Medical image super-resolution reconstruction method based on multi-level attention guidance

Abstract

The invention discloses a medical image super-resolution reconstruction method based on multi-level attention guidance, and the method comprises the following steps: S10, constructing a deep learning network model based on a generative adversarial network architecture, which comprises a generator and a discriminator; the generator is based on an improved U-Net architecture, a hierarchical attention module and a dual-path feature processing module are configured in an encoder and a decoder of the generator, the hierarchical attention module adopts different attention strategies according to network levels to consider structure and texture, and the dual-path feature processing module separates and processes low-frequency and high-frequency information; the generator further comprises a multi-level feature fusion module for integrating the multi-scale features of the decoder, and an attention guide up-sampling module for final enhancement and dimension raising. The discriminator adopts a spectrum normalization U-Net architecture and uses multi-scale features for matching; s20, training the network model by adopting a composite loss function comprising pixels, adversarial, perception and total variation loss; and S30, inputting the low-resolution image into the trained model, and outputting a high-resolution image. According to the method, through deep fusion of multi-level attention and multi-scale feature processing, the image restoration quality can be remarkably improved, the texture detail definition can be enhanced, the anatomical structure accuracy can be ensured, and the noise robustness can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image super-resolution reconstruction, and in particular to a medical image super-resolution reconstruction method based on multi-level attention guidance. Background Art

[0002] Medical imaging technology is an indispensable part of modern medicine, especially computed tomography (CT), which plays a vital role in clinical diagnosis. With the continuous advancement of CT technology, its clinical benefits in disease diagnosis and treatment planning have also been significantly improved, allowing doctors to more accurately understand the internal structure and abnormal changes of patients. However, despite the great success of CT technology, in reality, due to the constraints of many factors such as the acquisition equipment and environment, network transmission media and bandwidth, and the image degradation model itself, it is usually not possible to directly obtain ideal high-resolution (HR) CT images with sharp edges and no block blur. The balance between image resolution and radiation dose remains a major challenge facing the field of CT imaging.

[0003] To obtain high-quality diagnostic images while controlling radiation dose, Single Image Super-Resolution (SISR) reconstruction technology has emerged. SISR aims to recover the corresponding HR image from a low-resolution (LR) image, but this is a highly ill-posed machine vision problem because the mapping from LR to HR has multiple possibilities and the result is not unique. In addition, when the image is scaled to a large scale, it becomes extremely difficult to recover the lost high-frequency information (such as fine texture and edge details) from the LR image. Therefore, developing an algorithm that can effectively recover precise and detailed texture features and generate higher-resolution images has become a technical problem that needs to be solved urgently.

[0004] Current approaches to image super-resolution can be categorized into three main categories: interpolation-based methods, reconstruction constraint-based methods, and learning-based methods (especially deep learning). Interpolation-based methods, such as nearest neighbor interpolation, bilinear interpolation, and bicubic interpolation, are the most direct and simple. These methods are computationally fast, but they essentially rely on smooth fits to known pixel values, often resulting in a loss of high-frequency information in the reconstructed image. This results in blurry images lacking essential detail and failing to meet the clarity requirements for clinical diagnosis. Methods based on reconstruction constraints, such as the maximum a posteriori probability (MAP) method and projection onto convex sets (POCS), incorporate prior image information (such as gradient distribution and sparsity) as constraints during reconstruction. When the prior information accurately reflects image characteristics, these methods can restore relatively high-quality images. However, their effectiveness depends significantly on the accuracy of the prior model and the complexity of the solution process.

[0005] In recent years, super-resolution reconstruction methods based on deep learning have become a research hotspot and have made significant progress. A variety of models based on deep learning principles have been proposed, including SRCNN (Super-Resolution Convolutional Neural Network), ESPCN (Efficient Sub-Pixel Convolutional Neural Network), EDSR (Enhanced Deep Super-Resolution), SRGAN (Super-Resolution Generative Adversarial Network), ESRGAN (Enhanced SRGAN), and HAT (Hybrid Attention Transformer). SRCNN was the first to apply convolutional neural networks to SISR; ESPCN leveraged effective sub-pixel convolutional layers to achieve end-to-end upsampling learning; EDSR improved reconstruction performance by optimizing network depth and residual structure; SRGAN and its subsequent ESRGAN introduced the generative adversarial network (GAN) framework, significantly improving the visual realism and detailed texture quality of generated images through adversarial training; and models such as HAT further explored the potential of attention mechanisms in complex texture restoration. Although the above-mentioned deep learning-based super-resolution reconstruction method has made significant progress in natural images by utilizing deeper and more complex network structures, improving reconstruction accuracy and visual effects, in the field of CT images, which have higher requirements on image quality, more stringent requirements on details, and unique noise characteristics, how to effectively restore fine texture details with anatomical significance and ensure structural accuracy remains a major challenge.

[0006] On the other hand, deep learning techniques, particularly convolutional neural networks (CNNs), have been widely applied to many medical image processing tasks and have demonstrated great potential. The U-Net architecture, with its exceptional feature extraction and fusion capabilities, has achieved revolutionary breakthroughs in fields such as medical image segmentation, attracting widespread attention and demonstrating the feasibility of deep learning models in understanding and processing complex medical image structures.

[0007] In this context, existing general-purpose image super-resolution reconstruction methods, when directly applied to CT images, suffer from common problems such as low restored image quality, blurred critical diagnostic texture details, sensitivity to CT-specific noise, and the potential for generating anatomically inappropriate results. These challenges have become urgent technical challenges in this field. Summary of the Invention

[0008] The purpose of the present invention is to overcome the above-mentioned defects or problems existing in the background technology and provide a medical image super-resolution reconstruction method based on multi-level attention guidance, which can improve image quality, enhance the clarity of texture details and reduce the impact of noise.

[0009] To achieve the above object, the present invention adopts the following technical solutions:

[0010] Technical Solution 1: A medical image super-resolution reconstruction method based on multi-level attention guidance, which includes the following steps: S10: constructing a deep learning network model, which is constructed based on a generative adversarial network architecture, including a generator and a discriminator; the generator is constructed based on an improved U-Net architecture, and includes: an encoder, which encodes and reduces the dimension of input features through a number of downsampling units; and a decoder, which decodes and increases the dimension of the encoded features through a number of upsampling units; wherein, a hierarchical attention module and a dual-path feature processing module are configured in the downsampling unit of the encoder and the upsampling unit of the decoder; the hierarchical attention module adopts an attention calculation structure that focuses on processing spatial position information or an attention calculation structure that focuses on processing channel dimension information according to the depth of the network level in which it is located; the dual-path feature processing module includes a path for processing low-frequency information and a path for processing high-frequency information, and the outputs of the two paths are fused through a gating unit; the The generator also includes a multi-level feature fusion module and an attention-guided upsampling module; the multi-level feature fusion module is connected to the output ends of multiple upsampling units of the decoder, for receiving and fusing the output features from the multiple upsampling units; the attention-guided upsampling module is connected after the multi-level feature fusion module or constitutes the final output stage of the decoder, which includes a composite attention module and a PixelShuffle upsampling layer; the discriminator is based on a spectral normalized U-Net architecture, and its output layer is used to evaluate the authenticity probability of the input image and use its intermediate layer features for multi-scale feature matching; S20: train the deep learning network model, and use a composite loss function including mean absolute error loss, adversarial loss, perceptual loss and total variation loss terms to optimize the training of the generator and discriminator to obtain a super-resolution reconstruction model; S30: take the low-resolution image as input, and output the super-resolution reconstruction model obtained by S20 training to output a super-resolution image.

[0011] Technical Solution 2 based on Technical Solution 1: When the hierarchical attention module is in the shallow layer of the generator network, the attention calculation structure that focuses on processing spatial position information includes a spatial attention calculation unit for combining position encoding information, and the channel dimensions of the Q matrix and K matrix used to calculate attention are set to 1 / 4 of the number of input channels; when it is in the deep layer of the generator network, the attention calculation structure that focuses on processing channel dimension information includes a unit for performing attention calculation based on the channel dimension, and uses matrix transposition operation to reduce computational complexity.

[0012] Technical Solution 3 based on Technical Solution 1: The dual-path feature processing module, its path for processing low-frequency information is composed of a depth-separable convolution layer; its path for processing high-frequency information is composed of a deformable convolution layer and a SiLU activation function; the gating unit is composed of a 1x1 convolution layer, which is used to process the spliced low-frequency path features and the high-frequency path features to generate fusion weights.

[0013] Technical Solution 4 based on Technical Solution 1: The multi-level feature fusion module uses a deconvolution layer to implement the upsampling and alignment operation of the feature maps output by multiple upsampling units of the decoder; its network structure for calculating the fusion weight includes a global average pooling unit and a 1x1 convolution layer, followed by a Sigmoid activation function; its fusion operation is achieved by channel-dimensional splicing of the weight-adjusted feature maps and subsequent 3x3 convolution processing.

[0014] Technical Solution 5 based on Technical Solution 1: The attention-guided upsampling module includes a composite attention module that sequentially connects a channel attention sub-module and a spatial attention sub-module, and the spatial attention sub-module uses a 3x3 convolution layer to generate a spatial attention map; the PixelShuffle upsampling layer is used to increase the feature map resolution by a predetermined multiple.

[0015] Technical Solution 6 based on Technical Solution 1: The discriminator applies spectral normalization after each convolution layer in its spectral normalization U-Net architecture; the multi-scale feature matching is performed using the global features finally output by the encoder of the discriminator and the local features output by the intermediate layer of the decoder.

[0016] Technical Solution 7 based on Technical Solution 1: The loss function of the generator during training is: in, is the mean absolute error function, To counter the loss function, is the perceptual loss function, is the total variation loss function.

[0017] Technical Solution 8 based on Technical Solution 1: the encoder and decoder of the generator each include a predetermined number of layers; the number of instances of the hierarchical attention module and the dual-path feature processing module configured in each downsampling unit of the encoder is [2, 2, 1, 1] in sequence according to the layer depth; the number of instances of the combination module configured in each upsampling unit of the decoder is [1, 1, 2, 2] in sequence according to the layer depth.

[0018] Technical Solution 9 based on Technical Solution 1: In the generator, the encoder and decoder realize the transmission and integration of cross-level feature information through the multi-level feature fusion module.

[0019] Technical solution 10 based on technical solution 1: before step S10, it also includes step S01: data preprocessing; the data preprocessing includes acquiring pairs of low-resolution images and high-resolution images to form a training data set; the low-resolution images are simulated and generated by performing a downsampling operation on the corresponding high-resolution images; the training data of the discriminator includes the high-resolution images as real samples, and the super-resolution images generated by the generator based on the corresponding low-resolution images as pseudo samples.

[0020] From the above description of the present invention, it can be seen that compared with the prior art, the present invention has the following beneficial effects:

[0021] Technical Solution 1 provides a medical image super-resolution reconstruction method based on multi-level attention guidance. Through the specific structural design and component configuration of its deep learning network model, it can significantly improve the quality of super-resolution reconstruction of computed tomography images. The hierarchical attention module integrated within the generator adopts a differentiated attention calculation structure that focuses on spatial or channel information at different levels of the network encoder and decoder, enabling the model to discriminately process feature information at different scales. When processing shallow features with high spatial resolution, the attention mechanism focusing on spatial information helps accurately capture and maintain local and global spatial dependencies between pixels, which is crucial for preserving the contours, boundaries, and relative positions of anatomical structures, thereby improving the geometric accuracy of the reconstructed image. When processing deep features with high semantic information, the attention mechanism focusing on channel information can more effectively analyze the interrelationships between different feature channels, helping to more accurately model and reconstruct complex tissue textures and internal structural features. This hierarchical division of labor in the attention strategy enables the generator to simultaneously optimize the quality of structure preservation and texture generation.

[0022] Furthermore, the generator's dual-path feature processing module explicitly decomposes the feature information flow into two processing paths: low-frequency and high-frequency. The low-frequency path focuses on processing the overall image structure and smooth regions, helping to maintain the macromorphological stability of the reconstructed image; the high-frequency path focuses on extracting and enhancing image details such as edges and textures. Fusion of these two information paths via a gating unit allows the network to adaptively determine the degree of preservation and enhancement of structural and detail information based on the input features, effectively improving the ability to recover fine textures while maintaining structural integrity. Combined with a multi-level feature fusion module, this module receives and fuses the output features from multiple upsampling units at different decoder levels. Because the decoder outputs at different levels contain information at different scales and levels of abstraction, this multi-level fusion mechanism integrates effective information from different scales into the final reconstruction, avoiding the information loss that may result from fusing features at a single scale. Finally, the attention-guided upsampling module utilizes a composite attention module to recalibrate the feature maps across channels and spatial dimensions before generating the final high-resolution image, enhancing the representation of key diagnostic features. Resolution is then increased through the efficient upsampling layer, PixelShuffle. This series of structural designs for detail recovery and information integration work together to enable this method to generate high-resolution CT images with richer details and clearer textures.

[0023] Furthermore, this method employs a generative adversarial network architecture, equipped with a specific spectrally normalized U-Net discriminator and a multi-scale feature matching strategy. The GAN's adversarial training mechanism forces the generator to learn to produce outputs that are indistinguishable from real high-resolution CT images, thereby improving the visual realism and naturalness of the reconstructed images. Spectral normalization enhances the stability of discriminator training, enabling it to more reliably learn the data distribution of real images. The discriminator utilizes its U-Net architecture to extract features at different levels for multi-scale matching, assessing not only the overall image realism but also imposing constraints on the realism of structural and texture details at different scales. The composite loss function employed during training explicitly incorporates objectives for optimizing pixel-level accuracy, enhancing visual realism, and suppressing noise and artifacts. This training framework, combining an advanced discriminator architecture, multi-scale constraints, and multi-objective optimization, results in images generated by this method that exhibit superior overall quality, realism, and suppression of potential noise and artifacts.

[0024] Therefore, the medical image super-resolution reconstruction method based on multi-level attention guidance provided by the present invention optimizes the anatomical structure perception through the hierarchical attention module, enhances the ability to restore fine texture details through dual-path processing, multi-level fusion and attention-guided upsampling, and improves the image authenticity, quality and robustness through a specific GAN training framework and a composite loss function. Logically, it can effectively solve the key problems such as structural distortion, blurred details and low quality faced by CT image super-resolution reconstruction in the background technology, thereby obtaining high-resolution reconstruction results with accurate geometry, clear texture and better visual effects.

[0025] Technical Solution 2 further defines the specific computational structure of the hierarchical attention module at different network levels. By employing an attention computation structure focused on spatial position information in shallow networks processing high-resolution feature maps, combined with position encoding information, the model can more accurately model the spatial dependencies between pixels, thereby enhancing its ability to preserve geometric features such as anatomical contours and edges. Simultaneously, channel-dimensional compression of the Q and K matrices at this stage effectively captures spatial information while controlling computational complexity. In deep networks processing low-resolution, high-semantic feature maps, an attention computation structure focused on channel-dimensional information is employed. Through matrix transposition operations, correlations between different feature channels can be efficiently analyzed, which is crucial for accurately characterizing the intrinsic properties of complex tissue textures. The transposition operation significantly reduces the complexity of attention computation at this stage. This specific attention computation structure, customized based on hierarchical characteristics, enables the hierarchical attention mechanism to more efficiently and specifically optimize geometric structure preservation and texture detail representation, further improving the accuracy and efficiency of overall reconstruction.

[0026] Technical Solution 3 defines the specific implementation of the dual-path feature processing module. Its low-frequency path is composed of depthwise separable convolutional layers. This type of convolution effectively extracts spatial features while having a low computational cost, making it suitable for processing and preserving the overall structure and smooth area information of an image. Its high-frequency path is composed of deformable convolutional layers and SiLU activation functions. Deformable convolution can adaptively adjust the convolution kernel sampling position based on the input features, and is particularly good at processing fine textures and edge details of irregular shapes and complex deformations; the SiLU activation function provides excellent nonlinear expression capabilities. The gated unit uses a 1x1 convolutional layer to process the spliced two-path features to generate fusion weights, which is a computationally efficient mechanism that can effectively fuse channel information. This specific network layer selection and structural design, tailored for low-frequency and high-frequency information, enables the dual-path feature processing module to more effectively separate and optimize structure preservation and detail extraction, thereby further enhancing the performance of the reconstructed image in fine texture recovery while also taking into account computational efficiency.

[0027] Technical Solution 4 refines the specific implementation method of the multi-level feature fusion module. The deconvolution layer is used for upsampling alignment, which can restore the feature maps from different levels of the decoder to a uniform spatial resolution in a learnable way. The network structure used to calculate the fusion weight adopts a global average pooling unit and a 1x1 convolution layer, followed by a Sigmoid activation function, which constitutes a lightweight channel attention mechanism that can adaptively generate channel dimension weights based on the global information of each feature map to adjust the contribution of feature maps at each level to the final fusion result. The final fusion operation is achieved through channel dimension splicing and subsequent 3x3 convolution processing. This is a standard and effective feature fusion method that can effectively integrate multi-scale feature information after weight adjustment. This specific multi-level feature fusion implementation mechanism provides a structured and learnable way to integrate cross-scale feature information, which helps to more comprehensively utilize the features of each level generated by the decoder, thereby further improving the detail richness and overall quality of the reconstructed image.

[0028] Technical Solution 5 specifically defines the internal structure of the attention-guided upsampling module. The module contains a composite attention module that sequentially connects the channel attention submodule and the spatial attention submodule. This serial structure can refine and enhance features in the channel dimension and spatial dimension in turn; the spatial attention submodule uses a 3x3 convolution layer to generate a spatial attention map, which can effectively use local neighborhood information to determine important areas in space. After CBAM processing, the PixelShuffle upsampling layer is used for resolution improvement. PixelShuffle is a computationally efficient upsampling method that can effectively avoid the checkerboard effect and is often used in high-quality image generation tasks. This specific implementation method of combining CBAM feature enhancement with PixelShuffle efficient upsampling enables more accurate enhancement and retention of important feature information at the critical stage of resolution improvement, thereby further improving the clarity and detail sharpness of the final output high-resolution image.

[0029] The discriminator structure and its training strategy for technical solution six are further clarified. In the spectral normalization U-Net architecture of the discriminator, spectral normalization is applied after each convolutional layer to ensure the Lipschitz continuity of the entire discriminator network, which is crucial for stabilizing the training process of the generative adversarial network and can effectively prevent mode collapse and improve convergence. At the same time, it is clarified that multi-scale feature matching utilizes the global features finally output by the discriminator encoder and the local features output by the intermediate layer of the decoder. This strategy enables the discriminator to not only evaluate the authenticity of the overall image, but also constrain the authenticity of the structure and texture details of the local area. Through these specific limitations on the discriminator structure and training strategy, it is possible to more effectively guide the generator to learn to generate outputs that are highly similar to real high-resolution CT images in both global structure and local details, thereby improving the realism and quality of the reconstruction results.

[0030] Technical Solution 7 provides the specific weight settings for each loss in the composite loss function. These specific weight values reflect a tuned optimization strategy in which pixel-level accuracy is given the highest priority, while giving moderate attention to perceptual similarity and generated realism, and the total variation loss provides very slight smoothing regularization. The explicit perceptual loss term is calculated based on the VGG16 network, and specific implementation details are provided. This specific weight configuration is designed to finely balance the performance of the reconstructed image in multiple aspects such as pixel fidelity, structural texture realism, visual perception quality, and smoothness according to the requirements of the CT image super-resolution task, in order to achieve the best comprehensive reconstruction effect.

[0031] Technical Solution 8 limits the number of instances of combination units, consisting of hierarchical attention modules and dual-path feature processing modules, at different levels in the generator encoder and decoder. Each downsampling unit in the encoder contains [2, 2, 1, 1] combination units, in descending order of layer depth, while each upsampling unit in the decoder contains [1, 1, 2, 2] combination units, in descending order. This uneven module distribution reflects an optimized design for computing resources and network capacity. Having more processing units in the shallow layers of the encoder and the deep layers of the decoder means investing more computing power and model complexity in these critical stages for more comprehensive feature extraction or reconstruction. Having fewer processing units in the deep layers of the encoder and the shallow layers of the decoder helps control the number of parameters and computational cost of the overall model. This specific module distribution aims to optimize resource allocation based on the processing requirements of different network layers, achieving more efficient feature learning and image reconstruction within a limited computational budget.

[0032] Technical Solution Nine clarifies the method of cross-level information transmission and integration in the generator described in Claim 1. It points out that the cross-level feature information transmission and integration between the encoder and the decoder is achieved through the multi-level feature fusion module. This is in contrast to the direct, point-to-point jump connection used in the traditional U-Net architecture. The use of the multi-level feature fusion module as the main cross-level information interaction mechanism means that the final feature reconstruction does not only rely on the encoder features of a single corresponding level, but can integrate feature information from multiple different levels of the decoder. This more comprehensive information fusion method enables the decoder to utilize richer contextual information when reconstructing high-resolution features, which helps to better coordinate the generation of global structure and local details, thereby possibly obtaining better reconstruction effects than traditional jump connections, especially when dealing with complex structures and textures.

[0033] Technical solution ten defines the data preprocessing steps required before executing step S10. This step includes obtaining paired low-resolution CT images and high-resolution CT images to form a training data set, wherein the LR image is simulated and generated by performing a downsampling operation on the HR image. This paired data preparation method provides the necessary input and target pairs for the supervised learning of the generator. Generating LR images by simulated downsampling is a commonly used and effective training data construction method when there is a lack of real paired data. In addition, this step clarifies that the training data of the discriminator contains real samples and fake samples. The definition of such true and false samples is the basis for implementing generative adversarial network training. Therefore, the data preprocessing step defined in this claim provides a standardized and necessary data basis for the effective training of subsequent deep learning network models, and is a prerequisite for implementing the entire super-resolution reconstruction method. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0035] Figure 1 A schematic diagram of a deep learning network model in a medical image super-resolution reconstruction method based on multi-level attention guidance according to an embodiment of the present invention;

[0036] Figure 2 for Figure 1 Schematic diagram of the generator;

[0037] Figure 3 for Figure 2 Schematic diagram of the backbone network;

[0038] Figure 4for Figure 3 Schematic diagram of the mid-level attention module;

[0039] Figure 5 for Figure 3 Schematic diagram of the dual-path feature processing module;

[0040] Figure 6 for Figure 2 Schematic diagram of the attention-guided upsampling module;

[0041] Figure 7 for Figure 2 Schematic diagram of the final convolution module;

[0042] Figure 8 The results of the medical image super-resolution reconstruction method based on multi-level attention guidance involved in the embodiment of the present invention are shown as follows Figure 1 ;

[0043] Figure 9 The results of the medical image super-resolution reconstruction method based on multi-level attention guidance involved in the embodiment of the present invention are shown as follows Figure 2 ;

[0044] Figure 10 The results of the medical image super-resolution reconstruction method based on multi-level attention guidance involved in the embodiment of the present invention are shown as follows Figure 3 . DETAILED DESCRIPTION

[0045] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are preferred embodiments of the present invention and should not be regarded as excluding other embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0046] In the claims, description and drawings of the present invention, if the terms "include", "have" and their variations are used, they are intended to mean "including but not limited to".

[0047] Example

[0048] An embodiment of the present invention relates to a medical image super-resolution reconstruction method based on multi-level attention guidance. This method is particularly targeted at CT images in the field of medical imaging and aims to address the shortcomings of existing super-resolution reconstruction technology in CT images, such as low restoration quality, unclear texture details, and significant noise influence. It is hoped to provide an efficient image reconstruction method that can not only improve image restoration quality and enhance the clarity of texture details, but also reduce the influence of noise.

[0049] The method comprises the following steps:

[0050] S10: Build a deep learning network model based on the generative adversarial network architecture, including a generator and a discriminator;

[0051] The generator is built based on an improved U-Net architecture and includes: an encoder that encodes and reduces the dimension of input features through several downsampling units; and a decoder that decodes and increases the dimension of the encoded features through several upsampling units;

[0052] A hierarchical attention module and a dual-path feature processing module are configured in the downsampling unit of the encoder and the upsampling unit of the decoder; the hierarchical attention module adopts an attention calculation structure that focuses on processing spatial position information or an attention calculation structure that focuses on processing channel dimension information according to the depth of the network layer in which it is located; the dual-path feature processing module includes a path for processing low-frequency information and a path for processing high-frequency information, and the outputs of the two paths are fused through a gating unit;

[0053] The generator also includes a multi-level feature fusion module and an attention-guided upsampling module; the multi-level feature fusion module is connected to the output ends of the multiple upsampling units of the decoder, and is used to receive and fuse the output features from the multiple upsampling units; the attention-guided upsampling module is connected after the multi-level feature fusion module or constitutes the final output stage of the decoder, which includes a composite attention module and a PixelShuffle upsampling layer;

[0054] The discriminator is based on a spectral normalized U-Net architecture, whose output layer is used to evaluate the authenticity probability of the input image and uses its intermediate layer features for multi-scale feature matching;

[0055] S20: training the deep learning network model, optimizing the generator and the discriminator using a composite loss function including mean absolute error loss, adversarial loss, perceptual loss, and total variation loss to obtain a super-resolution reconstruction model;

[0056] S30: Take the low-resolution image as input, and output the super-resolution image through the super-resolution reconstruction model trained by S20.

[0057] Wherein, before step S10, step S01 is also included: data preprocessing; the data preprocessing includes acquiring pairs of low-resolution images and high-resolution images to form a training data set; the low-resolution images are simulated and generated by performing a downsampling operation on the corresponding high-resolution images; the training data of the discriminator includes the high-resolution images as real samples, and the super-resolution images generated by the generator based on the corresponding low-resolution images as pseudo samples.

[0058] Specifically, refer to Figure 1 , the deep learning network model includes a generator and a discriminator. Figure 2 In the deep learning network model constructed in step S10, the generator includes a backbone network composed of an encoder, a decoder, and a multi-level feature fusion module, followed by an attention-guided upsampling module and a final convolution module. Figure 7 The final convolutional module consists of a convolutional layer, an activation function layer, and a convolutional layer. The activation function layer uses the LeakyRelu activation function. The low-resolution image passes through the backbone network, the attention-guided upsampling module, and the final convolutional module to obtain a super-resolution image.

[0059] Reference Figure 3 , which shows the structure of the backbone network of the generator, which is built based on the improved U-Net architecture, including an encoder path on the left and a decoder path on the right. In the hierarchical attention module and the dual-path feature processing module configured in the downsampling unit of the encoder and the upsampling unit of the decoder, when the hierarchical attention module is in the shallow layer of the generator network, the attention calculation structure that focuses on processing spatial position information includes a spatial attention calculation unit for combining position encoding information, and the channel dimensions of the Q matrix and the K matrix used for calculating attention are set to 1 / 4 of the number of input channels; when it is in the deep layer of the generator network, the attention calculation structure that focuses on processing channel dimension information includes a unit for performing attention calculation based on the channel dimension, and uses matrix transposition operations to reduce computational complexity. The dual-path feature processing module, wherein the path for processing low-frequency information is composed of a depth-wise separable convolution layer; the path for processing high-frequency information is composed of a deformable convolution layer and a SiLU activation function; the gating unit is composed of a 1x1 convolution layer, which is used to process the spliced low-frequency path features and the high-frequency path features to generate a fusion weight.

[0060] Specifically, the encoder portion of the backbone network includes several downsampling units, of which four are present in this embodiment. These downsampling units are responsible for performing step-by-step encoding and spatial dimensionality reduction on the input low-resolution CT image features. These downsampling units are equipped with a hierarchical feature density (VDTA) module and a dual-path feature processing module (GDFN). Each downsampling unit of the encoder contains a combination of VDTA and GDFN processing units (2, 2, 1, 1) in order of hierarchical depth, enabling efficient extraction and transformation of features at different scales.

[0061] Reference Figure 4 , wherein the hierarchical attention module (VDTA) configured in the encoder downsampling unit adopts different computational structures depending on the depth of the network layer. In the shallow layer (such as layers 1-3), the attention computation structure that focuses on processing spatial position information is adopted. Its core computation can be expressed as follows:

[0062] In this formula, Q, K, V are query, key, and value matrices, P is the position encoding matrix, ⊙ is the channel-by-channel multiplication, and d k is the channel dimension of the key / query (e.g. set to 1 / 4 of the number of input channels). This structure is designed to accurately preserve spatial structure information.

[0063] In the deep layers (such as layers 4-5), an attention calculation structure that focuses on processing channel dimension information is adopted. Its core calculation can be expressed as follows: In this formula, Q T and K are the query and key matrices after transposition. This structure effectively analyzes organizational characteristics through channel-dimensional attention calculation and reduces computational complexity.

[0064] Reference Figure 5 The dual-path feature processing module (GDFN) configured in the encoder downsampling unit includes a path for processing low-frequency information and a path for processing high-frequency information. The low-frequency path is composed of a depth-separable convolution layer to efficiently extract global structural features. The high-frequency path is composed of a deformable convolution layer and a SiLU activation function to flexibly capture fine texture and edge details. The features F output by the two paths are: low and F high After splicing, a gating unit composed of a 1x1 convolution layer and a Sigmoid activation function is used to generate the gating weight G, and the final output is F GDFN :

[0065] G=σ(Conv 1×1 (F low ||F high ))

[0066] FGDFN =G⊙F high +(1-G)⊙F low This module effectively separates and adaptively fuses structure and detail information.

[0067] The decoder part in the backbone network also follows the upsampling path design of the improved U-Net architecture and includes several (for example, 4 according to this embodiment) upsampling units. These upsampling units are responsible for decoding and spatial dimensioning the features transmitted by the encoder (integrated through a multi-level feature fusion module). Similar to the encoder, these upsampling units are also equipped with a hierarchical attention module (VDTA) and a dual-path feature processing module (GDFN), and their structures are similar to those in the encoder. Each upsampling unit of the decoder includes [1,1,2,2] VDTA and GDFN combination processing units in sequence according to the hierarchical depth (from deep to shallow).

[0068] The multi-level feature fusion module uses a deconvolution layer to perform upsampling and alignment of feature maps output by multiple upsampling units in the decoder. The network structure used to calculate fusion weights includes a global average pooling unit and a 1x1 convolution layer followed by a sigmoid activation function. The fusion operation is achieved by channel-dimensional splicing of the weighted feature maps and subsequent 3x3 convolution processing. Furthermore, in the generator, the encoder and decoder are connected through the multi-level feature fusion module to achieve the transmission and integration of cross-level feature information.

[0069] The multi-level feature fusion module is the key to the improved U-Net architecture of this invention, replacing the traditional skip connection. It is connected to the output of multiple (for example, all) upsampling units of the decoder, receiving these feature maps containing different scales and different semantic levels. For the feature map from layer l, the module first uses the deconvolution layer to perform 2 (L-l) times upsampling, aligning its spatial size with the shallowest layer of the decoder (or target output size), and obtaining Then, a lightweight network (including global average pooling GAP, 1x1 convolution and Sigmoid activation function) is used to transform each aligned feature map. Calculate an adaptive fusion weight Finally, the weighted feature maps of all levels are spliced in the channel dimension, and the final feature fusion and channel adjustment are performed through a 3x3 convolution layer to obtain the final output features of the backbone network. This module can produce richer and more refined feature expressions than traditional U-Net skip connections through comprehensive and adaptive weighted fusion of multi-level information of the decoder, which serves as the input for subsequent processing (attention-guided upsampling and final convolution). Indicates channel dimension splicing.

[0070] Overall, the backbone network in the embodiment of the present invention forms a basic U-Net skeleton through an encoder and a decoder, and deeply integrates a hierarchical attention module (VDTA) based on a hierarchical adaptive adjustment strategy and a dual-path feature processing module (GDFN) for fine-grained processing of structures and details. At the same time, an innovative multi-level feature fusion module (MDF) replaces the traditional skip connection to integrate the multi-scale information of the decoder, and finally outputs a high-quality feature map to be sent to subsequent modules for high-resolution image generation.

[0071] The high-quality feature map F output by the backbone network fuse , is input to the attention-guided upsampling module in the generator. Figure 7 , which shows the structure of the attention-guided upsampling module. The module comprises a composite attention module connected in series with a channel attention submodule and a spatial attention submodule. The spatial attention submodule uses a 3x3 convolutional layer to generate a spatial attention map. The PixelShuffle upsampling layer is used to increase the feature map resolution by a predetermined multiple. This module aims to enhance key features through the attention mechanism and perform final high-resolution upsampling.

[0072] Reference Figure 6 , the attention-guided upsampling module first includes a convolutional layer (Conv) for preliminary processing and channel adjustment of the input fusion features. After the preliminary convolution processing, the feature map is sent to the PixelShuffle upsampling layer. The PixelShuffle layer increases the resolution of the feature map to the target high-resolution size according to the preset upsampling factor (for example, for 4x super-resolution, it may include two consecutive 2x PixelShuffle operations, or a direct 4x operation). This layer improves the spatial resolution by rearranging the information in the channel dimension of the low-resolution feature map. It is an efficient upsampling method that can reduce artifacts. Immediately following the PixelShuffle layer, a PReLU activation function layer (Parametric Rectified Linear Unit) is connected to introduce nonlinearity to the upsampled features.

[0073] Next, the upsampled and activated feature maps are fed into a composite attention module (CBAM). The CBAM module sequentially consists of a channel attention submodule and a spatial attention submodule. The channel attention submodule learns the importance weights of different channels through global pooling and a shared multi-layer perceptron (MLP) and recalibrates the feature channels. The spatial attention submodule generates a spatial attention map through a pooling operation along the channel dimension followed by a 3x3 convolutional layer. This map indicates which spatial regions are more important. Through this serial channel and spatial attention mechanism, the CBAM module can adaptively refine and enhance the feature map, highlighting the key information related to reconstructing high-quality images.

[0074] Finally, the high-resolution feature map (output of CBAM) processed by the attention-guided upsampling module is fed into the final convolution module. This module is responsible for converting the high-resolution feature map into the final super-resolution image I SR . The final convolution module can be composed of a series of convolution layers and activation functions. In a specific implementation of this embodiment, it can include a convolution layer (Conv) for further integrating features; then a LeakyReLU activation function (LeakyRectified Linear Unit) to provide nonlinearity; and finally a convolution layer (Conv) that usually has a smaller convolution kernel (such as 3x3) and adjusts the number of output channels to the number of channels of the target image (for example, 3 channels for RGB images). This module completes the mapping from deep features to final pixel values, generating the final super-resolution CT image I SR .

[0075] In the deep learning network model constructed in step S10, in addition to the generator, a discriminator is also included. The discriminator is intended to distinguish between the super-resolution image generated by the generator and the real high-resolution CT image, so as to drive the generator to produce more realistic results. The discriminator adopts a spectral normalization U-Net architecture. In the discriminator, the input super-resolution image is the image generated by the generator, and the high-resolution image is the image after the provided real image is magnified four times. The discriminator determines whether the two images belong to the same image and returns the corresponding probability value. Its structure includes an encoder path (downsampling) and a decoder path (upsampling), and fuses features of different levels through skip connections, which is similar to the U-Net skeleton of the generator, but usually the number of layers and channels can be different.

[0076] In which, spectral normalization is applied after each convolution layer in the spectral normalization U-Net architecture of the discriminator; the multi-scale feature matching is performed using the global features finally output by the encoder of the discriminator and the local features output by the intermediate layer of the decoder.

[0077] The discriminator's encoder path consists of multiple downsampling units, each of which includes a convolutional layer (Conv), a spectral normalization layer (SN), and a LeakyReLU activation function. Spectral normalization is applied to each convolutional layer to help stabilize the GAN training process. The decoder path consists of multiple upsampling units, including an upsampling operation (Upsample), feature concatenation (Concatenate) with the features of the corresponding encoder layer, a convolutional layer (Conv), a spectral normalization layer (SN), and a LeakyReLU activation function. To achieve multi-scale feature matching, the training process utilizes the global features output by the discriminator encoder (representing the judgment of the entire image) and the local features output by the intermediate layers of the decoder (representing the judgment of details and texture at different scales). These extracted features are used to calculate additional loss terms (such as L1 or L2 loss) and added to the generator's overall loss to constrain the authenticity of the generated image at different feature levels. The final output layer of the discriminator may include global average pooling, convolutional layer / fully connected layer and Sigmoid activation function, and output a probability value between 0 and 1, indicating the probability that the input image is a real image.

[0078] When training the deep learning network model in step S20, the generator and the discriminator are optimized simultaneously by adversarial training. The core of the training process is to minimize a composite loss function L SR , which is used to guide the optimization of the generator. Specifically, the loss function of the generator during training is: in, is the mean absolute error function, To counter the loss function, is the perceptual loss function, is the total variation loss function.

[0079] Among them, the Mean Absolute Error (MAE) is a commonly used loss function in regression tasks, which is used to measure the average absolute difference between the model's predicted value and the actual observed value. In the field of medical image reconstruction, MAE has significant advantages over the Mean Square Error (MSE). First, by calculating the average value of the absolute difference, MAE can effectively deal with outliers in medical images and improve the model's robustness to noise and artifacts. Secondly, compared with MSE, MAE pays more attention to preserving the details and texture of the image, which helps to maintain the clarity of small structures and edges in medical images, and provide more reliable visual information for medical professionals. Its specific formula is shown below, where y i represents the mean of the i-th sample, n is the number of samples, Represents the predicted value of the i-th sample:

[0080] Adversarial loss (Generative Adversarial Network Loss, GAN Loss) is a loss function in generative adversarial networks. Adversarial loss promotes the model to learn to generate more realistic samples by adjusting the adversarial process between the generator and the discriminator. The main reason for using adversarial loss in the model of this embodiment is that this training mechanism helps to improve the performance of the generator, enabling it to more accurately restore the detailed features of medical images, thereby achieving more outstanding reconstruction effects in the field of medical imaging. The specific formula is as follows, where G is the parameter of the generator, D is the parameter of the discriminator, x is a sample from the real data distribution, z is a noise sample from the latent space, pdata(x) is the distribution of real data, and p z (z) is the distribution in the latent space:

[0081] Perceptual loss (PerceptualLoss) is a key loss function that aims to quantify the similarity between generated images and real images at a higher semantic level by comparing their feature expressions in pre-trained deep neural networks. In tasks such as medical image reconstruction, this loss function plays an important role. It can comprehensively evaluate the similarity between generated images and real images, while paying attention to pixel-level details and image semantic information. This embodiment uses the feature representation of perceptual loss in the pre-trained VGG16 network to guide network learning, so as to effectively improve the quality of generated images and better preserve the key structures and information in medical images. Perceptual loss is indispensable in medical image reconstruction tasks, and provides the model of this embodiment with more powerful performance and more reliable results. The specific formula is shown below, where W i,j and H i,j Represent the dimensions of the corresponding feature maps of the generated images and real images in the VGG network:

[0082] Total variation loss

[13] is a loss function in image processing tasks, which aims to make the generated image have a smoother appearance. In medical image reconstruction, considering the particularity of medical images, it is necessary to accurately control noise and artifacts. Therefore, the introduction of total variation loss can effectively reduce the influence of noise and make the generated medical image smoother. This smoothing effect helps to improve the visual quality of medical images while reducing noise that may interfere with medical image diagnosis. In addition, the introduction of total variation loss also helps to retain key detail information in medical images, ensuring that the generated image is more accurate in expressing small structures and lesion areas. For a two-dimensional image I, the specific formula of its total variation loss is as follows: TV(I) = ∑ i,j |I i+1,j -I i,j |+|I i,j+1 -I i,j |.

[0083] To further illustrate the effectiveness of the above-mentioned multi-level attention-based image super-resolution reconstruction method, a specific embodiment and related comparative examples are provided below.

[0084] The experiment selected several public and commonly used CT image datasets, including QingLungCT, Covid_19, and KaggleLune datasets. These datasets cover different pathological conditions and imaging conditions, and can comprehensively evaluate the performance and generalization ability of the model.

[0085] In this embodiment, all models were trained for a total of 2000 rounds on the NVIDIA GeForce RTX 3090 GPU. The batch size (batch_size) of each model was set to 32, and the Adam optimizer was used for parameter optimization. In terms of training input, this embodiment used cropped high-resolution (HR) images with a size of 128*128 pixels. For comparison, this embodiment selected methods such as EDSR, ESPCN, SRGAN, SRResNet, SRGAN and ESRGAN for comparison. Among them, SRGAN and SRResNet both adopted a 7-layer residual network structure, while in ESRGAN, this embodiment selected a more complex 16-layer Residual in Residual Dense Block (RRDB) network structure. The final performance evaluation of all models is based on the two indicator values of peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) calculated on the YCbCr color brightness channel of the image. This choice helps to more comprehensively and accurately evaluate the performance of the model in the image reconstruction task. Among them, PSNR mainly measures the difference in pixel values between the reconstructed image and the original high-resolution image. The higher the value, the smaller the distortion. SSIM measures the similarity between images from three aspects: brightness, contrast, and structure. The closer the value is to 1, the more similar the structure is, which is more consistent with human visual perception.

[0086] After training, this embodiment obtains the corresponding super-resolution reconstruction neural network model. Under the condition of four times magnification, the PSNR and SSIM values of this model and other super-resolution reconstruction models are shown in the following table:

[0087] Table 1: PSNR values of different methods on various datasets

[0088]

[0089] Table 2: SSIM values of different methods on various datasets

[0090]

[0091] The analysis of the above experimental results shows that the CT image super-resolution reconstruction method based on multi-level attention proposed in the embodiment of the present invention shows significant performance advantages over the current mainstream and classic methods on the three different CT datasets tested (QingLungCT, Covid_19, KaggleLune). Figures 8 to 10 , respectively show the reconstruction results of this embodiment and other reconstruction methods in three different CT data sets, where Ours represents the super-resolution image corresponding to the reconstruction method designed in this embodiment.

[0092] First of all, from the perspective of the peak signal-to-noise ratio (PSNR) indicator (Table 1), PSNR measures the fidelity of the reconstructed image and the original high-resolution image at the pixel value level. The data show that the method of the present invention has achieved the highest PSNR value on all three data sets. For example, it reached 41.54dB on the KaggleLune data set, 37.31dB on the Covid_19 data set, and 35.87dB on the QingLungCT data set, which are significantly higher than other comparison methods. This directly proves that the present method can more accurately restore the pixel information of the image, effectively solves the problem of "low restoration quality" mentioned in the background technology, and has outstanding performance in ensuring geometric accuracy. The pixel-level difference between the generated image and the real high-resolution image is smaller.

[0093] Secondly, from the perspective of the structural similarity index (SSIM) indicator (Table 2), SSIM focuses more on evaluating the similarity of images in terms of structure, brightness and contrast, which is more in line with the perception of the human visual system. The method of the present invention also obtained the highest SSIM value on all three data sets, for example, it reached 0.9855 on KaggleLune, 0.9616 on Covid_19, and 0.9320 on QingLungCT. The high SSIM value means that the image reconstructed by this method can better maintain the structural information of the original image, which directly responds to the problem in the background technology about the need to restore "clear texture details" and avoid "unreasonable anatomical structure". By better maintaining the local structure (such as edges, texture direction) and contrast relationship of the image, the image generated by this method is visually closer to the original high-resolution image, reduces the sense of blur, and ensures the coherence and rationality of the reconstructed anatomical structure (such as organ boundaries, vascular networks).

[0094] Considering the two indicators of PSNR and SSIM, the method of the present invention achieves optimal performance in both pixel-level fidelity and structural similarity. This shows that the method successfully balances the requirements of geometric accuracy and texture authenticity mentioned in the background technology. High PSNR ensures the accurate recovery of the basic structure, while high SSIM ensures the authenticity of the structural details and visual perception. This comprehensive improvement is the result of the synergistic effect of its internal specific modules: the hierarchical attention module helps to maintain the accuracy of the anatomical structure (contributing to the high PSNR and structural terms in SSIM), dual-path processing, multi-level fusion and attention-guided upsampling enhance the recovery of fine texture details (contributing to high SSIM and visual clarity), and the GAN framework and composite loss function jointly drive the improvement of overall image quality and realism (contributing to both indicators).

[0095] In addition, this method shows consistent superiority on different types of datasets, which to some extent reflects its robustness and generalization ability. Whether it is lung CT (QingLungCT, Covid_19) or abdominal CT (KaggleLune includes spleen, etc.), this method can achieve the best results. This indirectly shows that it has good adaptability to different anatomical regions and different image characteristics (possibly including different noise levels), and relatively reduces sensitivity to specific data distributions, thus partially solving the problem of "being greatly affected by noise" mentioned in the background technology.

[0096] In summary, the experimental results, through quantitative comparisons of PSNR and SSIM metrics, as well as observations of actual reconstructed super-resolution images, fully demonstrate the significant superiority of the method proposed in this embodiment of the present invention over existing techniques. This method can generate high-resolution CT images with greater pixel fidelity, more similar structures, and clearer details. It effectively addresses key issues identified in prior art, such as low restoration quality, blurred texture details, and structural distortion, demonstrating its significant potential for improving CT image quality and assisting clinical diagnosis.

[0097] The above description and embodiments are intended to explain the scope of protection of the present invention, but do not constitute a limitation thereto. Modifications, equivalent substitutions, or other improvements to the embodiments of the present invention or portions thereof that can be obtained by a person of ordinary skill in the art through logical analysis, reasoning, or limited experimentation based on the teachings of the present invention or the above embodiments, combined with common knowledge, ordinary technical knowledge in the field, and / or prior art, should all be included within the scope of protection of the present invention.

Claims

1. A medical image super-resolution reconstruction method based on multi-level attention guidance, characterized by: The following steps are involved: S10: Build a deep learning network model based on the generative adversarial network architecture, including a generator and a discriminator; The generator is constructed based on an improved U-Net architecture and includes: an encoder that encodes and reduces the dimensionality of input features through a number of downsampling units; and a decoder that decodes and increases the dimensionality of the encoded features through a number of upsampling units; wherein, a hierarchical attention module and a dual-path feature processing module are configured in the downsampling units of the encoder and the upsampling units of the decoder; the hierarchical attention module adopts an attention calculation structure that focuses on processing spatial position information or an attention calculation structure that focuses on processing channel dimension information according to the depth of the network layer in which it is located; the dual-path feature processing module includes a path for processing low-frequency information and a path for processing high-frequency information, and the outputs of the two paths are fused through a gating unit; The generator also includes a multi-level feature fusion module and an attention-guided upsampling module; the multi-level feature fusion module is connected to the output ends of multiple upsampling units of the decoder, and is used to receive and fuse the output features from the multiple upsampling units; the attention-guided upsampling module is connected after the multi-level feature fusion module or constitutes the final output stage of the decoder, which includes a composite attention module and a PixelShuffle upsampling layer; the discriminator is based on a spectral normalized U-Net architecture, and its output layer is used to evaluate the authenticity probability of the input image and use its intermediate layer features for multi-scale feature matching; S20: training the deep learning network model, optimizing the generator and the discriminator using a composite loss function including mean absolute error loss, adversarial loss, perceptual loss, and total variation loss to obtain a super-resolution reconstruction model; S30: Take the low-resolution image as input, and output the super-resolution image through the super-resolution reconstruction model trained by S20.

2. The medical image super-resolution reconstruction method based on multi-level attention guidance according to claim 1, characterized in that: The hierarchical attention module, when it is in the shallow layer of the generator network, the attention calculation structure focusing on processing spatial position information includes a spatial attention calculation unit for combining position encoding information, and the channel dimension of the Q matrix and K matrix used for calculating attention is set to 1 / 4 the number of input channels; When it is in the deep layer of the generator network, the attention calculation structure that focuses on processing channel dimension information includes a unit for performing attention calculation based on the channel dimension, and uses matrix transposition operations to reduce computational complexity.

3. The medical image super-resolution reconstruction method based on multi-level attention guidance according to claim 1, characterized in that: The dual-path feature processing module, wherein the path for processing low-frequency information is composed of a depth-wise separable convolution layer; the path for processing high-frequency information is composed of a deformable convolution layer and a SiLU activation function; the gating unit is composed of a 1x1 convolution layer, which is used to process the spliced low-frequency path features and the high-frequency path features to generate a fusion weight.

4. The medical image super-resolution reconstruction method based on multi-level attention guidance according to claim 1, characterized in that: The multi-level feature fusion module uses a deconvolution layer to implement the upsampling and alignment operation of the feature maps output by multiple upsampling units of the decoder; the network structure for calculating the fusion weight includes a global average pooling unit and a 1x1 convolution layer, followed by a Sigmoid activation function; The fusion operation is achieved by splicing the channel dimension of the weighted feature map and subsequent 3x3 convolution processing.

5. The medical image super-resolution reconstruction method based on multi-level attention guidance according to claim 1, characterized in that: The attention-guided upsampling module includes a composite attention module that sequentially connects a channel attention sub-module and a spatial attention sub-module, and the spatial attention sub-module uses a 3x3 convolution layer to generate a spatial attention map; the PixelShuffle upsampling layer is used to increase the feature map resolution by a predetermined multiple.

6. The medical image super-resolution reconstruction method based on multi-level attention guidance according to claim 1, characterized in that: The discriminator applies spectral normalization after each convolution layer in its spectral normalization U-Net architecture; the multi-scale feature matching is performed using the global features finally output by the encoder of the discriminator and the local features output by the intermediate layer of the decoder.

7. The medical image super-resolution reconstruction method based on multi-level attention guidance according to claim 1, characterized in that: The loss function of the generator during training is: in, is the mean absolute error function, To counter the loss function, is the perceptual loss function, is the total variation loss function.

8. The medical image super-resolution reconstruction method based on multi-level attention guidance according to claim 1, characterized in that: The encoder and decoder of the generator each include a predetermined number of layers; the number of instances of the hierarchical attention module and the dual-path feature processing module configured in each downsampling unit of the encoder is [2, 2, 1, 1] in order of layer depth; the number of instances of the combination module configured in each upsampling unit of the decoder is [1, 1, 2, 2] in order of layer depth.

9. The medical image super-resolution reconstruction method based on multi-level attention guidance according to claim 1, characterized in that: In the generator, the encoder and decoder realize the transmission and integration of cross-level feature information through the multi-level feature fusion module.

10. The medical image super-resolution reconstruction method based on multi-level attention guidance according to claim 1, characterized in that: Before step S10, the method further includes step S01: data preprocessing; the data preprocessing includes acquiring pairs of low-resolution images and high-resolution images to form a training data set; the low-resolution images are simulated and generated by performing a downsampling operation on the corresponding high-resolution images; the training data of the discriminator includes the high-resolution images as real samples, and the super-resolution images generated by the generator based on the corresponding low-resolution images as pseudo samples.

Images