Super-resolution reconstruction method and system applied to 0.35 T MRI image

Abstract

The invention provides a super-resolution reconstruction method and system applied to a 0.35 T MRI image, and relates to the field of image processing. The method comprises the following steps: constructing a target training data set based on a FastMRI public data set; according to the target training data set, a lightweight Transform module and a CBAM attention module are adopted to establish and train a super-resolution reconstruction model; a composite loss function is constructed based on the perception loss, the pixel loss and the adversarial loss, and the composite loss function is adopted to optimize the super-resolution reconstruction model; obtaining a 0.35 T MRI image, dividing the 0.35 T MRI image into a plurality of image blocks, and performing recovery processing by using the super-resolution reconstruction model; fusing the restored image blocks by adopting an image fusion algorithm to obtain a target super-resolution image; according to the method, the resolution, details and texture information of the 0.35 T MRI image are remarkably improved, and more reliable image support is provided for clinical diagnosis.

Description

A super-resolution reconstruction method and system for 0.35T MRI images Technical Field

[0001] This invention relates to the field of image processing technology, and specifically to a super-resolution reconstruction method and system for 0.35T MRI images. Background Technology

[0002] Magnetic resonance imaging (MRI), as a core technology in modern medical imaging diagnosis, possesses irreplaceable clinical value in neonatal brain development assessment and early detection of brain injury due to its non-invasive nature, lack of ionizing radiation, and excellent soft tissue resolution. Currently, MRI equipment technology shows a clear trend of field strength differentiation. While high-field systems (≥1.5T) offer good signal-to-noise ratio and anatomical imaging capabilities, their imaging parameters are mostly optimized based on adult standards, making it difficult to fully adapt to the physiological characteristics of neonatal brain tissue, such as high water content and incomplete structural differentiation. This often results in problems such as low gray-white matter contrast and difficulty in structural resolution. In contrast, low-field MRI systems optimized for infants and young children, such as 0.35T and 0.55T, demonstrate significant advantages in practical applications. Their imaging parameters are more adapted to the characteristics of neonatal brain tissue, while avoiding the penetration depth limitations of ultrasound imaging and the ionizing radiation risks associated with CT examinations. They exhibit higher sensitivity and safety, particularly in areas such as gray-white matter boundary imaging and longitudinal brain development tracking studies.

[0003] However, due to the physical limitations of magnetic field strength, 0.35T MRI images generally suffer from quality problems such as low signal-to-noise ratio, insufficient contrast, poor spatial resolution, and significant artifact interference, seriously affecting their practicality and reliability in accurate diagnosis. With the gradual promotion and increasing clinical penetration of dedicated 0.35T pediatric MRI equipment, optimizing the image quality of 0.35T MRI has become one of the important directions in medical imaging research.

[0004] Traditional image upscaling methods, such as those based on K-space zero-filling and bicubic interpolation, can expand image size, but they are often accompanied by problems such as edge blurring, texture attenuation, and structural artifacts, which may even mislead image interpretation. To overcome this bottleneck, super-resolution reconstruction methods centered on image degradation modeling have been gradually developed in recent years. These methods typically introduce prior knowledge and construct optimization objectives based on a clear understanding of the degradation mechanism of low-resolution images, and achieve high-quality image restoration by solving an inverse problem. Among them, sparse representation reconstruction methods use an Lp norm minimization strategy to obtain sparse coding under an overcomplete dictionary, thereby improving image reconstruction quality. However, the reconstruction effect of this method is highly dependent on the construction and generalization ability of the dictionary, and it still suffers from insufficient detail reconstruction and limited resolution improvement in high magnification scenarios.

[0005] In recent years, the rapid development of deep learning technology has provided a new solution for high-quality reconstruction of low-field MRI images, especially Generative Adversarial Networks (GANs), which have demonstrated excellent performance in tasks such as image denoising, artifact suppression, and K-space undersampling completion. Compared with traditional methods, deep learning models have advantages such as end-to-end training, adaptive learning of image degradation features, and fast reconstruction speed, making them an important research direction for low-field MRI image reconstruction. However, despite their great technological potential, deep learning methods still face the following three key challenges in practical applications: 1) The performance of deep learning models largely depends on sufficient and high-quality training data, especially paired high- and low-resolution images or high- and low-field image pairs; in clinical practice, acquiring registered images of the same subject under high-field and low-field MRI systems is costly, and paired data is extremely scarce due to ethical approvals, patient compliance, and differences in imaging standards; low-field MRI generally suffers from low spatial resolution, poor signal-to-noise ratio (SNR), and insufficient contrast due to limited magnetic field strength. 2) Although deep networks possess powerful feature extraction and nonlinear modeling capabilities, they may still encounter problems such as incomplete detail restoration, artifact enhancement, and error amplification when faced with severely degraded original images. Simultaneously, low-quality input data increases the uncertainty of model training, making the model more prone to overfitting and difficult to robustly generalize to complex clinical cases. 3) While many deep reconstruction models in the research have achieved good results on publicly available datasets, their transfer to actual clinical systems still faces many obstacles. On the one hand, the network model parameters are large, resulting in high inference speed and computational resource consumption, which is not conducive to integration into low-power edge devices or rapid imaging workflows. On the other hand, some models lack interpretability, making it difficult to meet the stringent requirements of traceability and security in medical imaging. Furthermore, the current lack of unified evaluation metrics and standardized procedures makes the model's adaptability under multi-center, multi-device, and multi-disease conditions unclear. Summary of the Invention

[0006] The purpose of this invention is to provide a super-resolution reconstruction method and system for 0.35T MRI images. By introducing multi-scale feature extraction, a lightweight Transformer module and a CBAM attention module, and combining perceptual loss, pixel loss and adversarial loss to construct a hybrid optimization target, it can effectively enhance and restore details of low-field images without the need for real high and low field paired data, providing important technical support for its in-depth application in neonatal brain imaging diagnosis.

[0007] To achieve the above objectives, the present invention proposes the following technical solution: Firstly, a super-resolution reconstruction method for 0.35T MRI images is proposed, comprising: constructing a target training dataset based on the FastMRI public dataset, using a composite degradation strategy to process the image center plane of several selected and normalized high-field images; establishing and training a super-resolution reconstruction model using a lightweight Transformer module and a CBAM attention module based on the target training dataset; constructing a composite loss function based on perceptual loss, pixel loss, and adversarial loss, and optimizing the super-resolution reconstruction model using the composite loss function; acquiring 0.35T MRI images, dividing the 0.35T MRI images into several image blocks, and performing super-resolution restoration processing on each image block using the super-resolution reconstruction model; and using an image fusion algorithm to fuse the super-resolution restored image blocks to obtain the target super-resolution image.

[0008] Furthermore, the main body of the super-resolution reconstruction model is an SRCNN neural network model. The structure of the SRCNN neural network model includes a feature extraction layer, a nonlinear mapping layer, and a reconstruction layer. The feature extraction layer introduces a multi-scale feature extraction mechanism, setting convolutional kernels of different scales in parallel branches to expand the receptive field. Within each branch, an equivalent replacement strategy is adopted, using two cascaded 3×3 convolutions to approximate a 5×5 convolution and three cascaded 3×3 convolutions to approximate a 7×7 convolution. The outputs of each branch are concatenated in the channel dimension and added to the residual connection to alleviate gradient vanishing in deep networks and improve expressive power and computational efficiency. A lightweight Transformer module is introduced between the feature extraction layer and the nonlinear mapping layer. The lightweight Transformer module models the global correlation between regions within the image and performs secondary filtering on the features extracted by the feature extraction layer. A CBAM attention module is introduced between the nonlinear mapping layer and the reconstruction layer. The CBAM attention module jointly models channel attention and spatial attention, enabling the SRCNN neural network model to adaptively focus on key regions in the image.

[0009] Furthermore, based on the target training dataset, the training parameters for training the super-resolution reconstruction model are as follows: the target training dataset is divided into training set, validation set, and test set according to a data ratio of 7:2:1; the initial learning rate is set to 1e-4; the super-resolution reconstruction model is trained for 500 epochs, and the learning rate is reduced to half of its original value every 100 epochs; the Adam optimizer is used for training.

[0010] Furthermore, a composite loss function is constructed based on perceptual loss, pixel loss, and adversarial loss, specifically: the perceptual loss... The formula used to measure the difference between the reconstructed image and the real image in perceptual space is as follows: Among them, I SR For the reconstructed image output by the network, I HR For the real image that is paired with it, ∅ l (﹒) represents the feature mapping of the pre-trained feature extraction network at layer l, where C, H, and W are the number of channels, height, and width of layer l; the pixel loss... The formula used to ensure consistency between the reconstructed image and the real image in pixel space is as follows: Where N is the number of image pairs, and i represents the index of the i-th sample pair. and These represent the i-th reconstructed image and its corresponding ground truth image, respectively; the adversarial loss An adversarial learning strategy is used to guide the generator to produce reconstructed images with enhanced realism, as shown in the following formula: The formula for calculating the constructed composite loss function L is as follows: ; where λ1, λ2 and λ3 are the weight functions of each sub-loss.

[0011] Furthermore, the process of constructing the target training dataset based on the FastMRI public dataset and using a composite degradation strategy to process and normalize several high-field images at their image center levels is as follows: at least 2000 original high-field brain images are selected from the FastMRI public dataset, and the image center levels of the high-field images are normalized; a real 0.35T image is introduced as a reference, and downsampling degradation is used to process each normalized high-field image to generate a first simulated image simulating a decrease in spatial resolution; a Gaussian filter with a standard deviation of 0.8~1.2 is used to blur the first simulated image to obtain a second simulated image; additive Gaussian white noise is added to the second simulated image, and frequency truncation is used to simulate the effect of high-frequency loss in k-space to obtain a simulated low-field image; the simulated low-field image is paired with each high-field image to obtain several image pairs of high-field images and corresponding simulated low-field images, and the target training dataset is constructed from the image pairs.

[0012] Furthermore, before dividing the 0.35T MRI image into several image blocks, the method further includes: designing an adaptive filtering method based on directional statistical analysis to perform structure-preserving smoothing on the 0.35T MRI image.

[0013] Furthermore, after image fusion of the image blocks after super-resolution restoration, the process also includes: using an image enhancement method based on Retinex theory to repair the target super-resolution image.

[0014] Secondly, a super-resolution reconstruction system for 0.35T MRI images is proposed, comprising: a screening and construction module for constructing a target training dataset based on the FastMRI public dataset and employing a composite degradation strategy to screen and normalize several high-field images at the image center level; a training module for building and training a super-resolution reconstruction model based on the target training dataset using a lightweight Transformer module and a CBAM attention module; an optimization module for constructing a composite loss function based on perceptual loss, pixel loss, and adversarial loss, and using the composite loss function to optimize the super-resolution reconstruction model; an acquisition and reconstruction module for acquiring 0.35T MRI images, dividing the 0.35T MRI images into several image blocks, and performing super-resolution restoration processing on each image block using the super-resolution reconstruction model; and an image fusion module for performing image fusion on the super-resolution restored image blocks using an image fusion algorithm to obtain the target super-resolution image.

[0015] Furthermore, the super-resolution reconstruction model established by the training module is primarily an SRCNN neural network model. The SRCNN neural network model includes a feature extraction layer, a nonlinear mapping layer, and a reconstruction layer. The feature extraction layer introduces a multi-scale feature extraction mechanism, setting parallel branches of convolutional kernels at different scales to expand the receptive field. Within each branch, an equivalent substitution strategy is employed, using two cascaded 3×3 convolutions to approximate a 5×5 convolution and three cascaded 3×3 convolutions to approximate a 7×7 convolution. The outputs of each branch are concatenated along the channel dimension and added to the residual connections to alleviate gradient vanishing in deep networks and improve expressive power and computational efficiency. A lightweight Transformer module is introduced between the feature extraction layer and the nonlinear mapping layer. This lightweight Transformer module models the global correlation between regions within the image and performs secondary filtering on the features extracted by the feature extraction layer. A CBAM attention module is introduced between the nonlinear mapping layer and the reconstruction layer. This CBAM attention module, by jointly modeling channel attention and spatial attention, enables the SRCNN neural network model to adaptively focus on key regions in the image.

[0016] Thirdly, an electronic device is proposed, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the electronic device enables the above-described super-resolution reconstruction method for 0.35T MRI images.

[0017] As can be seen from the above technical solutions, the technical solutions of the present invention provide beneficial effects: The super-resolution reconstruction method and system for 0.35T MRI images disclosed in the present invention includes the following steps: Based on the FastMRI public dataset, a high-quality image set is first constructed from the public high-field MRI data, and a corresponding simulated low-field image is generated through a composite degradation strategy to form a high-resolution and low-resolution image pair, thus constructing a target training dataset; Based on the target training dataset, a super-resolution reconstruction model is established and trained using a lightweight Transformer module and a CBAM attention module; A composite loss function is constructed based on perceptual loss, pixel loss, and adversarial loss, and the super-resolution reconstruction model is optimized using the composite loss function; 0.35T MRI images are acquired, and the 0.35T images are segmented. MRI images are divided into several image patches and then restored using a super-resolution reconstruction model. An image fusion algorithm is then used to fuse the restored image patches to obtain the target super-resolution image. Specifically, a target training dataset is constructed using this invention to supervise the super-resolution reconstruction model, improving its ability to recover high-quality details from low-quality images. The super-resolution reconstruction model integrates a lightweight Transformer module and a CBAM attention module, thereby enhancing its feature representation capabilities and capturing global dependencies. A composite loss function is used, jointly considering reconstruction error, perceptual loss, and adversarial loss, to optimize the mapping from low-resolution to high-resolution images. Dividing 0.35T MRI images into several image patches not only facilitates processing larger images and reduces memory usage but also enables multi-threaded processing, improving image processing efficiency. This invention achieves a significant improvement in resolution, detail, and texture information of 0.35T MRI images, providing more reliable image support for clinical diagnosis.

[0018] The super-resolution reconstruction method for 0.35T MRI images disclosed in this invention introduces multi-scale feature extraction, a lightweight Transformer module, and a CBAM attention module. It combines perception, pixel, and adversarial loss to construct a hybrid optimization target, achieving effective enhancement and detail restoration of low-field images without requiring real high- and low-field paired data. This provides important technical support for its in-depth application in neonatal brain imaging diagnosis.

[0019] It should be understood that all combinations of the foregoing concepts and the additional concepts described in more detail below can be considered part of the inventive subject matter of this disclosure, provided that such concepts do not contradict each other.

[0020] The foregoing and other aspects, embodiments, and features of the teachings of the present invention will be more fully understood from the following description in conjunction with the accompanying drawings. Other additional aspects of the invention, such as features and / or beneficial effects of exemplary embodiments, will become apparent from the following description or may be learned through practice of specific embodiments according to the teachings of the present invention. Attached Figure Description

[0021] The accompanying drawings are not drawn to scale according to a real reference. In the drawings, each identical or nearly identical component shown in the various figures can be denoted by the same reference numeral. For clarity, not every component is labeled in each figure. Embodiments of various aspects of the invention will now be described by way of example and with reference to the accompanying drawings, in which: Figure 1 is a flowchart of the super-resolution reconstruction method for 0.35T MRI images disclosed in this invention; Figure 2 is a flowchart of constructing the target training dataset disclosed in this invention; Figure 3 is a framework diagram of the super-resolution reconstruction model disclosed in this invention; Figure 4 is a comparison diagram of the super-resolution reconstruction results of 0.35T MRI images using the method of this invention and conventional methods; Figure 5 is a comparison diagram of the super-resolution reconstruction results of 0.35T MRI images using the method of this invention and other deep learning methods; Figure 6 is a super-resolution reconstruction result diagram of other sequence images scanned by the method of this invention on a 0.35T magnetic resonance system; Figure 7 is a structural block diagram of the super-resolution reconstruction system for 0.35T MRI images disclosed in this invention; Figure 8 is a schematic diagram of an electronic device disclosed in an embodiment of this invention. Detailed Implementation

[0022] 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, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used herein should have the ordinary meaning understood by those skilled in the art.

[0023] The terms "first," "second," and similar words used in the specification and claims of this patent application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, unless the context clearly indicates otherwise, the singular forms of "an," "a," or "the," etc., do not indicate a quantity limitation, but rather indicate the presence of at least one. Terms such as "comprising" or "including" mean that the element or object preceding "comprising" encompasses the features, integrals, steps, operations, elements, and / or components listed following "comprising" or "including," and do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or collections thereof.

[0024] The low signal-to-noise ratio, insufficient contrast, poor spatial resolution, and significant artifact interference of 0.35T MRI images severely hinder their application in accurate diagnosis. Existing technologies employ various methods to optimize image quality; however, traditional image magnification methods suffer from edge blurring, texture attenuation, and structural artifacts; super-resolution reconstruction methods have limitations in detail reconstruction and resolution enhancement; and deep learning methods face data constraints and unclear adaptability. Therefore, this invention aims to propose a super-resolution reconstruction method and system for 0.35T MRI images, improving existing super-resolution reconstruction models, significantly enhancing the resolution of low-field MRI images, and identifying image details and texture information.

[0025] The super-resolution reconstruction method and system for 0.35T MRI images disclosed in this invention will be further described in detail below with reference to the specific embodiments shown in the accompanying drawings.

[0026] Referring to Figure 1, the super-resolution reconstruction method for 0.35T MRI images disclosed in this invention includes the following steps: Step S102, based on the FastMRI public dataset, a target training dataset is constructed from the image center plane of several selected and normalized high-field images using a composite degradation strategy; Step S104, based on the target training dataset, a super-resolution reconstruction model is established and trained using a lightweight Transformer module and a CBAM attention module; Step S106, a composite loss function is constructed based on perceptual loss, pixel loss, and adversarial loss, and the super-resolution reconstruction model is optimized using the composite loss function; Step S108, 0.35T MRI images are acquired, divided into several image blocks, and each image block is subjected to super-resolution restoration processing using the super-resolution reconstruction model; typically, 0.35T MRI images are stored in binary format, containing image header information and data portions. During reading, the header information is skipped by locating the data offset, and the image data portion is directly read for color space conversion, normalization, and other processing; the size of each image block is determined by specific parameters, wherein the 0.35T... The MRI image is divided into several image blocks. This is to facilitate the processing of larger images and reduce memory usage, and also to facilitate multi-threaded processing by the processor and improve image processing efficiency. In step S110, an image fusion algorithm is used to fuse the image blocks after super-resolution restoration to obtain the target super-resolution image.

[0027] This invention addresses the super-resolution processing task of 0.35T MRI images, improving upon the SRCNN model. First, a high-quality target training dataset is constructed: high-quality data is selected from the FastMRI public dataset, and the image center layer is normalized and uniformly sized to 256×256. A composite degradation method is used to simulate 0.35T MRI image degradation, and real 0.35T MRI images are introduced for similarity verification and optimization. Based on this dataset, a super-resolution reconstruction model architecture combining a lightweight Transformer module and a CBAM attention module is constructed, denoted as the LFGAN model. The lightweight Transformer module captures global information and long-range dependencies, while the CBAM attention module focuses on key regions through channel-spatial attention, reconstructing the high-resolution output image. A composite loss function is used to optimize the quality of the reconstructed high-resolution image. Ultimately, the method improves the resolution, detail, and texture information of low-field MRI images, providing more reliable image support for brain imaging diagnosis.

[0028] Referring to Figure 2, the process of constructing a target training dataset based on the FastMRI public dataset and employing a composite degradation strategy to normalize the image center layers of several selected and normalized high-field images includes: Step S1021, selecting at least 2000 original high-field brain images from the FastMRI public dataset; to ensure data consistency, the image center layers of the high-field images are selected for normalization; simultaneously, the image size is uniformly set to 256×256; Step S1022, introducing a real 0.35T image as a reference, and using a downsampling degradation method to normalize the images. The processed high-field images are further processed to generate a first simulated image with reduced spatial resolution. The purpose of introducing a real 0.35T image as a reference is to perform similarity checks and optimizations between the simulated and real images at multiple levels, thereby enhancing the realism and medical applicability of the simulated low-field image. The downsampling degradation method involves first interpolating and reducing the image size before restoring it to its original size, resulting in a decrease in simulated spatial resolution. In step S1023, a Gaussian filter with a standard deviation of 0.8~1.2 is used to blur the first simulated image to obtain a second simulated image. The calculation formula is as follows: Formula (1) where I down For the downsampled image, I blur This represents the second simulated image after Gaussian blurring, with a size similar to I. down Consistent size, G σ Represents a standard two-dimensional Gaussian kernel. σ∈[0.8,1.2] is the standard deviation of the kernel, and * represents the convolution operation.

[0029] Step S1024: Additive white Gaussian noise is added to the second simulated image, and frequency truncation is used to simulate the effect of high-frequency loss in k-space, thus obtaining a simulated low-field image. Adding additive white Gaussian noise can be used to simulate more obvious noise interference in low-field systems. The calculation formula is as follows: Formula (2) where, This indicates that the mean is 0 and the standard deviation is 0. Gaussian white noise, I noise This represents a simulated low-field image with added noise.

[0030] Step S1025: Pair the simulated low-field image with each high-field image to obtain several image pairs of high-field images and corresponding simulated low-field images, and construct the target training dataset from the image pairs.

[0031] As shown in Figure 3, the super-resolution reconstruction model established in this invention is an improvement on the SRCNN neural network model. The SRCNN neural network model uses a three-layer convolutional structure to achieve end-to-end mapping from low-resolution images to high-resolution images. Its basic structure includes: a feature extraction layer, a nonlinear mapping layer, and a reconstruction layer. Among them, the feature extraction layer introduces a multi-scale feature extraction mechanism, setting convolutional kernels of different scales in parallel branches to expand the receptive field. An equivalent substitution strategy is adopted within the branches, using two cascaded 3×3 convolutions to approximate a 5×5 convolution and three cascaded 3×3 convolutions to approximate a 7×7 convolution. The outputs of each branch are concatenated in the channel dimension and added to the residual connection to alleviate the gradient vanishing of deep networks and improve expressive power and computational efficiency. In the embodiment, two cascaded 3×3 convolutions are equivalent to a 5×5 convolution and three cascaded 3×3 convolutions are equivalent to a 7×7 convolution. Because the SRCNN neural network model uses a single-scale 3×3 convolution kernel, the receptive field is small and it is difficult to fully capture MRI. The complex tissue structures, boundaries, and texture features in images limit their performance in medical image super-resolution reconstruction; the equivalent replacement strategy improves the receptive field while taking into account network computational efficiency and training stability.

[0032] In low-field MRI images, the low magnetic field strength often leads to problems such as low signal-to-noise ratio, blurred structural edges, and insufficient contrast, resulting in a large amount of long-range dependencies and global structural information that are difficult to model. Traditional convolutional neural networks, due to the limitations of their local receptive fields, struggle to fully capture the global semantic relationships between different regions in the image, thus restricting their performance in low-field image super-resolution reconstruction. To enhance the network's ability to model global dependencies such as long-range structures, edge contexts, and non-local textures in low-field images, a lightweight Transformer module is introduced after the feature extraction layer. This module can effectively model the global correlations between regions within the image and perform secondary filtering on the features extracted by the feature extraction layer, thereby extracting more discriminative feature representations in complex tissues, blurred structures, and noisy backgrounds, thus improving super-resolution reconstruction performance.

[0033] The lightweight Transformer module introduced in this invention includes the following key steps:

[0034] The input 0.35T MRI image is flattened into a sequence according to the channel dimension, that is, a tensor of shape C×H×W is converted into an input sequence of N×C, where N=H×W. Then, the image is projected into a fixed-dimensional embedding space through linear transformation, which serves as the input of a lightweight Transformer module.

[0035] Three independent learnable linear layers are used to transform the input sequence to generate query, key, and value matrices, which are then used for subsequent attention calculations. The formula is as follows: Formula (3) where X is the input sequence, W Q W K W V This is the learnable parameter matrix.

[0036] For the low contrast and low SNR issues present in 0.35T MRI, attention helps to address long-distance dependencies on cross-regional structural associations, such as the cortical-medullic boundary and sulcus orientation. Self-attention mechanisms are used to calculate the correlations between locations in the sequence, enhancing the ability to model long-distance dependencies. Its basic form is scaled dot-product attention (Attention(Q,K,V),) calculated as follows: Formula (4) where d k The dimension of the key vector is used for scaling to stabilize the gradient; the multi-head attention mechanism can capture dependency information from multiple subspaces in parallel, further improving modeling capabilities.

[0037] Two feedforward layers, each containing the non-linear activation function GELU, are applied independently to the features at each location to enhance the model's expressive power; the form is as follows: Formula (5) In this formula, x represents a single location feature that has been preprocessed. W1+b1 is the first linear layer, which raises the dimension of the feature and fully mixes contrast information in the channel dimension, such as gray matter, cerebrospinal fluid and lesions, to provide capacity for image detail and texture enhancement. The nonlinear activation function used is GELU, which is more stable than ReLU under the condition of low field SNR, avoiding the loss of structural information. W2+b2 is the second linear layer, which is used to back-project the reconstructed features into the original dimension and align them with the multi-head attention output, so that subsequent layers can continue to model.

[0038] Additionally, the lightweight Transformer module includes residual connections and layer normalization (LayerNorm) to improve training stability and information flow.

[0039] This lightweight Transformer module serializes 0.35T MRI slices and significantly improves the network's ability to resolve blurred tissues and weak boundary regions in low-field MRI images by modeling the nonlocal structure and long-range dependencies of the images. This provides richer contextual semantic features for subsequent super-resolution reconstruction and exhibits better reconstruction results in scenarios with low signal-to-noise and weakened structures.

[0040] To enhance the model's focus on key anatomical regions and effective contrast, this invention introduces a CBAM attention module between nonlinear mapping and image reconstruction. This module, through joint modeling of channel attention and spatial attention, enables the network to adaptively focus on more critical regions and feature channels in the image. Specifically, the CBAM attention module first extracts channel features through global average pooling and max pooling, and uses a shared fully connected network to generate channel attention weights, enhancing useful feature channels. Subsequently, it calculates a spatial attention map based on pooling features in the spatial dimension, guiding the network to focus on key areas such as lesion areas, corticomedullary boundaries, sulci, and periventricular regions. CBAM enhances the response to important structures in low-field images with extremely low computational overhead, significantly improving boundary sharpness and detail reconstruction under conditions of weak contrast, blur, and ghosting / motion interference, and complements the long-range modeling of the lightweight Transformer.

[0041] The proposed super-resolution reconstruction model employs supervised learning to enhance its ability to recover high-quality details from low-quality images. The super-resolution reconstruction model is run on a 64-bit Windows 10 operating system, configured with 48GB of RAM, an NVIDIA GeForce 5070ti GPU with 16GB of video memory, and an i5 14600KF core processor. The development environment used is Python 3.7 and PyTorch 1.9.0 for building, training, and validating the model. During training, the target training dataset is divided into training, validation, and test sets in a 7:2:1 ratio. The initial learning rate is set to 1e-4, and the model is trained for 500 epochs, with the learning rate halved every 100 epochs. The Adam optimizer is used for training. During specific training, this model is trained in the PyTorch framework and deployed on an embedded platform using NCNN and Vulkan solutions. This allows the use of the parallel computing capabilities of GPUs, which significantly improves image processing speed. Combined with GPU acceleration, efficient inference processing is achieved, with a single image processing time of 20-30ms. This significantly improves the image super-resolution inference speed and meets the real-time imaging and reconstruction requirements of low-field MRI equipment.

[0042] This invention employs the NCNN framework as the deep learning inference engine for image processing tasks. NCNN is a high-efficiency deep learning inference framework designed specifically for mobile terminals and embedded devices, featuring cross-platform compatibility, high performance, and low resource consumption. By introducing the NCNN framework, deep learning models can run efficiently on low-power computing platforms, thereby effectively improving the overall processing efficiency of MRI image super-resolution reconstruction tasks. Furthermore, this invention introduces Vulkan as an accelerated computing interface in the image processing process. Vulkan has low-level access capabilities to GPU resources, fully leveraging hardware computing performance. In this invention, Vulkan is used to accelerate the image processing flow, achieving efficient parallel computing operations and enabling distributed parallel processing even in multi-GPU environments, thereby significantly improving the execution speed and system throughput of image super-resolution tasks.

[0043] In specific implementation, step S106 above constructs a composite loss function based on perceptual loss, pixel loss, and adversarial loss to optimize the super-resolution reconstruction model, thereby further improving the reconstruction quality of low-field MRI images and comprehensively constraining the image's performance in terms of structure, detail, and realism. The specific construction process of the composite loss function is as follows: the perceptual loss... The formula used to measure the difference between the reconstructed image and the real image in perceptual space is as follows: Formula (6) where I SR For the reconstructed image output by the network, I HR For the real image that is paired with it, ∅ l (﹒) represents the feature mapping of the pre-trained feature extraction network at layer l, where C, H, and W are the number of channels, height, and width of layer l; the pixel loss... The formula used to ensure consistency between the reconstructed image and the real image in pixel space is as follows: Formula (7) where N is the number of image pairs, and i represents the index of the i-th sample pair. and These represent the i-th reconstructed image and its corresponding ground truth image, respectively; the adversarial loss An adversarial learning strategy is used to guide the generator to produce reconstructed images with enhanced realism, as shown in the following formula: The formula for calculating the composite loss function L constructed by formula (8) is as follows: Formula (9).

[0044] Wherein, λ1, λ2 and λ3 are the weight functions of each sub-loss; λ1 ensures the consistency between the reconstruction and the high-resolution ground truth in the pixel space; λ2 is used to compensate for insufficient low-field contrast and emphasize the consistency between the corticomedullary boundary and the detailed structure; λ3 suppresses the risk of artifacts caused by adversarial attacks.

[0045] In the embodiments disclosed in this invention, after acquiring the low-field MRI image and before dividing the low-field MRI image into several image blocks, step S108 further includes: designing an adaptive filtering method based on directional statistical analysis to perform structure-preserving smoothing processing on the low-field MRI image to suppress background noise while retaining key tissue boundary information, as shown in the following formula; Formula (10) where I(q) represents the intensity of the input 0.35T MRI image at pixel q; I filtered(p) This represents the smoothing result at pixel p used to suppress low-field noise while preserving the boundaries of the anatomical structure; W dir (p,q) represents the directional similarity weight, which suppresses cross-boundary smoothing and enhances smoothing along the structural orientation based on the dominant direction near p; W int (p,q) represents the pixel intensity similarity weight, measuring the closeness of gray levels between p and q, and is used to suppress smooth transitions caused by abrupt changes in intensity under low field noise; Z P This is the normalization factor.

[0046] Furthermore, after using an image fusion algorithm in step S110 to fuse the image blocks after super-resolution restoration to obtain the target super-resolution image, the method further includes: using an image enhancement method based on Retinex theory to repair the target super-resolution image, so as to enhance the detail expression of the reconstructed image, and focus on repairing overexposed areas or abnormal brightness areas in the image, thereby improving the visual quality and diagnostic reliability of the image.

[0047] Figure 4 shows a comparison of the super-resolution reconstruction results of 0.35T MRI images using the super-resolution reconstruction method of this invention and the traditional method, with a magnification of 4×4. Figures 1a-1d represent the original 0.35T MRI images, 2a-2d represent the reconstruction results using the traditional interpolation method, 3a-3d represent the reconstruction results using the iterative back-projection method, and 4a-4d represent the reconstruction results of this invention. While the traditional interpolation method can expand the image size to some extent, it is essentially a local weighted calculation based on pixel grayscale values ​​and cannot introduce new structural information. Therefore, the reconstruction results often exhibit blurred images, severely smoothed textures, and significant attenuation of edge details, showing significant deficiencies in both structural accuracy and perceptual quality. The iterative back-projection method can compensate for some detail loss to a certain extent by modeling the image degradation process and repeatedly updating the reconstructed image. Compared with traditional interpolation methods, the results of this method have clearer edge contours and more natural texture distribution. However, at high magnification, there are still problems such as slow convergence speed, limited detail recovery and some artifact residue. The super-resolution reconstruction method of this invention is superior to other methods in terms of overall visual quality, edge preservation ability and structural restoration.

[0048] Figure 5 shows a comparison of the super-resolution reconstruction results of 0.35T MRI images using the super-resolution reconstruction method of this invention and commonly used deep learning methods, with a magnification of 4×4. Figures 1A-1D represent the original 0.35T MRI images, 2A-2D represent the reconstruction results using the SRCNN model, 3A-3D represent the reconstruction results using the VDSR model, and 4A-4D represent the reconstruction results of this invention. The results show that the super-resolution reconstruction model trained by this invention outperforms the other two models in terms of detail recovery and noise reduction in low-field MRI images.

[0049] As shown in Figure 6, t1~t3 are the original images of the T2W_FSE, T1W_FLASH, and T2_FLAIR sequences obtained from a 0.35T MRI system scan, while T1~T3 are the image reconstruction results of the T2W_FSE, T1W_FLASH, and T2_FLAIR sequences obtained from a 0.35T MRI system scan according to the present invention. The results show that the super-resolution reconstruction method for 0.35T MRI images disclosed in this invention is also applicable to the super-resolution reconstruction of other sequence images obtained from a 0.35T MRI system scan.

[0050] Based on the same inventive concept as the above-described method embodiments, this application also provides a super-resolution reconstruction system applied to 0.35T MRI images. Figure 7 shows a schematic diagram of the super-resolution reconstruction system framework. As shown in Figure 7, the super-resolution reconstruction system includes the following modules: a screening and construction module, used to construct a target training dataset based on the FastMRI public dataset and employing a composite degradation strategy to screen and normalize several high-field images at the image center level; a training module, used to establish and train a super-resolution reconstruction model based on the target training dataset using a lightweight Transformer module and a CBAM attention module; an optimization module, used to construct a composite loss function based on perceptual loss, pixel loss, and adversarial loss, and use the composite loss function to optimize the super-resolution reconstruction model; an acquisition and reconstruction module, used to acquire low-field MRI images, divide the low-field MRI images into several image blocks, and perform super-resolution restoration processing on each image block using the super-resolution reconstruction model; and an image fusion module, used to perform image fusion on the super-resolution restored image blocks using an image fusion algorithm to obtain the target super-resolution image.

[0051] The steps of the system for implementing the super-resolution reconstruction method for 0.35T MRI images disclosed in the above embodiments have already been described and will not be repeated here.

[0052] For example, the main body of the super-resolution reconstruction model established by the training module is an SRCNN neural network model. The structure of the SRCNN neural network model includes a feature extraction layer, a nonlinear mapping layer, and a reconstruction layer. The feature extraction layer introduces a multi-scale feature extraction mechanism to expand the receptive field. A lightweight Transformer module is introduced between the feature extraction layer and the nonlinear mapping layer. The lightweight Transformer module models the global correlation between different regions within the image and performs secondary filtering on the features extracted by the feature extraction layer. A CBAM attention module is introduced between the nonlinear mapping layer and the reconstruction layer. The CBAM attention module enables the SRCNN neural network model to adaptively focus on key regions in the image by jointly modeling channel attention and spatial attention.

[0053] The training parameters for the super-resolution reconstruction model are as follows: the target training dataset is divided into training, validation, and test sets according to a data ratio of 7:2:1; the initial learning rate is set to 1e-4; the super-resolution reconstruction model is trained for 500 epochs, and the learning rate is reduced to half its original value every 100 epochs; the Adam optimizer is used for training.

[0054] For example, the screening and construction module, based on the FastMRI public dataset, is an execution unit that uses a composite degradation strategy to construct the target training dataset from the image center level of several high-field images that have been screened and normalized. This includes: a selection unit, used to screen at least 2000 original high-field brain images from the FastMRI public dataset and perform normalization on the image center level of the selected high-field images; a generation unit, used to introduce real 0.35T images as references and process each normalized high-field image using a downsampling degradation method to generate a first simulated image simulating spatial resolution reduction; a blurring unit, used to blur the first simulated image using a Gaussian filter with a standard deviation of 0.8-1.2 to obtain a second simulated image; a noise addition unit, used to add additive Gaussian white noise to the second simulated image and simultaneously use frequency truncation to simulate the effect of high-frequency loss in k-space to obtain a simulated low-field image; and a pairing unit, used to pair the simulated 0.35T images with each high-field image to obtain several image pairs of high-field images and corresponding simulated 0.35T images, and to construct the target training dataset from these image pairs.

[0055] For example, the super-resolution reconstruction system for 0.35T MRI images disclosed in this invention further includes, before the reconstruction module divides the 0.35T MRI image into several image blocks, the following steps: designing an adaptive filtering method based on directional statistical analysis to perform structure-preserving smoothing processing on the 0.35T MRI image. Additionally, after the image fusion module obtains the target super-resolution image, an image enhancement method based on Retinex theory is used to repair the target super-resolution image, enhance the detail expression of the reconstructed image, and focus on repairing overexposed or abnormally bright areas in the image.

[0056] In summary, this invention addresses the super-resolution processing task of 0.35T MRI images by improving the SRCNN model to obtain the LFGAN model. Compared with existing technologies, this model not only has better overall visual quality, edge preservation ability, and structural restoration accuracy for 0.35T low-field images, but also has stronger image detail recovery capabilities. Furthermore, this invention can also be applied to the super-resolution reconstruction of other sequence images obtained from 0.35T MRI system scans, thus having a wider range of applications and greater practicality.

[0057] Based on the same inventive concept as the above method embodiments, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the electronic device implements the super-resolution reconstruction method for 0.35T MRI images described in the above embodiments.

[0058] In one embodiment, the electronic device may be a server. In this embodiment, the structure of the electronic device may be as shown in FIG8, including a memory 201, a communication module 203, and one or more processors 202.

[0059] The memory 201 is used to store computer programs executed by the processor 202. The memory 201 may mainly include a program storage area and a data storage area. The program storage area may store the operating system and programs required to run instant messaging functions, etc.; the data storage area may store various instant messaging information and operation instruction sets, etc.

[0060] Memory 201 may be volatile memory, such as random-access memory (RAM); memory 201 may also be non-volatile memory, such as read-only memory, flash memory, hard disk drive (HDD), or solid-state drive (SSD); or memory 201 may be any other medium capable of carrying or storing a desired computer program having the form of instructions or data structures and accessible by a computer, but is not limited thereto. Memory 201 may be a combination of the above-described memories.

[0061] The processor 202 may include one or more central processing units (CPUs) or digital processing units, etc. The processor 202 is used to implement the aforementioned audio data processing method when calling a computer program stored in the memory 201.

[0062] The communication module 203 is used to communicate with terminal devices and other servers.

[0063] This application embodiment does not limit the specific connection medium between the memory 201, communication module 203, and processor 202. In Figure 8, the memory 201 and processor 202 are connected via a bus 204, which is depicted by arrows. The connection methods between other components are merely illustrative and not intended to be limiting. The bus 204 can be an address bus, data bus, control bus, etc. For ease of description, only one arrow is used in Figure 8, but this does not imply that there is only one bus or one type of bus.

[0064] Based on the same inventive concept as the above-described method embodiments, embodiments of the present invention also provide a computer-readable storage medium for storing a computer program. When the computer program is run on a computer, it enables the electronic device to implement the super-resolution reconstruction method for 0.35T MRI images described in the above embodiments. The computer-readable storage medium can be a readable signal medium or a readable storage medium. A readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a magnetic disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0065] Based on the same inventive concept as the above-described method embodiments, embodiments of the present invention also provide a computer program product comprising a computer program that, when run on an electronic device, causes the electronic device to perform the steps of the super-resolution reconstruction method for 0.35T MRI images according to various exemplary embodiments of this application described above. The program product may take the form of any combination of one or more readable media. These computer program commands may be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the commands executed by the processor of the computer or other programmable data processing device produce means for implementing the functions specified in one or more flowcharts and / or one or more block diagrams.

[0066] While the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Those skilled in the art can make various modifications and refinements without departing from the spirit and scope of the invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A super-resolution reconstruction method for 0.35T MRI images, characterized in that, include: Based on the FastMRI public dataset, a target training dataset was constructed by using a composite degradation strategy to analyze the image center plane of several selected and normalized high-field images. Based on the target training dataset, a super-resolution reconstruction model was established and trained using a lightweight Transformer module and a CBAM attention module. A composite loss function was constructed based on perceptual loss, pixel loss, and adversarial loss, and the super-resolution reconstruction model was optimized using the composite loss function. 0.35T MRI images were acquired, divided into several image patches, and super-resolution reconstruction was performed on each image patch using the super-resolution reconstruction model. An image fusion algorithm was used to fuse the super-resolution restored image patches to obtain the target super-resolution image.

2. The super-resolution reconstruction method for 0.35T MRI images according to claim 1, characterized in that, The super-resolution reconstruction model is based on an SRCNN neural network model, which includes a feature extraction layer, a nonlinear mapping layer, and a reconstruction layer. The feature extraction layer introduces a multi-scale feature extraction mechanism, setting parallel branches of convolutional kernels at different scales to expand the receptive field. Within each branch, an equivalent substitution strategy is used, approximating a 5×5 convolution with two cascaded 3×3 convolutions and approximating a 7×7 convolution with three cascaded 3×3 convolutions. The outputs of each branch are concatenated along the channel dimension and added to the residual connections to alleviate gradient vanishing in deep networks and improve expressive power and computational efficiency. A lightweight Transformer module is introduced between the feature extraction layer and the nonlinear mapping layer. This module models the global correlation between regions within the image and performs secondary filtering on the features extracted by the feature extraction layer. A CBAM attention module is introduced between the nonlinear mapping layer and the reconstruction layer. This module jointly models channel attention and spatial attention, enabling the SRCNN neural network model to adaptively focus on key regions in the image.

3. The super-resolution reconstruction method for 0.35T MRI images according to claim 2, characterized in that, Based on the target training dataset, the training parameters for training the super-resolution reconstruction model are as follows: the target training dataset is divided into training set, validation set and test set according to the data ratio of 7:2:1; the initial learning rate is set to 1e-4; the super-resolution reconstruction model is trained for 500 epochs, and the learning rate is reduced to half of the original value every 100 epochs; the Adam optimizer is used for training.

4. The super-resolution reconstruction method for 0.35T MRI images according to claim 1, characterized in that, A composite loss function is constructed based on perceptual loss, pixel loss, and adversarial loss, specifically: the perceptual loss... The formula used to measure the difference between the reconstructed image and the real image in perceptual space is as follows: Among them, I SR For the reconstructed image output by the network, I HR For the real image that is paired with it, ∅ l (﹒) represents the feature mapping of the pre-trained feature extraction network at layer l, where C, H, and W are the number of channels, height, and width of layer l; the pixel loss... The formula used to ensure consistency between the reconstructed image and the real image in pixel space is as follows: Where N is the number of image pairs, and i represents the index of the i-th sample pair. and These represent the i-th reconstructed image and its corresponding ground truth image, respectively; the adversarial loss An adversarial learning strategy is used to guide the generator to produce reconstructed images with enhanced realism, as shown in the following formula: The formula for calculating the constructed composite loss function L is as follows: ; where λ1, λ2 and λ3 are the weight functions of each sub-loss.

5. The super-resolution reconstruction method for 0.35T MRI images according to claim 1, characterized in that, The process of constructing a target training dataset based on the FastMRI public dataset and employing a composite degradation strategy to normalize and filter several high-field images at their image center levels is as follows: At least 2000 original high-field brain images are selected from the FastMRI public dataset; the image center levels of these high-field images are then normalized; real 0.35T images are introduced as a reference, and downsampling degradation is used to process each normalized high-field image to generate a first simulated image simulating a decrease in spatial resolution; a Gaussian filter with a standard deviation of 0.8–1.2 is used to blur the first simulated image to obtain a second simulated image; additive Gaussian white noise is added to the second simulated image, and frequency truncation is used to simulate the effect of high-frequency loss in k-space to obtain a simulated low-field image; the simulated low-field image is paired with each high-field image to obtain several image pairs of high-field images and corresponding simulated low-field images, and the target training dataset is constructed from these image pairs.

6. The super-resolution reconstruction method for 0.35T MRI images according to claim 1, characterized in that, Before dividing the 0.35T MRI image into several image blocks, the method further includes: designing an adaptive filtering method based on directional statistical analysis to perform structure-preserving smoothing on the 0.35T MRI image.

7. The super-resolution reconstruction method for 0.35T MRI images according to claim 1, characterized in that, After image fusion of the image blocks after super-resolution restoration, the process also includes: using an image enhancement method based on Retinex theory to repair the target super-resolution image.

8. A super-resolution reconstruction system for 0.35T MRI images, characterized in that, include: A screening and construction module is used to construct a target training dataset based on the FastMRI public dataset, using a composite degradation strategy to screen and normalize several high-field images at the image center level. The system comprises the following modules: a training module for building and training a super-resolution reconstruction model using a lightweight Transformer module and a CBAM attention module based on the target training dataset; an optimization module for constructing a composite loss function based on perceptual loss, pixel loss, and adversarial loss, and using this composite loss function to optimize the super-resolution reconstruction model; an acquisition and reconstruction module for acquiring 0.35T MRI images, dividing them into several image patches, and performing super-resolution restoration processing on each image patch using the super-resolution reconstruction model; and an image fusion module for fusing the super-resolution restored image patches using an image fusion algorithm to obtain the target super-resolution image.

9. The super-resolution reconstruction system for 0.35T MRI images according to claim 8, characterized in that, The super-resolution reconstruction model established by the training module is based on an SRCNN neural network model. The SRCNN model comprises a feature extraction layer, a nonlinear mapping layer, and a reconstruction layer. The feature extraction layer introduces a multi-scale feature extraction mechanism, setting parallel branches of convolutional kernels at different scales to expand the receptive field. Within each branch, an equivalent substitution strategy is employed, using two cascaded 3×3 convolutions to approximate a 5×5 convolution and three cascaded 3×3 convolutions to approximate a 7×7 convolution. The outputs of each branch are concatenated along the channel dimension and added to the residual connections to alleviate gradient vanishing in deep networks and improve expressive power and computational efficiency. A lightweight Transformer module is introduced between the feature extraction layer and the nonlinear mapping layer. This module models the global correlation between regions within the image and performs secondary filtering on the features extracted by the feature extraction layer. A CBAM attention module is introduced between the nonlinear mapping layer and the reconstruction layer. This module jointly models channel attention and spatial attention, enabling the SRCNN neural network model to adaptively focus on key regions in the image.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it causes the electronic device to implement the super-resolution reconstruction method for 0.35T MRI images as described in any one of claims 1 to 7.

Images