Method for accelerating high-quality MRI (Magnetic Resonance Imaging) image reconstruction based on TC-KANReccon model

Abstract

The invention relates to an accelerated high-quality MRI (Magnetic Resonance Imaging) image reconstruction method based on a TC-KANReccon model. The method comprises the following steps: S1, collecting and preprocessing image data; s2, an accelerated high-quality MRI image reconstruction model based on a TC-KANReccon model is built; and S3, designing a noise cutting strategy of linear reduction of the cutting threshold. In each step of the sampling process, a linear decreasing cutting strategy is adopted according to the current progress of the sampling step. The strategy is applied to noise estimation generated by an MF-UKAN diffusion trunk so as to generate an MRI image with a better visual effect; s4, repeatedly training a diffusion model and a VAE module by taking the preprocessed MRI image and the full-sampling k-space data as input, optimizing network parameters, and continuously performing iterative optimization; and S5, inputting the under-sampled k-space data into the trained accelerated high-quality MRI image reconstruction model based on the TC-KANReccon model, and applying a matched noise cutting strategy of linear reduction of a cutting threshold to finally generate a high-quality MRI image.

Description

Technical Field

[0001] The invention relates to an accelerated high-quality MRI image reconstruction method based on a TC-KANRecon model, and belongs to the technical field of image restoration and reconstruction. Background Art

[0002] Magnetic resonance imaging (MRI) is an advanced medical imaging technique that occupies an important position in clinical diagnosis and research with its high resolution and multiple contrast mechanisms. MRI can provide precise anatomical and functional information, making it an important tool for diagnosing neurological diseases, cardiovascular diseases, and cancer. Compared with other imaging methods, MRI has significant advantages such as no radiation and high soft tissue contrast, which makes it more and more widely used in fields such as neuroimaging, cardiac imaging, and oncology. However, the application of MRI also faces many challenges, especially its long acquisition time, which not only increases the discomfort of patients but also limits the efficiency of the machine. Long scanning time may make it difficult for patients to remain still during the examination, resulting in motion artifacts and affecting image quality. This problem is particularly prominent in the elderly and children because it is more difficult for them to remain still for a long time. In addition, long scans also increase medical costs and reduce the turnover efficiency of equipment. Therefore, in order to improve the efficiency of MRI, researchers are constantly exploring ways to accelerate MRI acquisition and reconstruction in order to shorten the scan time while maintaining image quality.

[0003] The acquisition process of traditional MRI is slow, mainly because it takes a long time to fully acquire k-space data. K-space is the frequency domain representation of MRI images, and the final image can be obtained by inverse Fourier transforming the k-space data. In order to obtain high-quality images, it is usually necessary to fully acquire the entire k-space, which leads to a longer scan time. This extended acquisition time not only puts a burden on the patient, but may also cause image artifacts due to patient movement, reducing the accuracy of diagnosis. To solve this problem, undersampling k-space data has become a common acceleration method. This method undersamples the signal in the k-space and restores the signal through a post-reconstruction algorithm.

[0004] Although undersampling shortens the scanning time by reducing the amount of collected data, it violates the Nyquist-Shannon sampling theorem and easily introduces aliasing artifacts. In order to obtain high-quality images under undersampling, researchers proposed compressed sensing (CS) technology. CS effectively alleviates the aliasing artifacts caused by undersampling by formulating image reconstruction as an optimization problem and introducing the assumptions of sparsity and incoherence.

[0005] Although CS technology has improved the problems caused by undersampling to a certain extent, its complexity and limitations limit its widespread application. CS algorithms usually require multiple iterative calculations, which not only increases the calculation time, but also places high demands on computing resources. In addition, the CS method is highly dependent on experience in the selection of hyperparameters, and requires personalized adjustments under different scanning conditions, which increases the difficulty of application. In order to further improve the acceleration capability of MRI, researchers have explored other methods, such as low-rank constraints, adaptive sparse modeling, and parallel imaging technology. These methods improve the quality and speed of image reconstruction through different technical means, but there is still the problem of aliasing artifacts under high acceleration factors. Parallel imaging technology can improve the acquisition speed and image quality to a certain extent by using multiple receiving coils to collect signals at the same time. However, the effect of parallel imaging depends on the configuration of the coil and the selection of imaging parameters, and also requires experienced operators to debug.

[0006] In recent years, the introduction of artificial intelligence technology in the field of magnetic resonance imaging has brought new hope for MRI acceleration and reconstruction. By training deep learning models, researchers can achieve high-quality image reconstruction based on undersampled k-space data. By learning a large amount of sample data, deep learning models can automatically extract features and optimize parameters, greatly reducing the complexity of manual parameter adjustment. At the same time, deep learning methods can complete image reconstruction in a short time, significantly shortening the reconstruction time. However, although existing methods such as diffusion models have promoted the advancement of image reconstruction technology to a certain extent, they often face problems such as noise sensitivity, insufficient structural information retention, and weak robustness.

[0007] In general, the current technical problems are as follows:

[0008] 1) Artifact and noise suppression: Artifacts (such as motion artifacts, metal artifacts, chemical shift artifacts, etc.) and noise that often appear in MRI images seriously affect the image quality, and existing reconstruction technologies often find it difficult to effectively suppress these adverse factors. 2) Reconstruction speed and computational complexity: High-quality MRI reconstruction usually requires a long computing time and complex algorithms, which limits its application in scenarios that require rapid imaging (such as emergency medical situations). How to improve the reconstruction speed while ensuring image quality is an urgent problem to be solved. 3) Detail preservation and resolution improvement: The resolution and detail preservation of MRI images are crucial for accurate diagnosis. However, existing reconstruction technologies sometimes find it difficult to maintain or improve the details and resolution of the image while reducing noise and artifacts. 4) Model robustness and stability: MRI reconstruction models need to show good robustness and stability when facing different noise levels, changes in scanning parameters, or patient-specific conditions to ensure the reliability and consistency of the reconstruction results. 5) Patient safety and comfort: During MRI scanning, patients need to remain still for a long time and may be affected by magnetic fields and radio frequency radiation. How to improve patient comfort and safety while ensuring image quality is also one of the issues that need to be considered in MRI reconstruction technology. Summary of the invention

[0009] In order to overcome the shortcomings of existing research, the present invention provides an accelerated high-quality MRI image reconstruction method based on the TC-KANRecon model, which includes the MF-UKAN framework and its supporting noise trimming strategy. The method automatically generates high-resolution, high-definition and texture-detailed MRI images by inputting undersampled k-space data.

[0010] A method for accelerating high-quality MRI image reconstruction based on the TC-KANRecon model includes the following steps:

[0011] Step 1: Collect and preprocess image data;

[0012] S1.1. The datasets we used come from two public datasets, namely the fastMRI dataset and the SKM-TEA dataset.

[0013] 1) The fastMRI dataset contains raw K-space data and DICOM image data of the knee, brain, prostate, and chest. We used the single-coil knee dataset, which includes 1,172 subjects, each with about 35 slices, of which 973 subjects (about 34,055 slices) were used for training, and 199 subjects (about 6,965 slices) were used for evaluation.

[0014] 2) The SKM-TEA dataset is DICOM image data generated by scanning 155 patients under two Tesla (3T) GE MR750 scanners. We used the single-coil data selected from them, including 515 subjects, each with approximately 160 slices, of which 129 subjects (approximately 20,640 slices) were used for training, and 26 subjects (approximately 4,160 slices) were used for evaluation.

[0015] S1.2. Regarding data preprocessing, before training, we uniformly cropped all the data used to a size of 256*256.

[0016] Step 2: Build an accelerated high-quality MRI image reconstruction model based on the TC-KANRecon model. The overall model framework of the accelerated high-quality MRI image reconstruction method based on the TC-KANRecon model mainly includes three key components: a variational auto-encoder (VAE) module for encoding and decoding image data to reduce computational power, a conditional encoder module MC-Model for receiving conditional images to enhance image directional generation, and a backbone MF-UKAN that integrates the KAN network and the MF Model module to better predict the diffusion of noise.

[0017] MF-UKAN integrates the KAN network and MFModel module, which can balance the relationship between noise elimination and image structure information preservation, greatly improving the preservation of structural information and detail features in the reconstructed image. The core idea is to replace the multilayer perceptron (MLP) in the traditional U-Net network architecture with the KAN network, while combining the multi-head attention mechanism and scalar factors to perform fine-grained control of backbone and jump features to enhance the robustness of the model in complex noise environments. Specifically, a k-layer KAN can be described by the nesting of multiple KAN layers, and the mathematical expression is as follows:

[0018]

[0019] where Φ i Represents the i-th layer of the entire KAN network. Each KAN layer has n in dimensional input and n out Output, Φ = {φ q，p}, p = 1, 2, ..., n in ,q=1,2…,n out , Φ: contains n in *n out The calculation results of the KAN network from the kth layer to the k+1th layer can be expressed by the matrix:

[0020]

[0021] On the basis of introducing the KAN network, we also introduced the multi-head attention mechanism and two key scalar modulation factors. These two factors - backbone feature scaling factor (b l ) and the skip feature scaling factor (s l ) - designed to precisely adjust the flow of information: b l Aims to enhance the backbone feature map x l to strengthen its expressive power to consolidate its core role in denoising tasks; and s l is responsible for moderately attenuating the jump feature map h l The influence of b can be avoided to avoid excessive interference of high-frequency information, thus ensuring that the semantic features of the backbone network are not marginalized. l We adopt an adaptive adjustment strategy, which dynamically adjusts the scaling factor based on statistical information such as sample mean to achieve the main feature map x l Intelligent amplification:

[0022]

[0023] Among them, x l，i , i represents the feature map x l C represents the i-th channel of x l The total number of channels in α l represents the backbone factor, b l is a scalar constant. When optimizing the backbone feature modulation strategy, although strengthening the backbone features can significantly improve the denoising performance of the model, this process is often accompanied by potential damage to the high-frequency details of the image, resulting in excessive smoothing of the synthesized image. In order to delicately balance this contradiction, we adopted a feature scaling strategy: selectively apply the scaling operation to the feature map x l In half of the channels, the backbone features that are crucial for denoising are enhanced, while the high-frequency detail information in the other half of the channels is cleverly preserved, thus finding an elegant balance between denoising and preserving image details:

[0024]

[0025] For s l We further use spectral modulation in the Fourier domain to selectively reduce the low-frequency components of the jump features to further alleviate the over-smooth texture problem caused by enhanced denoising. The specific operations are as follows:

[0026]

[0027] Among them, FFT and IFFT are Fourier transform and inverse Fourier transform respectively. ⊙ represents pixel-level multiplication, β l，i is a Fourier mask designed as a function of the magnitude of the Fourier coefficients to achieve a frequency-dependent scaling factor s l . r is the radius. r thresh is the threshold frequency. We then concatenate and fuse the enhanced skip feature map with the finely modulated backbone feature map and use it as the input to subsequent layers in the U-Net architecture.

[0028] In addition, regarding the MC-Model module, it can better control the directional generation of MRI images. We use the fully sampled k-space information as a conditional guide. First, we use the fully sampled k-space information X obs Perform inverse Fourier transform and then merge it with the MRI image. This process is shown as follows:

[0029]

[0030] Where IFFT stands for Inverse Fourier Transform, Represents a splicing operation, Indicates the concatenated conditional information, and then Add noise to get Finally, in order to keep the input consistent with the diffusion backbone, we use an additional convolution module to compress We use a convolutional layer with a kernel size of 3, a stride of 1, and a padding of 1 for initial compression, followed by a RELU and group normalization layer. This sequence is performed three times, and finally a feature map with the same shape as the diffusion trunk is obtained, which is then passed to the subsequent module. The final feature vector z is expressed as:

[0031]

[0032] Where Conv() represents a convolution operation. Unlike alternative approaches that utilize cross-attention, concatenation, or CLIP image encoders to extract and transfer high-level semantic information in images, we directly use the encoder stage of MF-UKAN, which consists of 3 downsampling blocks and 2 Tok-KAN modules. Each convolutional block consists of a convolutional layer, a batch normalization layer, and a ReLU activation function with a kernel size of 3, a stride of 1, and a padding of 1. This module receives the fully sampled k-space information and MRI image information after Fourier transformation, and then transmits the resulting latent features through the downsampling blocks and intermediate blocks, carefully recording all the intermediate feature maps, and finally introducing these recorded feature maps into the upsampling segment of the diffusion backbone.

[0033] Step 3: Noise cropping strategy with linearly decreasing cropping threshold. The cropping threshold is crucial in balancing image quality and diversity during the generation process. A higher cropping threshold retains more predicted noise and variation, resulting in generated images with diverse styles and details, but it may also introduce undesirable noise or blur, affecting the overall image quality. On the contrary, a lower cropping threshold reduces noise and produces clearer and more stable images, but over-restriction can lead to a lack of detail and diversity, making the image look too smooth. Therefore, choosing an appropriate cropping threshold is crucial to achieve the best balance between image quality and diversity. This strategy allows the cropping threshold to be flexibly adjusted according to the training progress of the model and the stage of image generation. In the early stages of diffusion model sampling, when the model generates a fuzzy, low-resolution initial image structure from pure noise, a larger cropping threshold is needed to increase the degrees of freedom and reduce information loss, which helps capture the overall shape. As sampling proceeds and the image resolution and details increase, the cropping threshold should be gradually reduced to limit noise and enhance the fineness and quality of the image. Therefore, the linear model formula we proposed is:

[0034] s=-ω·t+b.

[0035] This linear model allows the cropping threshold s to decrease linearly as the sampling step t increases. This dynamic cropping strategy preserves the diversity of the image in the initial stage of diffusion model sampling and preserves the details and precision of the image as the sampling goes deeper. Therefore, it produces images that are more consistent with real-world images. When the model is trained, at each step of the sampling process, a linearly decreasing cropping strategy is adopted according to the current progress of the sampling step. This strategy is applied to the noise estimate generated by the MF-UKAN diffusion backbone to produce MRI images with better visual effects.

[0036] Step 4: Use the preprocessed MRI images and full-sampled k-space data as input to repeatedly train the diffusion model and VAE module, optimize the network parameters, and continuously perform iterative optimization;

[0037] Step 5: Use the undersampled k-space data to input into the trained accelerated high-quality MRI image reconstruction model based on the TC-KANRecon model, and apply the corresponding noise trimming strategy with linear reduction of the trimming threshold to finally generate a high-quality MRI image.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] By adopting the above technical solution, the present invention truly realizes the medical demand of automatically generating high-quality magnetic resonance images based on undersampled k-space data, specifically including:

[0040] 1. Significantly improve reconstruction efficiency: Through the optimization algorithm of deep learning, the present invention greatly shortens the acquisition and reconstruction time of MRI images, effectively solving the problems of long acquisition time and poor patient comfort in traditional MRI technology. This improvement makes MRI examinations more efficient, reduces patient waiting time, and improves the overall efficiency of medical services.

[0041] 2. Maintain and improve image quality: The multi-head attention mechanism and scalar modulation factor design within the MF-UKAN framework enable the model to accurately retain the structural information of the image while removing noise, ensuring the high resolution and clarity of the reconstructed image. This is particularly important for clinical diagnosis, as it can help doctors more accurately identify lesion details and improve the accuracy and reliability of diagnosis.

[0042] 3. Enhanced robustness in complex environments: The model of the present invention shows excellent performance in processing complex MRI data such as high noise and low sampling rate, effectively overcoming the limitations of traditional methods under harsh imaging conditions. This feature enables MRI technology to be applied to more diverse clinical scenarios.

[0043] 4. Reduce image detail loss: The noise cropping strategy intelligently adjusts the cropping interval according to the sampling steps, effectively avoiding the problem of image detail loss that may be caused by traditional cropping methods. This innovation not only enriches the visual features of the image, but also improves the overall appearance and diagnostic value of the image. 5. Improve patient experience and safety: Due to the shortened MRI examination time, patients do not need to remain still for a long time, reducing discomfort and anxiety during the examination. At the same time, it avoids the potential risks of long-term exposure to magnetic fields, improving the safety of the examination process and the comfort of patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0045] Figure 1 It is a schematic diagram of the process of the present invention;

[0046] Figure 2 The overall model architecture of the method for accelerating high-quality MRI image reconstruction based on the TC-KANRecon model of the present invention;

[0047] Figure 3 This is a flow chart of the noise clipping strategy with linear reduction of the clipping threshold of the present invention.

[0048] Figure 4 The renderings are generated for the model with 4x accelerated undersampled k-space data as input. The left column shows the contrast renderings of the real knee MRI images under the fastMRI dataset and the MRI images reconstructed based on the method of the present invention, and the right column shows the contrast renderings of the real knee MRI images under the SKM-TEA dataset and the MRI images reconstructed based on the method of the present invention. DETAILED DESCRIPTION

[0049] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0050] The present invention provides a method for accelerating high-quality MRI image reconstruction based on the TC-KANRecon model. Figure 1 As shown, the specific operations are as follows:

[0051] Step 1: Data collection and preprocessing. The data sets used in this embodiment come from two public data sets, namely the fastMRI data set and the SKM-TEA data set. The fastMRI data set contains the original K-space data and DICOM image data of the knee, brain, prostate, and chest data. We used the single-coil knee data set, which includes 1,172 subjects, each subject has about 35 slices, of which 973 subjects (about 34,055 slices) were used for training, and 199 subjects (about 6,965 slices) were used for evaluation. The SKM-TEA data set is DICOM image data generated by scanning 155 patients under two Tesla (3T) GEMR750 scanners. We used the single-coil data selected from it, including 515 subjects, each subject has about 160 slices, of which 129 subjects (about 20,640 slices) were used for training, and 26 subjects (about 4,160 slices) were used for evaluation. Regarding data preprocessing, before training, we uniformly cropped all the data used to a size of 256*256.

[0052] Step 2: If Figure 2As shown in the figure, a network model based on the TC-KANRecon model is built to accelerate the reconstruction of high-quality MRI images. Specifically, the overall model framework of the method for accelerating the reconstruction of high-quality MRI images based on the TC-KANRecon model mainly includes three key components: a VAE module for encoding and decoding image data to reduce computational power, a conditional encoder module MC-Model for receiving conditional images to enhance image directional generation, and a backbone MF-UKAN that integrates the KAN network and the MF Model module to better predict the diffusion of noise.

[0053] MF-UKAN integrates the KAN network and MFModel module, which can balance the relationship between noise elimination and image structure information preservation, greatly improving the preservation of structural information and detail features in the reconstructed image. The core idea is to replace the multilayer perceptron (MLP) in the traditional U-Net network architecture with the KAN network, while combining the multi-head attention mechanism and scalar factors to perform fine-grained control of backbone and jump features to enhance the robustness of the model in complex noise environments. Specifically, a k-layer KAN can be described by the nesting of multiple KAN layers, and the mathematical expression is as follows:

[0054]

[0055] where Φ i Represents the i-th layer of the entire KAN network. Each KAN layer has n in dimensional input and n out Output, Φ = {φ q，p}, p = 1, 2, ..., n in ,q=1,2…,n out , Φ: contains n in *n out The calculation results of the KAN network from the kth layer to the k+1th layer can be expressed by the matrix:

[0056]

[0057] On the basis of introducing the KAN network, we also introduced the multi-head attention mechanism and two key scalar modulation factors. These two factors - backbone feature scaling factor (b l ) and the skip feature scaling factor (s l ) - designed to precisely adjust the flow of information: b l Aims to enhance the backbone feature map x l to strengthen its expressive power to consolidate its core role in denoising tasks; and s l is responsible for moderately attenuating the jump feature map h lThe influence of b can be avoided to avoid excessive interference of high-frequency information, thus ensuring that the semantic features of the backbone network are not marginalized. l We adopt an adaptive adjustment strategy, which dynamically adjusts the scaling factor based on statistical information such as sample mean to achieve the main feature map x l Intelligent amplification:

[0058]

[0059] Among them, x l，i Represents the feature map x l C represents the i-th channel of x l The total number of channels in α l represents the backbone factor, b l is a scalar constant. When optimizing the backbone feature modulation strategy, although strengthening the backbone features can significantly improve the denoising performance of the model, this process is often accompanied by potential damage to the high-frequency details of the image, resulting in excessive smoothing of the synthesized image. In order to delicately balance this contradiction, we adopted a feature scaling strategy: selectively apply the scaling operation to the feature map x l In half of the channels, the backbone features that are crucial for denoising are enhanced, while the high-frequency detail information in the other half of the channels is cleverly preserved, thus finding an elegant balance between denoising and preserving image details:

[0060]

[0061] For s l We further use spectral modulation in the Fourier domain to selectively reduce the low-frequency components of the jump features to further alleviate the over-smooth texture problem caused by enhanced denoising. The specific operations are as follows:

[0062]

[0063] Among them, FFT and IFFT are Fourier transform and inverse Fourier transform respectively. ⊙ represents pixel-level multiplication, β l，i is a Fourier mask designed as a function of the magnitude of the Fourier coefficients to achieve a frequency-dependent scaling factor s l . r is the radius. r thresh is the threshold frequency. We then concatenate and fuse the enhanced skip feature map with the finely modulated backbone feature map and use it as the input to subsequent layers in the U-Net architecture.

[0064] In addition, regarding the MC-Model module, it can better control the directional generation of MRI images. We use the fully sampled k-space information as a conditional guide. First, we use the fully sampled k-space information X obsPerform inverse Fourier transform and then merge it with the MRI image. This process is shown as follows:

[0065]

[0066] Where IFFT stands for Inverse Fourier Transform, Represents a splicing operation, Indicates the concatenated conditional information, and then Add noise to get Finally, in order to keep the input consistent with the diffusion backbone, we use an additional convolution module to compress We use a convolutional layer with a kernel size of 3, a stride of 1, and a padding of 1 for initial compression, followed by a RELU and group normalization layer. This sequence is performed three times, and finally a feature map with the same shape as the diffusion trunk is obtained, which is then passed to the subsequent module. The final feature vector z is expressed as:

[0067]

[0068] Where Conv() represents a convolution operation. Unlike alternative approaches that utilize cross-attention, concatenation, or CLIP image encoders to extract and transfer high-level semantic information in images, we directly use the encoder stage of MF-UKAN, which consists of 3 downsampling blocks and 2 Tok-KAN modules. Each convolutional block consists of a convolutional layer, a batch normalization layer, and a ReLU activation function with a kernel size of 3, a stride of 1, and a padding of 1. This module receives the fully sampled k-space information and MRI image information after Fourier transformation, and then transmits the resulting latent features through the downsampling blocks and intermediate blocks, carefully recording all the intermediate feature maps, and finally introducing these recorded feature maps into the upsampling segment of the diffusion backbone.

[0069] Step 3: If Figure 3 As shown in the figure, a noise clipping strategy with a linear decrease in the clipping threshold is introduced. The linear model formula we proposed is s = -ω·t+b. This linear model allows the clipping threshold s to decrease linearly as the sampling step t increases. This dynamic clipping strategy preserves the diversity of the image in the initial stage of diffusion model sampling and retains the details and accuracy of the image as the sampling goes deeper. Therefore, the images it produces are more consistent with real-world images. When the model training is completed, at each step of the sampling process, a linear decreasing clipping strategy is adopted according to the current progress of the sampling step. This strategy clips the noise estimate generated by the MF-UKAN diffusion backbone to produce MRI images with better visual effects.

[0070] Step 4: All experiments of the present invention are implemented in Pytorch2.0, cuda11.8 framework, and trained on two NVIDIA A800 GPUs. Before training, we center-cropped the fastMRI dataset and SKM-TEA dataset to 256*256. The training process includes VAE components and diffusion components, where the VAE component uses a batch size of 8 and a training time of 48h, while the batch size of the diffusion component is set to 32 and the training time is 36h. Both components are optimized using the Adam optimizer. The learning rate of the VAE component is 0.0001, and the learning rate of the diffusion component is 0.000025.

[0071] Step 5: In the model inference stage, the undersampled k-space data is input, and the network model based on the trained TC-KANRecon model is used to accelerate the reconstruction of high-quality MRI images, and the matching noise trimming strategy with linear reduction of the trimming threshold is applied. Finally, a high-quality, high-resolution MRI image is generated. Figure 4 The model generates effect diagrams when 4 times accelerated undersampled k-space data is used as input. The left column is an example of the comparison effect diagram of the real knee MRI image under the fastMRI dataset and the MRI image reconstructed based on the method of the present invention, and the right column is an example of the comparison effect diagram of the real knee MRI image under the SKM-TEA dataset and the MRI image reconstructed based on the method of the present invention. It can be seen that the MRI image generated by the model is very similar to the real MRI image, and the structural details and texture features can be clearly seen from the enlarged detail comparison diagram, which shows that the present invention can generate high-quality MRI images.

[0072] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions and variations of these embodiments are made without departing from the principles and spirit of the present invention, and still fall within the scope of protection of the present invention.

Claims

1. A method for accelerating high-quality MRI image reconstruction based on the TC-KANRecon model, characterized in that: The following steps are involved: Step 1: Data collection and preprocessing: collect DICOM image data from the fastMRI dataset and SKM-TEA dataset, and perform preprocessing; Step 2: Build a network model based on the TC-KANRecon model to accelerate high-quality MRI image reconstruction. The overall framework of the network model mainly includes three key components: a variational autoencoder module for encoding and decoding image data to reduce computational power, a conditional encoder module MC-Model for receiving conditional images to enhance image directional generation, and a backbone MF-UKAN that integrates the KAN network and the MF Model module to better predict the diffusion of noise; Step 3: A noise clipping strategy with a linearly decreasing clipping threshold allows the clipping threshold s to decrease linearly with the increase of the sampling step t. This preserves the diversity of the image in the initial stage of the diffusion model sampling and preserves the details and accuracy of the image as the sampling goes deeper. The resulting image is more consistent with the real-world image. When the model training is completed, a linearly decreasing clipping strategy is adopted at each step of the sampling process according to the current progress of the sampling step. This strategy is applied to the noise estimate generated by the MF-UKAN diffusion backbone to produce MRI images with better visual effects. Step 4: Use the preprocessed MRI images and full-sampled k-space data as input to repeatedly train the diffusion model and VAE module, optimize the network parameters, and continuously perform iterative optimization; Step 5: Use the undersampled k-space data to input the trained accelerated high-quality MRI image reconstruction model based on the TC-KANRecon model, and apply the corresponding noise trimming strategy with linear reduction of the trimming threshold to finally generate a high-quality MRI image.

2. The method for accelerating high-quality MRI image reconstruction based on the TC-KANRecon model according to claim 1, characterized in that: The step 2 specifically includes: MF-UKAN integrates the KAN network and the MF Model module, which can balance the relationship between noise elimination and image structure information preservation, and improve the preservation of structural information and detail features in the reconstructed image. The k-layer KAN can be described by the nesting of multiple KAN layers. The mathematical expression is as follows: where Φ i Represents the i-th layer of the entire KAN network. Each KAN layer has n in dimensional input and n out Output, Φ = {φ q，p }, p = 1, 2, ..., n in ,q=1,2...,n out , Φ contains n in *n out The learnable activation function φ, the calculation results of the KAN network from the kth layer to the k+1th layer can be expressed by the matrix: Based on the introduction of the KAN network, a multi-head attention mechanism and two key scalar modulation factors are introduced. These two factors are the backbone feature scaling factor b l With skip feature scaling factor s l ——Designed to precisely adjust the flow of information: b l Aims to enhance the backbone feature map x l to strengthen its expressive power to consolidate its core role in denoising tasks; and s l is responsible for moderately attenuating the jump feature map h l The influence of b is avoided to avoid excessive interference of high-frequency information and to ensure that the semantic features of the backbone network are not marginalized. l The setting adopts an adaptive adjustment strategy, which dynamically adjusts the scale factor based on statistical information such as the sample mean to achieve the main feature map x l Intelligent amplification: Among them, x l，i Represents the feature map x l The i-th channel of l The total number of channels in , α l represents the backbone factor, b l is a scalar constant. When optimizing the backbone feature modulation strategy, a feature scaling strategy is adopted: the scaling operation is selectively applied to the feature map x l In half of the channels, the backbone features that are crucial for denoising are enhanced, and the high-frequency detail information in the other half of the channels is cleverly retained, finding an elegant balance between denoising and retaining image details: For s l , spectral modulation is used in the Fourier domain to selectively reduce the low-frequency components of the jump features to further alleviate the over-smooth texture problem caused by enhanced denoising. The specific operations are as follows: Among them, FFT and IFFT are Fourier transform and inverse Fourier transform respectively, ⊙ represents pixel-level multiplication, β l，i is a Fourier mask designed as a function of the magnitude of the Fourier coefficients to achieve a frequency-dependent scaling factor s l , r is the radius, r thresh is the threshold frequency, and the enhanced jump feature map is concatenated and fused with the finely modulated backbone feature map and used as the input of subsequent layers in the U-Net architecture.

3. The method for accelerating high-quality MRI image reconstruction based on the TC-KANRecon model according to claim 2, characterized in that: The MC-Model module in step 2 can better control the directional generation of MRI images, using the fully sampled k-space information as a conditional guide, first converting the fully sampled k-space information X obs Perform inverse Fourier transform and then merge it with the MRI image. This process is shown as follows: Where IFFT stands for Inverse Fourier Transform, Represents a splicing operation, Indicates the concatenated conditional information, and then Add noise to get Keeping the same input as the diffusion backbone, an additional convolutional module is used to compress A convolutional layer with a kernel size of 3, a stride of 1, and a padding of 1 is used for initial compression, followed by a RELU and group normalization layer. This sequence is performed three times, and finally a feature map with the same shape as the diffusion trunk is obtained, which is then passed to the subsequent module. The final feature vector z is expressed as: Where Conv() represents the convolution operation. Unlike alternative methods that use cross-attention, concatenation or CLIP image encoders to extract and transmit high-level semantic information in images, the encoder stage of MF-UKAN is directly used. The encoder stage of MF-UKAN consists of 3 downsampling blocks and 2 Tok-KAN modules. Each convolution block consists of a convolution layer, a batch normalization layer and a ReLU activation function with a kernel size of 3, a stride of 1, and a padding of 1. This module receives the fully sampled k-space information and MRI image information after Fourier transformation, and then transmits the resulting latent features through the downsampling blocks and intermediate blocks, and carefully records all intermediate feature maps. Finally, these recorded feature maps are introduced into the upsampling segment of the diffusion backbone.

4. The method for accelerating high-quality MRI image reconstruction based on the TC-KANRecon model according to claim 1, characterized in that: The step three specifically includes: The strategy allows the cropping threshold to be flexibly adjusted according to the training progress of the model and the stage of image generation. In the early stage of diffusion model sampling, when the model generates a fuzzy, low-resolution initial image structure from pure noise, a larger cropping threshold is needed to increase the degrees of freedom and reduce information loss, which helps to capture the overall shape. As sampling progresses and the image resolution and details increase, the cropping threshold should be gradually reduced to limit noise and enhance the fineness and quality of the image. The proposed linear model formula is: s=-ω·t+b. The linear model allows the cropping threshold s to decrease linearly with the increase of sampling steps t. The dynamic cropping strategy retains the diversity of images in the initial stage of diffusion model sampling and retains the details and accuracy of images as sampling deepens. The generated images are more consistent with real-world images. After the model training is completed, a linear decreasing cropping strategy is adopted in each step of the sampling process according to the current progress of the sampling step.

Images