Arbitrary scale MRI super-resolution reconstruction method based on frequency domain-texture joint learning

Abstract

The application discloses a frequency domain-texture joint learning arbitrary scale MRI super-resolution reconstruction method, relates to the technical field of medical image super-resolution reconstruction, and comprises the following steps: step 1, constructing a data set; step 2, constructing a cooperative feature extraction module MB-CFE; step 3, constructing a texture injection type implicit upsampling module TIM; step 4, adding image domain features and frequency domain features element by element to obtain joint features, and inputting the joint features and texture features into the TIM module together, so that pixel reconstruction of an arbitrary real multiple is completed in a period space, and a high-resolution image is output; step 5, constructing a hybrid loss function, and training a network; and step 6, inputting an image to be reconstructed into the network, and outputting a high-resolution reconstructed image. The frequency domain-texture joint learning arbitrary scale MRI super-resolution reconstruction method provided by the application solves the problems of insufficient texture detail recovery and single super-resolution scale in the prior art.

Description

Technical Field

[0001] This invention relates to the field of medical image super-resolution reconstruction technology, and in particular to an arbitrary-scale MRI super-resolution reconstruction method based on frequency domain-texture joint learning. Background Technology

[0002] Magnetic resonance imaging (MRI) technology, with its inherent advantages of no ionizing radiation and high soft tissue contrast, has become a core imaging tool for clinical analysis of anatomical and pathological features. Taking brain tumors as an example, accurate diagnosis requires images that clearly distinguish the tumor infiltration margin, necrotic core, and edema zone; however, high-resolution scans are too time-consuming, and the risk of artifacts introduced by patient movement is high. Clinically, this forces the use of low-resolution rapid sampling, resulting in the loss of high-frequency data in k-space and a decrease in spatial resolution. Therefore, performing super-resolution reconstruction of MRI images and restoring high-frequency details without increasing scan time has become an urgent need to improve diagnostic accuracy.

[0003] Existing models have achieved certain results in MRI super-resolution reconstruction tasks, but there are still key unresolved issues such as insufficient restoration of texture details and single super-resolution scale. Summary of the Invention

[0004] The purpose of this invention is to provide an arbitrary-scale MRI super-resolution reconstruction method based on frequency domain-texture joint learning, which solves the problems of insufficient texture detail recovery and single super-resolution scale in existing technologies.

[0005] To achieve the above objectives, this invention provides an arbitrary-scale MRI super-resolution reconstruction method based on joint frequency domain-texture learning, comprising the following steps:

[0006] Step 1: Construct the dataset by dividing it into a training set and a test set;

[0007] Step 2: Construct the collaborative feature extraction module MB-CFE, which includes image domain branch, frequency domain branch and texture branch, to obtain image domain features, frequency domain features and texture features. Extract complementary features from spatial structure, frequency distribution and local details respectively to achieve more robust feature fusion.

[0008] Step 3: Construct the Texture Injection Implicit Upsampling Module (TIM);

[0009] Step 4: Add the image domain features and frequency domain features element by element to obtain the joint features. Input the joint features and texture features into the TIM module to complete the reconstruction of pixels of any real multiple in the periodic space and output a high-resolution image.

[0010] Step 5: Construct a hybrid loss function and train the network consisting of the MB-CFE module and the TIM module;

[0011] Step 6: Input the low-resolution MRI image to be reconstructed into the trained network, and output a high-resolution reconstructed image.

[0012] Therefore, this invention adopts the above-mentioned frequency domain-texture joint learning arbitrary scale MRI super-resolution reconstruction method, constructs a multi-branch collaborative feature extraction module (MB-CFE), deploys image domain, frequency domain and texture three branches in parallel, and couples TIM implicit upsampling, which can complete 1.1~30.9× arbitrary real number super-resolution in a single forward inference, no longer limited to fixed integer multiples and the limitations of multiple model switching; the TIM module, through periodic coordinate modeling and cell constraints, modulates and then decodes to ensure zero drift of macroscopic structures such as brain regions and tumor boundaries, while independently injecting high-frequency texture, significantly alleviating the "structural smoothness-texture loss" disconnect; the texture branch adopts a progressive channel design of 1 to 64, reducing the number of parameters compared to conventional U-Net, and implicit MLP decoding replaces traditional interpolation, improving inference speed. A single 12 GB GPU can process 512×512×3 slices in real time, meeting the real-time clinical needs; during the training phase, a hybrid loss of 0.7 L1 and 0.3 LPIPS is used to simultaneously constrain pixel accuracy and perceptual quality, and can freely switch between continuous magnifications without retraining.

[0013] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0014] Figure 1 This is an overall flowchart of the arbitrary-scale MRI super-resolution reconstruction method based on frequency domain-texture joint learning of the present invention.

[0015] Figure 2 This is a diagram of the overall network architecture in an embodiment of the present invention;

[0016] Figure 3 This is an architectural diagram of the image domain branch in an embodiment of the present invention;

[0017] Figure 4 This is an architecture diagram of the RCANBlock block according to an embodiment of the present invention;

[0018] Figure 5 This is an architectural diagram of the frequency domain branch in an embodiment of the present invention;

[0019] Figure 6 This is an architecture diagram of the texture branch in an embodiment of the present invention;

[0020] Figure 7 These are the test results of the embodiments of the present invention on the BraTS dataset;

[0021] Figure 8 The present invention provides test results at different resolutions on the BraTS dataset, including error maps, based on embodiments of the present invention.

[0022] Figure 9 This is a comparison chart of the texture extraction method of this invention and the mainstream method Swin-Transformer at a magnification of 2.

[0023] Figure 10 This is a comparison chart of different loss functions for embodiments of the present invention. Detailed Implementation

[0024] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0025] Please see Figure 1-2 An arbitrary-scale MRI super-resolution reconstruction method based on frequency domain-texture joint learning includes the following steps:

[0026] Step 1: Construct the dataset by dividing it into a training set and a test set.

[0027] S11. Using the BraTS 2021 public dataset, 1251 registered 3D imaging data were downloaded from the multimodal MRI benchmark library. Each registered 3D imaging data contained four registered 3D imaging modalities: T1, T1ce, T2, and FLAIR. The T2 modality was used as the super-resolution reconstruction target. The T2 modality contained 1251 samples, each of which was three-dimensional data with dimensions of 240×240×155 and a voxel size of 1m. 3 The samples and spatial dimensions of the T1, T1ce, and FLAIR modes are consistent with those of the T2 mode, and are used for subsequent experimental verification.

[0028] S12. Each sample in the T2 mode is sequentially divided into 155 240×240 axial slices along the depth direction, resulting in... HR slices; downsampling was performed using bilinear interpolation, where the scaling factor... This generates the corresponding 120×120 LR slice;

[0029] S13. Pair the 240×240 HR slices with the corresponding 120×120 LR slices one by one to form one-to-one registered LR-HR two-dimensional slice pairs to obtain the dataset. Divide the dataset into training set and test set in a ratio of 8:2, and ensure that all slices of the same patient belong only to the training set or test set to prevent data leakage.

[0030] Step 2: Construct the collaborative feature extraction module MB-CFE, which includes image domain branch, frequency domain branch and texture branch, to obtain image domain features, frequency domain features and texture features. Extract complementary features from spatial structure, frequency distribution and local details respectively to achieve more robust feature fusion.

[0031] S21, please refer to Figure 3-4 An image domain branch is constructed using a Residual Channel Attention Network (RCAN) to accurately extract the structural-semantic features of slices in the T2 modality; the image domain branch includes 2 convolutional layers, 4 RCANBlocks, and a global residual.

[0032] S211, The initial convolutional layer slices the input LR layer using 64 3×3 convolutional kernels. Mapping to 64-dimensional features yields The feature map, while maintaining its spatial dimensions, is used to extract underlying texture and edge information. This indicates the input low-resolution MRI slice. This represents the real number field, indicating that the data's numeric type is real. Indicates the height of the slice. Indicates the width of the slice;

[0033] S212. Deep residual channel attention units are constructed by sequentially stacking four RCANBlock blocks to enhance feature representation step by step. Each RCANBlock block sequentially performs 3×3 convolution, ReLU, 3×3 convolution, adaptive average pooling, two layers of 1×1 convolution (64→4→64), and Sigmoid to generate channel weights. The channel weights are multiplied with local features channel by channel, and then added to the input residual of the RCANBlock block to output deep features, thereby achieving channel-level attention enhancement for key structures such as tumor edges in the image.

[0034] S213. The depth features output after stacking 4 RCANBlock blocks are uniformly integrated by 3×3 convolution, which corrects the loss of details in the stacking process and keeps the feature dimension and spatial resolution unchanged.

[0035] S214. Introduce a global residual connection, and add the feature map obtained in S211 to the integrated output obtained in S213 element-wise to obtain the image domain features. This provides a texture-semantic complementary enriched representation, ensuring that high-frequency details are fully transmitted to subsequent reconstruction modules and effectively mitigating gradient degradation issues.

[0036] S22, please refer to Figure 5 Frequency domain branching of the input LR slice A two-dimensional fast Fourier transform is performed to extract features in the frequency domain (k-space) through convolution, and then the features are returned to the image domain through inverse Fourier transform. The high-frequency gain mechanism in k-space is utilized to explicitly enhance tumor edges and fine textures, which complements the image domain branch.

[0037] S221. Slice the input LR. Performing a two-dimensional Fast Fourier Transform (FFT) maps the gray intensity distribution in the image domain to k-space data in the frequency domain, obtaining a complex matrix in k-space, thus realizing the conversion from the spatial domain to the frequency domain;

[0038] S222, Stack the real and imaginary parts of the complex matrix along the channel. The tensor, through 128 3×3 convolutions (each group with a kernel size of 3×3×2), performs frequency domain feature upscaling in the real domain, and... The tensor is expanded to 128 dimensions, and then split into two parts according to the "first 64 dimensions - last 64 dimensions", which serve as the real and imaginary parts of 64 sets of complex features, respectively, thereby enhancing the local spectral correlation.

[0039] S223. Perform inverse Fourier transform (IFFT) on each of the 64 groups of complex features, and remap the enhanced spectrum back to the image domain to obtain... Frequency domain characteristics Since each group of convolutional kernels focuses on different frequency directions and bandwidths in k-space, the 64 feature maps after inverse transformation each highlight edges and tumor boundaries at different scales and in different directions, forming a rich high-frequency detail basis, which complements the semantic features of the image domain branch and provides key frequency domain priors for subsequent fusion and reconstruction.

[0040] S23, please refer to Figure 6 The texture branch includes a differential texture extraction unit, a U-Net network unit, and a lesion attention unit; texture maps are generated through a Gaussian difference mechanism, and a lesion attention mechanism is embedded based on the U-Net structure to accurately capture high-frequency texture and detail features of T2 slices;

[0041] S231. Construct a differential texture extraction unit to process the input LR slices. Perform Gaussian difference, through An adaptive 5×5 Gaussian kernel performs Gaussian blurring, which is then subtracted pixel-by-pixel from the input image to obtain a texture map that focuses on high-frequency details, used to highlight lesion edges and microstructures. The standard deviation of the Gaussian kernel, also known as the smoothing coefficient of Gaussian blur, is expressed as:

[0042] ;

[0043] In the formula, For the output texture map , This is a Gaussian blur function; this operation can effectively separate the low-frequency structural contours and high-frequency texture details of an image, laying the foundation for the subsequent extraction of key texture features such as lesion edges and micro-tissues.

[0044] S232. Construct a U-Net network unit with a mirror structure of "three-level encoding-three-level decoding";

[0045] The encoder's logical flow is as follows: In the first-level encoder, a 3×3 convolution with padding=1 and ReLU are used to... The texture map is losslessly mapped to It captures the finest texture primitives at full resolution; in the second-level encoder, it compresses the space to [a specific value] using 3×3 convolutions with stride=2 and ReLU. And expand the channels to 32, the output feature dimension is This achieves "spatial redundancy reduction + local context aggregation"; in the third-level encoder, it preserves space redundancy through 3×3 convolutions with padding=1 and ReLU. Size, increase the number of channels to 64, the output feature dimension is Further abstracting high-frequency texture patterns completes the "compression-expansion" feature extraction at the encoding end; where padding is the number of pixels to fill the edge region of the input feature map before the convolution operation, and stride is the number of pixels to move each time the convolution kernel slides on the input feature map;

[0046] The decoder performs inverse recovery in a symmetrical manner. The process logic is as follows: In the first-stage decoder, the features are upsampled back using a 3×3 transposed convolution (stride=2, padding=1, output_padding=1, where output_padding is the number of extra pixels added to the edges of the generated output feature map after the transposed convolution operation) and ReLU. By reducing the number of channels to 32, the output feature dimension is... The first stage quickly reconstructs the geometric structure; the second stage decoder compresses the channels to 16 using 3×3 convolution and ReLU for fine-tuning edge details; the third stage decoder adjusts the channels back to 64 using 3×3 convolution and ReLU, and outputs the final product. Texture feature map;

[0047] Thus, the network discards spatial redundancy and expands channel capacity layer by layer during the encoding stage, and restores spatial details and refines channel semantics layer by layer during the decoding stage. This not only preserves deep high-frequency abstraction, but also has complete spatial resolution, providing a high-quality feature base with both details and semantics for the subsequent lesion attention module.

[0048] S233. Enhance key region texture features through lesion attention units: This involves enhancing the decoder output... Using the texture feature map as input, local correlation features are extracted through a 3×3 convolution kernel with padding=1; the local correlation features are compressed into a larger value through adaptive max pooling. The channel-level response vectors are used to focus on the strong response features of the lesion region; the channel weight distribution is adjusted by a 1×1 convolution kernel, and the channel attention weights in the 0-1 interval are generated by combining the sigmoid activation function; finally, the channel attention weights are multiplied by the channel dimension of the texture feature map output by the decoder to obtain the texture features. This enables the differential enhancement of texture features in lesion areas and the suppression of redundant features in background areas, thereby improving the clinical relevance of texture features.

[0049] Step 3: Construct the Texture Injection Implicit Upsampling Module (TIM).

[0050] S31, Based on target super-resolution scale Calculate the mesh size of the SR slice ,in For any real number and , Indicates the height of the SR slice. This indicates the width of the SR slice, determines the mesh size, and generates a normalized global coordinate matrix. ,in This represents the horizontal coordinate of a pixel in an SR slice. Representing the vertical coordinates of pixels in an SR slice, the global coordinate matrix C serves as the "baseline map" for all subsequent feature calculations, ensuring that any... The uniqueness of the pixel spatial location in the lower HR slice and the consistency with the global structure avoid anatomical area shifts in scenarios such as brain MRI;

[0051] S32, Regarding the global coordinate matrix Perform a cycle mapping to convert infinite coordinates into repeating cells; the calculation expression is:

[0052] ;

[0053] ;

[0054] In the formula, express The local coordinate offset of the direction, the result is The relative offset ratio within "integer coordinate units"; express The local coordinate offset of the direction, in The local relative offset ratio of the direction; and ; This function represents the floor function, which takes the largest integer not greater than the input value (e.g., floor(2.7)=2, floor(-0.3)=-1); it normalizes the period of the coordinates, provides a uniform period scale for high-frequency parameterization, and suppresses boundary jumps.

[0055] S33, Through The periodic activation function maps the coordinates Embedded into a 64-dimensional implicit space, we obtain a coordinate embedding vector. (Extended to 64 dimensions), the calculation expression is:

[0056] ;

[0057] In the formula, For frequency coefficients, This embedding vector can enhance the coordinates and subsequent frequency intensity. The characteristic correlation improves the accuracy of periodic structure representation;

[0058] S34. Define the cell unit: The cell unit size is 4×4 pixels. The cell unit size is 8×8 pixels; local normalized coordinates are generated within each cell unit. The local normalized coordinates are mapped to cell features through a fully connected layer (2D input coordinates, 64D output features). This feature can constrain the periodic deviation of the local area under the action of frequency parameters, thus avoiding structural distortion;

[0059] S35, using bilinear interpolation to... Extended to the size of SR slices This provides detail constraints for local pixels, in conjunction with the global coordinate matrix. ( Generate a complementary architecture for a "global-local" dual-coordinate system; for subsequent amplitude weighting ( Regional differential weighting and texture feature overlay lay the foundation for local accuracy, avoiding "local blurring" caused by global periodic features.

[0060] Step 4: Add the image domain features and frequency domain features element by element to obtain the joint features. Input the joint features and texture features into the TIM module to complete the reconstruction of pixels of any real multiple in the periodic space and output a high-resolution image.

[0061] S41. Image domain features Frequency domain characteristics Element-by-element addition yields the fusion feature The calculation expression is:

[0062] ;

[0063] In the formula, The output shape is Within the same representation space, it achieves lossless complementarity between spatial semantics and k-space frequency information, and serves as the main input branch of the TIM module; texture features Instead of entering the initial integration phase, but with Parallel input into the TIM module avoids high-frequency textures being diluted by global features, achieving joint learning and reconstruction of "structural framework + texture details";

[0064] S42, TIM module periodic modulation and arbitrary real-size reconstruction, calling the global coordinate matrix. With cell characteristics Perform pixel-level coupling;

[0065] S421. Extract frequency control parameters: for fused features Perform two separate 3×3 convolutions (padding=1) to obtain the frequency intensity. and amplitude weight The calculation expressions are as follows:

[0066] ;

[0067] ;

[0068] In the formula, The "frequency intensity map" used to define periodic features maps the 64-channel fused feature map to a 64-channel frequency intensity map with the same number of channels through a 3×3 convolution with padding=1. The output shape is... ; An amplitude map providing "dynamic weights" for periodic features is then expanded into a 128-channel amplitude-weighted map via another 3×3 convolution with padding=1. The output shape is... ;

[0069] S422, Generation and weighting of periodic features;

[0070] S4221, Upsampling: If , frequency intensity and amplitude weight All were obtained by bilinear interpolation to the size of the SR slice. and The spatial dimensions are changed from those of the LR slice. Extended to SR slices ;like If so, then bicubic interpolation is used;

[0071] S4222, Basic Periodic Characteristics: Through Cosine Trigonometric Functions and sine trigonometric functions Even and odd frequency components are captured separately, and the coordinates generated in step 3 are embedded into the global coordinate matrix along with the upsampled frequency intensity. Pixel-wise multiplication yields position-weighted frequency features. These position-weighted frequency features are then combined with cell features. Pixel-by-pixel summation yields a 64-dimensional periodic feature after local-global modulation. The 64-dimensional periodic feature is then subjected to... and The operation generates 64-dimensional even frequency components. and 64 Vic frequency components ,Will and After splicing, a 128-dimensional fundamental periodic feature is obtained. The calculation expressions are as follows:

[0072] ;

[0073] ;

[0074] ;

[0075] In the formula, Used to standardize cycle range; Indicates feature concatenation operation; and All ;

[0076] S4223, Amplitude Weighting: Weighting the upsampled amplitude... With basic cycle characteristics By multiplying elements together, the periodic characteristics of different regions are differentiated and adjusted to enhance the regional periodicity. The calculation expression is as follows:

[0077] ;

[0078] In the formula, The final amplitude-weighted periodic feature is the result of modulating the basic periodic feature with amplitude weights, which retains both frequency domain / position information and incorporates the "strength and weakness weights" of the texture.

[0079] S423. Calculate joint features: Combine texture features After expanding the dimension to 128 through 3×3 convolution and interpolating to the size of the SR slice, we obtain... ;Will and By adding pixels one by one, the "structure + texture" fusion is achieved, outputting a 128-dimensional joint feature. It combines global periodic structure with local texture details, and the calculation expression is:

[0080] ;

[0081] S424, HR Image Reconstruction (Pixel-level Mapping): The 128-dimensional joint features are transformed into a single-channel SR image through "convolutional dimensionality reduction + fully connected mapping" to complete the final reconstruction;

[0082] S4241, Convolutional Dimensionality Reduction: A 3×3 convolutional layer with padding=1 compresses the 128-dimensional joint features to 64 dimensions, resulting in... Feature maps, which reduce subsequent computation while preserving key features, are output.

[0083] S4242, Pixel-wise MLP: For dimensionality reduction The feature map is subjected to two consecutive 1×1 convolutional layers (64→16→1) to independently regress the single-channel grayscale value for each pixel location, achieving a point-to-point mapping from LR features to HR space and outputting a high-resolution image. ,in It is a real number, representing the number of channels, which is 1.

[0084] Step 5: Construct a hybrid loss function and train the network consisting of the MB-CFE module and the TIM module. Please refer to [link / reference]. Figure 2 .

[0085] S51, Through The norm loss function and the perceptual loss function LPIPS are used to construct a hybrid loss function, which is then called hybrid loss. The calculation expression is:

[0086] ;

[0087] In the formula, express Norm loss function, also known as mean absolute error loss; Represents the perceptual loss function;

[0088] The core of the loss function is to calculate the average of the absolute differences between the "model predicted value (super-resolution result)" and the "true label (high-resolution image)". Assume the super-resolution result output by the model is:

[0089] ;

[0090] In the formula, The total number of pixels in the image. For a single pixel value;

[0091] True high-resolution image labels are:

[0092] ;

[0093] In the formula, For each pixel, the corresponding real value of the high-resolution reference image is given.

[0094] The mathematical formula for loss is:

[0095] ;

[0096] The perceptual distance is calculated after extracting features using a pre-trained VGG-16 algorithm; in medical scenarios, structural accuracy is prioritized, therefore... Weight 0.7; LPIPS weight 0.3 is used to balance visual perception quality;

[0097] S52. Set the training strategy: Use the Adam optimizer, with an initial learning rate of 1×10⁻⁶. -4 The model weights were attenuated by 0.5 times at the 40th and 60th epochs; the early stopping tolerance was set to 8 epochs, and the minimum PSNR improvement threshold was 0.05 dB; the model weights and visualization results were saved every 2 epochs.

[0098] S53. Iterative optimization process: Input the low-resolution image and the corresponding high-resolution reference image into the network, obtain the reconstruction result by forward propagation, calculate the hybrid loss and backpropagate; record the batch PSNR and SSIM simultaneously, and accumulate the average epoch value to monitor convergence and trigger early stopping until all training cycles are completed.

[0099] Step 6: Input the low-resolution MRI image to be reconstructed into the trained network, and output a high-resolution reconstructed image.

[0100] Stability verification

[0101] The method of this invention performs double super-resolution reconstruction on the input low-resolution image to obtain a magnified ×2SR image. The ×2SR image is then compared with the corresponding high-resolution HR reference image. (See also...) Figure 7 The ×2SR images are highly consistent with the reference images in maintaining the sharpness of tumor edges and the details of tissue texture, verifying the effectiveness of the invention in real clinical data.

[0102] To systematically evaluate the stability of this invention at arbitrary real magnification, please refer to Table 1, which presents the reconstruction quality indicators of this invention at different magnifications. Within the range of 1 to 30.9×, the PSNR only gradually decreases from 37.01 dB to 34.13 dB, SSIM remains ≥0.955, LPIPS ≤0.071, the MSE increase is less than twofold, and the image exhibits no jagged edges or blockiness. Please refer to... Figure 8 Furthermore, absolute error plots (|SR-HR|) of the same slice at three magnifications: ×2, ×3.5, and ×30.9 are provided; observation Figure 8 It can be seen that the error distribution is uniform and there are no visible artifacts; Figure 8 Together with Table 1, this invention demonstrates that it offers controllable error, structural fidelity, and realistic texture across a wide scale range, providing a highly reliable, retrain-free, plug-and-play practical solution for MRI super-resolution in clinical settings.

[0103] Table 1. Image reconstruction quality indicators of the present invention at different magnifications

[0104]

[0105] Performance comparison verification

[0106] The same case slice from the BraTS 2021 test set was selected and tested under the same conditions as the existing pre-trained model. Please refer to Table 2, which presents the comparison results of PSNR, SSIM, and LPIPS at integer multiples of ×2, ×3, ×4, and ×8. The loss function uses a weighted fusion of L1 and LPIPS losses. At a ×30.9 multiple, only this invention can output data in a single forward pass; other methods require multiple interpolation and splicing, and therefore are not included. The data in Table 2 show that this invention significantly outperforms EDSR and Real-ESRGAN at all comparable multiples, and maintains a PSNR of 34.13 dB even at the extreme 30.9×, validating its wide-scale, high-fidelity clinical practical value.

[0107] Table 2 Comparison of super-resolution performance at different integer multiples on the BraTS 2021 validation set

[0108]

[0109] Network configuration comparison and verification

[0110] In the design of the texture extraction branch, this invention adheres to the principle of "lightweight first": it uses a progressively 1→64-channel U-Net to replace the current mainstream Swin-Transformer, and embeds lesion attention gating. Since the loss function used by Swin-Transformer is generally L1, the L1 loss function is used here for experimentation. Please refer to Table 3. The data in Table 3 shows that in the texture extraction subtask, the lightweight U-Net is sufficient, balancing speed, memory usage, and interpretability, and is superior to Swin-Transformer.

[0111] Please see Figure 9 In Swin, significant bright red spots appear at the lesion edges (easily misinterpreted as false positives), while U-Net shows a uniform yellow color overall, offering superior subjective fidelity. Correspondingly, please refer to Table 4. Objective indicators show that U-Net's PSNR is 0.42–0.8 dB lower than Swin's (average approximately 0.68 dB) at magnifications of ×2 to ×6; the SSIM difference is less than 0.3%; and although the MSE is 9.7%–20.3% higher than Swin's, U-Net achieves superior visual fidelity with less than half the computational cost. Therefore, the U-Net branch provides a preferred lightweight solution that balances efficiency and accuracy for super-resolution MRI at any scale.

[0112] Table 3 Comparison of different network configurations for texture extraction branch

[0113]

[0114] Table 4 Comparison of Objective Indicators for Texture Branch Networks

[0115]

[0116] Loss function optimization effect verification

[0117] The network configuration comparison above has confirmed that, under the same L1 supervision, the lightweight U-Net's PSNR is only about 0.42~0.8 dB lower than Swin-Transformer (average about 0.68 dB), while the number of parameters is reduced by half. To compensate for this slight difference without introducing any additional network parameters, this experiment verifies that while keeping the U-Net skeleton unchanged, the training loss is upgraded from L1 to a joint constraint of 0.7L1 + 0.3LPIPS. See Table 5; the new loss increases PSNR by 0.22~0.26 dB at ×3, ×4, and ×5 ratios, decreases MSE by about 5%, and maintains SSIM above 0.964. Meanwhile, please refer to... Figure 10The error heatmap shows a significant reduction in high-brightness artifacts at the lesion edges, with visual smoothness superior to that of Swing. Thanks to the perceptual constraints of LPIPS, U-Net achieves objective metrics comparable to Swing and superior subjective quality without increasing parameters or latency, verifying that the L1+LPIPS combined loss can provide a dual-excellent solution for super-resolution MRI at any scale, offering both "metrics" and "visual" advantages.

[0118] Table 5. Comparison of reconstruction performance of different loss functions at various magnifications.

[0119]

[0120] Therefore, this invention adopts the above-mentioned frequency domain-texture joint learning arbitrary scale MRI super-resolution reconstruction method, constructs a multi-branch collaborative feature extraction module (MB-CFE), deploys image domain, frequency domain and texture three branches in parallel, and couples TIM implicit upsampling, which can complete 1.1~30.9× arbitrary real number super-resolution in a single forward inference, no longer limited to fixed integer multiples and the limitations of multiple model switching; the TIM module, through periodic coordinate modeling and cell constraints, modulates and then decodes to ensure zero drift of macroscopic structures such as brain regions and tumor boundaries, while independently injecting high-frequency texture, significantly alleviating the "structural smoothness-texture loss" disconnect; the texture branch adopts a progressive channel design of 1 to 64, reducing the number of parameters by more than 40% compared with conventional U-Net, and implicit MLP decoding replaces traditional interpolation, improving the inference speed by no less than 35%, and a single 12 GB GPU can process 512×512×3 slices in real time, meeting the real-time clinical needs; during the training phase, a hybrid loss of 0.7 L1 and 0.3 LPIPS is used to constrain pixel accuracy and perception quality at the same time, and can freely switch between continuous magnifications without retraining.

[0121] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. An arbitrary scale MRI super-resolution reconstruction method of frequency domain-texture joint learning, characterized in that, The method comprises the following steps: Step 1, constructing a data set, dividing the data set into a training set and a test set; Step 2, constructing a collaborative feature extraction module MB-CFE, including an image domain branch, a frequency domain branch and a texture branch, to obtain image domain features, frequency domain features and texture features; Step 3, constructing a texture injection type implicit upsampling module TIM; S31, determining a grid size based on the target super-resolution scale Computing the grid size of the SR slice wherein is an arbitrary real number and , denotes the height of the SR slice, denotes the width of the SR slice, determining a grid size, and generating a normalized global coordinate matrix wherein denotes the horizontal coordinate of a pixel in the SR slice, denotes the vertical coordinate of a pixel in the SR slice; S32, to the global coordinate matrix Performing period mapping, converting infinite coordinates into repeating units; the calculation expression is: ； ； wherein represents a local coordinate offset in the direction of represents a local coordinate offset in the direction of represents a floor function, and ; S33、by The periodic activation function maps the coordinates to a 64-dimensional latent space, obtaining coordinate embedding vectors The calculation expression is: ； In the formula, is a frequency coefficient, ; S34, define cell units: when the cell unit size is 4x4 pixels, when the cell unit size is 8x8 pixels; generate local normalized coordinates inside each cell unit map the local normalized coordinates to cell features through a fully connected layer ; S35. generating the global-local bi-coordinate system by bilinear interpolation extending to the size of the SR slice , the global coordinate matrix generating the complementary architecture of the "global-local" bi-coordinate system; Step 4, element-wise adding the image domain features and the frequency domain features to obtain joint features, and inputting the joint features and the texture features into the TIM module, to complete pixel reconstruction of any real multiple in the period space and output a high-resolution image; Step 5, constructing a hybrid loss function to train the network composed of the MB-CFE module and the TIM module; Step 6, inputting a low-resolution MRI image to be reconstructed into the trained network to output a high-resolution reconstructed image.

2. The method of claim 1, wherein, The specific process of step 1 is as follows: S11, downloading from a multi-modal MRI benchmark library Example registered 3D imaging data, each registered 3D imaging data containing four registered 3D imaging modalities, respectively T1, T1ce, T2 and FLAIR; taking the T2 modality as the super-resolution reconstruction target; the T2 modality contains samples, each sample being three-dimensional data with a dimension of 240x240x155 and a voxel of 1m 3 , isotropic; The sample number and spatial dimension of the T1, T1ce and FLAIR modalities are consistent with those of the T2 modality; S12, each sample in the T2 modality is sequentially split into 155 axial slices of 240x240 in the depth direction, obtaining HR slices; down-sampling is performed by bilinear interpolation, wherein the scaling factor , to generate the corresponding 120x120 LR slices; S13, pairing the 240*240 HR slices and the corresponding 120*120 LR slices one by one to form a one-to-one registered LR-HR two-dimensional slice pair, and obtaining a data set; dividing the data set into a training set and a test set according to a ratio of 8:2, and at the same time, making all slices of the same patient belong to the training set or the test set to prevent data leakage.

3. The method of claim 2, wherein, The specific process of step 2 is as follows: S21, constructing an image domain branch through a residual channel attention network to extract structure-semantic features of the slice under the T2 modality; the image domain branch comprises two convolution layers, four RCANBlock and a global residual; S211, the initial convolutional layer maps the input LR slice to a 64-dimensional feature by 64 3x3 convolutional kernels The feature map keeps the spatial dimension unchanged, extracting the low-level texture and edge information. represents the input low-resolution MRI slice, represents the real number field, represents the height of the slice, represents the width of the slice;​ S212, constructing a deep residual channel attention unit by sequentially stacking four RCANBlock blocks to enhance feature expression step by step; Each RCANBlock block sequentially performs 3*3 convolution, ReLU, 3*3 convolution, adaptive average pooling, two layers of 1*1 convolution and Sigmoid to generate channel weights, multiplies the channel weights with local features channel by channel, and then adds the residual error of the input of the RCANBlock block to output deep features, so as to realize channel-level attention enhancement on key structures in the image; S213, uniformly integrating the deep features output by the four RCANBlock blocks through 3*3 convolution to correct the detail loss in the stacking process and keep the feature dimension and spatial resolution unchanged; S214, introducing global residual connection, element-wise adding the feature map obtained in S211 and the integrated output obtained in S213 to obtain an image domain feature ; S22, frequency domain branch to input LR slice performing two-dimensional fast Fourier transform, extracting features in frequency domain by convolution, returning to image domain by inverse Fourier transform; S221. Slice the input LR. Perform a two-dimensional fast Fourier transform to map the gray intensity distribution in the image domain to k-space data in the frequency domain, and obtain a complex matrix in k-space. S222, stack the real part and the imaginary part of the complex matrix along the channel stack Tensor, the dimension of the tensor is increased to 128 by 128 3x3 convolutions in the real domain, and the dimension of the tensor is increased to 128 by 128 3x3 convolutions in the real domain Tensor, the dimension of the tensor is increased to 128 by 128 3x3 convolutions in the real domain, and the dimension of the tensor is increased to 128 by 128 3x3 convolutions in the real domain S223, inverse Fourier transform is performed on the 64 groups of complex features one by one, and the enhanced frequency spectrum is remapped back to the image domain to obtain the frequency domain features ; S23, the texture branch comprises a differential texture extraction unit, a U-Net network unit and a lesion attention unit; S231, construct a differential texture extraction unit, input the LR slice performing a Gaussian difference by performing a Gaussian blur operation by an adaptive 5x5 Gaussian kernel, and then subtracting the input image pixel by pixel to obtain a texture map, wherein denotes the standard deviation of the Gaussian kernel; the expression is: ； wherein is the output texture, is a Gaussian blur function; S232, constructing a U-Net network unit with a mirror structure of "three-level encoding-three-level decoding"; The logic flow of the encoder is as follows: in the first level encoder, the texture map of is losslessly mapped to by 3*3 convolution with padding=1 and ReLU, and the texture primitives are captured at full resolution; in the second level encoder, the space is compressed to by 3*3 convolution with stride=2 and ReLU, and the channels are expanded to 32, and the output feature dimension is ; in the third level encoder, the size of is kept by 3*3 convolution with padding=1 and ReLU, and the channels are upgraded to 64, and the output feature dimension is ; wherein padding represents the number of pixels filled in the edge area of the input feature map before the convolution operation, and stride represents the number of pixels moved each time the convolution kernel slides on the input feature map. The decoder inversely recovers in a symmetrical manner, and the flow logic is as follows: in the first-level decoder, the features are up-sampled back to the original size through 3*3 transpose convolution and ReLU , the channel is reduced to 32, and the output feature dimension is , the geometry is reconstructed; in the second-level decoder, the channel is compressed to 16 through 3*3 convolution and ReLU, and the edge details are adjusted; in the third-level decoder, the channel is adjusted back to 64 through 3*3 convolution and ReLU, and the texture feature map of is output S233. Enhance texture features through lesion attention units: This involves enhancing the texture features output by the decoder. Using the texture feature map as input, local correlation features are extracted through a 3×3 convolution kernel with padding=1; the local correlation features are compressed into a larger value through adaptive max pooling. The channel-level response vector is generated; the channel weight distribution is adjusted using a 1×1 convolution kernel, and the channel attention weights in the 0-1 interval are generated by combining the Sigmoid activation function; the channel attention weights are multiplied by the channel dimension of the texture feature map output by the decoder to obtain the texture features. .

4. The method of claim 3, wherein, The specific process of step 4 is as follows: S41、the image domain feature and the frequency domain feature element-wise addition to obtain the fusion feature The calculation expression is: ； In the formula, The output shape is ; S42, TIM module periodic modulation and reconstruction of arbitrary real number size, call global coordinate matrix with cell features pixel-level coupling is performed; S421, extracting frequency control parameters; S422, period feature generation and weighting; S423, calculating joint features; S424, HR image reconstruction.

5. The method of claim 4, wherein, In S421: on the fused features Perform two 3x3 convolutions with padding=1 each to get frequency intensity and amplitude weight The expression is calculated as follows: ； 。 6. The arbitrary scale MRI super-resolution reconstruction method of frequency domain-texture joint learning according to claim 5, characterized in that, In S422: S4221, up-sampling: if the frequency intensity and the amplitude weight are bilinearly interpolated to the size of the SR slice, resulting in and the spatial size is expanded from the of the LR slice to the of the SR slice; if tricubic interpolation is used. S4222, base period feature: through cosine trigonometric function and sine trigonometric function Capture even and odd frequency components respectively, embed the coordinates generated in step 3 into the global coordinate matrix and the frequency intensity after upsampling Pixel by pixel multiplication, get position weighted frequency feature, combine position weighted frequency feature and cell feature Pixel by pixel addition, get 64-dimensional period feature after local-global modulation, respectively execute and Operation, generate 64-dimensional even frequency component and 64-dimensional odd frequency component , after splicing and Get 128-dimensional base period feature , the calculation expression is respectively: ； ； ； In the formula, for normalizing the period range; denotes a feature stitching operation; S4223, amplitude weight weighting: the up-sampled amplitude weight with the base period characteristics multiply by element, the expression is calculated as: ； In the formula, represents the period feature after amplitude weighting.

7. The arbitrary scale MRI super-resolution reconstruction method of frequency domain-texture joint learning according to claim 6, characterized in that, In S423: texture features Upsample to 128 dimensions via 3x3 convolution and interpolate to the size of the SR slice to get ; add and pixel by pixel, output 128-dimensional joint features , the calculation expression is: 。 8. The arbitrary scale MRI super-resolution reconstruction method of frequency domain-texture joint learning according to claim 7, characterized in that, In S424: the 128-dimensional joint features are converted into a single-channel SR image through "convolution dimension reduction + full connection mapping", and the reconstruction is completed; S4241、Convolutional Dimension Reduction: 128-dimensional joint features are compressed to 64-dimensional by a 3x3 convolutional layer with padding=1, resulting in feature maps; S4242, Pixel-wise MLP: on the reduced dimensionality The feature map is sequentially executed with two layers of 1x1 convolution, outputting high-resolution images wherein denotes a channel number of 1.

9. The method of claim 8, wherein, The specific process of step 5 is as follows: S51、 by The mixed loss function is constructed by the norm loss function and the perceptual loss function LPIPS. Then the mixed loss The calculation expression of the mixed loss function is: ； In the formula, denotes norm loss function; denotes perception loss function; S52, set training strategy: use Adam optimizer, initial learning rate 1x10 -4 At 40th epoch and 60th epoch, decay by 0.5 times; set early stop patience to 8 epochs, minimum PSNR improvement threshold to 0.05 dB; save model weights and visualization results every 2 epochs; S53、Iterative optimization process: input low-resolution image and corresponding high-resolution reference image into the network, calculate the hybrid loss and back propagation after forward reconstruction result; record batch PSNR and SSIM synchronously, and accumulate epoch average value for monitoring convergence and triggering early stop until completing the whole training period.

Images