Method for super-resolution reconstruction based on low-resolution magnetic resonance images

Abstract

The application discloses a super-resolution reconstruction method based on low-resolution nuclear magnetic resonance images and belongs to the technical field of medical image processing, and comprises the following steps: firstly, high-resolution nuclear magnetic resonance images are input into a data preprocessing module for simulating low-resolution images; secondly, the preprocessed nuclear magnetic resonance images are input into an interpolation upsampling module to adjust the size and resolution of the images to be consistent with those of high-definition images; and then the upsampled low-resolution images are input into a three-dimensional convolutional neural network model based on a residual module, and a feature encoder, a feature decoder and a regression head are combined to realize accurate prediction of the intensity of each voxel of the high-resolution nuclear magnetic resonance images. The application has the beneficial effect that it better adapts to the distribution characteristics of clinical low-definition MRI, effectively processes MRI data with non-fixed axial resolution, and the model can recover the details of the anatomical structure of the brain region and also reconstruct the high-frequency details of the skull and face.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a super-resolution reconstruction method based on low-resolution nuclear magnetic resonance images. Background Technology

[0002] Magnetic Resonance Imaging (MRI) is one of the most commonly used medical imaging techniques in clinical and research settings. Its imaging quality directly affects the accuracy of lesion localization, quantitative assessment, and disease diagnosis. However, due to limitations such as equipment performance, scan time, and patient cooperation, the actual MRI data often exhibits anisotropic structures, meaning that intra-slice resolution is high, while inter-slice resolution is low, and inter-slice resolution is not a fixed parameter. For example, slice thicknesses commonly seen in clinical scans may be 4mm or 6mm, or even greater. This uneven resolution not only affects the detailed presentation of anatomical structures but also results in significant deficiencies in the reconstruction of the overall three-dimensional structure of the skull, thus impacting the effectiveness of MRI-based three-dimensional structural reconstruction and visualization. Figure 1 As shown.

[0003] In recent years, with the continuous development of artificial intelligence and deep learning technologies, deep learning-based image generation and super-resolution algorithms have been introduced into the field of medical image processing. Image super-resolution algorithms can improve the resolution of image data, enhance the visual appeal of low-resolution images, and restore high-frequency details that traditional interpolation methods cannot provide. However, existing methods still have several shortcomings: First, existing methods generally rely on paired low-resolution and high-resolution datasets for supervised training, but such data is difficult to obtain in actual clinical settings; second, existing methods usually assume that axial resolution is a fixed parameter, making it difficult to handle dynamic, non-fixed resolution data in real-world scenarios; third, many methods only focus on the reconstruction of brain regions, neglecting the super-resolution reconstruction of the whole brain and skull structures, resulting in distortions in 3D rendering and facial contour restoration, failing to meet the needs of clinical and research applications for the realism of the entire head structure; finally, existing inventions often ignore contrast degradation in real low-resolution data, such as the contrast degradation between the ventricles and gray matter, leading to significant performance degradation of the model in real-world scenarios. Summary of the Invention

[0004] The purpose of this invention is to provide a super-resolution reconstruction method based on low-resolution MRI images. This method is trained in a self-supervised training framework that does not require paired data, effectively processes MRI data with non-fixed axial resolution, restores the details of brain region anatomical structures, and reconstructs high-frequency details of the skull and face. In the preprocessing stage, it simulates the contrast degradation of real low-resolution data to better adapt to the distribution characteristics of clinical low-resolution MRI.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] The super-resolution reconstruction method based on low-resolution nuclear magnetic resonance images includes the following steps:

[0007] S1. Data preprocessing: High-resolution MRI data is obtained by reading and standardizing the data. Then, the coronal, sagittal or transverse planes are randomly selected as the depth axis. One-dimensional Gaussian low-pass filtering and equal-interval downsampling are applied sequentially along the depth axis. The contrast is then adjusted by γ transformation. Finally, low-resolution MRI data with the same size as the high-resolution MRI data is obtained by interpolation and upsampling. Synthetic low-resolution-real high-resolution training data pairs are constructed.

[0008] S2. Model Training: The training data from step S1 is input into a 3D convolutional neural network model based on residual modules. The 3D convolutional neural network model includes an encoder, a bottleneck layer, a decoder, and a regression head. The encoder is used for feature extraction and downsampling. The bottleneck layer uses a two-layer spatiotemporal feature aggregation module (Video Swin Transformer Block, VSTB) to realize global modeling and fusion of features within and between slices. The decoder realizes feature restoration through upsampling, skip connections, and residual convolution modules. The regression head outputs high-resolution prediction results through 3D convolution. During the training process, a composite loss function is used to constrain the difference between the prediction results and the real high-resolution NMR data.

[0009] S3. Super-resolution reconstruction: Input the low-resolution MRI image to be processed into the trained model and output the corresponding super-resolution reconstruction result.

[0010] Furthermore, in step S1, the volume data is read and normalized using the NIfTI format. The qform / sform affine matrix in the header file is parsed to obtain the anisotropic dimensions and physical coordinate system information of the voxels. Intensity normalization is performed, and the voxel intensities after removing extreme values ​​are truncated. The truncated intensities are then linearly mapped to the [0, 1] interval and converted to 32-bit floating-point numbers, as shown in the following formula: In the formula, I(x) represents the voxel values ​​of the input low-resolution MRI, pα is the α% quantile, pβ is the β% quantile, and Ω is the voxel domain.

[0011] Furthermore, in step S1, the coronal plane, sagittal plane, or cross-section is randomly selected as the depth axis. The coronal plane, sagittal plane, or cross-section is extracted with equal probability. The array axis permutation only changes the data organization and processing direction, without changing the definition and affine relationship of the image physical coordinate system.

[0012] Furthermore, in step S1, the γ transform is applied only to the smoothed and downsampled low-resolution NMR data, and linear stretching and quantile truncation correction are added after the γ transform.

[0013] Furthermore, in step S2, the three-dimensional convolutional neural network model adopts an encoder-decoder symmetric architecture. Each layer of the encoder uses a residual convolution module, which contains two three-dimensional convolution operations. After convolution, the three-dimensional instance normalization and leaky rectified linear unit activation functions are sequentially connected.

[0014] Furthermore, in step S2, VSTB consists of a window partitioning structure, a spatiotemporal feature modeling structure, and a window restoration structure. The input of the window partitioning structure is the feature map output from the last layer of the encoder, and the feature map is divided into multiple independent windows of the same size. The spatiotemporal feature modeling structure consists of paired window self-attention structures and moving window self-attention structures, and global feature fusion is achieved through layer normalization, self-attention calculation, and MLP operation.

[0015] Furthermore, in step S2, the decoder concatenates the feature maps of the corresponding encoder at each level along the feature dimension via skip connections.

[0016] Furthermore, in step S2, the formula for calculating the composite loss function is as follows:

[0017] In the formula, Mean square error, For structural similarity loss, and These are the weighting coefficients. For high-resolution prediction data, Real high-resolution data; the formula for calculating SSIM is: In the formula, for The mean, express The mean, express standard deviation express standard deviation express and The covariances are c1=0.01 and c2=0.09.

[0018] The present invention has the following advantages:

[0019] 1. It is trained under a self-supervised training framework that does not require paired data, effectively processing MRI data with non-fixed axial resolution, and restoring the details of brain region anatomy while also reconstructing high-frequency details of the skull and face.

[0020] 2. By simulating the contrast degradation of real low-resolution data in the preprocessing stage, the model trained by this invention can better adapt to the distribution characteristics of clinical low-resolution MRI, thereby improving the reliability and clinical applicability of the reconstruction results.

[0021] 3. The model has both nonlinear expression capability and three-dimensional volume data feature extraction capability in terms of structure, and performs well in terms of parallel computing efficiency, prediction accuracy and reconstruction robustness. It can effectively restore the detailed information and overall structure of low-resolution nuclear magnetic resonance images. Attached Figure Description

[0022] Figure 1 This is a comparison image of low-resolution MRI, high-resolution MRI, and their 3D rendering effects.

[0023] Figure 2 This is a diagram illustrating the data preprocessing workflow.

[0024] Figure 3 This is a schematic diagram of a three-dimensional convolutional neural network model.

[0025] Figure 4 This is a schematic diagram of the residual convolution module structure.

[0026] Figure 5 This is a schematic diagram of the VSTB structure.

[0027] Figure 6 This is a comparison chart of the brain region volume difference between the method of this invention and the existing SynthSR and the gold standard. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0029] Therefore, the following detailed description of the 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.

[0030] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other.

[0031] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0032] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of this invention is in use, or the orientation or positional relationship commonly understood by those skilled in the art. They are only used for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention. In addition, the terms "first," "second," etc., are only used to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0033] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0034] refer to Figure 2-6 As shown, one embodiment of the present invention is as follows:

[0035] The super-resolution reconstruction method based on low-resolution nuclear magnetic resonance images includes the following steps:

[0036] S1. Data Preprocessing:

[0037] In real-world scenarios, due to the difficulty in obtaining paired low-resolution and high-resolution data, only high-resolution data is often available. Therefore, the core objective of data preprocessing is to generate low-resolution MRI (LR) data that conforms to the characteristics of real clinical scenarios by degrading high-resolution MRI (HR) volume data, constructing a "synthetic low-resolution - real high-resolution" training data pair for model training. Since reconstructing a high-resolution skull contour is one of the important objectives of this invention, and skull dissection is time-consuming, this embodiment does not perform skull dissection during preprocessing, achieving high-resolution skull and facial reconstruction by the model. The MRI dataset used in this embodiment comes from Hefei Fourth People's Hospital and the First Affiliated Hospital of the University of Science and Technology of China, containing 1328 high-resolution MRI cases. All data are in DICOM format and were converted to NIfTI format using the dcm2nii tool for easier subsequent processing. The scans in this dataset were all performed along the coronal plane, and the data shape is uniformly 256×256×188. Before model training, all data is randomly shuffled and divided into training, validation, and test sets in a 7:2:1 ratio, which are used for model training, hyperparameter tuning, and performance evaluation, respectively.

[0038] like Figure 2 As shown, high-resolution NMR volume data is first uniformly read and standardized. The NIfTI format is used to read the volume data, and the qform / sform affine matrices in the header file are parsed to accurately obtain the anisotropic dimensions and physical coordinate system information of the voxels, ensuring the consistency of physical spatial relationships in subsequent processing. To eliminate voxel intensity drift caused by different devices and scanning scenarios, intensity normalization is performed on the original volume data: first, extreme outliers are removed from the data, and quantile truncation is performed using the 0.5% quantile (pα) and 99.5% quantile (pβ) as truncation thresholds. Then, the truncated intensity is linearly mapped to the [0, 1] interval and converted to 32-bit floating-point numbers to ensure the numerical stability of subsequent transformations. The formula is as follows:

[0039] In the formula, I(x) represents the voxel values ​​of the input low-resolution MRI, pα is the α% quantile, pβ is the β% quantile, and Ω is the voxel domain. While ensuring that the affine matrix and voxel spacing metadata are not lost, aberrant voxels such as NaN (non-numeric) and Inf (infinity) in the data are repaired or masked in-situ to ensure the robustness and reproducibility of the preprocessing chain.

[0040] To realistically simulate the imaging geometry of low-resolution anisotropic acquisitions in clinical settings, a depth axis for LR volume data degradation is randomly selected from three views (coronal, sagittal, or transverse) using equal probability sampling. By permuting the array axes, only the data organization and processing direction are changed, without altering the definition and affine relationship of the image's physical coordinate system. This covers different acquisition schemes commonly used in clinical settings, forming a multi-view degradation distribution and improving the model's generalization ability to arbitrary incident angles and anisotropic conditions.

[0041] A one-dimensional separable Gaussian low-pass filter is applied along the selected depth axis to simulate axial blurring caused by increased interslice spacing in clinical thick-slice scanning and to suppress aliasing artifacts during downsampling. The Gaussian kernel function is defined as follows: In the formula, k is the size of the Gaussian kernel. To smooth the standard deviation, t source For the original layer thickness, t target The target layer thickness is [value]. After filtering, the sampling factor r is [value]. d (in this embodiment r) d Random integers or decimals between 4 and 6 (with support for non-integer factor interpolation resampling) are selected to perform equidistant downsampling on the depth axis, simulating the loss of interlayer details caused by increased layer thickness. During this process, the original in-plane resolution and pixel spacing are kept unchanged on the two horizontal axes (in the plane orthogonal to the depth axis), resulting in anisotropic LR volume data that is "sharp in-plane and blurred axially".

[0042] To simulate the tissue contrast degradation (such as reduced contrast between ventricles and gray-white matter) and darker mid-gray levels commonly found in low-resolution clinical MRI data, a γ transform was applied to the smoothed and downsampled LR volume data, defined as: In the formula, I norm For normalized MRI volume data, This represents the low-resolution data after γ-transformation, where γ is the contrast factor, randomly sampled from the interval [1.6, 2.0]. When γ > 1, high gray levels are compressed, and the contrast between gray and white matter is weakened, resulting in smoother tissue edges, consistent with the subjective perception under low-dose, fast-sequence, or thick-layer acquisition. To prevent information loss due to excessive darkening, a slight linear stretching and quantile truncation correction is added after γ-transformation to ensure that the dynamic range of the data still covers the main tissue signals. It is important to note that γ-transformation only applies to LR volume data; HR volume data remains unchanged to enhance the approximation of the true distribution and ensure the stability of the target domain during supervised learning.

[0043] To meet the requirement of aligning the input and output shapes of the subsequent model, a linear interpolation method is used to upsample the axially downsampled LR volume data, aligning its axial layer count with the desired output data. The linear interpolation formula is: In the formula, Let z be the spatial coordinates, and z be the layer index interpolated between z1 and z2. The relationship between z, z1, and z2 is as follows: z1 <z<z2, For point voxel values, For point The voxel values ​​are then calculated. In batch training scenarios, alternatively, the anisotropy of the LR volume data samples can be preserved during preprocessing. The model's interpolation and upsampling module then uses trilinear or B-spline interpolation to restore the data to the same voxel grid size as the HR volume data during training. This preprocessing yields "synthetic low-resolution data" generated from high-resolution data, thus constructing training data pairs. Ultimately, the output samples of the data preprocessing module are all low-resolution-high-resolution data pairs. .

[0044] S2, Model Training:

[0045] The training data from step S1 is input into the residual module-based 3D convolutional neural network model, such as... Figure 3 As shown, the 3D convolutional neural network model is an encoder-decoder symmetrical architecture, including an encoder, bottleneck layer, decoder, and regression head. The 3D convolutional neural network model (3D CNN) serves as the core architecture, with the encoder, decoder, skip connections, and regression head all containing 3D convolutional / residual modules, which take preprocessed LR volume data as input. Output the corresponding high-resolution prediction results and with the original high-resolution image Y HR As a supervisory truth value.

[0046] The encoder, used for feature extraction and downsampling, consists of five layers, each containing a residual convolution module. Internally, each residual convolution module performs two 3D convolution (Conv3D) operations, followed by 3D instance normalization (IN3D) and a Leaky Rectified Linear Unit (LeakyReLU) activation function. The computation flow of the residual convolution module is as follows: Figure 4 As shown. Initial number of channels C 0 =64. With each downsampling level, the feature map side length is halved, and the number of channels doubles. Therefore, the number of channels at each level is 64, 128, 256, 512, and 1024, respectively. The calculation process of the residual module is as follows: In the formula, F l-1 As a feature of the next higher level, W l For the weights of Conv3D, The operator represents a 3D convolution operation. The calculation process for residual connections is as follows: In the formula, x represents the data input to the residual convolution module. This represents the above-mentioned 3D convolution, instance normalization, and nonlinear activation function operations.

[0047] Since MRI volumetric data is a three-dimensional data type composed of several two-dimensional slices, feature fusion between voxels in different regions can significantly affect the super-resolution effect. Traditional bottleneck layers based on three-dimensional convolutional layers are limited by the local receptive field of the convolution operator, and can only aggregate 3... 3 The traditional Transformer-based self-attention architecture, while capable of modeling global feature information in a single step, introduces significant computational overhead. Therefore, this application adds two VSTB layers after the encoder to achieve global modeling and fusion of intra-slice and inter-slice features. This minimizes computational overhead while modeling global feature information, thus reducing the model's computational demands. Unlike traditional MRI super-resolution methods, modeling inter-slice features improves the accuracy of brain tissue recognition. The VSTB's input is the output of the encoder's last layer, and the output is a feature map aggregating intra-slice and inter-slice features. Two identical VSTB layers form the bottleneck layer, employing a self-attention mechanism to model the temporal information of the ultrasound sequence. This extracts the correlation information between sequences from a global perspective, resulting in highly parallel computation and avoiding gradient vanishing or exploding issues, thus better aggregating features from multiple slices.

[0048] This invention employs two VSTB layers at the bottleneck layer of the encoder, which serve only as auxiliary feature enhancement units. These units model and fuse global features within and between slices of the deep convolutional features output by the encoder. Without altering the 3DCNN backbone architecture, this supplements long-range dependency modeling capabilities, improving the continuity and detail accuracy of brain anatomical structures. This is fundamentally different from traditional methods that use Transformers as the backbone of the entire network. The core advantages of the two-layer VSTB are: it retains the ability of 3D CNN to accurately model local anatomical structures, edge details, and voxel intensity distribution in medical images, making it more suitable for MRI super-resolution reconstruction tasks; it uses a lightweight SwinTransformer structure only in the bottleneck layer to enhance features, with a computational cost far lower than the pure Transformer architecture, resulting in faster inference, lower GPU memory usage, and easier clinical deployment; the ability of 3D CNN to preserve local anatomical edges, voxel intensity, and fine structures in MRI is far superior to Transformer on medical datasets with limited scale, avoiding problems such as blurred details and structural distortion in low-resolution, anisotropic MRI; and it can fully preserve high-frequency details of the skull and face shape while maintaining high-precision reconstruction of brain region structures, meeting the needs of clinical 3D visualization.

[0049] VSTB consists of a window partitioning structure, a spatiotemporal feature modeling structure, and a window restoration structure.

[0050] The input to the window partitioning structure is the feature map output from the last layer of the encoder, and the feature map is divided into 16 partitions of size 4. 8 8. Consistent independent windows reduce the complexity of subsequent calculations.

[0051] The spatiotemporal feature modeling structure consists of paired window self-attention structures (3D W-MSA) and moving window self-attention structures (3D SW-MSA). Attention calculation between independent windows is achieved through the moving window self-attention structure, thereby expanding the local receptive field to an approximate global receptive field.

[0052] Each self-attention structure sequentially performs layer normalization (LN), self-attention computation, residual connection, layer normalization (LN), multilayer perceptron (MLP) operation, and residual connection, as follows: Figure 5 As shown. In the window self-attention structure, the input feature map, composed of 16 independent windows, first passes through a layer normalization structure, then enters the 3D W-MSA structure, where self-attention is calculated for each independent window. The feature map after attention calculation is added to the original input feature map to form a residual connection. The feature map is then passed through a layer normalization module and an MLP module to further increase the nonlinearity of the feature map, and then added to the original feature map again to form a residual structure. The 3D W-MSA operation process is as follows:

[0053]

[0054]

[0055]

[0056]

[0057] In the formula, d k Let K be the feature dimension, and K be the feature dimension. T Let K be the transpose matrix, Attn window This is the attention feature map output by an independent window, where Q is the Query matrix, K is the Key matrix, V is the Value matrix, and W is the Key matrix. Q W K W V There are three transformation matrices with different weights. This is a feature map of an independent window. The feature map calculated by the window self-attention module is input into the moving window self-attention module. The process is the same as that of the window self-attention module, but the self-attention operator is replaced by the moving window self-attention structure instead of the window self-attention structure.

[0058] The window restoration structure reassembles the window feature maps after spatiotemporal feature modeling into a complete feature map, and outputs a feature representation that aggregates global information.

[0059] The decoder and encoder have a symmetrical structure with five layers. Each layer doubles the spatial resolution of the feature map and halves the number of channels through upsampling or 3D deconvolution. Simultaneously, skip connections are used to concatenate the features from each layer of the decoder with the corresponding features from the encoder along the feature dimension, as shown in the formula:

[0060] In the formula, U l For the upsampled feature fusion output of the l-th term, F l F represents the l-th layer feature output by the encoder. L-1 The features of the (l-1)th layer output by the encoder. ⊕ indicates an upsampling operation, and ⊕ indicates concatenation along the feature dimension. The concatenated features are also processed by a residual convolution module (Conv3D+IN3D+LeakyReLU) to enhance nonlinear representation capabilities.

[0061] At the end of the decoder, a three-dimensional convolutional regression head is set up to compress the number of feature map channels to 1 and output the predicted high-resolution volume data.

[0062] During training, a composite loss function is used to constrain the difference between the predicted results and the actual high-resolution MRI data. This joint constraint enables the model to simultaneously guarantee pixel-level accuracy and structural similarity, thereby obtaining high-quality MRI super-resolution reconstruction results. The formula for the composite loss function is: In the formula, This refers to the mean squared error (MSE). The loss is based on the Structure Similarity Index Measure (SSIM). and These are the weighting coefficients. For high-resolution prediction data, Real high-resolution data. The formula for calculating SSIM is: In the formula, for The mean, express The mean, express standard deviation express standard deviation express and The covariances are c1=0.01 and c2=0.09.

[0063] S3, Super-resolution Reconstruction:

[0064] The low-resolution MRI image to be processed (must be in NIfTI format and have undergone intensity normalization preprocessing consistent with the training data) is input into the trained model. The encoder extracts features, the bottleneck layer fuses global information, the decoder restores the spatial resolution, and the regression head outputs the prediction results, finally obtaining the super-resolution reconstructed image.

[0065] To restore resolution and anatomical structure, a hybrid architecture employing 3D CNN as the primary method and VSTB lightweight feature enhancement is used. This architecture better preserves local anatomical edges, voxel intensity, and fine structures from MRI on medical datasets with limited scale, outputting continuous voxel intensity maps and exhibiting stronger robustness on low-resolution clinical data. To verify the clinical usability and superiority of this method, the neuroimaging analysis tool FreeSurfer was used to segment the reconstructed data and real high-resolution data into regions of interest (RoIs) for each brain region, and the volume of each brain region was calculated. The results were compared with the existing super-resolution method SynthSR, with the relative error of each brain region volume as the evaluation metric. The results are as follows: Figure 6 As shown in the comparison results, the average volume relative error of the method of the present invention in cortical structures is 12.90%, significantly lower than the 19.32% of the SynthSR method. In key brain regions such as the left and right nuclei accumbens, right amygdala, right hippocampus, left amygdala, left hippocampus, right globus pallidus, right thalamus, left globus pallidus, right putamen, left diencephalon, left putamen, and right diencephalon, the volume relative error of the method of the present invention is smaller than that of the SynthSR method, indicating that the reconstruction results of the present invention are closer to real high-definition data in terms of anatomical structure quantification accuracy. Furthermore, because the present invention does not perform cranial dissection during the preprocessing stage and simulates real contrast degradation through γ-transformation, the reconstruction results can not only accurately restore the details of the anatomical structures of brain regions but also completely reconstruct the high-frequency details of the skull and face. There is no significant distortion in 3D rendering and facial contour restoration, which better meets the needs of clinical and scientific research for the realism of the entire head structure.

[0066] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A super-resolution reconstruction method based on low-resolution nuclear magnetic resonance images, characterized in that, Includes the following steps: S1. Data preprocessing: High-resolution MRI data is obtained by reading and standardizing the data. Then, the coronal, sagittal or transverse planes are randomly selected as the depth axis. One-dimensional Gaussian low-pass filtering and equal-interval downsampling are applied sequentially along the depth axis. The contrast is then adjusted by γ transformation. Finally, low-resolution MRI data with the same size as the high-resolution MRI data is obtained by interpolation and upsampling. Synthetic low-resolution-real high-resolution training data pairs are constructed. S2, Model Training: The training data from step S1 is input into a 3D convolutional neural network model based on residual modules. The 3D convolutional neural network model includes an encoder, a bottleneck layer, a decoder, and a regression head. The encoder is used for feature extraction and downsampling. The bottleneck layer uses a two-layer spatiotemporal feature aggregation module to realize global modeling and fusion of features within and between slices. The decoder realizes feature restoration through upsampling, skip connections, and residual convolution modules. The regression head outputs high-resolution prediction results through 3D convolution. During the training process, a composite loss function is used to constrain the difference between the prediction results and the real high-resolution NMR data. S3. Super-resolution reconstruction: Input the low-resolution MRI image to be processed into the trained model and output the corresponding super-resolution reconstruction result.

2. The super-resolution reconstruction method based on low-resolution nuclear magnetic resonance images according to claim 1, characterized in that, In step S1, the volume data is read and normalized using NIfTI format. The qform / sform affine matrix in the header file is parsed to obtain the anisotropic dimensions and physical coordinate system information of the voxels. Intensity normalization is performed, and the voxel intensities after removing extreme values ​​are truncated. The truncated intensities are then linearly mapped to the [0, 1] interval and converted to 32-bit floating-point numbers, as shown in the following formula: In the formula, I(x) represents the voxel values ​​of the input low-resolution MRI, pα is the α% quantile, pβ is the β% quantile, and Ω is the voxel domain.

3. The super-resolution reconstruction method based on low-resolution nuclear magnetic resonance images according to claim 1, characterized in that, In step S1, the coronal plane, sagittal plane, or cross-section is randomly selected as the depth axis. The coronal plane, sagittal plane, or cross-section is extracted with equal probability. The array axis permutation only changes the data organization and processing direction, without changing the definition and affine relationship of the image physical coordinate system.

4. The super-resolution reconstruction method based on low-resolution nuclear magnetic resonance images according to claim 1, characterized in that, In step S1, the γ transformation is applied only to the smoothed and downsampled low-resolution NMR data, and linear stretching and quantile truncation correction are added after the γ transformation.

5. The super-resolution reconstruction method based on low-resolution nuclear magnetic resonance images according to claim 1, characterized in that, In step S2, the three-dimensional convolutional neural network model adopts an encoder-decoder symmetric architecture. Each layer of the encoder uses a residual convolution module, which contains two three-dimensional convolution operations. After convolution, the three-dimensional instance normalization and leaky rectified linear unit activation functions are sequentially connected.

6. The super-resolution reconstruction method based on low-resolution nuclear magnetic resonance images according to claim 1, characterized in that, In step S2, the spatiotemporal feature aggregation module consists of a window partitioning structure, a spatiotemporal feature modeling structure, and a window restoration structure. The input of the window partitioning structure is the feature map output by the last layer of the encoder, and the feature map is divided into multiple independent windows of the same size. The spatiotemporal feature modeling structure consists of paired window self-attention structures and moving window self-attention structures. Global feature fusion is achieved through layer normalization, self-attention calculation, and MLP operation.

7. The super-resolution reconstruction method based on low-resolution nuclear magnetic resonance images according to claim 5, characterized in that, In step S2, the decoder at each level splices the feature map of the corresponding encoder at the corresponding level along the feature dimension through skip connections.

8. The super-resolution reconstruction method based on low-resolution nuclear magnetic resonance images according to claim 1, characterized in that, In step S2, the formula for calculating the composite loss function is as follows: In the formula, Mean square error, For structural similarity loss, and These are the weighting coefficients. For high-resolution prediction data, Real high-resolution data; The formula for calculating SSIM is: In the formula, for The mean, express The mean, express standard deviation express standard deviation express and The covariances are c1=0.01 and c2=0.09.

Images