An anisotropic MRI super-resolution method based on a one-step diffusion model

Abstract

This invention relates to an anisotropic MRI super-resolution method based on a one-step diffusion model. The method includes: step 1, data preprocessing; step 2, pre-training a diffusion model; step 3, training a one-step diffusion model; and step 4, testing the one-step diffusion model. The superior technical effects of this invention are: it proposes a one-step diffusion model that can generate high-resolution MR volumetric data at a relatively fast speed; by designing a training strategy based on variational fractional distillation to align the diffusion fractions of two teacher models, the student diffusion model can capture the distribution mapping relationship between random noise and high-resolution MR data, thereby eliminating the iterative sampling process required by traditional diffusion models; it proposes noise-aware contrastive learning to further reduce the difference between the generated sample distribution and the reference data distribution; and the method of this invention can quickly achieve high-quality three-dimensional MRI super-resolution and alleviate the memory pressure on diffusion models when processing three-dimensional MR data.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical image processing, and particularly relates to an anisotropic MRI super-resolution method based on a one-step diffusion model. BACKGROUND

[0002] Magnetic resonance imaging (MRI) is an important clinical imaging method, which has the advantages of non-ionizing radiation and non-invasive. However, due to the limitations of scanning equipment and the inherent physical properties of MRI, it takes a lot of time to obtain high-resolution data. At the same time, accelerating the imaging process will result in a low signal-to-noise ratio of the MR image. In order to ensure the imaging quality and shorten the imaging time, a series of thick slices on the imaging plane (x-y plane) are often collected in the clinic, and are stacked along the z-axis as the final three-dimensional imaging result. A larger slice thickness will cause a serious anisotropy of voxels, that is, a high resolution in the x-y plane and a low resolution in the x-z and y-z planes. The MR volume data with anisotropic spatial resolution is difficult to provide accurate human tissue information, thereby affecting early diagnosis and subsequent processing.

[0003] In recent years, anisotropic MRI super-resolution methods based on deep learning have developed rapidly. Most existing methods use convolutional neural networks or Transformers to construct an end-to-end network architecture. The network aims to learn the mapping relationship between low-resolution MR data and high-resolution MR data. In the training stage, the super-resolution result and the original high-resolution label are used to calculate the loss function, so as to optimize the network parameters. However, the end-to-end anisotropic super-resolution method is difficult to accurately model the distribution characteristics of MR data, and is prone to over-smoothing and loss of high-frequency details.

[0004] For example, the patent application file with the Chinese invention patent application number CN201980018799.8 discloses a system and method for generating thin image slices from thick image slices, which uses a combination of multiple three-dimensional convolutional layers and linear activation functions for feature extraction, and finally obtains a residual map between the low-resolution input and the high-resolution label. Although the residual connection can accelerate convergence and preserve the original low-resolution spatial information, the model still cannot capture the accurate high-resolution data distribution, so that incorrect texture structures may be recovered in the test stage.

[0005] For example, the patent application file with the Chinese invention patent application number CN202111429939.6 discloses a head three-dimensional MRI super-resolution reconstruction method. By combining convolutional neural networks, Transformers, and fully connected layers, the three-dimensional super-resolution method can simultaneously focus on spatial information in different slice directions and effectively integrate three axial acquisition data at the output end. However, the multi-head attention mechanism in the Transformer and the fully connected layer have high computational complexity, which can cause high training costs and slow inference speed. In addition, the insufficient training of the traditional Transformer can cause the network to fail to learn high-quality distribution mapping relationships, resulting in blurred high-frequency details in the super-resolution results.

[0006] For another example, the patent application file with the Chinese invention patent application number CN202410580260.4 discloses a high-resolution three-dimensional isotropic fetal brain MRI reconstruction method. The method includes a motion correction network and a three-dimensional MR volume reconstruction network for simultaneously performing motion parameter matrix estimation, slice registration, and thick slice to thin slice super-resolution. By alternately optimizing the two networks, the method can accurately eliminate motion artifacts and provide high-resolution three-dimensional MR volume data. However, the three-dimensional MR volume reconstruction network receives three-dimensional data directly spliced from the corrected two-dimensional slices, which destroys the continuity between adjacent slices and can cause a large distribution difference between the final three-dimensional super-resolution result and the real three-dimensional high-resolution data.

[0007] Based on the problem that existing disclosed invention patent applications cannot accurately model isotropic high-resolution MR volume distribution, the present application uses the high-quality distribution fitting capability of the diffusion model to perform anisotropic MRI super-resolution. By taking low-resolution MR data as a condition, the diffusion model can restore high-resolution content in the z-axis direction. However, existing diffusion model-based methods still have the following problems:

[0008] First, the diffusion model needs multiple iterations to generate high-resolution MR volume data. Some methods use fast sampling algorithms to reduce the number of iteration steps in the inverse sampling process, and when the number of iteration steps is too small, the quality of the generated result will decrease significantly. Deploying the diffusion model in the latent space can also speed up the generation. However, it is worth noting that the performance of the latent space diffusion model is severely limited by the pre-trained autoencoder. For recovery tasks, the autoencoder cannot fully utilize the prior information of the low-resolution data, which can easily cause the generated result to be over-blurred. Although the above methods can speed up the inference speed of the diffusion model to some extent, they still have an iterative denoising process and do not meet the imaging requirements in clinical scenarios.

[0009] Secondly, when training directly on three-dimensional volume data, the diffusion model produces a large amount of memory occupation. Some methods choose to train a two-dimensional diffusion model on the x-y plane, and stack the slices in order to form the final MR volume data in the inference stage. However, the two-dimensional diffusion model ignores the correlation between adjacent slices, and usually shows weaker generation ability than the three-dimensional diffusion model. SUMMARY

[0010] Based on the problems or defects of the prior art described above, the present application proposes an anisotropic MRI super-resolution method based on one-step diffusion model, which can generate isotropic high-resolution MR volume data at a faster inference speed, and effectively alleviate the demand for computing resources of three-dimensional diffusion model by using sub-volume based training method.

[0011] The anisotropic MRI super-resolution method based on one-step diffusion model, comprising the following steps:

[0012] Step 1, data preprocessing stage: pre-process the high-resolution MR volume data in the training set and test set, simulate low-resolution MR volume data with different slice thicknesses, apply random overlap cropping operation to crop the high-resolution volume data and low-resolution volume data along the z-axis, and construct sub-volume training data;

[0013] Step 2, pre-training diffusion model stage: pre-train a diffusion model capable of performing anisotropic MRI super-resolution task through denoising diffusion loss, learn the prior distribution of high-resolution MR data, and provide initial weights for related network training in one-step diffusion model;

[0014] Step 3, training one-step diffusion model stage: use the trained diffusion model to construct one-step diffusion student model and two teacher diffusion models, perform variational score distillation by constraining the consistency of the prediction results of the two teacher models, so that the student model has one-step generation ability, use the prediction results of the student model and the high-resolution sub-volume data as positive sample pairs, and use Gaussian noise as negative sample, perform noise-aware contrastive learning to improve the accuracy of the one-step generation ability of the student model, and after optimizing the student model, fine-tune one of the teacher models using denoising diffusion loss;

[0015] Step 4, test one-step diffusion model stage: non-overlapping cropping along the z-axis is performed on the low-resolution MR volume in the test set to obtain a series of sub-volume data, which is input into the trained one-step diffusion student model in order together with random Gaussian noise, and the maximum time step is used to sample step to obtain high-resolution sub-volume prediction results, and the inferred high-resolution sub-volume data is stacked together to form the final high-resolution MR volume data.

[0016] ​Furthermore, step 1 specifically includes:

[0017] Step 1.1: Divide the high-resolution MR volume data into training and testing sets, and process them to meet the size requirements of U-Net upsampling and downsampling;

[0018] Step 1.2: Downsample the high-resolution MR volume data along the z-axis. The downsampling factor is consistent with the slice thickness. The downsampled volume data is restored to the same size as the original high-resolution volume data by repeated interpolation operations. The average value of the overall pixels of the z-axis slice is taken to simulate low-resolution MR volume data.

[0019] Step 1.3: Randomly select cropping positions in the z-axis direction and apply random overlap cropping to the high-resolution MR volume data and low-resolution MR volume data in the training set with a fixed number of slices to obtain sub-volume data with partially overlapping anatomical content in the z-axis direction.

[0020] Furthermore, step 2 specifically includes:

[0021] Step 2.1, based on the forward diffusion process, it is known that when the starting high-resolution sub-volume data... When there is a clear definition, it is possible to obtain any The forward Gaussian distribution at time t is as follows (1):

[0022] ... (1),

[0023] in, for At this moment The forward noise addition result, These are fixed hyperparameters. The identity matrix is ​​derived from the expression for the forward Gaussian distribution. The mean is The variance is Then it can be solved by reparameterization. As shown in equation (2), equation (2) is the formula for calculating forward noise addition:

[0024] ... (2),

[0025] in, , is with Random Gaussian noise of the same size It will happen over time It decreases as it increases, when hour, It is pure Gaussian noise;

[0026] Step 2.2, divide the data at different time steps With low-resolution subvolume data Connect by channel, and by time Together, they serve as input to the 3D U-Net, upon which the 3D U-Net inference is based. Additive Gaussian noise at time t That is, iterative optimization is performed using the following loss function, as shown in equation (3):

[0027] ... (3),

[0028] in, Represents a 3D U-Net. express The optimization parameters in Indicates that it is used for optimization Denoising diffusion loss, This indicates the calculation of mean square error;

[0029] Step 2.3, when the maximum number of iterations is reached, Stop training and save the optimized weights. .

[0030] Furthermore, step 3 specifically includes:

[0031] Step 3.1: Construct three diffusion models based on 3D U-Net: one one-step diffusion student model. A trainable teacher diffusion model A teacher diffusion model that cannot be trained and with the completed training initialization , and ;

[0032] Step 3.2, obtain At the maximum time step Noise prediction results As shown in equation (4):

[0033] ... (4),

[0034] in, Representing random Gaussian noise, combined with the forward noise addition formula, yields the super-resolution result generated in one step. As shown in equation (5):

[0035] ... (5);

[0036] Step 3.3, for Applying the forward noise formula as and The input yields the following equation (6):

[0037] ... (6),

[0038] in, for At time step The forward noise addition results are as follows. , To represent random Gaussian noise, calculate according to the following formula (7). and At different time steps Noise prediction results and As a diffusion fraction:

[0039] ... (7),

[0040] Variational distillation loss Intended to and The difference between them is passed to Then the gradient It can be expressed as the following formula (8):

[0041] ... (8),

[0042] therefore, The specific calculation formula is as follows (9):

[0043] ...(9),

[0044] in, This indicates the gradient clipping operation, which is... The value is considered constant, and no gradient is passed to the network by any element inside it;

[0045] Step 3.4, calculate according to the following formula (10). exist Forward noise addition results :

[0046] ... (10),

[0047] choose Encoding path As a feature extractor, it is used to perform noise-aware contrastive learning, where the anchor point and the positive sample are respectively , At the same time step The results of forward diffusion and Negative samples are Then compare the losses It can be represented as:

[0048] ... (11),

[0049] in, express The set of selected coding layer numbers, Indicates from The Features extracted from the layers, As a weighting factor, Indicates L2 distance calculation;

[0050] Step 3.5, will and Linear combination is used as the total loss function for training. :

[0051] ... (12),

[0052] in, Represents the total loss function. For hyperparameters;

[0053] Step 3.6, when completed During optimization, calculation exist Forward noise addition results :

[0054] ... (13),

[0055] in, Represents random Gaussian noise.

[0056] application right Make minor adjustments:

[0057] ... (14);

[0058] Step 3.7, iterative alternating optimization and Training stops when the maximum number of iterations is reached, and the optimized one-step diffusion model weights are saved. .

[0059] Furthermore, step 4 specifically includes:

[0060] Step 4.1: Divide the low-resolution MR volume data in the test set into non-overlapping sub-volume data along the z-axis on an average basis;

[0061] Step 4.2, Load the saved one-step diffusion model Randomly generate Gaussian noise of the same size as the sub-volume, and input it sequentially along with the low-resolution sub-volume. ,by For the time step, the corresponding subvolume super-resolution result is obtained as follows (15):

[0062] ... (15),

[0063] in, Indicates the generated first High-resolution MR subvolume data Represents random Gaussian noise. This represents low-resolution MR subvolume data;

[0064] Step 4.3: Stack them sequentially along the z-axis. The final high-resolution MR volume data is obtained by calculating the following formula (16). :

[0065] ... (16),

[0066] in, This indicates a stacking operation along the z-axis. This indicates the total number of sub-volumes.

[0067] Compared to existing anisotropic MRI super-resolution techniques based on diffusion models, the superior technical effects of the anisotropic MRI super-resolution method based on a one-step diffusion model described in this invention are as follows:

[0068] 1. The anisotropic MRI super-resolution method based on a one-step diffusion model described in this invention proposes a one-step diffusion model that can generate high-resolution MR volume data at a relatively fast speed. By designing a training strategy based on variational fractional distillation to align the diffusion fractions of two teacher models, the student diffusion model can capture the distribution mapping relationship between random noise and high-resolution MR data, thereby eliminating the iterative sampling process required by traditional diffusion models.

[0069] 2. The anisotropic MRI super-resolution method based on a one-step diffusion model described in this invention proposes noise-aware contrastive learning to further reduce the difference between the distribution of generated samples and the distribution of reference data. By narrowing the distance between high-resolution MR data and the forward diffusion results of generated samples at different feature levels, the one-step diffusion model can learn a more accurate one-step generation capability.

[0070] 3. The anisotropic MRI super-resolution method based on a one-step diffusion model described in this invention is implemented on sub-volume data during both the training and testing phases, reducing the memory requirements of the diffusion model when processing 3D MR data; by applying a random overlap cropping operation along the z-axis during the training phase, high-resolution content on the xy-plane is effectively preserved, while ensuring the smoothness of the xz-plane and yz-plane. Attached Figure Description

[0071] Figure 1 This is a schematic diagram of the flow of the anisotropic MRI super-resolution method based on a one-step diffusion model as described in this invention.

[0072] Figure 2 This is a schematic diagram of the random overlap cropping operation and the pre-trained diffusion model stage architecture of the anisotropic MRI super-resolution method based on a one-step diffusion model described in this invention.

[0073] Figure 3 This is a schematic diagram of the three-dimensional U-Net network structure shown in the anisotropic MRI super-resolution method based on a one-step diffusion model described in this invention.

[0074] Figure 4 This is a schematic diagram illustrating the training stage architecture of the one-step diffusion model in the anisotropic MRI super-resolution method based on the one-step diffusion model described in this invention. Figure 4 a is a schematic diagram of one-step diffusion generation. Figure 4 b is a schematic diagram of variational fractional distillation. Figure 4 c. A schematic diagram of noise perception contrastive learning.

[0075] Figure 5 This is a schematic diagram of the test one-step diffusion model stage architecture for the anisotropic MRI super-resolution method based on the one-step diffusion model described in this invention. Detailed Implementation

[0076] To make the objectives, technical solutions, and superior technical effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The specific embodiments of this invention are only used to illustrate the invention and to more clearly demonstrate its design, features, and technical content, and do not constitute any limitation on the invention. Example

[0077] like Figure 1 As shown, the anisotropic MRI super-resolution method based on a one-step diffusion model includes the following steps:

[0078] Step 1, Data Preprocessing Stage: Using T1-weighted and T2-weighted brain datasets respectively, the dataset and test set are proportionally divided, and the MR volume size is adjusted to meet the size requirements of U-Net upsampling and downsampling. This simulates thick slice low-resolution MR volume data. A random overlap cropping operation is applied to the high-resolution and low-resolution volume data along the z-axis to construct sub-volume data for training. Using the sub-volumes divided along the z-axis as training data not only reduces the GPU memory requirements for optimizing the 3D model but also further preserves high-resolution information in the xy-plane. Furthermore, the random overlap cropping operation ensures a smooth transition in the z-axis direction of the final volume reconstruction result. The specific steps are as follows:

[0079] Step 1.1, the voxel size of both T1-weighted and T2-weighted brain data is [missing information]. For the T1 weighted brain dataset, 1338 MR volume data points were used as the training set and 200 MR volume data points were used as the test set. The size of each MR volume data point was reduced to 128×128×128 through downsampling. For the T2 weighted brain dataset, 547 MR volume data points were randomly selected as training samples and the remaining 20 MR volume data points were used as test samples. Unnecessary zero-value backgrounds were cropped to obtain MR volume data points with a size of 256×256×128.

[0080] Step 1.2: The processed high-resolution volumetric data is downsampled by 4x and 8x along the z-axis. Interpolation is performed based on adjacent slices to restore the original size. The average pixel value of the z-axis slice is then calculated to simulate a... and Low-resolution MR volumetric data of slice thickness;

[0081] Step 1.3: In the z-axis direction, randomly overlap and crop the low-resolution MR volume data, retaining 8 consecutive xy-plane slices as low-resolution sub-volume data. The same operation is applied to the high-resolution MR volume data to obtain high-resolution sub-volume data, reducing the dimensionality of the sub-volume data from... Switch to This yields T1-weighted subvolume data of size 8×128×128 and T2-weighted subvolume data of size 8×256×256. Figure 2 This demonstrates a random overlapping cropping operation;

[0082] Step 2, Pre-training the Diffusion Model: The diffusion model is pre-trained using low-resolution MR subvolume data to obtain a diffusion model that performs anisotropic super-resolution, capturing the generation prior. When the pre-trained model can accurately predict noise, the one-step diffusion model obtains good initial weights, providing high-quality one-step super-resolution results at the starting point of training, thereby reducing training pressure. Figure 2 This is a schematic diagram of the pre-training diffusion model stage. The specific steps are as follows:

[0083] Step 2.1: For anisotropic super-resolution tasks, the forward process of the diffusion model is considered as processing high-resolution sub-volume data. The process of gradually adding noise, and The noise addition results at each moment and The noise addition results at each time step are related, i.e., they satisfy the following Markov chain form:

[0084] ,

[0085] in, Indicates the maximum time step, when hour, It is pure Gaussian noise. These are fixed hyperparameters used to represent variance, and they change with... Increases with the increase, set here. , ,

[0086] According to the Markov recursion, any Forward noise addition results at time 1 All can pass It is obtained directly and satisfies the following Gaussian distribution:

[0087] ,

[0088] in, , ,

[0089] From the expression of the forward Gaussian distribution, we can obtain The mean is The variance is The following equation can be solved directly by reparameterization. :

[0090] ,

[0091] The above formula is the forward noise addition formula, where, This represents random Gaussian noise;

[0092] Step 2.2, using a 3D U-Net as a noise predictor, where low-resolution sub-volume data As conditional information, with Connect by channel, and The 3D U-Net, fed into the network, consists of four downsampling operations and four upsampling operations. Downsampling reduces the feature map size to half its original size, while upsampling increases it to twice its original size. The input convolutional layers are used for fusion. and Both the encoding and decoding stages use residual blocks containing two convolutional layers for feature extraction. Features from different levels are passed from the upsampling path to the downsampling path via skip connections. Features of the same size are connected by channel and fused with the residual blocks in the upsampling path. The time step... After dimensionality expansion and multilayer perceptron encoding, the input is fed into the cross-attention layer at each stage, assigning different attention weights to different noisy images, and the output predicts the noise. Figure 3 This is a network structure diagram of a 3D U-Net. The noise prediction capability of the 3D U-Net is determined by the denoising diffusion loss function. guide:

[0093] ,

[0094] in, Represents a 3D U-Net. express The optimization parameters in The optimization target is , This indicates the calculation of mean square error;

[0095] Use the Adam optimizer to perform gradient descent, learning rate Perform iterative optimization, with the number of iterations set to... Set the batch size to 2;

[0096] Step 2.3, when the maximum number of iterations is reached, Stop training and save the optimized weights. ;

[0097] Step 3, Training the One-Step Diffusion Model: (After training is complete) The weights of the 3D U-Net are used to initialize the three diffusion models: one student diffusion model and two teacher diffusion models. Variational fractional distillation is performed by forcing consistency in the noise prediction results of the two teacher models. The student diffusion model learns how to directly map random noise distributions to high-resolution data distributions. Utilizing the sensitivity of the denoising 3D U-Net's encoding path to noise levels, noise-aware contrastive learning is performed to narrow the gap between the super-resolution results and the high-resolution labels at different time steps. The student diffusion model can fit a more accurate high-resolution data distribution, thereby generating higher-quality subvolume data. Figure 4 a to 4c are schematic diagrams of training a one-step diffusion model. The specific steps are as follows:

[0098] Step 3.1, construct three diffusion models based on 3D U-Net: each being a trainable one-step diffusion student model. A trainable teacher diffusion model A teacher diffusion model that cannot be trained ,in , and Weights saved during the pre-training phase initialization;

[0099] Step 3.2, obtain At the maximum time step Noise prediction results :

[0100] ,

[0101] in, This represents random Gaussian noise, because Predicting Gaussian noise, therefore It cannot be directly used as a one-step diffusion result. In order to reduce the domain gap with MR data, a forward noise addition formula is used to obtain the high-resolution sub-volume. The approximate solution is used as the result of a one-step diffusion. , Figure 4 a is a schematic diagram of one-step diffusion generation. The calculation process is as follows:

[0102] ;

[0103] Step 3.3, variational fractional distillation loss Aiming to utilize KL loss Align the two distributions and :

[0104] ,

[0105] in, This represents the target distribution to be fitted, and it has... , It is all The distribution of the sample does not need to be calculated directly. The following formula only requires the parameters. gradient :

[0106] ,

[0107] When on When applying the forward diffusion process, fractions and Diffusion models at different time steps and To approximate, therefore, it needs to be calculated according to the following formula. Forward noise addition results:

[0108] ,

[0109] in, , To represent random Gaussian noise, calculate according to the following formula. and At different time steps Noise prediction results and As a diffusion fraction:

[0110] ,

[0111] So, combining The calculation expression, Passed to gradient Calculated by the following formula:

[0112] ,

[0113] in, , Figure 4 b is a schematic diagram of variational fractional distillation, obtained based on the differential properties. The expression is as follows:

[0114] ,

[0115] in, This indicates the gradient clipping operation, which is... The value is considered constant, and no gradient is passed to the network by any element inside it;

[0116] Step 3.4, in noise perception contrastive learning, , The positive diffusion results are used as anchor points and positive samples. As a negative sample, it is necessary to calculate exist Forward noise addition results As shown in the following formula:

[0117] ,

[0118] use Encoding path Perform feature extraction and compare the losses. Guided one-step diffusion student model to perform contrastive learning, aiming to narrow the gap at different feature levels. and The distance between them, widen and The distance between them Figure 4 c is a schematic diagram of noise-perception contrastive learning. The expression is as follows:

[0119] ,

[0120] in, express The set of selected coding layer numbers, Indicates from The Features extracted from the layers, As a weighting factor, This indicates L2 distance calculation; it is set here. , ;

[0121] Step 3.5, and right When performing collaborative training, the total loss function... The expression is:

[0122] ,

[0123] in, Hyperparameters for balancing the proportions of the two loss functions;

[0124] Step 3.6, when When completing a training session, it is necessary to... Make fine adjustments to track The distribution changes, therefore, calculation Forward noise addition results As Input:

[0125] ,

[0126] in, Represents random Gaussian noise. As shown in the following formula right Fine-tuning:

[0127] ,

[0128] Use the Adam optimizer to perform gradient descent, learning rate Update, then... learning rate Update and set the number of training iterations to [number]. Set the batch size to 2;

[0129] Step 3.7, when the maximum number of iterations is reached, and Stop training and save the optimized weights. .

[0130] Step 4, Testing the One-Step Diffusion Model: Loading the trained model A one-step diffusion process is performed to generate high-resolution MR volumetric data. Specifically, the low-resolution MR volumetric data is divided into non-overlapping sub-volumes along the z-axis. This sub-volumes serve as conditional information to control the generation process. Noise samples of the same size as the sub-volumes are randomly generated and concatenated with the low-resolution sub-volumes along the channel, with the maximum time step. For the sampling step, input in sequence. High-resolution MR sub-volume data is generated, and these sub-volume data are stacked sequentially along the z-axis to obtain the final high-resolution MR volume data. Figure 5 To test the schematic diagram of the one-step diffusion model, the specific steps are as follows:

[0131] Step 4.1: Clip the low-resolution MR volumetric data in the test set along the z-axis into non-overlapping sub-volumetric data, and change the dimension from... Switch to For the T1-weighted brain dataset, the size of each low-resolution sub-volume data is 8×128×128, and for the T2-weighted brain dataset, the size of each low-resolution sub-volume data is 8×256×256.

[0132] Step 4.2: Randomly generate Gaussian noise of the same size as the sub-volume, as... The one-step diffusion target uses low-resolution MR subvolume data as control conditions, and noise and subvolume data are input sequentially. ,by Using time steps, the reconstructed high-resolution MR subvolume data is obtained as follows:

[0133] ,

[0134] in, Indicates the generated first High-resolution MR subvolume data Represents random Gaussian noise. This represents low-resolution MR subvolume data;

[0135] Step 4.3: Stack the generated high-resolution MR sub-volume data along the z-axis to obtain super-resolution results at the volumetric level. :

[0136] ,

[0137] in, This indicates a stacking operation along the z-axis. This represents the total number of sub-volumes. Restore dimensions to This results in the final high-resolution MR volumetric data.

[0138] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. An anisotropic MRI super-resolution method based on one-step diffusion model, comprising: Step 1, data preprocessing stage: pre-process high-resolution MR volume data in the training set and test set, simulate low-resolution MR volume data with different slice thickness, apply random overlap cropping operation to crop high-resolution volume data and low-resolution volume data along the z-axis, and construct sub-volume training data; Step 2, pre-training diffusion model stage: pre-train a diffusion model capable of performing anisotropic MRI super-resolution task through denoising diffusion loss, learn the prior distribution of high-resolution MR data, and provide initial weights for related network training in one-step diffusion model; Step 3, training one-step diffusion model stage: use the trained diffusion model to construct one-step diffusion student model and two teacher diffusion models, perform variational score distillation by constraining the consistency of the prediction results of the two teacher models, make the student model have one-step generation capability, use the prediction results of the student model and high-resolution sub-volume data as positive sample pairs, and use Gaussian noise as negative sample, perform noise-aware contrastive learning to improve the accuracy of the one-step generation capability of the student model, and after optimizing the student model, fine-tune one of the teacher models using denoising diffusion loss; Step 4, test one-step diffusion model stage: non-overlapping cropping along z-axis is performed on the low-resolution MR volume in the test set to obtain a series of sub-volume data, which is sequentially input into the trained one-step diffusion student model together with random Gaussian noise to obtain the predicted high-resolution sub-volume data at each time step For the sampling step, the high-resolution sub-volume prediction results are obtained, and the inferred high-resolution sub-volume data is stacked together to form the final high-resolution MR volume data.

2. The anisotropic MRI super-resolution method based on one-step diffusion model according to claim 1, step 1 specifically comprising: Step 1.1, divide the high-resolution MR volume data into training set and test set, and process it to meet the size of U-Net up-sampling and down-sampling; Step 1.2, down-sample the high-resolution MR volume data along the z-axis, the down-sampling factor is consistent with the slice thickness, restore the down-sampled volume data to the same size as the original high-resolution volume data through repeated interpolation operation, and take the mean value of the overall pixels of the z-axis slice to simulate low-resolution MR volume data; Step 1.3, randomly select the cropping position in the z-axis direction, apply random overlap cropping operation to the high-resolution MR volume data and low-resolution MR volume data in the training set with fixed slice number, to obtain sub-volume data with partially overlapping anatomical content in the z-axis direction.

3. The anisotropic MRI super-resolution method based on one-step diffusion model according to claim 1, wherein step 2 specifically comprises: Step 2.

1. From the forward diffusion process, we know that when the starting high-resolution subvolume data arbitrary forward Gaussian distribution at any time instant, as follows (1): …… (1)， wherein, is at the moment of the forward noise addition result, is a fixed hyperparameter, denotes a unit matrix, and the mean of the forward Gaussian distribution expression is , and the variance is , then the can be solved by reparameterization , as follows in equation (2), equation (2) being a forward noise addition calculation equation: …… (2)， wherein is a random Gaussian noise of the same size as , and will decrease as time increases, when , is a pure Gaussian noise; Step 2.2, the sub-volumes at different time steps with low-resolution sub-volume data connected by channels, and time together as the input of the three-dimensional U-Net, based on which the three-dimensional U-Net infers additive Gaussian noise at the time instant i.e. iteratively optimized by the following loss function, as shown in equation (3): …… (3)， in, Represents a 3D U-Net. express The optimization parameters in Indicates that it is used for optimization Denoising diffusion loss, This indicates the calculation of mean square error; Step 2.

3. When the maximum number of iterations is reached, Stop training and save the optimized weights .

4. The anisotropic MRI super-resolution method based on one-step diffusion model according to claim 1, wherein step 3 specifically comprises: Step 3.1: Construct three diffusion models based on 3D U-Net: one one-step diffusion student model. A trained teacher diffusion model A teacher diffusion model that cannot be trained and with the completed training initialization , and ; Step 3.2, obtaining the noise prediction result at the maximum time step as follows (4):​ …… (4)， wherein, represents a random Gaussian noise, and combining the forward noise adding formula, a one-step generated super-resolution result is obtained as the following formula (5): …… (5)； Step 3.3, for applying the forward noise adding formula as and input, the following formula (6) is obtained: …… (6)， wherein is the forward noisy result at time step , , denotes a random Gaussian noise, computed according to the following equation (7) and the noise prediction result at different time steps , and as a diffusion score: …… (7)， Variational distillation losses Intended to and The difference between them is passed to Then the gradient It can be expressed as the following formula (8): …… (8)， Thus, The specific calculation formula is the following formula (9). ……(9)， wherein, represents a gradient clipping operation, i.e. the values of are treated as constants, and all elements inside are not passing gradients to the network they belong to; Step 3.4, the forward noise addition result is calculated according to the following formula (10) In the forward noise addition result : …… (10)， selecting an encoding path as a feature extractor for performing noise-aware contrastive learning, wherein the anchor and the positive sample are , the forward diffusion result at the same time step and , and the negative sample is , then the contrastive loss can be expressed as: ……(11)， wherein, denotes a set of selected encoding layer numbers, denotes features extracted from the first layer, is a weighting factor, denotes L2 distance computation; Step 3.5, to with linear combination as the total loss function, to train : …… (12)， wherein, represents the total loss function, is a hyperparameter; Step 3.6, when the optimization of the is completed, the forward noise added result under is calculated …… (13)， wherein denotes a random Gaussian noise, Applications To Fine-tune: …… (14)； Step 3.7, Iterative Alternating Optimization and Stop training when the maximum number of iterations is reached and save the one-step diffusion model weights that have been optimized .

5. The anisotropic MRI super-resolution method based on one-step diffusion model according to claim 1, wherein step 4 specifically comprises: Step 4.1, divide the low-resolution MR volume data in the test set into non-overlapping sub-volume data along the z-axis. Step 4.2, load the saved one-step diffusion model , randomly generate Gaussian noise with the same size as the sub-volume, and input it in sequence with the low-resolution sub-volume , as the time step, the corresponding sub-volume super-resolution result is as follows formula (15): ​ …… (15)， wherein represents the generated first high-resolution MR sub-volume data, represents random Gaussian noise, represents low-resolution MR sub-volume data; Step 4.3, stacking in z-axis direction sequentially The final high-resolution MR volume data is calculated as follows (16) : …… (16)， wherein, represents sequentially stacking operations along the z-axis, represents the total number of sub-volumes.

Images