workflow is fully compatible with the MR-Linac clinical workflow, achieving minute-level processing from imaging to segmentation, significantly improving the workflow efficiency, system stability, and clinical feasibility of online adaptive radiotherapy.

[0073] Figure 2 This is a schematic diagram of the serial processing joint learning architecture in the embodiments of this application, as shown below. Figure 2 As shown, the acquired raw three-dimensional magnetic resonance data is sliced, and the slices are processed by the frequency domain Schrödinger bridge diffusion model to obtain the reconstructed MRI. The reconstructed MRI is then segmented by the DpnUNet segmentation model to obtain the segmentation result.

[0074] Figure 3 This is a schematic diagram of the operation flow of the serial processing joint learning model in the embodiments of this application, as shown below. Figure 3 As shown, 3D MRI data is split into multiple independent 2D slices; each 2D slice is independently and rapidly reconstructed to produce a high-resolution 2D MRI image; the reconstructed 2D image is segmented into target areas and organs at risk, and a 2.5D segmentation mask is output; all 2.5D segmentation masks are fused sequentially to generate a 3D segmentation result.

[0075] Figure 4 This is a schematic diagram of the magnetic resonance imaging reconstruction process in an embodiment of this application, as shown below. Figure 4 As shown, the flowchart on the left constitutes a typical pre-processing and planning flow for radiotherapy (or a similar medical procedure). The right side compares the traditional MRI imaging with the rapid MRI imaging flow of the present application embodiment: the existing technology requires about 3 minutes for full-duration imaging, while the present application embodiment only takes 42 seconds in total. The present application embodiment is divided into two stages: (1) Stage 1: 22-second imaging. Only 22 seconds are needed to acquire low-quality MRI. (2) Stage 2: MRI reconstruction. It takes 20 seconds to reconstruct the low-quality MRI into a high-quality MRI using the frequency domain Schrödinger bridge diffusion model.

[0076] Figure 5 This is a schematic diagram of the training frequency domain Schrödinger bridge diffusion model in an embodiment of this application, as shown below. Figure 5 As shown, the model receives an image input (the small image on the left) and a time step. The time step is processed by the temporal embedding module, converting it into a high-dimensional vector to incorporate temporal information into the network's feature representation. Encoder path (left): The input image first passes through a series of convolutional layers, such as "CIR7×7" and "CIR3×3". CIR stands for combined layers (e.g., convolution, instance normalization, and ReLU activation). Downsampled layers continuously reduce the spatial size (resolution) of the feature map while increasing the number of feature channels as the network depth increases, thereby extracting features at different levels of the image. Bottleneck / intermediate layers (middle): After two downsampling operations, the feature map enters the deeper part of the network. This contains two FFTResBlocks (Fast Fourier Transform Residual Blocks). This indicates that the model may incorporate frequency domain information when processing features, utilizing the Fast Fourier Transform to capture global context or specific frequency characteristics, and using residual connections to facilitate information flow. The results of the temporal embedding are integrated into these intermediate layers (represented by connecting lines), enabling the network to be aware of the current processing time step. The Z-transform module may be used to process or transform the time-varying noise embedding vector Z, or it may be a transformation operation related to the latent space. Decoder path (right): Upsampling layers progressively increase the spatial size of the feature maps for image reconstruction. Upsampling layers alternate with "CIR3×3" convolutional blocks. The final output layer consists of a "Conv7×7" convolution and a "Tanh" (hyperbolic tangent activation function). Tanh typically scales the output pixel values ​​to the range of -1 to 1.

[0077] This application's embodiments introduce preset noise scheduling. At random sampling time step Noisy latent variables are generated at the location Design a forward generator take over The noise is gradually removed, and finally the predicted clean and high-quality features are output. ,Right now Figure 5 middle Design a backward generator Receive clean features from the forward generator output Reconstructing with Noisy variables with consistent distribution .

[0078] In some embodiments, performing image reconstruction using a pre-trained forward generator in the frequency domain Schrödinger bridge diffusion model to generate a two-dimensional magnetic resonance image specifically includes:

[0079] Preset shallow feature extraction is performed on the two-dimensional magnetic resonance data slices to obtain initial latent features. The initial latent features are then combined with preset noise scheduling to form noisy latent variables.

[0080] The latent variables are input into the pre-trained forward generator, where a denoising process is performed on the latent variables through multiple cascaded residual blocks that integrate the Fourier convolution module, outputting clean latent features.

[0081] The clean latent features are upsampled and spatially mapped to generate the two-dimensional magnetic resonance image.

[0082] A pre-defined shallow feature extraction is performed on the input 2D magnetic resonance image data slices (i.e., undersampled, low-quality images). This is typically achieved through a few convolutional layers, the purpose of which is not to perform depth analysis, but to transform the raw image pixel data into a more representative and compact initial latent feature. This initial latent feature is then combined with a pre-defined noise schedule. Noise scheduling is one of the core concepts of diffusion models; it defines a schedule that controls how much noise should be added to the data at different inference steps (or time steps). Here, the noise corresponding to a specific time step is fused with the initial latent feature in a weighted manner to form a noisy latent variable. This process simulates placing the low-quality image in a well-defined "noise spectrum," setting a starting point for subsequent denoising. The noisy latent variable is then input into a pre-trained forward generator. The core of the generator is a deep network consisting of multiple cascaded residual blocks integrating Fourier convolutional modules. The noisy latent variable passes through these residual blocks sequentially. Each residual block is responsible for progressively removing a portion of the noise and optimizing the features. As data is passed through layers of the network, its feature representation is gradually purified, and noise is gradually stripped away. Finally, the output from the last residual block is the clean latent feature, a denoised high-level feature representation corresponding to the high-quality image. Within each residual block, spatial domain feature extraction (capturing local context through traditional convolution) and frequency domain feature enhancement (capturing global texture and high-frequency details through Fourier convolution modules) are performed in parallel. The Fourier convolution module transforms the features to the frequency domain for processing and then inversely transforms them back, fusing the result with the result of the spatial domain convolution path. This dual-path architecture ensures that the model, while denoising, can maximally recover and enhance the high-frequency details and edge information most easily lost during undersampling, which is crucial for achieving high-fidelity reconstruction. Skip connections within each residual block (directly adding the input to the output) ensure the stability of gradient flow, making it possible to train such a deep network and preventing the loss of important low-level information. The clean latent features are then upsampled and spatially mapped. This is typically achieved through a series of upsampling operations such as transposed convolutions or pixel shuffling, progressively enlarging the spatial size of the feature map to the resolution of the original image. Simultaneously, the final convolutional layer (typically using 1x1 convolution) maps the number of channels to 1 (for grayscale MRI images), completing the spatial mapping from high-dimensional features to a single image channel, ultimately generating a high-resolution two-dimensional magnetic resonance image suitable for clinical use. The entire forward generator reconstruction process is a refined, data-driven workflow that starts with low-quality input, sets a starting point through noise injection, performs precise denoising and feature enhancement via a deep space-frequency joint network, and finally upsamples and reconstructs a high-quality image. This process perfectly combines the distribution transformation advantages of Schrödinger's bridge theory with the detail recovery capabilities of Fourier convolution, thus achieving both fast and high-fidelity magnetic resonance image reconstruction.

[0083] In some embodiments, the denoising process performed on the latent variables to output clean latent features includes:

[0084] In the current residual block, spatial domain feature extraction and frequency domain feature enhancement are performed in parallel. The spatial domain feature extraction includes extracting spatial dimension features from the input features through a spatial domain convolutional layer to obtain a first output result. The frequency domain feature enhancement includes: performing a two-dimensional fast Fourier transform on the input features to convert them into frequency domain features; decomposing the frequency domain features into real components and imaginary components; performing convolutional modulation on the real components to generate a first modulation result; performing convolutional modulation on the imaginary components to generate a second modulation result; and reconstructing the first modulation result and the second modulation result into spatial domain features through an inverse Fourier transform to obtain a second output result.

[0085] The first output result and the second output result are fused to obtain intermediate features. The intermediate features are then added to the input features of the current residual block through a skip connection to form the output features of the current residual block, which are then used as the input features of the next residual block.

[0086] The output features of the last residual block are used as the clean latent features.

[0087] The current residual block receives input features from the previous block, which are simultaneously fed into two parallel processing paths: a spatial domain feature extraction path and a frequency domain feature enhancement path. The spatial domain feature extraction path aims to capture local texture, shape, and contour information in the image spatial domain (i.e., the pixel domain). The input features are processed through one or more spatial convolutional layers (such as standard 2D convolutional layers) to obtain a first output that preserves spatial local correlations. The frequency domain feature enhancement path aims to capture global structure, periodic patterns, and high-frequency details (such as organ edges and tissue boundaries) masked by noise in the frequency domain. The input features are subjected to a two-dimensional Fast Fourier Transform (FFT) to transform them from the spatial domain to the frequency domain, obtaining complex-valued frequency domain features. These complex-valued frequency domain features are then decomposed into real and imaginary components. The real part typically represents amplitude / global context information, while the imaginary part typically represents phase / detail texture information. Convolutional layers are used to extract and modulate the separated real and imaginary components, a crucial step for adaptively enhancing or suppressing different frequency components in the frequency domain. The modulated real and imaginary components are then recombined into a complex form, and an Inverse Two-Dimensional Fast Fourier Transform (IFFT) is performed to reconstruct them back into the spatial domain. This yields a second output rich in global information and enhanced high-frequency details. The first output from the spatial domain path and the second output from the frequency domain path are then fused. This is typically done through element-wise addition or channel concatenation followed by convolution. The local features extracted from the spatial domain are complementary to the enhanced global features from the frequency domain, forming more comprehensive intermediate features. Finally, the fused intermediate features are added element-wise to the initial input features of the current residual block. The output features generated after the above operations are then used as the input features for the next residual block. This process is repeated continuously, with each residual block further purifying and refining the features. When the feature sequence passes through the last residual block, its output features are considered to have sufficiently removed noise and recovered the essential information of the image. This final feature is defined as the clean latent feature, which will be used for subsequent upsampling and spatial mapping to generate the final high-resolution two-dimensional magnetic resonance image.

[0088] In some embodiments, the method further includes modulating the second output result, specifically including:

[0089] Generate a time-varying noise embedding vector, which is associated with the current time step of the frequency domain Schrödinger bridge diffusion model;

[0090] The time-varying noise embedding vector is input into a multilayer perceptron to generate scale parameters and displacement parameters.

[0091] The second output result is subjected to adaptive scaling and displacement compensation using the scale and displacement parameters.

[0092] Figure 6 This is a schematic diagram of the Fourier convolution module in an embodiment of this application. The workflow of FFTCoV (two-dimensional fast Fourier transform convolutional layer) is as follows: Figure 6 As shown on the left, for input features x A two-dimensional Fast Fourier Transform (FFT) is performed to transform the image from the spatial domain to the frequency domain, yielding frequency domain features. These features are then decomposed into real and imaginary components. The real component primarily contains amplitude and global context information, while the imaginary component is closely related to phase and detail texture. Separate convolutional layers (Conv) are used to modulate the real and imaginary components. These two convolutional layers do not share parameters to learn and enhance different types of frequency domain information separately. The modulation results in a first modulation result (real component) and a second modulation result (imaginary component). In the diffusion model, time step t represents the entire process from pure noise (t=0) to a clean image (t=1). The system has a learnable embedding table associated with the current time step. Based on this time step, the system queries or computes a fixed, low-dimensional time-varying noise embedding vector Z. This time-varying noise embedding vector Z is input into a lightweight multilayer perceptron (MLP). The output of this MLP is two vectors or tensors: a scaling parameter β and a shift parameter γ. The scaling parameter β and displacement parameter γ are used to modulate the features processed by the FFTConv module; β and γ perform adaptive scaling and displacement compensation on the second output result, and finally fuse the original input features through skip connections to output the enhanced spatial features. Finally, a skip connection of the input features is added before passing through FFTConv, and the output is then skip-connected with the input to form a Fourier residual block, preserving the original feature information.

[0093] In some embodiments, the method further includes training the forward generator by means of the following steps:

[0094] A training dataset is constructed, which contains multiple pairs of preset low-quality magnetic resonance images and paired high-quality magnetic resonance images. An initial bidirectional diffusion generation network containing a forward generator and a backward generator is constructed.

[0095] The low-quality magnetic resonance image is input into the forward generator to obtain a predicted high-quality image. The predicted high-quality image is then input into the backward generator to obtain a reconstructed noisy variable. A Schrödinger bridge loss function is calculated, which includes a first difference and a second difference. The first difference is the difference between the predicted high-quality image and the high-quality magnetic resonance image. The second difference is the second difference between the reconstructed noisy variable and the noisy variable obtained from the low-quality magnetic resonance image after a preset noise schedule. Based on the Schrödinger bridge loss function, the network parameters of the forward generator are updated through gradient backpropagation.

[0096] Repeat the previous step until the network parameters of the forward generator converge, thus obtaining the pre-trained forward generator.

[0097] A large-scale dataset needs to be collected, where each dataset contains a pair of paired magnetic resonance images. A pair consists of a low-quality magnetic resonance image and its paired high-quality magnetic resonance image. The low-quality image is typically obtained through simulated k-space undersampling of the high-quality image, or directly using a low-resolution, high-noise image obtained from a fast scan. The paired high-quality image, serving as the "ground truth" or "target," is typically a fully sampled, high signal-to-noise ratio standard scan image. Both images must be from the same anatomical site and scanned at the same time, ensuring perfect anatomical alignment so the model can learn a meaningful mapping. A complete, untrained bidirectional diffusion generative network is initialized, characterized by its bidirectional structure, comprising two main neural networks: a forward generator and a backward generator. The forward generator, the target model to be trained, learns to denoise the low-quality image and reconstruct it into a high-quality image. The backward generator, an auxiliary network, has the opposite task to the forward generator: learning how to reconstruct the corresponding noisy, low-quality variable from a high-quality image. The bidirectional diffusion generation network is trained iteratively, with each iteration consisting of a forward propagation and a gradient update. Forward generation: A batch of low-quality MRI images is taken from the dataset and input into the forward generator. The network computes and outputs its predicted high-quality images. Backward reconstruction: The predicted high-quality images (which may not be perfect at this point) output by the forward generator are input into the backward generator. The backward generator's task is to attempt to reconstruct a noisy variable that is similar in noise characteristics to the original low-quality input image, starting from this predicted image. First difference (output fidelity constraint): Directly compares the difference between the predicted high-quality image and the paired real high-quality images in the dataset. This is typically calculated using L1 or L2 norms (such as mean squared error). This is the most direct constraint, forcing the final output of the forward generator to be as close as possible to the real high-quality image at the pixel or feature level. It ensures the overall correctness of the generated images. Second difference (path consistency constraint): Compares the difference between the reconstructed noisy variable and the noisy variable obtained from the original low-quality image after a pre-defined noise schedule. "Noise scheduling on the original low-quality image" refers to adding noise to the original low-quality image according to the noise scheduling at the current time step, based on the theory of diffusion models, to obtain a noisy variable that theoretically conforms to the diffusion process. This is the essence of Schrödinger's bridge theory. It does not directly constrain the final output, but rather constrains the path of the entire generation process. It requires that the closed loop formed by the forward generator and the backward generator must be able to work consistently and reversibly. If the forward generator produces a "fake" image that deviates from the true data distribution, then the backward generator will have difficulty accurately reconstructing the initial noise distribution from this "fake" image. This difference will be exposed through a second difference and propagated back to correct the forward generator.The total loss of the Schrödinger Bridge loss function equals the sum of the first and second differences. The model aims to minimize this total loss. After calculating the Schrödinger Bridge loss, the gradient of the loss function with respect to the parameters of the forward generator network is calculated using the gradient backpropagation algorithm. Then, the optimizer is used to update only the parameters of the forward generator. In this training step, the parameters of the backward generator are usually kept constant or trained alternately using some strategy. The core is to ensure that the gradient signal is primarily used to optimize the final desired forward generator. The above iterative steps are repeated until the network parameters of the forward generator converge. Convergence is typically indicated by the loss function value decreasing to a low level and stabilizing, and by evaluation on an independent validation set, the image reconstruction quality no longer showing a significant improvement. At this point, training is complete, resulting in a high-performance pre-trained forward generator that can be used for practical fast MRI reconstruction tasks.

[0098] In some embodiments, the Schrödinger bridge loss function takes the following specific form:

[0099] ;

[0100] in, This represents the set of learnable parameters of the forward generator; The first step in representing the diffusion process i Each discrete time step takes values ​​in the range [0, 1]. This represents a noisy latent variable generated by combining initial latent features with noise scheduling. This represents the predicted high-quality image features output by the forward generator; Indicates the preset hyperparameters; The forward generator represents the noisy latent variables. Predicting high-quality image features under the given conditions The conditional probability distribution; Represents information entropy; This represents the KL divergence constraint formula.

[0101] in, High-quality fake images are generated by predicting after forward denoising. For the diffusion process i Each discrete time step belongs to a sequence. ( No noise (maximum noise) Noisy variables reconstructed for the backward generator It is the entropy regularization coefficient (hyperparameter), which balances the weights of the "distance term" and the "entropy term" to control the smoothness of the diffusion process; Time step weighting factor: The larger the value (the stronger the noise), the smaller the factor becomes, weakening the entropy constraint and adapting to the optimization needs of different time steps; It is the set of learnable parameters (such as neural network weights and biases) of the backward generator (denoising network). For time step Below, with To optimize the Schrödinger bridge loss function for the objective, Represents the joint distribution of generated fake images. The information entropy of the probability of fake images (the higher the entropy, the higher the uncertainty of the distribution, which is used to regularize the diffusion process and avoid overfitting). This is the KL divergence constraint formula, used to calculate the noisy variables reconstructed by the backward generator. High-quality fake images generated by prediction after forward denoising The average distance between them measures the characteristic differences between the two, making the diffusion process tend towards the "optimal transmission path," that is, the reconstructed noisy variable. Achieve true high-quality images through "accurate noise reduction". , For noisy variables High-quality fake images Pixel-level feature differences.

[0102] Entropy is a measure of randomness or uncertainty. Maximizing entropy means encouraging the model to produce more diverse and uncertain outputs. Without it, the model might find a very "clever" but singular mapping to minimize transmission costs (e.g., producing an average, blurry output for all inputs), a phenomenon known as "mode collapse." Entropy regularization prevents this degradation, ensuring that the reconstructed image has the rich texture and detail that a natural image should have, rather than being overly smooth or patterned. Schrödinger's bridge theory suggests that the path with maximum entropy under transmission cost constraints is the most likely, smoothest, and most natural path. This avoids the path getting stuck in some unreasonable extreme cases, making the entire denoising process more stable and the generated image more consistent with physical reality.

[0103] In some embodiments, the method further includes:

[0104] The joint probability distribution of fake images is calculated using the following formula:

[0105] ;

[0106] in, The forward generator represents the noisy latent variables. Predicting high-quality image features under the given conditions The conditional probability distribution; Represents the noisy latent variable The prior probability distribution.

[0107] in, Forward generator (parameters) Noisy variables generated -High-quality fake images The joint probability distribution of "" describes the noisy latent variables. and high-quality image features The probability of both occurring simultaneously. Given a noisy variable At that time, the forward generator predicts / generates high-quality fake images. The conditional probability distribution, For noisy variables The prior probability distribution (usually a diffusion process) Noise distribution at any given time, such as a Gaussian distribution.

[0108] The marginal probability distribution is calculated using the following formula:

[0109] ;

[0110] in, This represents a high-quality image generated by the forward generator. The marginal probability distribution. This is the target distribution that the model ultimately hopes to perfectly match, namely the data distribution of real high-quality MRI images. Represents the relationship with noisy variables Prior distribution The mathematical expectation, which represents the result in order to obtain The distribution needs to consider all possible noise inputs. (Based on its prior distribution) ), and then for each Calculate the distribution of conditions The mean. It will be the forward generator (conditional distribution) The performance of the final generated result and the overall quality of the marginal distribution In order to produce a high-quality, highly diverse output distribution, a good forward generator must perform well on all possible noisy inputs.

[0111] The above describes a fast magnetic resonance imaging reconstruction and segmentation method in the embodiments of this application. The computer system in the embodiments of this application will be described in detail below in conjunction with the above fast magnetic resonance imaging reconstruction and segmentation method.

[0112] Please see Figure 7This is a schematic diagram of an exemplary hardware structure of a computer system in an embodiment of this application.

[0113] In some embodiments, the computer system 700 includes a computer device, which may be a terminal device. The computer device includes a processor 701, a memory 702, a sensor module 703, a communication module 704, an input device 705, and an output device 706 connected via a system bus. The processor 701 of the computer device provides computing and control capabilities. The memory 702 of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database is used to store data.

[0114] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0115] In some embodiments of this application, a computer-readable storage medium is provided, including instructions that, when executed on a computer system 700, cause the computer system 700 to perform a fast magnetic resonance imaging reconstruction and segmentation method according to an embodiment of this application.

[0116] In some embodiments of this application, a computer program product is also provided, which, when run on a computer system 700, causes the computer system 700 to execute a fast magnetic resonance imaging reconstruction and segmentation method according to an embodiment of this application.

[0117] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0118] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive), etc.

[0119] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A fast magnetic resonance imaging reconstruction and segmentation method, characterized in that, include: Acquire raw three-dimensional magnetic resonance data collected by a magnetic resonance scanning device, and decompose the raw three-dimensional magnetic resonance data into multiple independent two-dimensional magnetic resonance data slices along the slice sequence; The two-dimensional magnetic resonance data slices are sequentially input into a pre-trained frequency domain Schrödinger bridge diffusion model. Image reconstruction is performed through the pre-trained forward generator in the frequency domain Schrödinger bridge diffusion model to generate a two-dimensional magnetic resonance image. The forward generator integrates a Fourier convolution module, which is used to perform frequency domain transformation and feature enhancement processing on the feature mapping during the image reconstruction process. Each of the two-dimensional magnetic resonance images is sequentially input into a pre-trained segmentation model to segment the target area and preset dangerous organs, generating a corresponding two-dimensional segmentation mask. All the two-dimensional segmentation masks are rearranged and stacked according to the slice order, and a three-dimensional fusion operation is performed to generate a three-dimensional segmentation result. The step of performing image reconstruction using a pre-trained forward generator in the frequency domain Schrödinger bridge diffusion model to generate a two-dimensional magnetic resonance image specifically includes: Preset shallow feature extraction is performed on the two-dimensional magnetic resonance data slices to obtain initial latent features. The initial latent features are then combined with preset noise scheduling to form noisy latent variables. The latent variables are input into the pre-trained forward generator, where a denoising process is performed on the latent variables through multiple cascaded residual blocks that integrate the Fourier convolution module, outputting clean latent features. The clean latent features are upsampled and spatially mapped to generate the two-dimensional magnetic resonance image. The process of performing denoising on the latent variables to output clean latent features includes: In the current residual block, spatial domain feature extraction and frequency domain feature enhancement are performed in parallel. The spatial domain feature extraction includes extracting spatial dimension features from the input features through a spatial domain convolutional layer to obtain a first output result. The frequency domain feature enhancement includes: performing a two-dimensional fast Fourier transform on the input features to convert them into frequency domain features; decomposing the frequency domain features into real components and imaginary components; performing convolutional modulation on the real components to generate a first modulation result; performing convolutional modulation on the imaginary components to generate a second modulation result; and reconstructing the first modulation result and the second modulation result into spatial domain features through an inverse Fourier transform to obtain a second output result. The first output result and the second output result are fused to obtain intermediate features. The intermediate features are then added to the input features of the current residual block through a skip connection to form the output features of the current residual block, which are then used as the input features of the next residual block. The output features of the last residual block are used as the clean latent features. The method further includes training the forward generator through the following steps: A training dataset is constructed, which contains multiple pairs of preset low-quality magnetic resonance images and paired high-quality magnetic resonance images. An initial bidirectional diffusion generation network containing a forward generator and a backward generator is constructed. The low-quality magnetic resonance image is input into the forward generator to obtain a predicted high-quality image. The predicted high-quality image is then input into the backward generator to obtain a reconstructed noisy variable. A Schrödinger bridge loss function is calculated, which includes a first difference and a second difference. The first difference is the difference between the predicted high-quality image and the high-quality magnetic resonance image. The second difference is the second difference between the reconstructed noisy variable and the noisy variable obtained from the low-quality magnetic resonance image after a preset noise schedule. Based on the Schrödinger bridge loss function, the network parameters of the forward generator are updated through gradient backpropagation. Repeat the previous step until the network parameters of the forward generator converge, thus obtaining the pre-trained forward generator.

2. The fast magnetic resonance imaging reconstruction and segmentation method according to claim 1, characterized in that, The method further includes modulating the second output result, specifically including: Generate a time-varying noise embedding vector, which is associated with the current time step of the frequency domain Schrödinger bridge diffusion model; The time-varying noise embedding vector is input into a multilayer perceptron to generate scale parameters and displacement parameters. The second output result is subjected to adaptive scaling and displacement compensation using the scale and displacement parameters.

3. The fast magnetic resonance imaging reconstruction and segmentation method according to claim 1, characterized in that, The specific form of the Schrödinger bridge loss function is as follows: ； in, This represents the set of learnable parameters of the forward generator; The first step in representing the diffusion process i Each discrete time step takes values ​​in the range [0, 1]. This represents a noisy latent variable generated by combining initial latent features with noise scheduling. This represents the predicted high-quality image features output by the forward generator; Indicates the preset hyperparameters; The forward generator represents the noisy latent variables. Predicting high-quality image features under the given conditions The joint probability distribution of ; Represents information entropy; This represents the KL divergence constraint formula.

4. The fast magnetic resonance imaging reconstruction and segmentation method according to claim 3, characterized in that, The method further includes: The joint probability distribution of fake images is calculated using the following formula: ； in, The forward generator represents the noisy latent variables. Predicting high-quality image features under the given conditions The conditional probability distribution; Represents the noisy latent variable The prior probability distribution.

5. A computer system comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-4.

6. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method described in any one of claims 1-4.

7. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method described in any one of claims 1-4.

Images