Diffusion model-based diffusion magnetic resonance image super-resolution reconstruction method

Abstract

A diffusion model-based diffusion magnetic resonance image super-resolution reconstruction method comprises the following steps of 1) performing downsampling and normalization processing on high-resolution diffusion magnetic resonance data, and constructing a low-resolution-high-resolution image pair and a corresponding T1 weighted image as training samples; (2) in the forward process of the diffusion model, Gaussian noise is gradually added and high-resolution and low-resolution image residuals are fused, so that a noise adding result approaches low-resolution image distribution; 3) constructing a network architecture based on a cross attention mechanism, and learning a mapping relation from dispersion data to high-resolution data under different noise levels by combining multi-scale feature fusion and structural prior information of a T1 weighted image; and 4) performing a reverse denoising process by using the trained network to realize super-resolution reconstruction of the dispersion magnetic resonance image. According to the method, the super-resolution reconstruction quality of the diffusion magnetic resonance image is remarkably improved by fusing the multi-modal prior information and the progressive up-sampling strategy.

Description

Technical Field

[0001] This invention relates to the field of medical imaging and diffusion magnetic resonance imaging (dMRI) image processing, and particularly to a method for super-resolution reconstruction of diffusion magnetic resonance images based on a diffusion model. Background Technology

[0002] Diffusion magnetic resonance imaging (dMRI), an important non-invasive imaging technique in neuroscience research, can reveal the microstructural features of white matter fibers by detecting the diffusion characteristics of water molecules in brain tissue. This technology plays a crucial role in brain connectomics research, diagnosis of neurodegenerative diseases, and brain tumor detection. However, in practical applications, the quality of dMRI is limited by several factors: First, high angular resolution imaging requires acquiring a large amount of data in the diffusion gradient directions, leading to a significant increase in scan time, which is often difficult to achieve in clinical settings; second, hardware limitations (such as gradient field strength and signal-to-noise ratio) and unavoidable head movements of the subject can adversely affect image quality; finally, while existing super-resolution techniques based on interpolation or traditional machine learning methods can improve image resolution to some extent, they still suffer from insufficient accuracy when dealing with complex fiber crossing regions and cannot guarantee the accuracy of microstructural indices calculation. These limitations severely restrict the widespread application of dMRI in clinical diagnosis and scientific research.

[0003] To address the aforementioned issues, deep learning-based super-resolution methods have demonstrated significant potential in recent years, particularly diffusion models, which have garnered widespread attention in medical image reconstruction due to their powerful generative capabilities and advantages in modeling high-dimensional data. However, most existing methods rely solely on low-resolution diffusion magnetic resonance imaging (DMRI) data for reconstruction, failing to fully utilize multimodal prior information (such as T1-weighted structural images). This results in error accumulation issues in tasks such as fiber orientation distribution estimation and fiber bundle imaging. Therefore, there is an urgent need for a DMRI super-resolution method that can efficiently fuse multimodal information, maintain the realism of microstructures, and significantly reduce data acquisition requirements to meet the pressing clinical and research demands for high-precision and high-efficiency imaging. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a diffusion-based super-resolution reconstruction method for diffusion magnetic resonance images. It designs a diffusion-based super-resolution reconstruction method for diffusion magnetic resonance imaging (DMI) data types, generating high-resolution DMI images from low-resolution DMI images.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A diffusion-based super-resolution reconstruction method for diffusion magnetic resonance images includes the following steps:

[0007] 1) The high-resolution diffusion magnetic resonance imaging data is downsampled in both the diffusion gradient dimension and the frequency space dimension, and the resulting low-resolution diffusion magnetic resonance imaging, high-resolution diffusion magnetic resonance imaging and corresponding T1-weighted images are normalized.

[0008] 2) During the forward propagation of the diffusion model, Gaussian noise is gradually added to the diffusion image, while the residuals of the low-resolution and high-resolution diffusion images are fused to make the noise-added result approximate the distribution characteristics of the low-resolution diffusion image.

[0009] 3) Construct a deep learning network based on the cross-attention mechanism, combine the structural prior information of T1 weighted images, and learn the mapping relationship between diffuse data and high-resolution data under different noise levels through a multi-scale feature fusion module;

[0010] 4) Use the trained network to perform diffusion model inverse recovery to achieve super-resolution reconstruction of low-resolution diffusion magnetic resonance imaging.

[0011] The technical concept of this invention is as follows: First, the original high-resolution diffusion magnetic resonance data is downsampled and normalized to construct low-resolution-high-resolution image pairs as training samples; based on the reverse denoising process of the diffusion model, the image resolution is gradually improved; a multi-scale feature fusion module is introduced into the network architecture, which combines the structural prior information provided by the T1-weighted image and enhances the feature representation capability through the cross-attention mechanism; finally, a progressive upsampling strategy is adopted to achieve super-resolution reconstruction of the diffusion magnetic resonance image.

[0012] Compared with the prior art, the beneficial effects of the present invention are:

[0013] 1. This invention, through progressive optimization of the diffusion model in both forward and reverse processes, combined with multi-scale feature fusion, can more accurately recover the detailed information of high-resolution diffusion magnetic resonance images, avoiding blurring or artifacts caused by traditional interpolation or convolutional neural network methods.

[0014] 2. This invention introduces structural prior information from T1-weighted images and achieves multimodal feature alignment through a cross-attention mechanism, ensuring that the diffuse image after super-resolution reconstruction remains highly consistent with the anatomical structure and improving the reliability of subsequent analysis.

[0015] 3. The denoising learning strategy based on the diffusion model of this invention can effectively handle low-resolution inputs with different noise levels, adapt to differences in clinical acquisition conditions, and has better generalization ability compared with traditional methods.

[0016] 4. The denoising learning strategy based on the diffusion model of this invention can effectively handle low-resolution inputs with different noise levels, adapt to differences in clinical acquisition conditions, and has better generalization ability compared with traditional methods. Attached Figure Description

[0017] Figure 1 This is the overall flowchart of the present invention.

[0018] Figure 2 This is an illustration of the downsampling of the original image signal x according to the present invention.

[0019] Figure 3 This is a framework diagram of the forward and reverse processes of constructing the diffusion model in this invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the specific implementation methods of this invention will be further derived and described below in conjunction with specific embodiments and accompanying drawings:

[0021] Reference Figures 1-3 A diffusion-based super-resolution reconstruction method for diffusion-based magnetic resonance images includes a forward diffusion process based on residual diffusion; a deep learning network based on a cross-attention mechanism; a cross-attention mechanism module; and a reverse reconstruction process, comprising the following steps:

[0022] 1) Downsampling preprocessing: The diffusion gradient dimension and frequency space dimension of the high-resolution diffusion magnetic resonance imaging are downsampled according to the sampling rate, and the low-resolution diffusion magnetic resonance imaging y, the high-resolution diffusion magnetic resonance imaging x, and the corresponding T1 weighted images obtained by downsampling are all normalized.

[0023] Downsampling is performed on the diffusion gradient dimension and frequency space dimension of high-resolution diffusion magnetic resonance imaging (diffuse magnetic resonance imaging) image x. First, in the diffusion gradient dimension, the gradient angle distribution is optimized using a mixed-integer linear programming (MILP) algorithm to achieve sparse sampling. Simultaneously, in the frequency space dimension, an undersampled fringe mask (e.g., ...) is applied after performing a Fourier transform on the diffusion image in each diffusion gradient direction. Figure 2 As shown in the figure, the downsampled data is finally obtained through inverse Fourier transform.

[0024] 2) During the forward propagation of the diffusion model, Gaussian noise is gradually added to the diffuse image x, while the residuals of the low-resolution diffuse image y and the high-resolution diffuse image x are weighted and fused to make the noisy result x T The distribution characteristics of y in the low-resolution diffuse image are gradually approximated.

[0025] In the forward diffusion process of the diffusion model, Gaussian noise is added, and the residuals e (i.e., e = yx) of the low-resolution diffuse image y and the high-resolution diffuse image x are fused to accelerate the diffusion sampling process. In a Markov chain of length T, their staggered differences e are progressively added, and the diffusion process is described as follows:

[0026] q(x t |x t-1 ,y)=N(x t x+α t e,κα t I),

[0027] Where {α1, α2, ..., α t} are given coefficients used to control the weights of the residual e, and satisfy the following condition: And α1=0, κ is a hyperparameter controlling the Gaussian noise weights, x t The diffusion image at time step t can be further obtained as the diffusion image at any time step x. T Distribution:

[0028]

[0029] In each diffusion step, we gradually perturb the previous state of the sample with Gaussian noise and the residual e. As the time step increases, the sample gradually approaches the distribution of y:

[0030] x t =x t-1 +α t e+κ∈,

[0031] Where, x t-1 Let be the diffuse image at time step t-1, e be the residual between the high-resolution and low-resolution diffuse images, κ be the hyperparameter controlling the weight of Gaussian noise, and ∈ be Gaussian noise that conforms to a standard normal distribution. As the time step increases, the sample will gradually approach the distribution of the low-resolution diffuse image.

[0032] 3) Introduce a deep learning network based on the cross-attention mechanism. Incorporate a multi-scale feature fusion module into the network architecture, and combine it with the structural prior information provided by T1-weighted images to train on diffuse data x with different levels of noise. t Mapping relationship between high-resolution diffuse data;

[0033] In this embodiment, a cross-attention mechanism is introduced, such as... Figure 1 As shown, through the embedding layer, the T1 image is projected into an intermediate representation z. T1 Then, at each time step, it is mapped to the intermediate layer of the network through a cross-attention layer. Figure 1(D)), where query Q and key K are calculated from the output of the embedding layer, while the value V is derived from the intermediate representation of x. t Calculated based on T1 cross-attention layer. It can be represented as:

[0034]

[0035] Among them, Conv Q Conv K and Conv V This indicates a 3×3 convolutional layer with a stride of 1.

[0036] A deep learning network based on a cross-attention mechanism is constructed and trained. This network utilizes pairs of low-resolution diffusion data y, high-resolution diffusion data x, and corresponding T1-weighted images to establish diffusion data x under different diffusion steps. t The mapping relationship with high-resolution diffuse data x, such as Figure 1 As shown, a cross-attention mechanism is embedded in the network module, and training is performed by minimizing the following loss function.

[0037] L=||xf θ (x t ,y,z T1 ,t)|| 2 ,

[0038] Where θ represents the network parameters, f θ (x t ,y,z T1 ,t) is the deep learning network constructed to predict x.

[0039] 4) By performing the reverse recovery process of the diffusion model based on the network trained in step 3), the super-resolution reconstruction results of low-resolution diffusion magnetic resonance imaging can be obtained.

[0040] In this embodiment, super-resolution reconstruction is achieved based on a trained diffusion model. Through the inverse denoising process of the diffusion model, the low-resolution diffusion magnetic resonance image is gradually reconstructed into a high-resolution image. The inverse process p θ (x 0:T |y,z T1 (Starting at x) T As the number of iterations increases, the samples gradually converge to the data distribution of the target high-resolution image. This process can be represented as:

[0041]

[0042] Among them, due to p θ (x t-1 |x t ,y,z T1The variance of p follows a standard normal distribution. θ (x t-1 |x t ,y,z T1 It can be approximated by its mean μ. θ (x t ,y,z T1 ,t) can be represented as:

[0043]

[0044] The mapping network f corresponding to each time step is trained in the fourth step. θ (x t ,y,z T1 The method (t) simultaneously processes the noisy image of the current step and the given prior information, i.e., the T1-weighted image, during the reverse reconstruction process to achieve super-resolution reconstruction of the diffuse magnetic resonance image.

[0045] In the simulation of this invention, since random noise is introduced during the reverse reasoning process, a multi-round combined reasoning strategy is adopted to stabilize the denoising results. Specifically, multiple rounds of reasoning are performed on the same input, and the average of the results is taken as the final output.

[0046] This invention compares and analyzes three image super-resolution algorithms and zero-filling (ZF) images: Unet, Vision Transformer (VIT), and Swin Transformer for MRI (Swin_mr). The performance of various methods is evaluated using two metrics: Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM). Table 1 describes the reconstruction results of diffused images under different frequency spatial dimension undersampling rates r_k and diffusion gradient dimension undersampling rates r_q, demonstrating that our dMRI image super-resolution strategy significantly improves upon the aforementioned methods. Table 1 shows the results of calculating PSNR and SSIM for the reconstructed images;

[0047]

[0048] Table 1

[0049] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for super-resolution reconstruction of diffusion magnetic resonance images based on a diffusion model, characterized in that, The method includes the following steps: 1) The high-resolution diffusion magnetic resonance imaging data is downsampled in both the diffusion gradient dimension and the frequency space dimension, and the resulting low-resolution diffusion magnetic resonance imaging, high-resolution diffusion magnetic resonance imaging and corresponding T1-weighted images are normalized. 2) During the forward propagation of the diffusion model, Gaussian noise is gradually added to the diffusion image, while the residuals of the low-resolution and high-resolution diffusion images are fused to make the noise-added result approximate the distribution characteristics of the low-resolution diffusion image. 3) Construct a deep learning network based on the cross-attention mechanism, combine the structural prior information of T1 weighted images, and learn the mapping relationship between diffuse data and high-resolution data under different noise levels through a multi-scale feature fusion module; 4) Use the trained network to perform diffusion model inverse recovery to achieve super-resolution reconstruction of low-resolution diffusion magnetic resonance imaging.

2. The method for super-resolution reconstruction of diffusion magnetic resonance images based on a diffusion model according to claim 1, characterized in that, In step 1), downsampling is performed in the diffusion gradient dimension and the frequency space dimension to construct low-resolution-high-resolution image pairs and corresponding T1-weighted images as training samples.

3. A method for super-resolution reconstruction of diffusion magnetic resonance images based on a diffusion model according to claim 1 or 2, characterized in that, In step 2), diffusion is performed using a linear combination of Gaussian noise and the residuals of high- and low-resolution diffused images, where the residual weighting coefficients satisfy... And α1 = 0, for any time step t, the diffuse image x t Its distribution is represented as: x t =x t-1 +a t e+κ∈ Where, x t-1 Let be the diffuse image at time step t-1, e be the residual between the high-resolution and low-resolution diffuse images, κ be the hyperparameter controlling the weight of Gaussian noise, and ∈ be Gaussian noise that conforms to a standard normal distribution. As the time step increases, the sample will gradually approach the distribution of the low-resolution diffuse image.

4. A method for super-resolution reconstruction of diffusion magnetic resonance images based on a diffusion model according to claim 1 or 2, characterized in that, In step 4), super-resolution reconstruction is achieved based on the trained diffusion model. Through the reverse denoising process of the diffusion model, the low-resolution diffusion magnetic resonance image is gradually reconstructed into a high-resolution image. This process is represented as follows: Among them, due to p θ (x t-1 |x t ,y,z T1 The variance of p follows a standard normal distribution. θ (x t-1 |x t ,y,z T1 It can be approximated by its mean μ. θ (x t ,y,z T1 ,t), is represented as: The mapping network f corresponding to each time step trained by 3) θ (x t ,y,z T1 The algorithm, t), processes the noisy image of the current step and the given prior information, i.e., the T1-weighted image, simultaneously during the reverse reconstruction process. It updates t = t-1 during iteration, repeats the steps until t = 0, and outputs x0 as the final high-resolution reconstructed image.

Images