Heart magnetic resonance image blind super-resolution reconstruction method and system based on continuous time condition diffusion model

Abstract

The invention discloses a heart magnetic resonance image blind super-resolution reconstruction method and system based on a continuous time condition diffusion model, and relates to the field of heart magnetic resonance image blind super-resolution reconstruction. Carrying out data enhancement and degradation processing on the heart magnetic resonance image in the heart magnetic resonance image data set to obtain a low-resolution image; constructing a continuous time condition diffusion model based on a cascade residual attention network feature extractor, a continuous time condition diffusion module, a hybrid parameterized score predictor and an image quality loss module; setting an optimizer, training parameters and hardware configuration to train continuous time conditional diffusion model parameters; a to-be-detected heart magnetic resonance source image is obtained, the continuous time condition diffusion model is input, a super-resolution reconstruction image is generated, and efficient and high-precision medical image super-resolution reconstruction is achieved by improving the time efficiency and the generation quality of the diffusion model.

Description

Technical Field

[0001] The present invention relates to the technical field of blind super-resolution reconstruction of cardiac magnetic resonance images, and in particular to a method and system for blind super-resolution reconstruction of cardiac magnetic resonance images based on a continuous-time conditional diffusion model. Background Art

[0002] Cardiac magnetic resonance imaging (CMRI) is an important non-invasive medical imaging technology with significant clinical value in the assessment of cardiovascular disease. High-resolution CMRI is crucial for the early detection of heart disease and improving diagnostic accuracy and efficiency. However, traditional diffusion models rely on iterative sampling of Markov chains, resulting in a very time-consuming super-resolution reconstruction process. Existing methods suffer from problems such as blurred details and unclear edges when dealing with complex degradation. Traditional methods do not adequately model the mapping relationship between low-resolution and high-resolution images, resulting in poor structural consistency between the generated results and the real images. Therefore, super-resolution technology that uses low-resolution images for high-resolution reconstruction has become a research hotspot in the field of cardiac imaging.

[0003] Image super-resolution reconstruction aims to restore high-resolution images from low-resolution images. This task has attracted widespread attention due to its ill-posedness and broad practical application prospects. In recent years, super-resolution reconstruction technology based on deep learning has significantly improved the clarity and detail expression of cardiac magnetic resonance images, and promoted the development of medical diagnosis and treatment towards a more precise and intelligent direction. However, the super-resolution methods based on deep learning in existing technologies are based on ideal degradation models, which are significantly different from the complex degradation processes in the real world and are difficult to cope with the complex degradation of real scenes. The super-resolution method based on generative model effectively deals with the ill-posedness problem by learning the distribution from low-resolution images to high-resolution images, but it still has the problems of high time cost and insufficient high-frequency feature extraction capabilities.

[0004] Therefore, a method and system for blind super-resolution reconstruction of cardiac magnetic resonance images based on a continuous-time conditional diffusion model are provided to solve the above problems. Summary of the Invention

[0005] In order to solve the above problems, the present invention provides a blind super-resolution reconstruction method and system for cardiac magnetic resonance images based on a continuous-time conditional diffusion model, which achieves efficient and high-precision super-resolution reconstruction of medical images by improving the time efficiency and generation quality of the diffusion model.

[0006] To achieve the above object, the present invention provides a method for blind super-resolution reconstruction of cardiac magnetic resonance images based on a continuous-time conditional diffusion model, which specifically includes the following steps:

[0007] S1: Obtain a cardiac magnetic resonance image dataset acquired by a medical scanner, perform data enhancement and degradation processing on the cardiac magnetic resonance source images in the cardiac magnetic resonance image dataset, and obtain a low-resolution image through bicubic downsampling;

[0008] S2: Construct a continuous-time conditional diffusion model based on a cascaded residual attention network feature extractor, a continuous-time conditional diffusion module, a hybrid parameterized score predictor, and an image quality loss module;

[0009] The low-resolution image obtained in S1 is subjected to feature extraction by a cascaded residual attention network feature extractor, and the obtained high-frequency features are used as the conditional input of the continuous-time conditional diffusion model;

[0010] The cardiac magnetic resonance source image is noisy and converted into a Gaussian distribution using the forward stochastic differential equation in the continuous-time conditional diffusion module;

[0011] Super-resolution reconstructed images are generated by inverse denoising using the probability flow ordinary differential equation in the continuous-time conditional diffusion module and dynamically adjusting the parameterization strategy using a hybrid parameterized score predictor.

[0012] Through the image quality loss module, the perceptual consistency between the generated super-resolution reconstructed image and the cardiac magnetic resonance source image is optimized by combining the score matching loss and the perceptual loss;

[0013] S3: Setting the optimizer, training parameters, and hardware configuration to train the continuous-time conditional diffusion model parameters obtained in S2 to obtain a trained continuous-time conditional diffusion model;

[0014] S4: Acquire the cardiac magnetic resonance source image to be measured, input it into the continuous-time conditional diffusion model trained in S3, and generate a super-resolution reconstructed image corresponding to the cardiac magnetic resonance source image to be measured.

[0015] Preferably, the data enhancement in S1 includes rotating, horizontally or vertically flipping, and scaling the cardiac magnetic resonance source image;

[0016] The degradation generation in S1 involves downsampling the cardiac magnetic resonance source image through an isotropic Gaussian blur kernel to generate a low-resolution image.

[0017] Preferably, the cascaded residual attention network feature extractor in S2 includes a shallow feature extraction module, a multi-layer cascaded residual attention block module and a feature fusion and upsampling module;

[0018] The shallow feature extraction module uses a 3×3 convolutional layer to extract shallow features of low-resolution images;

[0019] The cascaded residual attention block contains 10 cascaded residual blocks. The residual block consists of 5 channel attention modules and 5 non-local channel attention modules. The channel attention module generates channel weights through global average pooling to enhance important channel features. The non-local channel attention module captures long-range feature dependencies by calculating cross-channel global correlations.

[0020] The feature fusion and upsampling module adds shallow features to deep high-frequency features and upsamples them to the target resolution through pixel reorganization.

[0021] Preferably, the forward stochastic differential equation in S2 is expressed as:

[0022]

[0023] Where μ(y) is the mean value of the low-resolution image y upsampled by bicubic interpolation, σ 2 (y) Empirical variance of the training set, β(t) linearly increasing noise variance, β(0) = 0.1, β(T) = 20.

[0024] Preferably, the probability flow ordinary differential equation in S2 is expressed as:

[0025]

[0026] Among them, s θ Parameterize the score predictor for the mixture.

[0027] Preferably, the parameterization strategy of the hybrid parameterized score predictor in S2 includes ε parameterization, x0 parameterization and hybrid interpolation;

[0028] ε parameterizes the prediction noise component ε θ , used in high noise areas;

[0029] x0 parameterizes the predicted clean image component x0, which is used for low noise areas;

[0030] Hybrid interpolation dynamically adjusts the interpolation coefficient λ(t)=α(t) c (c∈[0.5,1.5]), balancing the two parameterizations is expressed as:

[0031] s θ (x,y,t)=λ(t)s θ,ε (x,y,t)+(1-λ(t))s θ,ε (x,y,t).

[0032] Preferably, the hybrid parameterized score predictor in S2 adopts the U-Net architecture, and the input is the noise image x t , time step encoding t e and conditional feature F cond, denoising is performed layer by layer through residual blocks and Mish activation function.

[0033] Preferably, the image quality loss module in S2 includes score matching loss, perceptual loss and total loss function;

[0034] The score matching loss minimizes the mean squared error between the predicted score and the target score;

[0035] Perceptual loss uses the pre-trained VGG-19 network to extract features and calculate the feature space distance between the generated image and the real high-resolution image;

[0036] The total loss function combines the score matching loss and the perceptual loss to optimize the perceptual consistency between the generated images and the real images.

[0037] Preferably, the optimizer in S3 uses an adaptive moment estimation algorithm, and the settings of training parameters include batch size, training rounds, diffusion time step and hardware configuration.

[0038] A blind super-resolution reconstruction system for cardiac magnetic resonance images based on a continuous-time conditional diffusion model, comprising a data preparation module, a continuous-time conditional diffusion model construction module, a training module, and a super-resolution reconstruction module;

[0039] The data preparation module obtains a cardiac magnetic resonance image dataset collected by a medical scanner, performs data enhancement and degradation processing on the cardiac magnetic resonance source images in the cardiac magnetic resonance image dataset, and obtains a low-resolution image through bicubic downsampling;

[0040] The continuous-time conditional diffusion model building module includes a cascaded residual attention network feature extractor, a continuous-time conditional diffusion module, a hybrid parameterized score predictor, and an image quality loss module;

[0041] The low-resolution image obtained by the data preparation module is subjected to feature extraction through the cascaded residual attention network feature extractor. The obtained high-frequency features are used as the conditional input of the continuous-time conditional diffusion model.

[0042] The cardiac magnetic resonance source image is noisy and converted into a Gaussian distribution using the forward stochastic differential equation in the continuous-time conditional diffusion module;

[0043] Super-resolution reconstructed images are generated by inverse denoising using the probability flow ordinary differential equation in the continuous-time conditional diffusion module and dynamically adjusting the parameterization strategy using a hybrid parameterized score predictor.

[0044] Through the image quality loss module, the perceptual consistency between the generated super-resolution reconstructed image and the cardiac magnetic resonance source image is optimized by combining the score matching loss and the perceptual loss;

[0045] The training module sets the optimizer, training parameters and hardware configuration to train the parameters of the continuous-time conditional diffusion model to obtain the trained continuous-time conditional diffusion model;

[0046] The super-resolution reconstruction module acquires the cardiac magnetic resonance source image to be measured, inputs the trained continuous-time conditional diffusion model, and generates a super-resolution reconstructed image corresponding to the cardiac magnetic resonance source image to be measured. Therefore, the present invention adopts the above-mentioned blind super-resolution reconstruction method for cardiac magnetic resonance images based on the continuous-time conditional diffusion model. The continuous-time conditional diffusion model, constructed by cascading a residual attention network feature extractor, a continuous-time conditional diffusion module, a hybrid parameterized score predictor, and an image quality loss module, solves the problems of long time consumption and insufficient quality and consistency of the generated images when using iterative sampling. This reduces the time consumption of the diffusion model in super-resolution reconstruction, improves the detail representation and structural consistency of the super-resolution reconstruction of cardiac magnetic resonance images, and enhances the model's robustness to complex degradation.

[0047] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 Schematic diagram of a flow chart of a method for blind super-resolution reconstruction of cardiac magnetic resonance images based on a continuous-time conditional diffusion model in the present invention;

[0049] Figure 2 Schematic diagram of the structure of the continuous-time conditional diffusion model in the present invention;

[0050] Figure 3 Schematic diagram of the structure of the cascaded residual attention network feature extractor in the present invention;

[0051] Figure 4 Schematic diagram of the residual block in the present invention;

[0052] Figure 5 This is a diagram showing the visual effects of different super-resolution reconstruction methods in super-resolution reconstruction of cardiac horizontal plane magnetic resonance images according to an embodiment of the present invention;

[0053] Figure 6 Graphs showing the visual effects of different super-resolution reconstruction methods in super-resolution reconstruction of cardiac coronary magnetic resonance images according to embodiments of the present invention;

[0054] Figure 7 1. It is a diagram showing the visual effects of different super-resolution reconstruction methods in super-resolution reconstruction of cardiac sagittal magnetic resonance images according to an embodiment of the present invention;

[0055] Figure 8Graphs showing the visual effects of different residual models in super-resolution reconstruction of cardiac magnetic resonance images according to an embodiment of the present invention;

[0056] Figure 9 Schematic diagram of ablation study results of different cascaded channel attention blocks and cascaded residual attention blocks in an embodiment of the present invention;

[0057] Figure 10 Graphs showing the visual effects of different parameterized models in an embodiment of the present invention. DETAILED DESCRIPTION

[0058] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0059] Unless otherwise defined, technical or scientific terms used in the present invention shall have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs.

[0060] The words “include” or “comprising” and similar words used in the present invention mean that the elements before the word include the elements listed after the word, and do not exclude the possibility of also including other elements. The orientation or position relationship indicated by the terms “inside”, “outside”, “upper”, “lower”, etc. is based on the orientation or position relationship shown in the accompanying drawings. It is only for the convenience of describing the present invention and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it cannot be understood as a limitation of the present invention. When the absolute position of the described object changes, the relative position relationship may also change accordingly. In the present invention, unless otherwise clearly stipulated and limited, the terms such as “attachment” should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral whole; it can be directly connected or indirectly connected through an intermediate medium, and it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances.

[0061] Example

[0062] The present invention provides a blind super-resolution reconstruction method for cardiac magnetic resonance images based on a continuous-time conditional diffusion model. Figure 1 As shown, the specific steps include:

[0063] S1: Obtain a cardiac magnetic resonance image dataset acquired by a medical scanner, perform data enhancement and degradation processing on the cardiac magnetic resonance source images in the cardiac magnetic resonance image dataset, and obtain a low-resolution image through bicubic downsampling;

[0064] The dataset uses the AMRGCardiac MRIAtlas cardiac MRI dataset acquired by the Auckland Magnetic Resonance Research Group's Siemens Avanto scanner and the Cardiac MRI dataset provided by Dr. Paul Babyn, Director of Radiology, and Dr. Shi Joon Yoo, Director of Cardiac Radiology at the Hospital for Sick Children in Toronto.

[0065] The data augmentation in this application involves rotating the original image by ±90°, flipping it horizontally / vertically, and scaling it by 0.8-1.2 times to expand the training samples in the dataset.

[0066] The degradation processing downsamples the high-resolution image through an isotropic Gaussian blur kernel to generate a low-resolution image that simulates the real scene, where the kernel size is 21×21 and the width range is [0.2, 4.0].

[0067] S2: If Figure 2 As shown in the figure, a continuous-time conditional diffusion model CCDM is constructed based on a cascaded residual attention network feature extractor, a continuous-time conditional diffusion module, a hybrid parameterized score predictor, and an image quality loss module;

[0068] The low-resolution image obtained in S1 is subjected to feature extraction by a cascaded residual attention network feature extractor, and the obtained high-frequency features are used as the conditional input of the continuous-time conditional diffusion model;

[0069] like Figure 3-4 As shown in Figure 2, the cascaded residual attention network feature extractor includes a shallow feature extraction module, a multi-layer cascaded residual attention block module, and a feature fusion and upsampling module; the shallow feature extraction module uses a 3×3 convolutional layer to extract the shallow features F of the low-resolution image. shallow The cascaded residual attention block contains 10 cascaded residual blocks. The residual block consists of 5 channel attention CA modules and 5 non-local channel attention NCA modules. The channel attention CA module generates channel weights through global average pooling to enhance important channel features. The non-local channel attention NCA module captures long-distance feature dependencies by calculating cross-channel global correlations. The feature fusion and upsampling module adds shallow features to deep high-frequency features and upsamples them to the target resolution through PixelShuffle. After the low-resolution image is input, the shallow features gradually pass through the residual attention block, weightedly fuse global and local information, and output high-frequency features F. cond Serves as a conditional input to the diffusion model.

[0070] The forward stochastic differential equation in the continuous-time conditional diffusion module is used to noise the cardiac magnetic resonance source image and transform it into a Gaussian distribution. The mean and variance of the data are maintained during the forward process, reducing the complexity of model training. The forward process is expressed as:

[0071]

[0072] Where μ(y) is the mean value of the low-resolution image y upsampled by bicubic interpolation, σ 2 (y) Empirical variance of the training set, β(t) linearly increasing noise variance, β(0) = 0.1, β(T) = 20.

[0073] Through inverse denoising of the probability flow ordinary differential equation in the continuous-time conditional diffusion module, combined with the hybrid parameterized fractional predictor to dynamically adjust the parameterization strategy, super-resolution reconstructed images are generated; high-resolution images are efficiently generated through the ordinary differential equation solver, reducing the traditional Markov chain iteration time.

[0074] The reverse sampling formula is expressed as:

[0075] Among them, s θ Parameterize the score predictor for the mixture.

[0076] Hybrid parameterized fractional predictor parameterization strategies include ε parameterization, x0 parameterization, and hybrid interpolation; ε parameterization predicts the noise component ε θ , used for high noise areas; x0 parameterized prediction of clean image component x0, used for low noise areas; hybrid interpolation dynamically adjusts the interpolation coefficient λ(t) = α(t) c (c∈[0.5,1.5]), balancing the two parameterizations is expressed as:

[0077] s θ (x,y,t)=λ(t)s θ,ε (x,y,t)+(1-λ(t))s θ,ε (x,y,t).

[0078] The hybrid parameterized score predictor uses the U-Net architecture and takes the noise image x as input. t , time step encoding t e and conditional feature F cond , denoising is performed layer by layer through residual blocks and Mish activation function.

[0079] Through the image quality loss module, the perceptual consistency between the generated super-resolution reconstructed image and the cardiac magnetic resonance source image is optimized by combining the score matching loss and the perceptual loss;

[0080] On the basis of the score matching loss of the continuous-time conditional diffusion model, the feature space distance loss based on the VGG network is added to directly optimize the perceptual consistency between the generated image and the real image.

[0081] The image quality loss module includes score matching loss, perceptual loss and total loss function; score matching loss Minimize the mean square error between the predicted score and the target score, expressed as:

[0082] Perceptual loss Use the pre-trained VGG-19 network to extract features and calculate the generated image SR(y) and the real high-resolution image x GT The feature space distance is expressed as:

[0083]

[0084] The total loss function combines the score matching loss and the perceptual loss to optimize the perceptual consistency between the generated image and the real image, which is expressed as:

[0085]

[0086] S3: Setting the optimizer, training parameters, and hardware configuration to train the continuous-time conditional diffusion model parameters obtained in S2 to obtain a trained continuous-time conditional diffusion model;

[0087] (1) Optimizer: Adam algorithm, set β1 = 0.9, β2 = 0.999, and the initial value of the learning rate is 1e-4.

[0088] (2) Training parameters:

[0089] 1) The batch size is 16 and the number of training rounds is 200.

[0090] 2) The diffusion time step is T=100, and the noise variance increases linearly according to VP-SDE.

[0091] (3) Hardware configuration: 2×NVIDIA GeForce RTX 3090Ti GPU.

[0092] S4: Obtain the cardiac magnetic resonance source image to be measured, input it into the continuous-time conditional diffusion model trained in S3, and generate a super-resolution reconstructed image corresponding to the cardiac magnetic resonance source image to be measured. The specific steps are as follows:

[0093] (1) Input low-resolution image: Preliminary upsampling to the target size through bicubic interpolation.

[0094] (2) Input continuous-time conditional diffusion model.

[0095] 1) Feature extraction: Input cascade residual attention network, output conditional feature F cond .

[0096] 2) Probability stream sampling:

[0097] Initial noise The inverse equation is solved using a Runge-Kutta ordinary differential equation solver with a tolerance of 1e-4 to generate high-resolution images.

[0098] (3) Output results: Optimize visual effects through post-processing, such as histogram equalization post-processing method.

[0099] Example 1

[0100] A blind super-resolution reconstruction system for cardiac magnetic resonance images based on a continuous-time conditional diffusion model, comprising a data preparation module, a continuous-time conditional diffusion model construction module, a training module, and a super-resolution reconstruction module;

[0101] The data preparation module obtains a cardiac magnetic resonance image dataset collected by a medical scanner, performs data enhancement and degradation processing on the cardiac magnetic resonance source images in the cardiac magnetic resonance image dataset, and obtains a low-resolution image through bicubic downsampling;

[0102] The continuous-time conditional diffusion model building module includes a cascaded residual attention network feature extractor, a continuous-time conditional diffusion module, a hybrid parameterized score predictor, and an image quality loss module;

[0103] The low-resolution image obtained by the data preparation module is subjected to feature extraction through the cascaded residual attention network feature extractor. The obtained high-frequency features are used as the conditional input of the continuous-time conditional diffusion model.

[0104] The cardiac magnetic resonance source image is noisy and converted into a Gaussian distribution using the forward stochastic differential equation in the continuous-time conditional diffusion module;

[0105] Super-resolution reconstructed images are generated by inverse denoising using the probability flow ordinary differential equation in the continuous-time conditional diffusion module and dynamically adjusting the parameterization strategy using a hybrid parameterized score predictor.

[0106] Through the image quality loss module, the perceptual consistency between the generated super-resolution reconstructed image and the cardiac magnetic resonance source image is optimized by combining the score matching loss and the perceptual loss;

[0107] The training module sets the optimizer, training parameters and hardware configuration to train the parameters of the continuous-time conditional diffusion model to obtain the trained continuous-time conditional diffusion model;

[0108] The super-resolution reconstruction module obtains the cardiac magnetic resonance source image to be measured, inputs the trained continuous-time conditional diffusion model, and generates a super-resolution reconstructed image corresponding to the cardiac magnetic resonance source image to be measured.

[0109] Example 2

[0110] The effectiveness of the blind super-resolution reconstruction method for cardiac magnetic resonance images based on the continuous-time conditional diffusion model provided in this application was verified as follows:

[0111] 1. Improved time efficiency:

[0112] This application adopts probability flow ordinary differential equation sampling, which reduces the reconstruction time by more than 15% compared with traditional diffusion models such as DBSR. The experimental data show that CCDM is 1.62 seconds and DBSR is 1.90 seconds.

[0113] Model runtime is crucial for super-resolution reconstruction of medical images. To this end, this example compares CCDM with IKC, SRDiff, and DBSR in terms of parameter count and runtime. As shown in Table 1 below, CCDM achieves the optimal PSNR while offering superior reconstruction time compared to DBSR. This demonstrates that CCDM significantly reduces computational complexity while maintaining high-quality cardiac MRI image reconstruction performance, thereby achieving an optimal balance between super-resolution reconstruction performance and model computational complexity.

[0114] Table 1

[0115]

[0116] 2. Image quality optimization:

[0117] (1) PSNR value is improved. In the horizontal plane dataset, when the blur kernel is 1.2, the PSNR value is improved to 32.57dB, which is 0.03dB higher than the existing optimal method DBSR 32.54dB.

[0118] To further evaluate the super-resolution reconstruction performance of CCDM, we compared it with the mainstream methods SRCNN, VDSR, DRRN, LapSRN, A 2 N, RCAN, SRMDNF, IKC, DASR, SRDiff, StableSR and DBSR are compared, and the results are shown in Table 2 below.

[0119] As can be seen from Table 2 below, when the Gaussian isotropic blur kernel is 1.2, for the horizontal plane dataset, when the magnification factor is 4, CCDM achieves a PSNR value of 32.57dB. Compared with other methods, CCDM has a significantly improved performance: it is significantly better than SRCNN, VDSR, DRRN, LapSRN, and A 2N,RCAN, SRMDNF, IKC, DASR, SRDiff, StableSR and DBSR are 4.12dB, 3.69dB, 3.46dB, 3.32dB, 3.20dB, 1.86dB, 1.83dB, 0.38dB, 0.19dB, 0.08dB, 0.31dB and 0.03dB higher, respectively. This shows that CCDM exhibits stronger robustness and generative ability in super-resolution reconstruction tasks under complex blur kernel degradation conditions.

[0120] It can be further observed that SRCNN, VDSR, DRRN, LapSRN, A 2 N and RCAN have low PSNR values ​​on cardiac magnetic resonance datasets in the horizontal, coronal, and sagittal planes. This is because these methods assume that low-resolution images are degraded based on bicubic downsampling. In contrast, SRMDNF, IKC, DASR, and DBSR are designed for complex degradation problems in the real world and can better handle low-resolution image degradation caused by blur kernels, achieving higher PSNR values ​​on the horizontal, coronal, and sagittal planes.

[0121] Table 2

[0122]

[0123]

[0124] (2) The perception indicators are significantly improved. The method provided in this application has an LPIPS of 0.0461 and an FID of 63.85.

[0125] To comprehensively evaluate CCDM's reconstruction performance on low-resolution cardiac MRI images, we compared it with other blind super-resolution reconstruction models and a diffusion model using the PSNR, SSIM, LPIPS, and FID image evaluation metrics. As shown in Table 3 below, CCDM achieved the lowest LPIPS and FID values, while also achieving the highest PSNR and SSIM values, fully demonstrating its superior performance in low-resolution cardiac MRI reconstruction. This demonstrates that CCDM not only produces images of high perceptual quality but also surpasses existing methods in structural similarity and pixel accuracy, demonstrating its strong practical application value.

[0126] Table 3

[0127]

[0128]

[0129] 3. Enhanced robustness:

[0130] This application can handle complex degradation under Gaussian isotropic blur kernel with a width range of [0.2, 4.0] and adapt to real medical imaging scenes.

[0131] Compared to these state-of-the-art methods, CCDM achieves the highest PSNR values ​​on cardiac magnetic resonance transverse, coronal, and sagittal planes, demonstrating significant performance improvements. Thanks to CCDM's conditional diffusion model network framework, its powerful generative capabilities not only enable more accurate prediction of blur kernels but also generate higher-quality super-resolution images, fully demonstrating its superiority under complex degradation conditions.

[0132] Figure 5-7 The visual effects comparison of different super-resolution reconstruction methods in super-resolution reconstruction of cardiac magnetic resonance images is shown. It can be seen that the cardiac magnetic resonance super-resolution images reconstructed by SRCNN and RCAN are relatively blurry, and it is difficult to clearly present important tissue details in the image. Although SRMDNF, IKC and DASR show good performance in the reconstruction process, the boundary details of their reconstructed images are still not clear enough. In contrast, the images reconstructed by StableSR, SRDiff, DBSR and CCDM are clearer. However, CCDM shows a smoother visual effect in the reconstruction, with less edge blur, and retains more image details, such as pathological details such as myocardial fibrosis and edema, which fully demonstrates its superiority in super-resolution reconstruction of cardiac magnetic resonance images.

[0133] Example 3

[0134] To further validate the effectiveness of CCDM, this example conducted ablation studies on the cascaded residual attention network feature extractor, the hybrid parameter score predictor, and the image quality loss module. These experiments were conducted on a dataset of cardiac magnetic resonance images in horizontal planes with a magnification factor of 4, evaluating the contribution of each module to super-resolution reconstruction performance. By analyzing model performance under different configurations, we can more clearly understand the role of each module in improving image quality, enhancing detail recovery, and accelerating the generation process, thereby verifying the impact of these designs on the final model performance.

[0135] CCDM-Res refers to the cascaded residual attention block without channel attention mechanism and non-local channel attention mechanism, CCDM CA refers to the cascaded residual attention block with channel attention mechanism, CCDM-NC refers to the cascaded residual attention block with non-local channel attention mechanism, CCDM refers to the cascaded residual attention block with channel attention mechanism and non-local channel attention mechanism, such as Figure 8The figure shows the visual effects of CCDM-Res, CCDM-CA, CCDM-NC and CCDM in super-resolution reconstruction of cardiac magnetic resonance images. It can be seen that CCDM combines the cascaded residual attention block of the channel attention mechanism and the non-local channel attention mechanism, which can reconstruct clearer and more detailed cardiac magnetic resonance super-resolution images than CCDM-Res, CCDM-CA and CCDM-NC.

[0136] Since the number of cascaded channel attention blocks and cascaded residual attention blocks has a significant impact on the reconstruction performance of CCDM, an ablation study on the number of these blocks in the cascaded residual attention network feature extractor is conducted. The results are shown in Figure 2. Figure 9 As shown, from Figure 9 It can be seen that as the number of cascaded channel attention blocks and cascaded residual attention blocks increases, the PSNR value of CCDM also increases accordingly. However, when the number of these blocks reaches a certain threshold, the PSNR value of CCDM no longer increases and even decreases slightly. Therefore, this application ultimately sets the number of cascaded channel attention blocks and cascaded residual attention blocks to 5 and 10 respectively to obtain the best super-resolution reconstruction performance.

[0137] In order to study the effects of different parameterization choices in the denoising network and image quality loss on cardiac magnetic resonance image reconstruction, this embodiment conducts an ablation study on the parameter selection. Figure 10 As shown in the figure, it can be seen that the model trained with Lquality can achieve better reconstruction performance than the model without Lquality, which fully illustrates the importance of image quality loss to super-resolution reconstruction based on the diffusion model.

[0138] Therefore, the present invention adopts the above-mentioned method and system for blind super-resolution reconstruction of cardiac magnetic resonance images based on the continuous-time conditional diffusion model, and extracts high-frequency features F from the low-resolution image through the cascaded residual attention network feature extractor. cond , as a conditional input to the diffusion model. The hybrid parameterized score predictor is combined with F cond , time-step encoding, and noisy images, dynamically adjusting the parameterization strategy. The probability flow ordinary differential equation utilizes a conditional score function to rapidly generate super-resolution reconstructed images. Image quality loss further constrains the consistency of the generated results with the source image. By improving the time efficiency and generation quality of the diffusion model, efficient and high-precision super-resolution reconstruction of medical images is achieved.

[0139] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solutions of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A blind super-resolution reconstruction method for cardiac magnetic resonance images based on a continuous-time conditional diffusion model, characterized in that: The specific steps include: S1: Obtain a cardiac magnetic resonance image dataset acquired by a medical scanner, perform data enhancement and degradation processing on the cardiac magnetic resonance source images in the cardiac magnetic resonance image dataset, and obtain a low-resolution image through bicubic downsampling; S2: Construct a continuous-time conditional diffusion model based on a cascaded residual attention network feature extractor, a continuous-time conditional diffusion module, a hybrid parameterized score predictor, and an image quality loss module; The low-resolution image obtained in S1 is subjected to feature extraction by a cascaded residual attention network feature extractor, and the obtained high-frequency features are used as the conditional input of the continuous-time conditional diffusion model; The cardiac magnetic resonance source image is noisy and converted into a Gaussian distribution using the forward stochastic differential equation in the continuous-time conditional diffusion module; Super-resolution reconstructed images are generated by inverse denoising using the probability flow ordinary differential equation in the continuous-time conditional diffusion module and dynamically adjusting the parameterization strategy using a hybrid parameterized score predictor. Through the image quality loss module, the perceptual consistency between the generated super-resolution reconstructed image and the cardiac magnetic resonance source image is optimized by combining the score matching loss and the perceptual loss; S3: Setting the optimizer, training parameters, and hardware configuration to train the continuous-time conditional diffusion model parameters obtained in S2 to obtain a trained continuous-time conditional diffusion model; S4: Acquire the cardiac magnetic resonance source image to be measured, input it into the continuous-time conditional diffusion model trained in S3, and generate a super-resolution reconstructed image corresponding to the cardiac magnetic resonance source image to be measured.

2. The method for blind super-resolution reconstruction of cardiac magnetic resonance images based on a continuous-time conditional diffusion model according to claim 1, characterized in that: The data augmentation in S1 includes rotating, horizontally or vertically flipping, and scaling operations on cardiac magnetic resonance source images; The degradation generation in S1 involves downsampling the cardiac magnetic resonance source image through an isotropic Gaussian blur kernel to generate a low-resolution image.

3. The method for blind super-resolution reconstruction of cardiac magnetic resonance images based on a continuous-time conditional diffusion model according to claim 1, characterized in that: The cascaded residual attention network feature extractor in S2 includes a shallow feature extraction module, a multi-layer cascaded residual attention block module, and a feature fusion and upsampling module; The shallow feature extraction module uses a 3×3 convolutional layer to extract shallow features of low-resolution images; The cascaded residual attention block contains 10 cascaded residual blocks. The residual block consists of 5 channel attention modules and 5 non-local channel attention modules. The channel attention module generates channel weights through global average pooling to enhance important channel features. The non-local channel attention module captures long-range feature dependencies by calculating cross-channel global correlations. The feature fusion and upsampling module adds shallow features to deep high-frequency features and upsamples them to the target resolution through pixel reorganization.

4. The method for blind super-resolution reconstruction of cardiac magnetic resonance images based on a continuous-time conditional diffusion model according to claim 1, characterized in that: The forward stochastic differential equation in S2 is expressed as: Where μ(y) is the mean of the low-resolution image y upsampled by bicubic interpolation, σ 2 (y) Empirical variance of the training set, β(t) linearly increasing noise variance, β(0) = 0.1, β(T) = 20.

5. The method for blind super-resolution reconstruction of cardiac magnetic resonance images based on a continuous-time conditional diffusion model according to claim 4, characterized in that: The probability flow ordinary differential equation in S2 is expressed as: Among them, s θ Parameterize the score predictor for the mixture.

6. The method for blind super-resolution reconstruction of cardiac magnetic resonance images based on a continuous-time conditional diffusion model according to claim 5, characterized in that: The parameterization strategies of the hybrid parameterized score predictor in S2 include ε parameterization, x0 parameterization, and hybrid interpolation; ε parameterizes the prediction noise component ε θ , used in high noise areas; x0 parameterizes the predicted clean image component x0, which is used for low noise areas; Hybrid interpolation dynamically adjusts the interpolation coefficient λ(t)=α(t) c (c∈[0.5,1.5]), balancing the two parameterizations is expressed as: s θ (x,y,t)=λ(t)s θ,ε (x,y,t)+(1-λ(t))s θ,ε (x,y,t)。 7. The method for blind super-resolution reconstruction of cardiac magnetic resonance images based on a continuous-time conditional diffusion model according to claim 6, characterized in that: The mixed parameterized score predictor in S2 adopts the U-Net architecture, and the input is the noise image x t , time step encoding t e and conditional feature F cond , denoising is performed layer by layer through residual blocks and Mish activation function.

8. The method for blind super-resolution reconstruction of cardiac magnetic resonance images based on a continuous-time conditional diffusion model according to claim 1, characterized in that: The image quality loss module in S2 includes score matching loss, perceptual loss and total loss function; The score matching loss minimizes the mean squared error between the predicted score and the target score; Perceptual loss uses the pre-trained VGG-19 network to extract features and calculate the feature space distance between the generated image and the real high-resolution image; The total loss function combines the score matching loss and the perceptual loss to optimize the perceptual consistency between the generated images and the real images.

9. The method for blind super-resolution reconstruction of cardiac magnetic resonance images based on a continuous-time conditional diffusion model according to claim 1, characterized in that: The optimizer in S3 uses an adaptive moment estimation algorithm. The settings of training parameters include batch size, training rounds, diffusion time step, and hardware configuration.

10. A system for blind super-resolution reconstruction of cardiac magnetic resonance images based on a continuous-time conditional diffusion model according to any one of claims 1 to 9, characterized in that: The blind super-resolution reconstruction system of cardiac magnetic resonance images based on the continuous-time conditional diffusion model includes a data preparation module, a continuous-time conditional diffusion model construction module, a training module and a super-resolution reconstruction module; The data preparation module obtains a cardiac magnetic resonance image dataset collected by a medical scanner, performs data enhancement and degradation processing on the cardiac magnetic resonance source images in the cardiac magnetic resonance image dataset, and obtains a low-resolution image through bicubic downsampling; The continuous-time conditional diffusion model building module includes a cascaded residual attention network feature extractor, a continuous-time conditional diffusion module, a hybrid parameterized score predictor, and an image quality loss module; The low-resolution image obtained by the data preparation module is subjected to feature extraction through the cascaded residual attention network feature extractor. The obtained high-frequency features are used as the conditional input of the continuous-time conditional diffusion model. The cardiac magnetic resonance source image is noisy and converted into a Gaussian distribution using the forward stochastic differential equation in the continuous-time conditional diffusion module; Super-resolution reconstructed images are generated by inverse denoising using the probability flow ordinary differential equation in the continuous-time conditional diffusion module and dynamically adjusting the parameterization strategy using a hybrid parameterized score predictor. Through the image quality loss module, the perceptual consistency between the generated super-resolution reconstructed image and the cardiac magnetic resonance source image is optimized by combining the score matching loss and the perceptual loss; The training module sets the optimizer, training parameters and hardware configuration to train the parameters of the continuous-time conditional diffusion model to obtain the trained continuous-time conditional diffusion model; The super-resolution reconstruction module obtains the cardiac magnetic resonance source image to be measured, inputs the trained continuous-time conditional diffusion model, and generates a super-resolution reconstructed image corresponding to the cardiac magnetic resonance source image to be measured.

Images