Blind isotropic MRI (Magnetic Resonance Imaging) super-resolution reconstruction method and system based on double-input degradation modeling

Abstract

The invention provides a blind isotropic MRI (Magnetic Resonance Imaging) super-resolution reconstruction method and system based on double-input degradation modeling, and belongs to the field of medical image super-resolution reconstruction. The problems of inaccurate degradation modeling, dependence on paired data and difficulty in adapting to multi-scale change in super-resolution reconstruction of existing medical images are solved. The method comprises the steps that non-paired isotropic MRI and anisotropic MRI are acquired, a two-dimensional image in the vertical axis direction is extracted from the isotropic MRI to serve as a high-resolution image, and a two-dimensional image in the vertical axis direction is extracted from the anisotropic MRI to serve as a low-resolution image; constructing a dual-input degradation network, wherein the dual-input degradation network comprises a generator and a discriminator; constructing a super-resolution reconstruction network; introducing a closed-loop feedback mechanism, and alternately freezing parameters of a degraded network and a reconstructed network; a scale sensing module based on conditional convolution is introduced into the degradation network. The method is mainly used in the magnetic resonance imaging field.

Description

Technical Field

[0001] This invention belongs to the field of medical image super-resolution reconstruction technology, and in particular relates to a blind isotropic MRI super-resolution reconstruction method based on dual-input degradation modeling. Background Technology

[0002] Magnetic resonance imaging (MRI) is a non-invasive medical imaging technique that uses the principle of nuclear magnetic resonance to acquire images of the internal structures of the human body. MRI has advantages such as being radiation-free, having high soft tissue contrast, and being able to perform multi-parameter imaging, making it an important tool for diagnosing brain diseases, preoperative planning, and evaluating treatment efficacy. However, the resolution of MRI is limited by factors such as scan time, signal-to-noise ratio, and equipment performance. In clinical practice, to shorten scan time, thick-slice scanning is often used, resulting in image resolution in the vertical axis being much lower than in the horizontal and longitudinal axes, forming anisotropic images. This anisotropy can seriously affect the accuracy of subsequent tasks such as 3D reconstruction, registration, segmentation, and quantitative analysis. Therefore, improving the resolution of anisotropic MRI in the vertical axis has become an urgent challenge in the field of medical imaging.

[0003] Image super-resolution, a classic problem in computer vision, aims to recover high-resolution images from low-resolution images. Traditional methods include interpolation, sparse representation, and edge priors, but their effectiveness is limited. In recent years, deep learning methods have made breakthrough progress in super-resolution tasks; however, most of these methods assume that the low-resolution image is obtained by bicubic downsampling of the high-resolution image, which does not conform to the actual degradation process in MRI. MRI degradation is affected by a variety of physical factors, such as slice thickness, point spread function, sampling trajectory, and coil sensitivity, making it difficult to model with simple mathematical models.

[0004] Super-resolution reconstruction of medical images faces greater challenges compared to super-resolution reconstruction of natural images. In clinical practice, it is difficult to obtain strictly paired high-resolution-low-resolution image sets. Interpolation methods commonly used to generate high-resolution-low-resolution image pairs cannot accurately reflect the complex degradation processes inherent in medical imaging data. Furthermore, the formation of low-resolution MRI images is influenced by multiple factors, which cannot be fully represented by structural information alone. To address these challenges, researchers have explored blind super-resolution methods suitable for medical imaging. These methods utilize neural networks to infer degradation information from unpaired data, achieving efficient reconstruction from low-resolution to high-resolution images without the need for paired training data.

[0005] Current research on MRI vertical axis degradation and reconstruction mainly focuses on single-input methods. These methods use anisotropic MRI as input, treating the horizontal and vertical axes as high-resolution directions and the vertical axis as a low-resolution direction. The aim is to construct a degradation model applicable to the vertical axis by studying the degradation characteristics of the horizontal and vertical axes. However, from the perspective of imaging physics, the image generation processes in the horizontal and vertical axes of MRI differ fundamentally from those in the vertical axis. Typically, signal reception is more concentrated and the signal-to-noise ratio is higher in the horizontal and vertical axes; while in the vertical axis, due to technical limitations, the voxel size is larger, and the volumetric effect is more significant. Therefore, degradation models constructed based on this often introduce systematic errors when applied to vertical axis reconstruction, ultimately affecting the results. Summary of the Invention

[0006] In view of this, the present invention aims to propose a blind isotropic MRI super-resolution reconstruction method based on dual-input degradation modeling, in order to solve the problems of inaccurate degradation modeling, reliance on paired data, and difficulty in adapting to multi-scale changes in existing medical image super-resolution reconstruction.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A blind isotropic MRI super-resolution reconstruction method based on dual-input degradation modeling, the method comprising: Step S1: Acquire unpaired isotropic and anisotropic MRI images. Extract the two-dimensional image along the vertical axis from the isotropic MRI as a high-resolution image, and extract the two-dimensional image along the vertical axis from the anisotropic MRI as a low-resolution image. Step S2: Construct a dual-input degradation network, which includes a generator and a discriminator; the generator takes a high-resolution image and a low-resolution image as input, simulates the image degradation process in the vertical axis direction, and generates a simulated low-resolution image; the discriminator is used to determine whether the input image is a real low-resolution image; Step S3: Construct a super-resolution reconstruction network, which takes a real low-resolution image or a simulated low-resolution image as input and reconstructs a high-resolution image through a deep neural network; Step S4: Introduce a closed-loop feedback mechanism, and optimize in stages by alternately freezing the parameters of the degenerate network and the reconstruction network, including the first stage of fixing the parameters of the reconstruction network to train the degenerate network and the second stage of fixing the parameters of the degenerate network to train the reconstruction network. Step S5: Introduce a scale-aware module based on conditional convolution into the degenerate network. The scale-aware module dynamically generates convolution kernel parameters according to the input scaling factor to adapt to the degenerate modeling requirements under different scaling factors.

[0008] Furthermore, a preferred embodiment is proposed, wherein constructing the dual-input degenerate network in step S2 includes: The generator is used to generate a simulated low-resolution image based on the high-resolution image and the low-resolution image; The discriminator includes a high-resolution image discriminator and a low-resolution image discriminator; The high-resolution image discriminator is used to determine whether the input image is a real high-resolution image; The low-resolution image discriminator is used to determine whether the input image is a real low-resolution image; The high-resolution image discriminator and the low-resolution image discriminator constitute a dual discriminator structure, which is used to enhance the realism and physical consistency of degradation modeling.

[0009] Furthermore, a preferred method is proposed, wherein the training process of the closed-loop feedback mechanism in step S4 includes: In the first training phase, the parameters of the reconstructed network are fixed, and the degenerate network is trained using a hybrid loss function, which includes: Point spread function constraint loss is used to constrain the shape of the point spread function during the degradation process; Adversarial loss is used to improve the realism of generated images; Reconstruction loss is used to ensure pixel consistency between the degraded image and the real image; Weight decay loss is used to control model complexity; In the second training phase, the parameters of the degraded network are fixed, and the reconstruction network is trained to recover high-resolution images from low-resolution images. Repeat the first and second training phases to form a closed-loop feedback loop, gradually optimizing the accuracy of degradation modeling and image reconstruction.

[0010] Furthermore, a preferred embodiment is proposed, wherein the point spread function constraint loss includes boundary loss and peak loss; The boundary loss is used to penalize non-zero values ​​of the point spread function at the boundary, ensuring that the point spread function at the image boundary tends to zero. The calculation formula is as follows:

[0011] The peak loss is used to constrain the peak position and symmetry of the point spread function, ensuring that the peak position of the point spread function coincides with the axis of symmetry. The calculation formula is as follows:

[0012] in, express The first in One element, express The first in One element, express The first in One element, express The first in One element, Represents the point spread function generated by the degenerate network. PSF length, express The first in One element, express The first in -1 element.

[0013] Furthermore, a preferred approach is proposed in which the degradation network and the reconstruction network are trained using a self-supervised learning method with unpaired isotropic and anisotropic MRI images, without the need for strictly paired high-resolution-low-resolution image pairs.

[0014] Furthermore, a preferred embodiment is proposed, wherein the scale-aware module in step S5 includes: Two expert convolutional kernels are used to represent different types of degradation patterns; A weight prediction network that takes a scaling factor as input and outputs a set of normalized combined weights; A dynamic convolution kernel generation unit is used to perform weighted fusion of expert convolution kernels according to the combined weights to generate a dynamic convolution kernel suitable for the current scale factor.

[0015] Furthermore, a preferred method is proposed, wherein the expert convolution kernel size is a standard 3×3; the weight prediction network consists of two fully connected layers, with a hidden layer dimension of 64, and a scale factor of [missing information]. As input, output a set of normalized combined weights. The final scale-adaptive convolution kernel The results were obtained through weighted combination calculations by expert panelists: , in, and This indicates two expert cores, initialized using the Kaiming initialization method.

[0016] Based on the same inventive concept, this invention also proposes a blind isotropic MRI super-resolution reconstruction system based on dual-input degradation modeling, the system comprising: The image acquisition unit is used to acquire unpaired isotropic MRI and anisotropic MRI, extract two-dimensional images in the vertical direction from the isotropic MRI as high-resolution images, and extract two-dimensional images in the vertical direction from the anisotropic MRI as low-resolution images. A dual-input degradation network construction unit is used to construct a dual-input degradation network, which includes a generator and a discriminator. The generator takes a high-resolution image and a low-resolution image as inputs, simulates the image degradation process in the vertical axis direction, and generates a simulated low-resolution image. The discriminator is used to determine whether the input image is a real low-resolution image. An image reconstruction unit is used to construct a super-resolution reconstruction network, which takes a real low-resolution image or a simulated low-resolution image as input and reconstructs a high-resolution image through a deep neural network. The joint optimization unit is used to introduce a closed-loop feedback mechanism. It optimizes the network in stages by alternately freezing the parameters of the degenerate network and the reconstruction network. The first stage fixes the parameters of the reconstruction network to train the degenerate network, and the second stage fixes the parameters of the degenerate network to train the reconstruction network. A scale-aware unit is used to introduce a conditional convolution-based scale-aware module into the degradation network. The scale-aware module dynamically generates convolution kernel parameters according to the input scaling factor to adapt to the degradation modeling requirements under different scaling factors.

[0017] Based on the same inventive concept, the present invention also proposes a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes a blind isotropic MRI super-resolution reconstruction method based on dual-input degradation modeling according to any one of the above.

[0018] Based on the same inventive concept, the present invention also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of a blind isotropic MRI super-resolution reconstruction method based on dual-input degradation modeling as described in any of the preceding claims.

[0019] Compared with existing technologies, this invention fundamentally abandons the indirect and physically incompatible approach of inferring the vertical axis from the horizontal / vertical axes, instead employing a dual-input direct modeling paradigm based on real vertical axis information. This includes: Unlike existing single-input methods, this invention, for the first time, uses both real-world vertical images from isotropic and anisotropic MRI as inputs to the degradation network. This design allows the degradation model's learning process to be directly based on the true image characteristics and physical degradation relationships along the vertical axis, rather than through indirect and inaccurate inferences from the horizontal / vertical axes with different imaging characteristics. By introducing physical constraints including point spread function boundary loss and peak loss, the generated degradation process is forced to conform to the physical priors of the MRI imaging system, thus ensuring the accuracy and realism of the degradation model from the outset.

[0020] This invention breaks through the limitations of traditional methods that involve degradation followed by reconstruction or independent training, constructing a closed-loop system where the degradation network and the super-resolution reconstruction network are trained and optimized alternately. By freezing the parameters of one network in stages while training the other, and iterating in a loop, the low-resolution images generated by the degradation model can be better processed by the reconstruction network. Simultaneously, the needs of the reconstruction network can guide the optimization of the degradation model. This collaborative optimization mechanism effectively reduces the accumulation of errors in a single network and the problem of model mismatch, promoting consistency in the structure and physical characteristics of the two networks, leading them together towards the global optimum.

[0021] To address the diverse slice thickness requirements in clinical practice, this invention innovatively integrates a scale-aware module into the degradation network. This module dynamically generates convolutional kernel parameters based on the input scaling factor, enabling the degradation model to perceive and adapt to degradation patterns under different scale factors. This is fundamentally different from existing methods that train independent models for each fixed scale or use fixed kernels, significantly enhancing the model's multi-scale generalization ability and clinical applicability.

[0022] By combining a closed-loop feedback mechanism and physical constraint loss, this invention ensures that the reconstructed high-resolution image not only closely approximates the real image in pixel intensity but also maintains a high degree of consistency in structural details and physical properties. This effectively avoids structural distortions or non-physical artifacts that may occur with traditional methods, making the reconstruction results more in line with the needs of clinical diagnosis. Attached Figure Description

[0023] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a system framework diagram of the blind isotropic MRI super-resolution reconstruction method based on dual-input degradation modeling described in this invention. Figure 2 This is a structural diagram of the scale sensing module described in this invention; Figure 3This is a comparison chart of Gaussian slice contour estimation for the blind isotropic MRI super-resolution reconstruction method based on dual-input degradation modeling described in this invention. The first row shows the results of this method, and the second row shows the results of the ESPRESO method. From left to right, they represent the results under scaling factors of 2.0, 3.5, and 4.9, respectively. Figure 4 This is a comparison chart of rectangular slice contour estimation in the blind isotropic MRI super-resolution reconstruction method based on dual-input degradation modeling described in this invention. The first row shows the results of this method, and the second row shows the results of the ESPRESO method. From left to right, they represent the results under scaling factors of 2.0, 3.5, and 4.9, respectively. Figure 5 The image shows a visual comparison of the super-resolution reconstruction results of the blind isotropic MRI super-resolution reconstruction method based on dual-input degradation modeling described in this invention. From left to right, the images represent low-resolution image, B-spline reconstruction result, SMORE reconstruction result, ESPRESO reconstruction result, reconstruction result of this method, and high-resolution image, respectively. Figure 6 The image shows a comparison of slice contour estimation results for the blind isotropic MRI super-resolution reconstruction method based on dual-input degradation modeling described in this invention, with and without a scale-aware module. From left to right, the images represent slice contour estimation with and without a scale-aware module under Gaussian slice contours, and slice contour estimation with and without a scale-aware module under rectangular slice contours. Figure 7 The image shows a comparison of the reconstruction results of the blind isotropic MRI super-resolution reconstruction method based on dual-input degradation modeling described in this invention with and without a scale-aware module. From left to right, the images represent low-resolution image, reconstruction result without module, reconstruction result with module, and high-resolution image, respectively. Figure 8 The image shows a comparison of slice contour estimation results under different reconstruction loss weight settings for the blind isotropic MRI super-resolution reconstruction method based on dual-input degradation modeling described in this invention. From left to right, the image represents slice contour estimation under high weight, medium weight, and low weight settings, respectively. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present invention can be combined with each other, and the described embodiments are only some embodiments of the present invention, not all embodiments.

[0025] Implementation Method 1: A blind isotropic MRI super-resolution reconstruction method based on dual-input degradation modeling, the method comprising: Step S1: Acquire unpaired isotropic and anisotropic MRI images. Extract the two-dimensional image along the vertical axis from the isotropic MRI as a high-resolution image, and extract the two-dimensional image along the vertical axis from the anisotropic MRI as a low-resolution image. Step S2: Construct a dual-input degradation network, which includes a generator and a discriminator; the generator takes a high-resolution image and a low-resolution image as input, simulates the image degradation process in the vertical axis direction, and generates a simulated low-resolution image; the discriminator is used to determine whether the input image is a real low-resolution image; Step S3: Construct a super-resolution reconstruction network, which takes a real low-resolution image or a simulated low-resolution image as input and reconstructs a high-resolution image through a deep neural network; Step S4: Introduce a closed-loop feedback mechanism, and optimize in stages by alternately freezing the parameters of the degenerate network and the reconstruction network, including the first stage of fixing the parameters of the reconstruction network to train the degenerate network and the second stage of fixing the parameters of the degenerate network to train the reconstruction network. Step S5: Introduce a scale-aware module based on conditional convolution into the degenerate network. The scale-aware module dynamically generates convolution kernel parameters according to the input scaling factor to adapt to the degenerate modeling requirements under different scaling factors.

[0026] The method proposed in this embodiment, by directly utilizing real high- and low-resolution image pairs along the vertical axis for modeling, can more accurately capture the degradation process unique to this direction, such as partial volumetric effects. Compared with existing single-input methods, the estimated point spread function profile is closer to the real physical model, thereby directly resulting in higher fidelity super-resolution reconstructed image quality, showing significant improvements in objective metrics such as PSNR, SSIM, and subjective visual perception.

[0027] The method proposed in this embodiment employs a self-supervised learning and unpaired data training strategy, enabling effective training without the need for difficult-to-obtain, strictly paired high- and low-resolution image pairs. This overcomes the stringent data preparation requirements of traditional deep learning methods, allowing the technology to be directly applied to real-world clinical anisotropic data from different hospitals and scanners, thus enhancing its practicality.

[0028] Thanks to the introduction of the scale-aware module, this implementation can effectively handle super-resolution tasks with different scale factors using only a single model, avoiding the tedious process of training and maintaining multiple models for each scale. This model can adaptively adjust the degradation kernel, achieving stable and superior reconstruction results for images with different resolution scales.

[0029] Implementation Method Two: This implementation method further defines the blind isotropic MRI super-resolution reconstruction method based on dual-input degradation modeling described in Implementation Method One. The construction of the dual-input degradation network in step S2 includes: The generator is used to generate a simulated low-resolution image based on the high-resolution image and the low-resolution image; The discriminator includes a high-resolution image discriminator and a low-resolution image discriminator; The high-resolution image discriminator is used to determine whether the input image is a real high-resolution image; The low-resolution image discriminator is used to determine whether the input image is a real low-resolution image; The high-resolution image discriminator and the low-resolution image discriminator constitute a dual discriminator structure, which is used to enhance the realism and physical consistency of degradation modeling.

[0030] Implementation Method 3: This implementation method further defines the blind isotropic MRI super-resolution reconstruction method based on dual-input degradation modeling described in Implementation Method 1. The training process of the closed-loop feedback mechanism in step S4 includes: In the first training phase, the parameters of the reconstructed network are fixed, and the degenerate network is trained using a hybrid loss function, which includes: Point spread function constraint loss is used to constrain the shape of the point spread function during the degradation process; Adversarial loss is used to improve the realism of generated images; Reconstruction loss is used to ensure pixel consistency between the degraded image and the real image; Weight decay loss is used to control model complexity; In the second training phase, the parameters of the degraded network are fixed, and the reconstruction network is trained to recover high-resolution images from low-resolution images. Repeat the first and second training phases to form a closed-loop feedback loop, gradually optimizing the accuracy of degradation modeling and image reconstruction.

[0031] Implementation Method 4: This implementation method further defines the blind isotropic MRI super-resolution reconstruction method based on dual-input degradation modeling described in Implementation Method 3. The point spread function constraint loss includes boundary loss and peak loss. The boundary loss is used to penalize non-zero values ​​of the point spread function at the boundary, ensuring that the point spread function at the image boundary tends to zero. The calculation formula is as follows:

[0032] The peak loss is used to constrain the peak position and symmetry of the point spread function, ensuring that the peak position of the point spread function coincides with the axis of symmetry. The calculation formula is as follows:

[0033] in, express The first in One element, express The first in One element, express The first in One element, express The first in One element, Represents the point spread function generated by the degenerate network. PSF length, express The first in One element, express The first in -1 element.

[0034] Implementation Method 5: This implementation method further defines the blind isotropic MRI super-resolution reconstruction method based on dual-input degradation modeling described in Implementation Method 1. The degradation network and reconstruction network are trained using a self-supervised learning method with unpaired isotropic and anisotropic MRI images, without the need for strictly paired high-resolution-low-resolution image pairs.

[0035] Implementation Method Six: This implementation method further defines the blind isotropic MRI super-resolution reconstruction method based on dual-input degradation modeling described in Implementation Method One. The scale-aware module in step S5 includes: Two expert convolutional kernels are used to represent different types of degradation patterns; A weight prediction network that takes a scaling factor as input and outputs a set of normalized combined weights; A dynamic convolution kernel generation unit is used to perform weighted fusion of expert convolution kernels according to the combined weights to generate a dynamic convolution kernel suitable for the current scale factor.

[0036] Implementation Method Seven: This implementation method further defines the blind isotropic MRI super-resolution reconstruction method based on dual-input degradation modeling described in Implementation Method Six. The expert convolution kernel size in this step is a standard 3×3; the weight prediction network consists of two fully connected layers with a hidden layer dimension of 64, and uses a scale factor... As input, output a set of normalized combined weights. The final scale-adaptive convolution kernel The results were obtained through weighted combination calculations by expert panelists: , in, and This indicates two expert cores, initialized using the Kaiming initialization method.

[0037] Implementation Method 8: A blind isotropic MRI super-resolution reconstruction system based on dual-input degradation modeling, the system comprising: The image acquisition unit is used to acquire unpaired isotropic MRI and anisotropic MRI, extract two-dimensional images in the vertical direction from the isotropic MRI as high-resolution images, and extract two-dimensional images in the vertical direction from the anisotropic MRI as low-resolution images. A dual-input degradation network construction unit is used to construct a dual-input degradation network, which includes a generator and a discriminator. The generator takes a high-resolution image and a low-resolution image as inputs, simulates the image degradation process in the vertical axis direction, and generates a simulated low-resolution image. The discriminator is used to determine whether the input image is a real low-resolution image. An image reconstruction unit is used to construct a super-resolution reconstruction network, which takes a real low-resolution image or a simulated low-resolution image as input and reconstructs a high-resolution image through a deep neural network. The joint optimization unit is used to introduce a closed-loop feedback mechanism. It optimizes the network in stages by alternately freezing the parameters of the degenerate network and the reconstruction network. The first stage fixes the parameters of the reconstruction network to train the degenerate network, and the second stage fixes the parameters of the degenerate network to train the reconstruction network. A scale-aware unit is used to introduce a conditional convolution-based scale-aware module into the degradation network. The scale-aware module dynamically generates convolution kernel parameters according to the input scaling factor to adapt to the degradation modeling requirements under different scaling factors.

[0038] Implementation Method Nine: A computer device according to this implementation method includes a memory and a processor. The memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes a blind isotropic MRI super-resolution reconstruction method based on dual-input degradation modeling according to any one of Implementation Methods One to Seven.

[0039] Implementation Method 10: A computer-readable storage medium according to this implementation method stores a computer program, which, when executed by a processor, performs the steps of a blind isotropic MRI super-resolution reconstruction method based on dual-input degradation modeling as described in any one of Implementation Methods 1 to 7.

[0040] Implementation Method 11, see below Figures 1 to 8 This embodiment describes a specific example of the blind isotropic MRI super-resolution reconstruction method based on dual-input degradation modeling described in Embodiment 1. It also serves to explain Embodiments 2 through 7. Specifically: The method proposed in this embodiment employs a dual-input architecture for vertical axis degradation modeling and super-resolution reconstruction of unpaired volumetric MRI. For example... Figure 1 As shown, the degradation branch takes a high-resolution image as input, dynamically estimates the vertical axis point spread function through a scale-aware degradation generator, and downsamples it to generate a low-resolution image. The reconstruction branch inputs the generated low-resolution image and the real low-resolution image into an RCAN-based reconstruction network to obtain a high-resolution output. These two branches operate in an iterative and collaborative manner by alternately freezing parameters and jointly optimizing. The degradation generator is trained by combining adversarial loss, physical constraint loss based on point spread function, and reconstruction loss to capture realistic imaging degradation features; simultaneously, the reconstruction network is optimized by adversarial loss and reconstruction loss to improve super-resolution accuracy. Together, they construct a closed-loop feedback mechanism to alleviate the accumulation of errors during training. To further improve performance, this method introduces a dual discriminator supervision mechanism: a high-resolution discriminator evaluates the structural fidelity between the reconstructed high-resolution image and the real high-resolution image, and a low-resolution discriminator evaluates the consistency of degradation patterns between the generated low-resolution image and the real low-resolution image.

[0041] Step 1: Construction of a dual-input blind isotropic super-resolution network framework In anisotropic MRI, the vertical axis resolution is significantly lower than the other two axes. Isotropic super-resolution reconstruction aims to restore these anisotropic bodies to isotropic MRI with consistent resolution across all dimensions. For isotropic super-resolution networks based on a single input of anisotropic MRI, degradation and reconstruction are learned only through the horizontal and vertical axes. Due to the feature differences between the horizontal and vertical axes and the vertical axis, this method utilizes real vertical axis information to guide the network in learning degradation and reconstruction. Unpaired high-resolution and low-resolution MRIs are constructed by extracting 2D slices along the vertical axis from isotropic and anisotropic MRIs respectively. After super-resolution reconstruction, the 2D slices of the anisotropic body are recombined to obtain the isotropic MRI. The 2D slices extracted from the isotropic body are denoted as... The 2D slices extracted from the anisotropic body are denoted as This method reconstructs a high-resolution image from a given low-resolution input, which can be represented as:

[0042] in, This represents the high-resolution image obtained after reconstruction. This represents the dual-input blind isotropic super-resolution network proposed in this method.

[0043] Similarly, this method utilizes features along the vertical axis to learn degradation, and the degradation fitting process can be represented as:

[0044] in, This represents the low-resolution image obtained through degenerate fitting. This represents the degenerate network in this method.

[0045] Step 2: Generative Adversarial Network for Degenerate Fitting Based on Vertical Axis Generative adversarial networks (GANs) are used to dynamically estimate slice contours. For example... Figure 1 As shown, learnable tensors Hierarchical features are generated through two convolutional layers (each with a 3×3 kernel). Each convolutional layer is followed by a ReLU activation function. The first convolutional layer outputs 256 channels, and the second convolutional layer reduces the number of channels in the feature map to 1. Simultaneously, the scale factor... With learnable tensors Common Input Scale Awareness Module The adaptive weight matrix is ​​obtained. This is then multiplied element-wise with the hierarchical features. This process can be formally represented as:

[0046]

[0047]

[0048] in, This indicates element-wise multiplication.

[0049] A single-layer convolution is applied to the output to reduce the channel dimension, and then normalization is performed using the Softmax function to ensure the estimated slice contours. It strictly conforms to the physical definition of the point spread function, avoiding artifact propagation caused by probability distribution distortion. This process can be expressed as:

[0050] To further constrain the estimation results and ensure symmetry, the estimated... The average of its flipped versions is obtained As the point spread function in low-resolution MRI imaging, the degradation process of MRI can therefore be expressed as:

[0051] in, This indicates a true high-resolution MRI. This represents the generated low-resolution MRI, and ⊙ represents the convolution operation along the vertical axis. The scaling factor along the vertical axis is... The downsampling function.

[0052] To fully utilize all information from both high-resolution and low-resolution images, this method employs a dual-discriminator approach to guide generator optimization, comprising one discriminator for distinguishing real high-resolution images and one discriminator for distinguishing real low-resolution images. The structures of the two discriminators are as follows: Figure 1 As shown, each layer consists of 5 convolutional layers, alternating between LeakyReLU activation functions. Spectral normalization is applied after each convolutional layer to stabilize the training process. Except for the last layer, all convolutional layers have 64 output channels; the first three layers have a kernel size of 3, and the last two have a kernel size of 1. The network output is converted into a probability map using the Sigmoid function for subsequent discriminator optimization.

[0053] Step 3: Design of a scale-aware module based on conditional convolution This module contains two key components: (1) a set of expert cores The size is standard 3×3, and the Kaiming uniform initialization method is used as the learnable parameter to provide a basis for representing multiple degradation modes; (2) A lightweight dynamic weight predictor is composed of two fully connected layers with a hidden layer dimension of 64, and the scale factor is used to predict the size of the hidden layer. As input, output a set of normalized combined weights. The final scale-adaptive convolution kernel The results were obtained through weighted combination calculations by expert panelists:

[0054] like Figure 2 As shown, the scale factor is converted into combined weights through a weight prediction network, which are used to synthesize a scale-aware dynamic filter. The filter operates on the embedding vector through standard convolution. Ω Generate an adaptive weight matrix . The calculation process can be expressed as:

[0055] in, This represents the convolution operation.

[0056] Step 4: Loss Function Design During the training phase of the degenerate network, the parameters of the reconstructed network are fixed, and a hybrid loss function is used to optimize the parameters of the degenerate network. This hybrid loss function includes point spread function constraint loss, adversarial loss, and reconstruction loss.

[0057] Point spread function loss includes boundary loss and peak weight loss The calculation formula is:

[0058]

[0059] Wherein, the point spread function is , length is , express The Each element.

[0060] Combating losses Represented as:

[0061]

[0062]

[0063] in, and These represent two discriminators. and These represent real high-resolution and low-resolution images, respectively. and These represent the generated high-resolution image and low-resolution image, respectively.

[0064] Reconstruction losses Represented as:

[0065] The complete loss function for the degradation phase is a weighted sum of the above terms:

[0066] During the network reconstruction training phase, the parameters of the degenerate network are fixed, and the loss function is expressed as:

[0067] This method was validated using multimodal isotropic 3D MRI data. The datasets included the ITKTubeTK dataset and the BraTS2013 dataset. Preprocessing of the isotropic 3D MRI included: (1) correcting intensity inhomogeneity using the N4ITK algorithm; and (2) normalizing white matter intensity to 1000. Considering that the commonly used slice thicknesses in clinical scanning are 2.0 mm, 3.5 mm, and 4.9 mm, the scale factor was set to {2.0, 3.5, 4.9}. To generate unpaired anisotropic 3D MRI, the degradation process was simulated by applying slice contour degradation in the vertical direction of the isotropic image and then downsampling according to the specified scale factor. Degradation modeling was performed using Gaussian and rectangular slice contours: for Gaussian contours, the half-peak width (FWHM) was set to 50% (large gap), 75% (medium gap), 100% (no gap), and 125% (overlap) of the slice spacing. For rectangular profiles, FWHM is set to the closest odd number corresponding to the four Gaussian conditions mentioned above.

[0068] The Adam optimizer was used to train both the degenerate and reconstruction networks, with the learning rate dynamically adjusted via single-cycle learning rate scheduling: the maximum learning rate was set to 0.01, the initial learning rate to 0.001, and the total number of iterations to 100,000. For the input anisotropic 3D MRI, low-resolution image patches (16×16) were extracted along the vertical axis; corresponding high-resolution image patches (16×16) were extracted from the isotropic 3D MRI according to the scale factor. + )),in This indicates the slice outline length, set to 21. The batch size is fixed at 128.

[0069] In the initial training phase, 10 rounds of supervised pre-training are performed: slice contours are predefined according to the corresponding scale factors; convolution and downsampling are performed on the vertical axis of the high-resolution image using both the predefined slice contours and the slice contours estimated by the degradation network to generate the real low-resolution image and the estimated low-resolution image; the degradation network is optimized by minimizing the L2 loss between the estimated low-resolution image and the real low-resolution image. Subsequently, the reconstruction network is trained to recover the high-resolution image from the real low-resolution image, and its reconstruction capability is initialized using L1 loss. The weights of the loss function are set to... .

[0070] Tables 1 and 3 compare the degradation network performance of the ESPRESO method and our proposed method on Gaussian slice contours and rectangular slice contours in the T1w and T2w modes of the ITKTubeTK dataset, respectively; Tables 2 and 4 show the experimental results on the BraTS2013 dataset. Overall, our proposed method outperforms the ESPRESO method in degradation simulation performance. Taking the T1w mode of the ITKTubeTK dataset as an example, a visual comparative analysis of the Gaussian slice contour estimation of our proposed method and the ESPRESO method is performed, such as... Figure 3 As shown. Figure 4 Compare the estimation results of the rectangular slice outline.

[0071] Table 1 Performance evaluation of the degenerate network in Gaussian slice contour scenes on the ITKTubeTK dataset.

[0072] The evaluation metric, FAE (mm) / SAE (mm), represents the average performance of the last row of each scale factor under different FWHM values.

[0073] Table 2 Performance evaluation of the degenerate network in Gaussian slice contour scenes on the BraTS2013 dataset.

[0074] Table 3 Performance evaluation of degenerate networks in rectangular slice contour scenarios on the ITKTubeTK dataset.

[0075] Table 4 Performance evaluation of the degenerate network in the rectangular slice outline scene of the BraTS2013 dataset.

[0076] The super-resolution reconstruction performance of the method proposed in this invention was compared with that of B-spline, SMORE, ESPRESO, and SynthSR methods. The results consistently show that the proposed method outperforms the benchmark methods in multiple super-resolution reconstruction quality metrics, as shown in Tables 5 and 6. Figure 5 The super-resolution reconstruction results of this method and the comparison methods are visualized.

[0077] Table 5. Comparison of super-resolution reconstruction performance under Gaussian slice contour degradation

[0078] The evaluation metric is PSNR (dB) / SSIM, which represents the average performance across all FWHM settings, with the last row of each scale factor group being the average performance.

[0079] Table 6. Comparison of super-resolution reconstruction performance under rectangular slice contour degradation

[0080] The performance results of the degraded network with and without the scale-aware module are shown in the table below, and the visualization results of the degraded network and the reconstructed network are as follows. Figure 5 and Figure 6 .

[0081] Table 7 Performance of degraded and reconstructed networks with and without scale-aware modules

[0082] The reconstruction loss weights in the degraded network are shown in the table below. Figure 7 The reconstruction results under different weight settings are visualized.

[0083] Table 8. Performance of degraded networks under different weight settings

[0084] Table 9. Reconstruction network performance under different weight settings

[0085] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, systems, or computer program products. Therefore, this disclosure can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0086] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this disclosure and not to limit its protection scope. Although this disclosure has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading this disclosure, they can still make various changes, modifications or equivalent substitutions to the specific implementation of the invention, but these changes, modifications or equivalent substitutions are all within the protection scope of the published pending claims.

Claims

1. A blind isotropic MRI super-resolution reconstruction method based on dual-input degradation modeling, characterized in that, The method includes: Step S1: Acquire unpaired isotropic and anisotropic MRI images. Extract the two-dimensional image along the vertical axis from the isotropic MRI as a high-resolution image, and extract the two-dimensional image along the vertical axis from the anisotropic MRI as a low-resolution image. Step S2: Construct a dual-input degradation network, which includes a generator and a discriminator; the generator takes a high-resolution image and a low-resolution image as input, simulates the image degradation process in the vertical axis direction, and generates a simulated low-resolution image; the discriminator is used to determine whether the input image is a real low-resolution image; Step S3: Construct a super-resolution reconstruction network, which takes a real low-resolution image or a simulated low-resolution image as input and reconstructs a high-resolution image through a deep neural network; Step S4: Introduce a closed-loop feedback mechanism, and optimize in stages by alternately freezing the parameters of the degenerate network and the reconstruction network, including the first stage of fixing the parameters of the reconstruction network to train the degenerate network and the second stage of fixing the parameters of the degenerate network to train the reconstruction network. Step S5: Introduce a scale-aware module based on conditional convolution into the degenerate network. The scale-aware module dynamically generates convolution kernel parameters according to the input scaling factor to adapt to the degenerate modeling requirements under different scaling factors.

2. The blind isotropic MRI super-resolution reconstruction method based on dual-input degradation modeling according to claim 1, characterized in that, The construction of the dual-input degenerate network in step S2 includes: The generator is used to generate a simulated low-resolution image based on the high-resolution image and the low-resolution image; The discriminator includes a high-resolution image discriminator and a low-resolution image discriminator; The high-resolution image discriminator is used to determine whether the input image is a real high-resolution image; The low-resolution image discriminator is used to determine whether the input image is a real low-resolution image; The high-resolution image discriminator and the low-resolution image discriminator constitute a dual discriminator structure, which is used to enhance the realism and physical consistency of degradation modeling.

3. The blind isotropic MRI super-resolution reconstruction method based on dual-input degradation modeling according to claim 1, characterized in that, The training process of the closed-loop feedback mechanism in step S4 includes: In the first training phase, the parameters of the reconstructed network are fixed, and the degenerate network is trained using a hybrid loss function, which includes: Point spread function constraint loss is used to constrain the shape of the point spread function during the degradation process; Adversarial loss is used to improve the realism of generated images; Reconstruction loss is used to ensure pixel consistency between the degraded image and the real image; Weight decay loss is used to control model complexity; In the second training phase, the parameters of the degraded network are fixed, and the reconstruction network is trained to recover high-resolution images from low-resolution images. Repeat the first and second training phases to form a closed-loop feedback loop, gradually optimizing the accuracy of degradation modeling and image reconstruction.

4. The blind isotropic MRI super-resolution reconstruction method based on dual-input degradation modeling according to claim 3, characterized in that, The point spread function constraint loss includes boundary loss and peak loss; The boundary loss is used to penalize non-zero values ​​of the point spread function at the boundary, ensuring that the point spread function at the image boundary tends to zero. The calculation formula is as follows: The peak loss is used to constrain the peak position and symmetry of the point spread function, ensuring that the peak position of the point spread function coincides with the axis of symmetry. The calculation formula is as follows: in, express The first in One element, express The first in One element, express The first in One element, express The first in One element, Represents the point spread function generated by the degenerate network. PSF length, express The first in One element, express The first in -1 element.

5. The blind isotropic MRI super-resolution reconstruction method based on dual-input degradation modeling according to claim 1, characterized in that, The degradation network and reconstruction network employ a self-supervised learning method, trained using unpaired isotropic and anisotropic MRI images, without requiring strictly paired high-resolution-low-resolution image pairs.

6. The blind isotropic MRI super-resolution reconstruction method based on dual-input degradation modeling according to claim 1, characterized in that, The scale-aware module in step S5 includes: Two expert convolutional kernels are used to represent different types of degradation patterns; A weight prediction network that takes a scaling factor as input and outputs a set of normalized combined weights; A dynamic convolution kernel generation unit is used to perform weighted fusion of expert convolution kernels according to the combined weights to generate a dynamic convolution kernel suitable for the current scale factor.

7. The blind isotropic MRI super-resolution reconstruction method based on dual-input degradation modeling according to claim 6, characterized in that, The expert convolution kernel size in the step is a standard 3×3; the weight prediction network consists of two fully connected layers, with a hidden layer dimension of 64, and a scale factor. As input, output a set of normalized combined weights. The final scale-adaptive convolution kernel The weighted combination calculation by expert cores yielded the following: ， in, and This indicates two expert cores, initialized using the Kaiming initialization method.

8. A blind isotropic MRI super-resolution reconstruction system based on dual-input degradation modeling, characterized in that, The system includes: The image acquisition unit is used to acquire unpaired isotropic MRI and anisotropic MRI, extract two-dimensional images in the vertical direction from the isotropic MRI as high-resolution images, and extract two-dimensional images in the vertical direction from the anisotropic MRI as low-resolution images. A dual-input degradation network construction unit is used to construct a dual-input degradation network, which includes a generator and a discriminator. The generator takes a high-resolution image and a low-resolution image as inputs, simulates the image degradation process in the vertical axis direction, and generates a simulated low-resolution image. The discriminator is used to determine whether the input image is a real low-resolution image. An image reconstruction unit is used to construct a super-resolution reconstruction network, which takes a real low-resolution image or a simulated low-resolution image as input and reconstructs a high-resolution image through a deep neural network. The joint optimization unit is used to introduce a closed-loop feedback mechanism. It optimizes the network in stages by alternately freezing the parameters of the degenerate network and the reconstruction network. The first stage fixes the parameters of the reconstruction network to train the degenerate network, and the second stage fixes the parameters of the degenerate network to train the reconstruction network. A scale-aware unit is used to introduce a conditional convolution-based scale-aware module into the degradation network. The scale-aware module dynamically generates convolution kernel parameters according to the input scaling factor to adapt to the degradation modeling requirements under different scaling factors.

9. A computer device, characterized in that: The system includes a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes a blind isotropic MRI super-resolution reconstruction method based on dual-input degradation modeling according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of a blind isotropic MRI super-resolution reconstruction method based on dual-input degradation modeling as described in any one of claims 1-7.

Images