MRI condition reconstruction method based on coil calibration image control

Abstract

The invention belongs to the technical field of medical imaging, and particularly relates to an MRI (Magnetic Resonance Imaging) condition reconstruction method based on coil calibration image control, which comprises the following steps of: acquiring a coil calibration image and undersampled MRI image data; inputting the coil calibration image into a condition control module to estimate a coil sensitivity distribution map; taking the coil calibration image as condition input, and reconstructing undersampled MRI image data in combination with a diffusion model; wherein the diffusion model comprises a forward diffusion process and a reverse diffusion process, image reconstruction is realized by gradually adding and removing noise, and detail reduction and definition of a reconstructed image are remarkably improved by combining the forward and reverse noise removal processes of the diffusion model with a condition control mechanism of a coil calibration image; the condition control module dynamically adjusts the influence of the coil calibration image on the reconstructed image, thereby effectively reducing the noise and artifacts in the image, and enabling the reconstructed image to be clearer and more accurate.

Description

Technical Field

[0001] The invention belongs to the technical field of medical imaging, and in particular relates to an MRI conditional reconstruction method based on coil calibration image control. Background Art

[0002] Since its inception, magnetic resonance imaging (MRI) technology has been an important tool in the field of medical imaging. With the continuous improvement of medical needs, fast and high-quality MRI imaging has become a research hotspot. However, traditional MRI imaging methods face the dilemma of long scanning time and difficulty in balancing image quality and imaging speed.

[0003] In recent years, the introduction of parallel imaging technology and compressed sensing theory has provided new ideas for accelerating MRI imaging. These methods shorten the scanning time by reducing the amount of sampled data, but at the same time they also bring about the problem of reduced image reconstruction quality. Especially in the case of highly undersampling, the reconstructed image often has problems such as noise, artifacts, and loss of details.

[0004] At present, the most advanced MRI reconstruction methods are mainly based on deep learning technology. These methods learn the mapping relationship from undersampled data to complete images by training neural networks. Although certain results have been achieved, there are still some key problems: first, these methods often use coil calibration data only to estimate coil sensitivity maps, and fail to make full use of the rich structural information contained therein; second, when processing highly undersampled data, the reconstruction quality is still unsatisfactory, especially in terms of preserving detailed structures and reducing artifacts; third, the existing methods are not flexible enough when processing input data of different contrast and resolution, and it is difficult to adapt to complex and changing clinical needs.

[0005] In addition, fast MRI sequences such as fast spin echo (FSE) and echo planar imaging (EPI) face greater challenges when applying existing reconstruction methods due to their special signal acquisition methods. These sequences often fail to embed sufficient internal calibration signals, resulting in further degradation of reconstruction quality.

[0006] In view of the above problems, a new MRI image reconstruction method is urgently needed, which can make full use of the information in the coil calibration image, improve the quality of the reconstructed image, and adapt to different imaging sequences and scanning conditions. The present invention is an innovative solution to these technical challenges. Summary of the invention

[0007] The MRI conditional reconstruction method based on coil calibration image control proposed in the present invention aims to solve the key problems existing in the prior art. Specifically, the present invention is committed to improving the quality and efficiency of MRI image reconstruction while enhancing the flexibility and adaptability of the method. By innovatively combining coil calibration images, diffusion models and conditional control mechanisms, the present method achieves high-quality reconstruction of undersampled MRI data.

[0008] Terminology explanation: Magnetic Resonance Imaging Neural Network (MRINN): A method of reconstructing magnetic resonance images using deep learning technology, which is different from the traditional compressed sensing method.

[0009] The present invention discloses an MRI conditional reconstruction method based on coil calibration image control, the method comprising the following steps:

[0010] acquiring a coil calibration image and undersampled MRI image data;

[0011] Inputting the coil calibration image into a condition control module to estimate a coil sensitivity distribution map;

[0012] Reconstructing the undersampled MRI image data using the coil calibration image as a conditional input in combination with a diffusion model;

[0013] The diffusion model includes a forward diffusion process and a reverse diffusion process, and image reconstruction is achieved by gradually adding and removing noise.

[0014] Preferably, the method further comprises:

[0015] The feature fusion of the coil calibration image and the under-sampled MRI image data is achieved through a cross-attention mechanism,

[0016] The cross attention mechanism is defined by the following formula:

[0017]

[0018] Where Q represents the query matrix, which represents the features of the MRI image; K and V represent the key and value matrices, which represent the features of the coil calibration image; d k Represents a scaling factor used to stabilize the attention distribution.

[0019] Preferably, the method further comprises:

[0020] Data consistency constraints are applied during the reconstruction process to ensure that the reconstruction results are consistent with the actual acquired K-space data.

[0021] The data consistency constraint is defined by the following formula:

[0022]

[0023] Among them, (y) is the K-space data actually acquired, is the combined operator of coil sensitivity and Fourier transform, representing the data acquisition process, is the conjugate transpose of the operator, which is used to ensure data consistency. λ is a tuning parameter that controls the strength of the data consistency constraint.

[0024] Preferably, the reverse diffusion process is defined by the following formula:

[0025]

[0026] in, is the noise adjustment coefficient, which controls the noise ratio in the reverse diffusion process, and z is the noise added in the reverse diffusion process.

[0027] Preferably, the method further comprises:

[0028] During the model training phase, the encoder parameters of the pre-trained diffusion model are frozen;

[0029] The conditional control module and decoder parameters are optimized to achieve feature fusion and image reconstruction.

[0030] Preferably, the method further comprises:

[0031] A combination of multiple loss functions is used for model training, including mean square error (MSE) loss, structural similarity (SSIM) loss, and perceptual loss.

[0032] Preferably, the method further comprises:

[0033] The features of the coil calibration image are adjusted using an adaptive control gate to avoid over-reliance on the calibration image and reduce artifacts in reconstruction.

[0034] Preferably, the coil calibration image and the under-sampled MRI image data may have different contrasts and resolutions.

[0035] Preferably, the method further comprises:

[0036] In the inference phase, the fully sampled calibration image and the undersampled target image are first input;

[0037] The features are fused through the cross-attention mechanism to generate a preliminary reconstruction result;

[0038] In combination with the data consistency constraint, further optimizing the reconstructed image;

[0039] Through multiple iterations of back diffusion and data consistency adjustment, high-quality MRI reconstructed images are finally generated.

[0040] Preferably, the method is applicable to fast spin echo FSE and echo planar imaging EPI fast MRI imaging sequences,

[0041] When the internal coil calibration signal cannot be embedded due to signal attenuation, the coil calibration image is used as a conditional input to improve the quality of the reconstructed image.

[0042] Beneficial effects of the present invention:

[0043] 1. Significantly improve the quality of MRI image reconstruction: Through the forward and reverse denoising process of the diffusion model, combined with the conditional control mechanism of the coil calibration image, the detail restoration and clarity of the reconstructed image are significantly improved.

[0044] 2. Reduce noise and artifacts in reconstruction: The condition control module dynamically adjusts the influence of the coil calibration image on the reconstructed image, effectively reducing the noise and artifacts in the image, making the reconstructed image clearer and more accurate.

[0045] 3. Applicable to different MRI imaging sequences: The method of the present invention is not only applicable to fast MRI imaging sequences such as fast spin echo (FSE) and echo planar imaging (EPI), but can also process image data of different resolutions and contrasts, and has strong versatility.

[0046] This method can be widely used in medical imaging equipment, medical image reconstruction software, telemedicine systems, and medical research and teaching fields, and can significantly improve the diagnostic quality and efficiency of MRI imaging. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a demonstration of the actual reconstruction effect of the method of the present invention. DETAILED DESCRIPTION

[0048] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation methods, structures, features and effects thereof are described in detail below in conjunction with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0049] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0050] Please refer to Figure 1The present invention provides a method for conditional reconstruction of magnetic resonance imaging (MRI) based on coil calibration image control. The method first acquires a coil calibration image and undersampled MRI image data. In clinical applications, this usually involves using an MRI scanner for rapid scanning to obtain a low-resolution coil calibration image and undersampled target area MRI data. For example, in brain imaging, a low-resolution calibration image covering the entire head may be acquired first, and then an undersampled scan of the brain area of ​​interest is performed.

[0051] The method of the present invention first requires data preparation. When performing MRI image reconstruction tasks, the quality and diversity of training data are crucial to the performance of the model. In order to train a model that can stably reconstruct high-quality MRI images in different clinical scenarios, a large amount of high-quality MRI data and corresponding coil calibration data need to be prepared. These data sets can be derived from public data sets, privately collected data, or generated through simulation technology. The data include the following:

[0052] High-resolution MRI image data: This is the target reconstructed image of the model and should contain MRI data with various contrasts and resolutions, such as T1, T2 and other sequences.

[0053] Low-resolution coil calibration data: Coil calibration data is mainly used in the conditional control module of the model to help correct artifacts and noise in MRI images. Low-resolution calibration images can provide coil sensitivity information and provide additional constraints for the model.

[0054] For example, the public fastMRI dataset is used. This dataset contains fully sampled high-resolution images and low-resolution coil calibration data, which is widely used in MRI image reconstruction research. This dataset provides different contrast data including T1 and T2 sequences, which can simulate common MRI imaging scenarios in clinical practice.

[0055] Or through private data collection: For some specific imaging needs or insufficient data in specific scenarios, high-resolution images and corresponding coil calibration data can be collected by MRI equipment. This kind of data has the advantage of being highly customized and suitable for model training in specific application scenarios.

[0056] Simulation data set generation: When actual data is insufficient, simulation software can be used to generate MRI images with noise and undersampling characteristics and corresponding coil calibration images. Although simulation data cannot completely replace real data, it can be used as a supplement in some cases.

[0057] After obtaining enough training data, the model needs to be initialized through a pre-trained diffusion model (such as DDPM, Denoising Diffusion Probabilistic Model). The diffusion model converts MRI images into noise through the forward diffusion process, thus laying the foundation for subsequent reverse diffusion denoising. The training of the diffusion model is mainly divided into two parts: forward diffusion and reverse diffusion:

[0058] Forward diffusion: gradually adding noise to the image until the image completely loses the original information and becomes pure noise.

[0059] Back diffusion: gradually denoise the image through the trained model to restore the image details.

[0060] The formula is as follows:

[0061]

[0062] Among them, x t is the image state at time step (t), β t is the noise parameter that increases with the time step (t) during the forward diffusion process, indicating the noise intensity, is a Gaussian (normal) distribution, and I is the unit matrix.

[0063] Each step in the diffusion process is described by the above formula, that is, part of the image information is lost in each step and a certain amount of random noise is added:

[0064]

[0065] Among them, z is the random noise sampled from the standard Gaussian distribution, which gradually masks the details of the image. Through forward diffusion, the model can gradually learn how to restore image details through the denoising process.

[0066] In a specific embodiment of the present invention, the fastMRI dataset is used to pre-train the diffusion model. In particular, based on its extensive high-resolution data, a basic model capable of handling a variety of MRI imaging conditions can be trained.

[0067] In the inference stage, the actual acquired MRI data is usually undersampled data, which needs to be compensated by combining fully sampled coil calibration images. The calibration images can be of the same contrast and resolution or different contrast and resolution.

[0068] The specific steps are as follows:

[0069] The MRI device is used to acquire low-resolution coil calibration data and undersampled target MRI data.

[0070] The calibration image is used in the condition control module to provide coil sensitivity and image structure information to help the model to accurately reconstruct.

[0071] In a specific embodiment of the present invention, collecting coil calibration images and MRI target data with the same contrast and resolution can improve reconstruction accuracy, which is particularly suitable for scenarios with high requirements for high resolution and details.

[0072] Collect calibration images of different contrast or low resolution. In this case, the model can dynamically adjust the influence of the calibration image through the conditional control module to avoid reconstruction artifacts caused by large image differences.

[0073] The base model is trained using the IXI or HCP datasets. Compared to fastMRI, these datasets may contain different scanning protocols and imaging sequences, providing more diversity for the model.

[0074] Next, the method inputs the coil calibration image into the condition control module to estimate the coil sensitivity distribution map. This step is crucial for accurately reconstructing MRI images because the coil sensitivity distribution directly affects the received signal strength. In practical applications, such as when using a multi-channel head coil for brain scanning, the sensitivity distribution of each channel may be different, and accurately estimating these distributions can significantly improve the quality of the reconstructed image.

[0075] Based on the diffusion model, in order to further improve the reconstruction accuracy and reduce noise, a conditional control module (ControlNet) is added. This module combines the feature information of the coil calibration image and the MRI target image, and uses the CrossAttention mechanism to achieve deep fusion of features, thereby improving the model's reconstruction ability for low-resolution and under-sampled data. Formula:

[0076] The formula of the cross attention mechanism is as follows:

[0077]

[0078] Among them, Q is the query matrix, which represents the characteristics of the MRI image, K and V are the key and value matrices, which represent the characteristics of the coil calibration image, and d k is a scaling factor used to stabilize the attention distribution. This mechanism can dynamically adjust the feature weights in the reconstruction process according to the similarity between the MRI image and the coil calibration data, enabling the model to flexibly handle different types of inputs.

[0079] In a specific embodiment of the present invention, a ControlNet module is added to the encoder part of the basic model, and a zero convolution layer is used to achieve feature fusion of the input image and the calibration image. Through the Cross Attention mechanism, the model can dynamically adjust the information flow between the two to improve the reconstruction quality. The features of the calibration image are adjusted using an adaptive control gate. Through this method, over-reliance on the calibration image can be avoided and artifacts in reconstruction can be reduced.

[0080] During the model training phase, the optimization goal of the model is to adjust the parameters of the conditional control module and the decoder part, while freezing the encoder part of the basic diffusion model. This ensures that the model can flexibly use the information of the coil calibration data through the conditional control module to improve the reconstruction quality of MRI images while maintaining the denoising ability of the basic diffusion model. The key training configuration is as follows:

[0081] Freeze the encoder layer of the base model: The base diffusion model has good feature extraction and denoising capabilities through pre-training. Therefore, we no longer adjust these parameters during training to reduce the computational burden and prevent overfitting.

[0082] Training condition control module and decoder: Focus on optimizing these modules so that they can effectively fuse the features in coil calibration data and undersampled MRI data to further improve image reconstruction.

[0083] The training parameters are set as follows:

[0084] Optimizer: Usually Adam or AdamW optimizer is used. Adam optimizer can speed up training convergence and make the model more robust by combining momentum and adaptive learning rate.

[0085] Epochs: Models usually require multiple rounds of iterative training to be fully optimized. It is recommended to perform 200 rounds of training on a large-scale dataset and use the performance on the validation set to evaluate whether early stopping is needed.

[0086] Batch size: Set an appropriate batch size based on hardware resources (such as GPU memory size). The recommended range is 8-32, and the specific value can be adjusted based on model performance and memory usage.

[0087] Loss function selection

[0088] Mean Square Error (MSE): It is used to measure the pixel difference between the reconstructed image and the real image to ensure that the model output is close to the real value.

[0089] Structural Similarity (SSIM) loss: used to measure the structural similarity between the reconstructed image and the real image and enhance the detail fidelity of the image.

[0090] Perceptual Loss: The image quality is measured by the high-level features of the pre-trained network to ensure that the reconstructed image is not only pixel accurate but also retains texture and details.

[0091] In actual training, a combination of multiple loss functions is often used to optimize the model. For example, MSE, SSIM and perceptual loss can be combined to ensure that the model can generate high-quality images at both the pixel level and the structural level.

[0092] The model was trained on the public fastMRI dataset. The Adam optimizer was used, the learning rate was set to 1×10-4, the batch size was 16, and the training was performed for 200 epochs. The loss function used a combination of MSE and SSIM, with weights of 1.0 and 0.5, respectively.

[0093] The model was trained on a privately collected dataset. The AdamW optimizer was used, the learning rate was set to 5×10-5, the batch size was 8, and 100 epochs were trained. A combination of MSE, SSIM, and perceptual loss was used, focusing on the structural recovery of the image.

[0094] On the simulation dataset, the batch size is set to 32 and trained for 150 epochs. SSIM and perceptual loss are used to optimize the reconstruction results, with a special focus on the restoration of details and textures.

[0095] By freezing the encoder of the base model and optimizing the conditional control module and decoder, the model is able to fully utilize the coil calibration data, improve the accuracy of MRI image reconstruction, and generate higher quality reconstructed images.

[0096] In the inference stage, the actual acquired MRI data is usually undersampled data, which needs to be compensated by combining the fully sampled coil calibration image. The calibration image can be of the same contrast and resolution or different contrast and resolution. The specific steps are as follows:

[0097] The MRI device is used to acquire low-resolution coil calibration data and undersampled target MRI data.

[0098] The calibration image is used in the condition control module to provide coil sensitivity and image structure information to help the model to accurately reconstruct.

[0099] Acquiring coil calibration images and MRI target data with the same contrast and resolution can improve reconstruction accuracy, especially for scenarios with high requirements for high resolution and details.

[0100] Collect calibration images of different contrast or low resolution. In this case, the model can dynamically adjust the influence of the calibration image through the conditional control module to avoid reconstruction artifacts caused by large image differences.

[0101] During the inference process, the model needs to combine data consistency constraints and back diffusion process to generate high-quality reconstructed images. Data consistency ensures that the reconstruction results of the model are consistent with the actual acquired K-space data, preventing the model from generating false image content; the formula is as follows:

[0102] Data consistency constraints:

[0103]

[0104] Wherein, (y) is the K-space data actually acquired. is the combined operator of coil sensitivity and Fourier transform, representing the data acquisition process. It is the conjugate transpose of the operator, which is used to ensure data consistency. λ: Adjustment parameter to control the strength of data consistency constraint.

[0105] Back diffusion process:

[0106]

[0107] in, is the noise adjustment coefficient, which controls the noise ratio in the reverse diffusion process. z is the noise added in the reverse diffusion process.

[0108] During the inference process, the fully sampled calibration image and the undersampled target image are first input. The model generates a preliminary reconstruction result by fusing features through the cross-attention mechanism. Then, the reconstructed image is further optimized in combination with the data consistency constraint to ensure that the generated image is consistent with the actual collected data.

[0109] Through multiple iterations of back diffusion and data consistency adjustment, high-quality MRI reconstructed images are finally generated.

[0110] This method gradually achieves high-quality MRI image reconstruction through multiple steps such as preparing training data, obtaining a basic diffusion model, adding a conditional control module, conducting model training, and combining data consistency for prediction.

[0111] The method of the present invention also includes applying data consistency constraints during the reconstruction process, which ensures that the reconstruction result is consistent with the actual acquired K-space data. This step is crucial to ensure the authenticity of the reconstructed image and reduce artifacts.

[0112] The data consistency constraint is defined by the following formula:

[0113]

[0114] Among them, (y) is the K-space data actually acquired, is the combined operator of coil sensitivity and Fourier transform, representing the data acquisition process, is the conjugate transpose of the operator, which is used to ensure data consistency. λ is an adjustment parameter that controls the strength of the data consistency constraint. Through this constraint, the model can ensure that the reconstructed image is consistent with the actual collected data while retaining the details of the image.

[0115] For example, in the application of fast spin echo (FSE) sequence, signal attenuation may lead to the lack of low-frequency information in K space. By applying data consistency constraints, the image artifacts caused by this can be effectively reduced, and the quality and credibility of the reconstructed image can be improved.

[0116] The reverse diffusion process of the present invention is defined by the following formula:

[0117]

[0118] in, is the noise adjustment coefficient, which controls the noise ratio in the reverse diffusion process, and z is the noise added in the reverse diffusion process.

[0119] In this formula, x t-1 and x t Represent the image status at time t-1 and time t respectively; represents the noise adjustment coefficient; z represents the noise sampled from the standard normal distribution. This process allows the model to gradually remove noise and restore image details.

[0120] In practical applications, such as dynamic cardiac MRI imaging, this back-diffusion process can effectively recover temporally resolved cardiac motion sequences from undersampled data. By carefully adjusting the α_{t-1} parameter, an optimal balance can be achieved between image quality and reconstruction speed.

[0121] The method of the present invention allows the coil calibration image and the undersampled MRI image data to have different contrasts and resolutions. This flexibility greatly expands the application range of the method and enables it to adapt to various clinical scenarios.

[0122] For example, in a multi-sequence MRI scan, a T1-weighted low-resolution calibration image can be used to assist in the reconstruction of a T2-weighted high-resolution target image. This capability is particularly useful in certain clinical applications, such as in the diagnosis of neurological diseases, where the complementary information provided by different contrast sequences can be fully utilized.

[0123] The method of the present invention is particularly suitable for fast MRI imaging sequences such as fast spin echo (FSE) and echo planar imaging (EPI). In these sequences, it is usually difficult to embed internal coil calibration signals due to signal attenuation. The present method effectively overcomes this limitation by using externally acquired coil calibration images as conditional input, significantly improving the quality of reconstructed images.

[0124] For example, in functional MRI (fMRI) studies, EPI sequences are often used to obtain information about rapidly changing brain activity. This method can improve spatial resolution and image quality while maintaining high temporal resolution, thereby enhancing the ability to detect activation of tiny brain regions.

[0125] In general, the MRI conditional reconstruction method based on coil calibration image control provided by the present invention realizes high-quality MRI image reconstruction by innovatively combining diffusion model, cross-attention mechanism and data consistency constraint. This method can not only improve image quality, but also significantly shorten MRI scanning time, which is of great significance for improving patient comfort and diagnostic efficiency. In addition, the flexibility and adaptability of this method enable it to be widely used in various MRI imaging scenarios, providing a powerful tool for clinical diagnosis and medical research.

[0126] It should be noted that the above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An MRI conditional reconstruction method based on coil calibration image control, characterized in that: The method comprises the following steps: acquiring a coil calibration image and undersampled MRI image data; Inputting the coil calibration image into a condition control module to estimate a coil sensitivity distribution map; Reconstructing the undersampled MRI image data using the coil calibration image as a conditional input in combination with a diffusion model; The diffusion model includes a forward diffusion process and a reverse diffusion process, and image reconstruction is achieved by gradually adding and removing noise.

2. The MRI conditional reconstruction method according to claim 1, characterized in that: The method further comprises: The feature fusion of the coil calibration image and the under-sampled MRI image data is achieved through a cross-attention mechanism, The cross attention mechanism is defined by the following formula: Where Q represents the query matrix, which represents the features of the MRI image; K and V represent the key and value matrices, which represent the features of the coil calibration image; d k Represents a scaling factor used to stabilize the attention distribution.

3. The MRI conditional reconstruction method according to claim 1, characterized in that: The method further comprises: Data consistency constraints are applied during the reconstruction process to ensure that the reconstruction results are consistent with the actual acquired K-space data. The data consistency constraint is defined by the following formula: Among them, (y) is the K-space data actually acquired, is the combined operator of coil sensitivity and Fourier transform, representing the data acquisition process, is the conjugate transpose of the operator, which is used to ensure data consistency. λ is a tuning parameter that controls the strength of the data consistency constraint.

4. The MRI conditional reconstruction method according to claim 1, characterized in that: The back diffusion process is defined by the following formula: in, is the noise adjustment coefficient, which controls the noise ratio in the reverse diffusion process, and z is the noise added in the reverse diffusion process.

5. The MRI conditional reconstruction method according to claim 1, characterized in that: The method further comprises: During the model training phase, the encoder parameters of the pre-trained diffusion model are frozen; The conditional control module and decoder parameters are optimized to achieve feature fusion and image reconstruction.

6. The MRI conditional reconstruction method according to claim 5, characterized in that: The method further comprises: A combination of multiple loss functions is used for model training, including mean square error (MSE) loss, structural similarity (SSIM) loss, and perceptual loss.

7. The MRI conditional reconstruction method according to claim 1, characterized in that: The method further comprises: The features of the coil calibration image are adjusted using an adaptive control gate to avoid over-reliance on the calibration image and reduce artifacts in reconstruction.

8. The MRI conditional reconstruction method according to claim 1, characterized in that: The coil calibration image and the undersampled MRI image data may have different contrast and resolution.

9. The MRI conditional reconstruction method according to claim 1, characterized in that: The method further comprises: In the inference phase, the fully sampled calibration image and the undersampled target image are first input; The features are fused through the cross-attention mechanism to generate a preliminary reconstruction result; In combination with the data consistency constraint, further optimizing the reconstructed image; Through multiple iterations of back diffusion and data consistency adjustment, high-quality MRI reconstructed images are finally generated.

10. The MRI conditional reconstruction method according to claim 1, characterized in that: The method is applicable to fast spin echo FSE and echo planar imaging EPI fast MRI imaging sequences. When the internal coil calibration signal cannot be embedded due to signal attenuation, the coil calibration image is used as a conditional input to improve the quality of the reconstructed image.

Images