A magnetic resonance image enhancement method based on noise utilization and generative adversarial network

Abstract

The application discloses a magnetic resonance image enhancement method based on noise utilization and a generative adversarial network, and belongs to the technical field of medical image processing. The method comprises the following steps: acquiring paired 3T and 7T magnetic resonance images and preprocessing to obtain 3T MRI image blocks; a three-dimensional generative adversarial network with a fusion noise utilization mechanism is constructed, the generator is based on V-net construction and fuses a feature-level adaptive noise injection mechanism and a 3D CBAM attention module, and the discriminator is a four-stage hierarchical 3D convolutional network; a total loss function composed of pixel, perception, gradient structure consistency and adversarial loss is designed; the network is trained in an alternating update mode; and a 7T-level high-quality image is output by inputting a 3T image to be processed into the trained generator. The application effectively realizes high-quality enhancement from low-field to high-field magnetic resonance images, significantly improves the signal-to-noise ratio and resolution of the images, and is suitable for clinical auxiliary diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing technology, specifically relating to a magnetic resonance image enhancement method based on noise utilization and generative adversarial networks. Background Technology

[0002] Magnetic Resonance Imaging (MRI), a non-invasive medical imaging technique, has been widely used for detecting changes in brain anatomy and function over the past 40 years. The magnetic field strength of clinical MRI has evolved from less than 0.5T initially to the widely used 3.0T, and the increasingly prevalent 7T. 7T high-field MRI, with its ultra-high resolution and excellent contrast-to-noise ratio (CNR) and signal-to-noise ratio (SNR), can clearly reveal anatomical details of the hippocampus subregion, cortical layered microstructures, and small lesions, providing a new perspective for precise imaging of neurological diseases such as epilepsy and Alzheimer's disease. Studies have shown that compared to the widely used 3T MRI, 7T MRI, after post-processing with a morphometric analysis program (MAP), can detect an additional 16.7% of epileptic lesions in patients who were negative on traditional 3T MRI.

[0003] However, the clinical adoption of 7T MRI is limited by high costs and the limited number of devices available globally, hindering its widespread use in medical institutions at all levels. Therefore, augmenting widely used 3T MRI images to 7T using computational techniques has become a necessary alternative and a significant challenge in the field of medical image processing.

[0004] Generative Adversarial Networks (GANs), an important branch of deep learning, generate data through a game between generative and discriminative models. Their variants, such as Conditional GANs, Cycle GANs, and 3D GANs, have shown great potential in medical image enhancement and cross-domain translation. However, existing GANs and their variants still cannot meet the stringent requirements of high-fidelity anatomical structures in medical image tasks: on the one hand, they are prone to ambiguity and are sensitive to input data; on the other hand, training models on two-dimensional slices ignores the spatial correlation between voxels, leading to tomographic artifacts in the coronal or sagittal planes of the reconstructed images, and the problem of maintaining spatial continuity in three-dimensional medical images has not yet been well resolved.

[0005] Noise injection is a classic and commonly used technique in machine learning, playing a positive role in enhancing model generalization ability, improving performance, and accelerating convergence, especially when training samples are limited or contaminated with noise. Its beneficial effects can be explained by the stochastic resonance effect. MRI data typically contains unavoidable noise such as Gaussian noise, speckle noise, Rayleigh noise, and motion artifacts. Theoretically, noise injection strategies have good adaptability. However, the effectiveness and application of noise injection in GAN-based 3T MRI to 7T MRI enhancement tasks have not been fully explored.

[0006] Existing deep learning-based 3T to 7T MRI conversion methods can be mainly divided into three categories: 7T-guided super-resolution methods, diffusion model-based image conversion methods, and GAN-based methods. Among them, although GAN-based methods are widely used, they still have shortcomings in areas such as fine reconstruction of anatomical structures and enhancement of small lesion features, making it difficult to meet the requirements of clinical diagnosis for image structural integrity and detail realism.

[0007] Therefore, there is an urgent need for a 3T MRI to 7T MRI conversion method that can balance spatial continuity, anatomical structure fidelity, and generation efficiency. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to address the shortcomings of the prior art by providing a magnetic resonance image enhancement method based on noise utilization and generative adversarial networks. This method is used to solve the technical problems of low structural fidelity and poor spatial continuity in the generation of 7T MRI images in the 3T to 7T magnetic resonance image (MRI) conversion task, and to achieve high-fidelity conversion from 3T MRI to 7T MRI, thereby providing high-field MRI images with diagnostic value for clinical use.

[0009] The present invention adopts the following technical solution: A magnetic resonance image enhancement method based on noise exploitation and generative adversarial networks includes the following steps: S1. Acquire paired 3T and 7T MRI images and perform preprocessing operations to obtain preprocessed 3T MRI image blocks; S2. Construct a three-dimensional generative adversarial network that integrates a noise utilization mechanism. The three-dimensional generative adversarial network includes a generator and a multi-scale discriminator. The generator is constructed based on the V-net network and integrates a feature-level adaptive noise injection mechanism and a 3D CBAM attention module. The multi-scale discriminator is a four-stage hierarchical 3D convolutional network structure. S3. Design a total loss function formed by a weighted linear combination of pixel loss function, perceptual loss function, gradient structure consistency loss function and generative adversarial loss function; S4. The generator and multi-scale discriminator are trained adversarially using an alternating update method, and the network parameters of the three-dimensional generative adversarial network are optimized through the total loss function. S5. Input the 3T MRI image to be processed into the trained generator and output the corresponding 7T high-quality magnetic resonance image.

[0010] Preferably, in step S1, the preprocessing operation includes the following steps in sequence: S101. The N4 off-field correction algorithm is used to perform off-field correction on 3T and 7T MRI images to eliminate intensity inhomogeneity; S102. Perform skull dissection on the off-field rectified image and remove non-brain tissue; S103. Spatially align the 3T image with its corresponding 7T image using affine transformation; S104. Perform histogram matching on the 3T and 7T images respectively to make the image intensity distribution consistent; S105. Linearly normalize the voxel intensity of the image and scale the voxel intensity to the [-1,1] interval. S106. Cut the paired 3T and 7T images into three-dimensional blocks of a preset size to obtain preprocessed 3T MRI image blocks.

[0011] Preferably, in step S2, the generator construction process includes: S201. The single-channel 3TMRI image block is expanded into a multi-channel feature map through the initial convolutional layer. At the same time, residual connections are added, and zero-mean Gaussian noise is injected into the multi-channel feature map to achieve feature-level adaptive noise injection. S202. The spatial resolution of the feature map after injecting noise is gradually reduced by the downsampling residual block of the encoder, while the number of channels is increased exponentially. S203. Embed a 3DCBAM attention module at the jump connection between the encoder end and the decoder to perform channel and spatial dimension attention weighting on the feature map. S204. The spatial resolution of the feature map is gradually improved by the upsampling residual block of the decoder, while the number of channels is reduced by a factor of two. During the upsampling process, the feature map of the corresponding layer of the encoder is stitched together. S205. After upsampling the feature map to its original size, a single-channel output is generated through a convolutional layer, and the Tanh activation function is used to restrict the pixel values ​​to the range of [-1, 1].

[0012] Preferably, in step S201, the specific process of feature-level adaptive noise injection is as follows: The input single-channel 3D MRI image tensor is input into the first layer feature extraction module of the generator, and the initial feature tensor is obtained through feature extraction. Zero-mean Gaussian noise is injected into the initial feature tensor to obtain the feature tensor after noise injection. The standard deviation of the Gaussian noise is the mapping result of the learnable parameters after being monotonically non-negatively constrained by the Softplus function, so that the noise intensity is adaptively adjusted as the model is trained.

[0013] Preferably, the encoder includes three downsampled residual blocks, each of which consists of a convolution operation, batch normalization, and a LeakyReLU activation function. The stride of the convolution operation is 2, and the number of feature map channels is 16, 32, 64, and 128, respectively. The decoder contains three upsampled residual blocks. Each upsampled residual block first doubles the feature map size and halves the number of channels through upsampling, and then concatenates it with the feature map of the corresponding layer of the encoder.

[0014] Preferably, the 3DCBAM attention module includes a channel attention module (CAM) and a spatial attention module (SAM). The feature map is processed sequentially by the CAM module and the SAM module, specifically as follows: After the feature map is input into the CAM module, it is processed by global average pooling and global max pooling to obtain two feature vectors. The two feature vectors are then input into an MLP structure with shared weights. The channel attention weights are obtained by passing the sigmoid activation function. Finally, the original feature map is multiplied and weighted element-wise to obtain the output of the CAM module. After the output of the CAM module is input into the SAM module, channel-dimensional pooling is first performed to obtain two feature maps. The two feature maps are concatenated and processed by a convolutional kernel and activated by a sigmoid function to obtain spatial attention weights. Then, the output of the CAM module is multiplied element-wise and weighted to obtain the final output of the 3DCBAM attention module.

[0015] Preferably, in step S2, the construction process of the multi-scale discriminator is as follows: The input layer of the multi-scale discriminator receives real 7TMRI images and generated 7TMRI images output by the generator, and inputs them into the feature extraction layer after standardization. The feature extraction layer adopts a four-stage downsampling convolution structure. Each stage consists of a three-dimensional convolution block, which includes convolution operation, batch normalization and LeakyReLU activation function. The kernel size of the first to third stages is 7×7×7, the kernel size of the fourth stage is 5×5×5, the convolution stride is 2, and the padding method is adaptive 3D padding. Each predictive convolutional module contains a 3D copy padding layer and a 3D convolutional layer, outputting the image authenticity judgment result at the corresponding scale.

[0016] Preferably, in step S3, the total loss function is:

[0017] in, , , , For hyperparameters, To counter the loss function, For pixel loss function, For the perceptual loss function, This is the gradient structure consistency loss function.

[0018] Preferably, in step S4, the specific parameters of the adversarial training are: The network model was built based on the PyTorch framework, with a total training duration of 400 epochs and a batch size of 10. The network weights were randomly initialized. The Adam optimizer was used for network optimization, with an initial learning rate of 0.0001. The initial learning rate was kept constant for the first 150 epochs, and then linearly decayed to zero for the remaining 250 epochs. The hyperparameters of the pixel loss function, perceptual loss function, gradient structure consistency loss function, and generative adversarial loss function in the total loss function were 1.0, 0.1, 1.0, and 0.1, respectively, and the noise intensity was set to 0.1. The gradient structure consistency loss function was calculated by using a 3DSobel convolution kernel to calculate the voxel gradients of the 3D image volume in the axial, sagittal, and coronal planes to obtain the gradient mapping of the image. This gradient mapping was then constructed by minimizing the distance between the gradient mappings of the generated 7TMRI image and the real 7TMRI image.

[0019] Secondly, embodiments of the present invention provide a magnetic resonance image enhancement system based on noise exploitation and generative adversarial networks, comprising: The data module is used to acquire pairs of 3T and 7T magnetic resonance images and perform preprocessing operations to obtain preprocessed 3T MRI image blocks; The network module is used to construct a three-dimensional generative adversarial network that integrates a noise utilization mechanism. The three-dimensional generative adversarial network includes a generator and a multi-scale discriminator. The generator is constructed based on the V-net network and integrates a feature-level adaptive noise injection mechanism and a 3D CBAM attention module. The multi-scale discriminator is a four-stage hierarchical 3D convolutional network structure. The function module is used to design the total loss function, which is a weighted linear combination of the pixel loss function, the perceptual loss function, the gradient structure consistency loss function, and the generative adversarial loss function. The training module is used to perform adversarial training on the generator and multi-scale discriminator using an alternating update method, and to optimize the network parameters of the 3D generative adversarial network through the total loss function. The output module is used to input the 3T MRI image to be processed into the trained generator and output the corresponding 7T high-quality magnetic resonance image.

[0020] Thirdly, a computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the above-described magnetic resonance image enhancement method based on noise exploitation and generative adversarial networks.

[0021] Fourthly, embodiments of the present invention provide a computer-readable storage medium including a computer program that, when executed by a processor, implements the steps of the above-described magnetic resonance image enhancement method based on noise exploitation and generative adversarial networks.

[0022] Fifthly, a chip includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the above-described magnetic resonance image enhancement method based on noise exploitation and generative adversarial networks.

[0023] In a sixth aspect, embodiments of the present invention provide an electronic device including a computer program, which, when executed by the electronic device, implements the steps of the above-described magnetic resonance image enhancement method based on noise exploitation and generative adversarial networks.

[0024] Compared with the prior art, the present invention has at least the following beneficial effects: A magnetic resonance imaging enhancement method based on noise exploitation and generative adversarial networks (GANs) is proposed. By acquiring and preprocessing paired 3T / 7T images, the spatial alignment and intensity consistency of the training data are ensured, laying the foundation for high-quality learning. The constructed 3D GAN, incorporating a noise exploitation mechanism, breaks away from traditional denoising methods by injecting noise as a regularization term into the feature domain, effectively preventing overfitting and improving the model's generalization ability. The weighted linear combination design of multiple loss functions not only focuses on pixel-level accuracy but also considers perceptual quality and structural integrity. An alternating update adversarial training strategy allows the generator and discriminator to co-evolve in a game-like process, ultimately achieving high-quality mapping from low-field 3T images to high-field 7T images. The generator is built on V-net and integrates feature-level adaptive noise injection and a 3D CBAM module. The discriminator adopts a multi-scale hierarchical structure, solving the problem of spatial correlation loss in 3D medical images from the bottom layer of the network architecture. This lays the foundation for subsequent optimization schemes, offering broad protection and essential technical features, and enabling high-quality conversion from 3T to 7T MRI as a whole.

[0025] Furthermore, through the orderly execution of steps such as N4 off-field correction, skull dissection, and spatial alignment, various interfering factors in MRI images are eliminated at the source. N4 off-field correction effectively solves the problem of image intensity inhomogeneity; skull dissection removes non-brain tissue, reducing the interference of invalid features on model training; spatial alignment achieved by affine transformation ensures the anatomical structure matching between 3T and 7T images; histogram matching and linear normalization unify the intensity distribution and numerical range of the images; and 3D block segmentation adapts to the input requirements of 3D networks.

[0026] Furthermore, the encoder-decoder architecture based on the V-net network retains V-net's advantages in spatial feature extraction for medical 3D image segmentation, adapting to the processing requirements of MRI 3D voxel data. Channel expansion and residual connections in the initial convolutional layers avoid the gradient vanishing problem in feature extraction. Feature-level adaptive noise injection introduces noise into the feature domain, which, unlike traditional input layer noise injection, enhances the model's feature learning ability and generalization. The synchronized adjustment of the encoder and decoder's resolution and channel count enables multi-scale feature extraction and fusion. The embedding of the 3D CBAM module at skip connections allows the model to adaptively focus on key anatomical features, while the use of the Tanh activation function ensures the reasonableness of the output image pixel values.

[0027] Furthermore, by designing the standard deviation of Gaussian noise as a learnable parameter constrained by the Softplus function, the noise intensity can be adaptively adjusted during the model training process. In the early stage of training, an appropriate amount of noise is used to improve the model's generalization ability, while in the later stage of training, noise is reduced to ensure the accuracy of the generated images. The stochastic resonance effect is used to fully leverage the positive role of noise in model training.

[0028] Furthermore, the symmetrical design of three downsampling residual blocks and three upsampling residual blocks ensures hierarchical matching in the feature extraction and reconstruction process. The convolution operation with a stride of 2 achieves an orderly reduction in spatial resolution, and the exponential increase or decrease in the number of channels allows the model to extract rich feature information at different scales. The Leaky ReLU activation function in the downsampling residual block solves the ReLU death problem, ensuring gradient propagation in deep networks. The concatenation of the feature maps with the corresponding layers of the encoder during upsampling achieves the fusion of shallow detailed features and deep semantic features, avoiding feature loss during the upsampling process.

[0029] Furthermore, the 2D CBAM is extended to a 3D structure, perfectly adapting to the processing requirements of MRI 3D voxel images and overcoming the deficiency of existing 2D attention modules in capturing the spatial correlation between voxels. The CAM module weights features along the channel dimension, allowing the model to focus on the importance of anatomical features corresponding to different channels. The SAM module focuses on key anatomical regions in the spatial dimension. The combination of pooling operations with MLP structures and convolutional kernels makes the calculation of attention weights more reasonable and accurate. Element-wise multiplication weighting enables fine-tuning of the feature map, allowing the model to adaptively focus on key clinical anatomical structures such as the hippocampal subregion and cortical laminar microstructures, significantly improving the clarity of small lesions and fine structures in the generated images.

[0030] Furthermore, the four-stage downsampling convolutional structure gradually extracts image features at different scales. Convolutional kernels of different sizes are adapted to the extraction requirements of features at different scales. Adaptive 3D filling ensures the regularity of feature map size. Multiple predictive convolutional modules output discrimination results at different scales, enabling the discriminator to more accurately identify the differences between generated images and real images.

[0031] Furthermore, adversarial loss, pixel loss, perceptual loss, and gradient structure consistency loss are fused together to address the issue that a single loss function cannot simultaneously address image pixel accuracy, structural consistency, and realism. Adversarial loss ensures the distribution consistency between the generated and real images; pixel loss achieves precise pixel-level matching; perceptual loss focuses on the high-level semantic feature similarity of the images; and gradient structure consistency loss ensures that the edges and structural gradients of the generated images are consistent with those of the real images. By adjusting the contribution weights of each loss function through hyperparameters, the optimization direction of the model can be flexibly adjusted according to training needs. This allows the model to simultaneously optimize image pixel accuracy, anatomical structural integrity, and visual realism during training, achieving a significant multi-dimensional improvement in the quality of generated images.

[0032] Furthermore, the implementation based on the PyTorch framework ensures the convenience and compatibility of network training. The 400-epoch training rounds balance training effectiveness and efficiency, while the batch size of 10 achieves a balance between training stability and computational resource utilization. The use of the Adam optimizer adaptively adjusts the learning rate, improving training convergence speed. The linear decay strategy of the learning rate ensures convergence speed in the early stages of training and improves model convergence accuracy in the later stages. Random initialization of network weights avoids the model getting trapped in local optima. Quantitative settings of various hyperparameters and noise intensity form a standardized model training process, allowing the model to converge stably to the optimal state. At the same time, the gradient structure consistency loss is calculated through 3D Sobel convolution, effectively preserving the three-dimensional edge details of MRI images and further improving the structural fidelity of the generated images.

[0033] It is understood that the beneficial effects of the second to sixth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here.

[0034] In summary, the method of this invention deeply integrates noise utilization with 3D GAN, solving core problems such as poor spatial continuity; the entire process of preprocessing, network construction, loss function, and training is standardized, balancing fidelity and generation efficiency, and the method and system correspond to each other, making it easy to be engineered and implemented.

[0035] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 Here is a diagram of the generator structure; Figure 3 Here is a diagram of the discriminator structure; Figure 4 This is a 3D CBAM module structure diagram; Figure 5 The figure shows the experimental results of this invention; Figure 6 Visual comparison for brain tissue segmentation; Figure 7 A schematic diagram of a computer device provided in an embodiment of the present invention; Figure 8 This is a block diagram of a chip provided according to an embodiment of the present invention.

[0037] Among them, 60. Computer equipment; 61. Processor; 62. Memory; 63. Computer program; 600. Electronic device; 610. Processing unit; 620. Storage unit; 6201. Random access memory unit; 6202. Cache memory unit; 6203. Read-only memory unit; 6204. Program / utility; 6205. Program module; 630. Bus; 640. Display unit; 650. Input / output interface; 660. Network adapter; 700. External device. Detailed Implementation

[0038] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] In the description of this invention, it should be understood that the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0040] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0041] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Additionally, the character " / " in this invention generally indicates that the preceding and following objects have an "or" relationship.

[0042] It should be understood that although terms such as first, second, third, etc., may be used in the embodiments of the present invention to describe the preset range, these preset ranges should not be limited to these terms. These terms are only used to distinguish the preset ranges from one another. For example, without departing from the scope of the embodiments of the present invention, the first preset range may also be referred to as the second preset range, and similarly, the second preset range may also be referred to as the first preset range.

[0043] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0044] The accompanying drawings illustrate various structural schematic diagrams according to embodiments disclosed in this invention. These drawings are not to scale, and some details have been enlarged for clarity, and some details may have been omitted. The shapes of the various regions and layers shown in the drawings, as well as their relative sizes and positional relationships, are merely exemplary and may deviate from reality due to manufacturing tolerances or technical limitations. Furthermore, those skilled in the art can design regions / layers with different shapes, sizes, and relative positions as needed.

[0045] This invention provides a magnetic resonance imaging enhancement method based on noise exploitation and generative adversarial networks (GANs). First, paired 3T and 7T MRI images undergo preprocessing such as field correction and spatial alignment, and are then segmented into 3D blocks. Next, a 3D GAN with trainable parameters and adaptive noise is constructed. The generator is based on V-net combined with a 3D CBAM module, and the discriminator has a multi-scale hierarchical structure. Simultaneously, multiple loss functions are designed and weighted to form the total loss. The generator and discriminator are alternately trained using an Adam optimizer. Finally, the 3T MRI image is input into the trained generator, outputting a high-quality 7T MRI image. This invention improves model robustness and image fidelity, providing high-value images for clinical diagnosis and demonstrating strong practicality.

[0046] Please see Figure 1 This invention discloses a magnetic resonance image enhancement method based on noise exploitation and generative adversarial networks, comprising the following steps: S1. Acquire paired 3T and 7T magnetic resonance images and perform preprocessing operations; The specific steps of magnetic resonance image preprocessing are as follows: S101. Apply the N4 off-field correction algorithm to all 3T and 7T MRI images to eliminate intensity inhomogeneities; S102. Use the FSL software package to perform skull dissection on all off-field corrected images and remove non-brain tissue. S103. Using the FSL FLIRT toolbox, spatially align the 3T image of each patient with its corresponding 7T image through affine transformation. S104. Apply histogram matching to the 3T and 7T images respectively to ensure that the intensity distribution is consistent across different data groups; S105. Linearly normalize the voxel intensity of each image and scale it to the [-1,1] interval; S106. Finally, the paired 3T and 7T images are cut into 64×64×64 three-dimensional blocks with an overlap step of 8.

[0047] S2. Construct a three-dimensional generative adversarial network that integrates noise utilization mechanisms, including a generator and a discriminator; Please see Figure 2 The generator structure based on the V-net network in generative adversarial networks is as follows: S201. The initial convolutional layer first expands the image from a 1-channel input to a 16-channel input and adds residual connections, then injects noise; Trainable parameter adaptive noise is injected into the feature domain output by the first-layer feature extraction module of the generator to achieve feature-level adaptive noise injection. The specific process is as follows: Let the input single-channel 3D MRI image tensor be... Where B is the batch size, and D, H, and W are the depth, height, and width of the image, respectively.

[0048] The first-layer feature extraction module of the generator consists of 3D convolution, batch normalization, and PReLU activation function in sequence. The image tensor is input into this module, and the initial feature tensor is obtained through feature extraction. , where K is the number of 3D convolution kernels in this layer.

[0049] Then, input the initial feature tensor Injecting zero-mean Gaussian noise, the resulting feature tensor is represented as: ,in To and A dimensionally perfectly matched unit Gaussian noise tensor. This represents the standard deviation of noise.

[0050] noise standard deviation Designed for learnable parameters The mapping results are processed by the Softplus function on the learnable parameters. Apply monotonically nonnegative constraints, the constraint formula is as follows: This allows the noise intensity to be adaptively adjusted during model training.

[0051] S202: The encoder progressively reduces the spatial resolution using three downsampled residual blocks, while the number of channels increases exponentially, successively to 16, 32, 64, and 128. Each downsampled residual block contains a convolution operation (with a stride of 2), batch normalization, and the LeakyReLU activation function. S203. The decoder consists of three upsampled residual blocks. Each block first doubles the feature map size and halves the number of channels through upsampling, and then concatenates it with the feature map of the corresponding layer of the encoder. S204. An attention module is introduced at the jump connection between the encoder end and the decoder, which combines channel attention and spatial attention. Please see Figure 4 The CBAM attention module consists of two parts: the channel attention module (CAM) and the spatial attention module (SAM). The CBAM module is extended from 2D to 3D to adapt to the input and embedded in the skip connections, as detailed below: Assume the input feature map is Where C, H, W, and D represent the number of channels, feature map height, feature map width, and feature map dimension, respectively. After entering the CAM module, this feature map undergoes global average pooling and global max pooling to obtain: , .

[0052] Processed feature map , The input consists of two fully connected layers sharing weights, i.e., an MLP structure. In short, channel attention is calculated as follows:

[0053] in, It is the sigmoid activation function. , These are the weight matrices for the first and second layers of the MLP structure, respectively.

[0054] Finally, use the obtained weight matrix The original feature map is reweighted to obtain the output of the CAM module. , in This indicates element-wise multiplication.

[0055] After the feature map is processed by the CAM module, it is passed to the SAM module. The feature map after processing by the CAM module... Upon entering the SAM module, channel-level optimization will be performed first: Then, the feature maps obtained in the previous step are combined through cascading. , ,get Process it with a 7×7 convolution kernel and activate it with the sigmoid activation function to obtain... ; Finally, the obtained weight matrix is ​​used. For the input feature map Reweighting: This is the final output of the attention module.

[0056] S205. After upsampling to the original size, a 1-channel output is generated through two convolutional layers, and the pixel value is restricted to [-1, 1] using the Tanh activation function.

[0057] Please see Figure 3 The discriminator is constructed as follows: The multi-scale discriminator is based on a relative average generative adversarial network design and adopts a four-stage hierarchical 3D convolutional network structure. It achieves multi-granularity image authenticity discrimination through progressive downsampling and multi-stage feature extraction. The specific structure is as follows: The discriminator's input layer receives two types of image input: real high-quality images ( , For the distribution of real high-quality images) and to generate high-quality images ( , The high-quality image distribution output by the generator is used as the input image. After standardization, the input image enters the feature extraction layer. The discriminator adopts a four-stage downsampling convolutional structure. Each stage consists of a 3D convolutional block, and each convolutional block includes convolution operations, batch normalization, and the Leaky ReLU activation function. The kernel size is 7×7×7 in the first to third stages and 5×5×5 in the fourth stage, with a stride of 2. The padding method is adaptive 3D padding calculated based on the kernel size and dilation rate. Each prediction convolutional module includes a 3D copy padding layer, a 3D convolutional layer, and outputs the discrimination result at the corresponding scale.

[0058] The training objective of the discriminator is to maximize the loss function. This function achieves accurate differentiation between two types of input images by superimposing two component losses. The specific formula is as follows:

[0059] in, This represents the output of the discriminator.

[0060] S3. Design a multi-loss function optimization model; Construct multi-channel loss functions, including pixel loss function, perceptual loss function, gradient structure consistency loss, and generative adversarial loss function. The gradient structure consistency loss is specifically constructed as follows: Let I be a 3D image volume. , and These represent the voxel gradients in the axial, sagittal, and coronal planes at the point, respectively. The value at that location, then , , ,in Represents 3D convolution operation. , and These represent standard 3D Sobel convolution kernels in different directions, each with a size of 3×3×3, and their specific form is as follows: , , , , , , , , .

[0061] Therefore, record For 3D images The gradient mapping at each voxel is then

[0062] Adversarial loss function:

[0063]

[0064] in, , These represent the input real 7T MRI image and the generator-generated 7T MRI image, respectively.

[0065] Pixel loss function:

[0066] in, Representing 3T MRI images, Represents generator, This represents a 7T MRI image generated by the generator.

[0067] Perceptual loss function:

[0068] Where J represents the total number of layers in the VGG neural network. This represents the feature map extracted from the j-th layer of the VGG neural network. A scaling factor is introduced. The ∈(0,1) parameter guides the generator to retain deeper feature details in the original image to a greater extent.

[0069] To ensure that the gradient magnitude of the generated image closely approximates the gradient magnitude of the real target image at the voxel level, and to guarantee that the structural strength of the generated image is consistent with that of the real image, the gradient magnitude of the generated image is minimized. Gradient mapping and corresponding real 7T image The gradient loss is defined by the distance between gradient mappings, i.e., the gradient structure consistency loss (GSCL) is defined as:

[0070] in, Represents a true 3T MRI image. This represents the true distribution of 3T MRI images. This represents high-quality 7T MRI images obtained from the generative model. This represents the expected value of the corresponding sample distribution.

[0071] The gradient is calculated using 3D Sobel convolution, which guides the generator to preserve edge details of brain MRI images.

[0072] The total loss function is defined as the adversarial loss function above. A weighted linear combination of pixel loss function, perceptual loss function, and gradient structure consistency loss function:

[0073] in, , , , This is a hyperparameter used to balance the contributions of various losses.

[0074] S4. Alternately update the generator and discriminator for adversarial training to optimize the network model; The network model was built based on the PyTorch deep learning framework and trained on a computing platform equipped with an NVIDIA GeForce RTX 4090 GPU. The training process lasted for 400 epochs with a batch size of 10.

[0075] The network weights are initialized randomly.

[0076] The Adam optimizer is used for network optimization, and the generator and discriminator are updated alternately for adversarial training. The network parameters are updated by optimizing the overall loss function.

[0077] The initial learning rate is set to 0.0001.

[0078] The learning rate scheduling strategy is as follows: the initial learning rate is kept constant for the first 150 epochs, and then linearly decays to zero over the remaining 250 epochs.

[0079] Hyperparameters , , , The values ​​were set to 1.0, 0.1, 1.0, and 0.1 respectively, and the noise intensity was set to 0.1.

[0080] S5. After training is complete, input the 3T MRI image to be processed into the trained generator, and output the corresponding high-quality 7T MRI image.

[0081] In one embodiment of the present invention, a magnetic resonance image enhancement system based on noise exploitation and generative adversarial networks is provided. This system can be used to implement the above-mentioned magnetic resonance image enhancement method based on noise exploitation and generative adversarial networks. Specifically, the magnetic resonance image enhancement system based on noise exploitation and generative adversarial networks includes a data module, a network module, a function module, a training module, and an output module.

[0082] The data module is used to acquire pairs of 3T and 7T magnetic resonance images and perform preprocessing operations to obtain preprocessed 3T MRI image blocks. The network module is used to construct a three-dimensional generative adversarial network that integrates a noise utilization mechanism. The three-dimensional generative adversarial network includes a generator and a multi-scale discriminator. The generator is constructed based on the V-net network and integrates a feature-level adaptive noise injection mechanism and a 3D CBAM attention module. The multi-scale discriminator is a four-stage hierarchical 3D convolutional network structure. The function module is used to design the total loss function, which is a weighted linear combination of the pixel loss function, the perceptual loss function, the gradient structure consistency loss function, and the generative adversarial loss function. The training module is used to perform adversarial training on the generator and multi-scale discriminator using an alternating update method, and to optimize the network parameters of the 3D generative adversarial network through the total loss function. The output module is used to input the 3T MRI image to be processed into the trained generator and output the corresponding 7T high-quality magnetic resonance image.

[0083] This invention provides a terminal device comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, graphics processing units (GPUs), tensor processing units (TPUs), digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions to achieve a corresponding method flow or corresponding function. The processor described in this embodiment can be used for the operation of a magnetic resonance image enhancement method based on noise exploitation and generative adversarial networks, including: Pairs of 3T and 7T MRI images were acquired and preprocessed to obtain preprocessed 3T MRI image patches. A three-dimensional generative adversarial network (GAN) with a noise exploitation mechanism was constructed. The GAN includes a generator and a multi-scale discriminator. The generator is constructed based on a V-net network and incorporates a feature-level adaptive noise injection mechanism and a 3D CBAM attention module. The multi-scale discriminator is a four-stage hierarchical 3D convolutional network structure. A total loss function was designed, which is a weighted linear combination of pixel loss function, perceptual loss function, gradient structure consistency loss function, and generative adversarial loss function. The generator and multi-scale discriminator were trained adversarially using an alternating update method, and the network parameters of the GAN were optimized through the total loss function. The 3T MRI image to be processed was input into the trained generator, and the corresponding high-quality 7T MRI image was output.

[0084] Please see Figure 6 The terminal device is a computer device. In this embodiment, the computer device 60 includes a processor 61, a memory 62, and a computer program 63 stored in the memory 62 and executable on the processor 61. When executed by the processor 61, the computer program 63 implements the magnetic resonance image enhancement method based on noise exploitation and generative adversarial networks in this embodiment. To avoid repetition, details are omitted here. Alternatively, when executed by the processor 61, the computer program 63 implements the functions of each model / unit in the magnetic resonance image enhancement system based on noise exploitation and generative adversarial networks in this embodiment. To avoid repetition, details are omitted here.

[0085] Computer device 60 can be a desktop computer, laptop, handheld computer, cloud server, or other computing device. Computer device 60 may include, but is not limited to, a processor 61 and a memory 62. Those skilled in the art will understand that... Figure 6 This is merely an example of computer device 60 and does not constitute a limitation on computer device 60. It may include more or fewer components than shown, or combine certain components, or different components. For example, computer device may also include input / output devices, network access devices, buses, etc.

[0086] The processor 61 may be a Central Processing Unit (CPU), or other general-purpose processors, graphics processing units (GPUs), tensor processing units (TPUs), digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0087] The memory 62 can be an internal storage unit of the computer device 60, such as a hard disk or memory of the computer device 60. The memory 62 can also be an external storage device of the computer device 60, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. equipped on the computer device 60.

[0088] Furthermore, the memory 62 may include both internal storage units of the computer device 60 and external storage devices. The memory 62 is used to store computer programs and other programs and data required by the computer device. The memory 62 can also be used to temporarily store data that has been output or will be output.

[0089] Please see Figure 7 The terminal device is an electronic device 600, which is manifested in the form of a general-purpose computing device. The components of the electronic device may include, but are not limited to: at least one processing unit 610, at least one storage unit 620, a bus 630 connecting different platform components (including storage unit 620 and processing unit 610), a display unit 640, etc.

[0090] The storage unit stores program code, which can be executed by the processing unit 610 to perform the steps described in the method section of this specification according to various exemplary embodiments of the present invention. For example, the processing unit 610 can perform actions such as... Figure 1 The steps are shown in the figure.

[0091] Storage unit 620 may include a readable medium in the form of a volatile storage unit, such as random access memory (RAM) 6201 and / or cache memory 6202, and may further include a read-only memory (ROM) 6203.

[0092] Storage unit 620 may also include a program / utility 6204 having a set (at least one) program module 6205, such program module 6205 including but not limited to: operating system, one or more application programs, other program modules and program data, each or some combination of these examples may include an implementation of a network environment.

[0093] Bus 630 can represent one or more of several types of bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the multiple bus structures.

[0094] Electronic device 600 can also communicate with one or more external devices 700 (e.g., keyboard, pointing device, Bluetooth device, etc.), and with one or more devices that enable a user to interact with electronic device 600, and / or with any device that enables electronic device 600 to communicate with one or more other computing devices (e.g., router, modem). This communication can be performed via input / output interface 650. Furthermore, electronic device 600 can also communicate with one or more networks (e.g., local area network, wide area network, and / or public network, such as the Internet) via network adapter 660. Network adapter 660 can communicate with other modules of electronic device 600 via bus 630. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with electronic device 600, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage platforms.

[0095] Example 4 This invention also provides a storage medium, specifically a computer-readable storage medium, which is a memory device in a terminal device for storing programs and data. It is understood that the computer-readable storage medium here can include both built-in storage media in the terminal device and extended storage media supported by the terminal device; it can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor, which can be one or more computer programs (including program code). More specific examples of the computer-readable storage medium include: an electrical connection with one or more wires, a portable disk, a hard disk, random access memory, read-only memory, erasable programmable read-only memory, optical fiber, portable compact disk read-only memory, optical storage device, magnetic storage device, or any suitable combination thereof.

[0096] Computer-readable storage media also include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium can also be any readable medium other than a readable storage medium that can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium can be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, radio frequency, etc., or any suitable combination thereof.

[0097] Program code for performing the operations of this invention can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can execute entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0098] One or more instructions stored in a computer-readable storage medium can be loaded and executed by a processor to implement the corresponding steps of the magnetic resonance image enhancement method based on noise exploitation and generative adversarial networks in the above embodiments; one or more instructions in the computer-readable storage medium are loaded by the processor and executed as follows: Pairs of 3T and 7T MRI images were acquired and preprocessed to obtain preprocessed 3T MRI image patches. A three-dimensional generative adversarial network (GAN) with a noise exploitation mechanism was constructed. The GAN includes a generator and a multi-scale discriminator. The generator is constructed based on a V-net network and incorporates a feature-level adaptive noise injection mechanism and a 3D CBAM attention module. The multi-scale discriminator is a four-stage hierarchical 3D convolutional network structure. A total loss function was designed, which is a weighted linear combination of pixel loss function, perceptual loss function, gradient structure consistency loss function, and generative adversarial loss function. The generator and multi-scale discriminator were trained adversarially using an alternating update method, and the network parameters of the GAN were optimized through the total loss function. The 3T MRI image to be processed was input into the trained generator, and the corresponding high-quality 7T MRI image was output.

[0099] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0100] 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. 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. Therefore, the following detailed description of the embodiments of the present 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 present invention without inventive effort are within the scope of protection of the present invention.

[0101] Please see Figure 5 The visualization analysis results of 7T MRI images synthesized from the model at each stage intuitively present the improvement effect of adding different modules on the synthesis effect, specifically including: 1) Multi-stage image visualization comparison: The original 3T MRI image, the real 7T MRI image, and the 7T MRI images output by the five stages of the model: basic V-Net, basic V-Net+3D CBAM module, basic V-Net+3D CBAM module+gradient structure consistency loss, basic V-Net+3D CBAM module+gradient structure consistency loss+fixed noise, and the complete model of this invention are displayed in sequence. The continuity of anatomical structure and the clarity of details of each stage image in the axial, sagittal and coronal planes are clearly compared. 2) Staged Absolute Error Map Analysis: For the model output results of the above four stages, the absolute error maps of the synthesized images and real 7T MRI images at each stage were calculated. The error maps were calculated based on real data normalized to the [0,1] interval and the transformation results, with the numerical range constrained to [0,1]. The lower values ​​(visually represented by light colors) represent smaller differences between the synthesized images and real images and higher fidelity of anatomical structures. By comparing the five sets of error maps, it can be directly verified that with the gradual addition of adversarial learning mechanism, 3D CBAM module, gradient structure consistency loss, and trainable Gaussian noise, the structural differences between the synthesized images and real 7T MRI images continuously decrease, and the integrity and consistency of three-dimensional anatomical structures are significantly improved.

[0102] To evaluate the enhancement of tissue contrast and its clinical applicability after cross-field intensity conversion, this study conducted quantitative and qualitative assessments based on brain tissue segmentation. The SPM software, commonly used in neuroimaging, was employed to segment all images into three core tissue regions: gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF). Please refer to [link to relevant documentation]. Figure 6 The segmentation results of the original 3T MRI, the real 7T MRI, and the 7T MRI synthesized from the complete model of this invention.

[0103] To quantitatively verify the brain tissue structure reconstruction capability of the 3TMRI to 7TMRI conversion model, the Dice Similarity Coefficient (DSC) was used to assess the segmentation consistency of three core brain tissues: white matter, gray matter, and cerebrospinal fluid. The DSC value ranged from [0,1]. The closer the value was to 1, the higher the overlap between the image being evaluated and the gold standard tissue morphology, and the better the anatomical structure fidelity.

[0104] Table 1. DSC and corresponding standard deviations for segments of white matter, gray matter, and cerebrospinal fluid.

[0105] Table 1 presents the DSC and corresponding standard deviations for segmentation of white matter, gray matter, and cerebrospinal fluid. The results show that the DSC indices of each tissue from the synthesized 7T MRI are highly consistent with the gold standard of the real 7T MRI and significantly superior to the results obtained from direct segmentation of the original 3T MRI. This demonstrates that the model proposed in this invention can effectively preserve and restore the structural features of brain tissue, and the synthesized images possess anatomical accuracy and high-quality imaging characteristics sufficient for medical analysis.

[0106] To further quantify and evaluate the accuracy of tissue contrast restoration in the 3TMRI to 7TMRI conversion task, this study used the white-to-gray matter signal ratio (WM / GM_SR) as a quantitative indicator to reflect the difference in voxel signal intensity between white and gray matter brain regions. The calculation formula is as follows: ,in The average signal intensity of all voxels within the white matter mask coverage area after brain tissue extraction from the target image. This represents the average voxel signal intensity of the corresponding region in the gray matter mask within the same image. It is calculated based on the effective brain voxels after skull dissection. (Original 3T MRI) (Real 7T MRI) (Synthetic 7TMRI) White-to-gray matter signal ratios from three datasets. Only effective voxels extracted from brain tissue were selected for calculation, excluding interference from cerebrospinal fluid, skull, and background noise regions to ensure the accuracy of the signal mean calculation. Paired Wilcoxon signed-rank tests were used for statistical comparisons between groups, with the significance level set at [value missing]. Using a within-subjects paired design, participants were separately paired between raw 3T and real 7T, and between synthetic 7T and real 7T. Paired tests were conducted to statistically verify the effectiveness of synthesized 7T MRI in approximating the signal characteristics of real 7T tissues compared to input 3T MRI.

[0107] Quantitative results showed that the original 3T group The value was 1.9546 ± 0.0634, for the actual 7T group. The value was 4.6035 ± 0.5689, and the synthesis was performed on a 7T group. The value was 4.9694 ± 0.3662. Paired Wilcoxon test results showed that there was a highly significant difference in the white-gray matter signal ratio between the original 3T and the real 7T (p = 0.0078); there was no significant difference between the synthesized 7T and the real 7T (p = 0.1484), indicating that the synthesized image can closely approximate the real 7T in terms of tissue signal ratio; at the same time, there was a highly significant difference between 3T and the synthesized 7T (p = 0.0078), confirming that the proposed conversion model can effectively realize the signal distribution transfer from 3T to 7T and complete the tissue contrast-preserving reconstruction of low-field to high-field images.

[0108] To quantify and verify the effectiveness of module-by-module optimization, a multi-dimensional quantitative evaluation was conducted on the five stages: "basic V-Net → adding 3D CBAM → adding gradient loss → adding fixed noise → adding trainable Gaussian noise," integrating four metrics: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), Mutual Information (MI), and Fraser Initial Distance (FID). (1) Peak signal-to-noise ratio (PSNR): measures pixel-level error. The higher the value, the higher the pixel accuracy of the synthesized image compared to the real 7T MRI. (2) Structural similarity (SSIM): assesses the consistency of three-dimensional anatomical structures. The closer the value is to 1, the better the edge structures in the axial, sagittal and coronal planes are matched. (3) Mutual information (MI): quantifies the degree of information overlap between image texture and detail; the higher the value, the more complete the 7T MRI features are preserved. (4) Fraser Initial Distance (FID): Evaluates the consistency of the distribution between the generated image and the real image. The lower the value, the stronger the realism and stability of the generated result.

[0109] By comparing the changes in the four indicators of the five-stage model, it can be clearly verified that with each new module, all evaluation indicators of the model show a positive optimization trend, and the complete model of this invention has the best performance, which fully demonstrates the effectiveness of the synergistic effect of each module in improving the conversion effect of 3T to 7T MRI.

[0110] Table 2. Impact of each module of the method of the present invention on model indices

[0111] Table 2 shows the impact of each module of the method of this invention on the model metrics. Experimental results show that the modules proposed in this paper enable the algorithm to outperform the basic model in multiple metrics, taking into account image details, edge sharpness, contrast, etc., resulting in better image visual effects.

[0112] To objectively verify the effectiveness and superiority of the method of the present invention, the classic traditional method histogram matching (HM) and the mainstream wavelet-based deep learning method (WATnet) were selected as control methods. The experimental results are shown in Tables 3 and 4.

[0113] Table 3 Comparison of evaluation indicators for different MRI image enhancement methods

[0114] As shown in Table 3, compared with HM and WATnet, the proposed model in this invention has improved in all four indicators, which fully demonstrates that the proposed model outperforms traditional methods and existing deep learning methods in terms of preserving image structural information, strengthening information correlation, and improving image visual fidelity, and has stronger image processing capabilities.

[0115] Table 4 Comparison of tissue segmentation accuracy of different MRI image enhancement methods

[0116] As shown in Table 4, the model proposed in this invention outperforms the HM and WATnet methods in segmentation accuracy across various tissues, achieving precise segmentation of key medical tissues. This provides more reliable data support for subsequent medical imaging diagnosis and lesion analysis, and meets the high-precision processing requirements of the medical field. Please refer to... Figure 6 The results of brain tissue segmentation using HM, WATnet, and the model proposed in this invention are presented visually.

[0117] The data in this invention comes from 42 patients with epilepsy. Each patient underwent 3T and 7T MRI scans at the same medical center (the Second Affiliated Hospital of Zhejiang University School of Medicine). Three-dimensional T1-weighted images were obtained for each patient, thus forming a paired dataset of 3T and 7T images for the same patient.

[0118] In summary, this invention presents a magnetic resonance imaging enhancement method based on noise utilization and generative adversarial networks (GANs), effectively solving the technical challenges of low structural fidelity and poor spatial continuity in the conversion from 3T to 7T MRI. Feature-level adaptive noise injection enhances the model's robustness and generalization ability; the 3D CBAM module allows the model to adaptively focus on key clinical anatomical structures; and the combination of multi-scale discriminators and multiple loss functions achieves multi-dimensional optimization of pixel accuracy, structural consistency, and realism in the generated images. Experiments show that the complete model of this invention significantly outperforms the baseline model in terms of SSIM, PSNR, and MI, while significantly reducing the FID index. The generated 7T MRI images clearly present fine anatomical features such as hippocampal subregions and cortical layered microstructures, and can detect more microscopic lesions in patients with negative 3T MRI findings. This invention does not rely on expensive 7T equipment and can enhance ordinary 3T MRI images to the 7T level, providing primary healthcare institutions with a way to acquire high-field-strength MRI images, significantly reducing the cost of accurate clinical imaging, and possessing extremely high clinical application value and promising prospects for widespread adoption.

[0119] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0120] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0121] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0122] In the embodiments provided by this invention, it should be understood that the disclosed devices / terminals and methods can be implemented in other ways. For example, the device / terminal embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0123] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0124] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0125] If the integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random-access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0126] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus, and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0127] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0128] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0129] The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. A magnetic resonance image enhancement method based on noise exploitation and generative adversarial networks, characterized in that, Includes the following steps: S1. Acquire paired 3T and 7T MRI images and perform preprocessing operations to obtain preprocessed 3T MRI image blocks; S2. Construct a three-dimensional generative adversarial network that integrates a noise utilization mechanism. The three-dimensional generative adversarial network includes a generator and a multi-scale discriminator. The generator is constructed based on the V-net network and integrates a feature-level adaptive noise injection mechanism and a 3DCBAM attention module. The multi-scale discriminator is a four-stage hierarchical 3D convolutional network structure. S3. Design a total loss function formed by a weighted linear combination of pixel loss function, perceptual loss function, gradient structure consistency loss function and generative adversarial loss function; S4. The generator and multi-scale discriminator are trained adversarially using an alternating update method, and the network parameters of the three-dimensional generative adversarial network are optimized through the total loss function. S5. Input the 3T MRI image to be processed into the trained generator and output the corresponding 7T high-quality magnetic resonance image.

2. The magnetic resonance image enhancement method based on noise exploitation and generative adversarial networks according to claim 1, characterized in that, In step S1, the preprocessing operations include, in sequence: S101. The N4 off-field correction algorithm is used to perform off-field correction on 3T and 7T MRI images to eliminate intensity inhomogeneity; S102. Perform skull dissection on the off-field rectified image and remove non-brain tissue; S103. Spatially align the 3T image with its corresponding 7T image using affine transformation; S104. Perform histogram matching on the 3T and 7T images respectively to make the image intensity distribution consistent; S105. Linearly normalize the voxel intensity of the image and scale the voxel intensity to the [-1,1] interval. S106. Cut the paired 3T and 7T images into three-dimensional blocks of a preset size to obtain preprocessed 3T MRI image blocks.

3. The magnetic resonance image enhancement method based on noise exploitation and generative adversarial networks according to claim 1, characterized in that, In step S2, the generator construction process includes: S201. The single-channel 3TMRI image block is expanded into a multi-channel feature map through the initial convolutional layer. At the same time, residual connections are added, and zero-mean Gaussian noise is injected into the multi-channel feature map to achieve feature-level adaptive noise injection. S202. The spatial resolution of the feature map after injecting noise is gradually reduced by the downsampling residual block of the encoder, while the number of channels is increased exponentially. S203. Embed a 3DCBAM attention module at the jump connection between the encoder end and the decoder to perform channel and spatial dimension attention weighting on the feature map. S204. The spatial resolution of the feature map is gradually improved by the upsampling residual block of the decoder, while the number of channels is reduced by a factor of two. During the upsampling process, the feature map of the corresponding layer of the encoder is stitched together. S205. After upsampling the feature map to its original size, a single-channel output is generated through a convolutional layer, and the Tanh activation function is used to restrict the pixel values ​​to the range of [-1, 1].

4. The magnetic resonance image enhancement method based on noise exploitation and generative adversarial networks according to claim 3, characterized in that, In step S201, the specific process of feature-level adaptive noise injection is as follows: The input single-channel 3D MRI image tensor is input into the first layer feature extraction module of the generator, and the initial feature tensor is obtained through feature extraction. Zero-mean Gaussian noise is injected into the initial feature tensor to obtain the feature tensor after noise injection. The standard deviation of the Gaussian noise is the mapping result of the learnable parameters after being monotonically non-negatively constrained by the Softplus function, so that the noise intensity is adaptively adjusted as the model is trained.

5. The magnetic resonance image enhancement method based on noise exploitation and generative adversarial networks according to claim 3, characterized in that, The encoder contains three downsampled residual blocks. Each downsampled residual block consists of a convolution operation, batch normalization, and a LeakyReLU activation function. The stride of the convolution operation is 2, and the number of feature map channels is 16, 32, 64, and 128, respectively. The decoder contains three upsampled residual blocks. Each upsampled residual block first doubles the feature map size and halves the number of channels through upsampling, and then concatenates it with the feature map of the corresponding layer of the encoder.

6. The magnetic resonance image enhancement method based on noise exploitation and generative adversarial networks according to claim 3, characterized in that, The 3DCBAM attention module includes a channel attention module (CAM) and a spatial attention module (SAM). The feature map is processed sequentially by the CAM module and the SAM module, and the specific process is as follows: After the feature map is input into the CAM module, it is processed by global average pooling and global max pooling to obtain two feature vectors. The two feature vectors are then input into an MLP structure with shared weights. The channel attention weights are obtained by passing the sigmoid activation function. Finally, the original feature map is multiplied and weighted element-wise to obtain the output of the CAM module. After the output of the CAM module is input into the SAM module, channel-dimensional pooling is first performed to obtain two feature maps. The two feature maps are concatenated and processed by a convolutional kernel and activated by a sigmoid function to obtain spatial attention weights. Then, the output of the CAM module is multiplied element-wise and weighted to obtain the final output of the 3DCBAM attention module.

7. The magnetic resonance image enhancement method based on noise exploitation and generative adversarial networks according to claim 1, characterized in that, In step S2, the construction process of the multi-scale discriminator is as follows: The input layer of the multi-scale discriminator receives real 7TMRI images and generated 7TMRI images output by the generator, and inputs them into the feature extraction layer after standardization. The feature extraction layer adopts a four-stage downsampling convolution structure. Each stage consists of a three-dimensional convolution block, which includes convolution operation, batch normalization and LeakyReLU activation function. The kernel size of the first to third stages is 7×7×7, the kernel size of the fourth stage is 5×5×5, the convolution stride is 2, and the padding method is adaptive 3D padding. Each predictive convolutional module contains a 3D copy padding layer and a 3D convolutional layer, outputting the image authenticity judgment result at the corresponding scale.

8. The magnetic resonance image enhancement method based on noise exploitation and generative adversarial networks according to claim 1, characterized in that, In step S3, the total loss function is: in, , , , For hyperparameters, To counter the loss function, For pixel loss function, For the perceptual loss function, This is the gradient structure consistency loss function.

9. The magnetic resonance image enhancement method based on noise exploitation and generative adversarial networks according to claim 1, characterized in that, In step S4, the specific parameters for the adversarial training are: The network model was built based on the PyTorch framework, with a total training duration of 400 epochs and a batch size of 10. The network weights were randomly initialized. The Adam optimizer was used for network optimization, with an initial learning rate of 0.0001. The initial learning rate was kept constant for the first 150 epochs, and then linearly decayed to zero for the remaining 250 epochs. The hyperparameters of the pixel loss function, perceptual loss function, gradient structure consistency loss function, and generative adversarial loss function in the total loss function were 1.0, 0.1, 1.0, and 0.1, respectively, and the noise intensity was set to 0.

1. The gradient structure consistency loss function was calculated by using a 3DSobel convolution kernel to calculate the voxel gradients of the 3D image volume in the axial, sagittal, and coronal planes to obtain the gradient mapping of the image. This gradient mapping was then constructed by minimizing the distance between the gradient mappings of the generated 7TMRI image and the real 7TMRI image.

10. A magnetic resonance image enhancement system based on noise exploitation and generative adversarial networks, characterized in that, include: The data module is used to acquire pairs of 3T and 7T magnetic resonance images and perform preprocessing operations to obtain preprocessed 3T MRI image blocks; The network module is used to construct a three-dimensional generative adversarial network that integrates a noise utilization mechanism. The three-dimensional generative adversarial network includes a generator and a multi-scale discriminator. The generator is constructed based on the V-net network and integrates a feature-level adaptive noise injection mechanism and a 3D CBAM attention module. The multi-scale discriminator is a four-stage hierarchical 3D convolutional network structure. The function module is used to design the total loss function, which is a weighted linear combination of the pixel loss function, the perceptual loss function, the gradient structure consistency loss function, and the generative adversarial loss function. The training module is used to perform adversarial training on the generator and multi-scale discriminator using an alternating update method, and to optimize the network parameters of the 3D generative adversarial network through the total loss function. The output module is used to input the 3T MRI image to be processed into the trained generator and output the corresponding 7T high-quality magnetic resonance image.

Images