Medical image reconstruction method and system based on deep learning

Abstract

The invention discloses a medical image reconstruction method and system based on deep learning, and the method comprises the steps: firstly employing a pre-training convolutional neural network to extract initial features for a low-quality image obtained by a medical imaging device, enhancing the edge details through gradient calculation and a channel attention mechanism, and carrying out the image reconstruction through an adversarial generative network. And then, the reconstructed image is corrected in combination with a physical imaging model and an equipment deviation mode, and the image quality is further improved through a multi-scale convolutional network and a residual refining network. And finally, a dual-path network is adopted to generate a heat map to locate a suspicious lesion area, marking is performed according to pathology priori knowledge, and a diagnosis auxiliary image is generated. The medical image quality can be effectively improved, and the visualization effect of the lesion area is enhanced.

Description

Technical Field

[0001] The present invention belongs to the field of medical imaging technology, and in particular relates to a medical image reconstruction method and system based on deep learning. Background Art

[0002] Medical image reconstruction is a key technology in the field of medical imaging, and is of great significance for subsequent doctors to improve diagnostic accuracy and treatment effects. High-quality medical images can clearly present the anatomical structure and lesion characteristics of the human body, providing a reliable basis for clinical decision-making. However, when dealing with incomplete and low-quality data, existing medical image reconstruction methods often find it difficult to balance the preservation of image details with the suppression of artifacts, resulting in limitations in the clinical application of reconstructed images. Traditional methods rely on manually designed filters or simple mathematical models, which are difficult to adapt to the needs of different imaging devices and complex pathological scenarios, and the reconstruction results often reduce the reference value due to noise, bias or distortion. In deep learning-based medical image reconstruction, the authenticity of the reconstructed image faces several core challenges. Image reconstruction requires restoring high-quality images from incomplete or low-quality data, but the noise and incompleteness of the input data make it difficult for the model to accurately learn the mapping relationship between image features and complete images. This directly leads to the possibility that the reconstructed image may lose key anatomical details or introduce artifacts that do not conform to physical laws. At the same time, due to the systematic deviations introduced by different imaging devices and parameter settings, the reconstructed image may show inconsistency between different devices, affecting the reliability of multicenter studies and remote diagnosis. In addition, medical images usually contain complex textures and lesion areas. The model needs to highlight the details of key areas while ensuring the overall image quality, which increases the complexity of the reconstruction task. Therefore, how to design a deep learning method that can accurately restore real and consistent medical images from incomplete data, while adaptively correcting device deviations and highlighting lesion areas, has become a key issue in the field of medical image reconstruction. Summary of the invention

[0003] In order to solve the above technical problems, the present invention provides a medical image reconstruction method and system based on deep learning. Among them, a medical image reconstruction method based on deep learning includes:

[0004] Low-quality image data is acquired based on medical imaging equipment, and initial features are extracted using a pre-trained convolutional neural network to generate a first feature set containing texture and structure information;

[0005] Performing gradient calculation processing on the first feature set to generate a second feature set including edge distribution features;

[0006] According to the comparison result of the edge detail intensity of the second feature set and the preset threshold, a channel attention mechanism is used to process and generate a third feature set;

[0007] Input the third feature set into a generative adversarial network to generate a preliminary reconstructed image, and optimize to generate a first reconstructed image after evaluating the feature distribution difference through a discriminator network;

[0008] Perform artifact correction processing on the first reconstructed image to generate a second reconstructed image that satisfies physical constraints;

[0009] Perform intensity correction on the second reconstructed image according to a pre-trained device bias pattern to generate a third reconstructed image;

[0010] Use a multi-scale convolutional network to perform feature extraction and fusion processing on the third reconstructed image to generate a fourth reconstructed image;

[0011] Input the fourth reconstructed image into a residual refinement network for detail enhancement processing to generate a final high-quality reconstructed image;

[0012] Extract feature data from the final high-quality reconstructed image, and use a dual-path network to generate a heat map and locate suspicious lesion areas;

[0013] Perform type annotation processing on the suspicious lesion areas according to pathological prior knowledge to generate a diagnostic assistance image.

[0014] Preferably, the process of generating the first feature set including texture and structure information includes:

[0015] Obtain low-quality image data output by a medical imaging device, and use a pre-trained convolutional neural network to extract initial features to generate a first feature set including texture and structure information;

[0016] Process the low-quality image data corresponding to the first feature set through adaptive histogram equalization to enhance the image contrast and obtain an image data set;

[0017] Use a pre-trained convolutional neural network to perform secondary feature extraction on the image data set to generate an information feature set including enhanced texture and structure information;

[0018] If the texture information entropy value of the information feature set is lower than a preset threshold, optimize the texture of the information feature set through a generative adversarial network to obtain an optimized feature set; otherwise, directly output the information feature set as the optimized feature set;

[0019] According to the optimized feature set, perform dimensionality reduction processing on the texture and structure information by using the principal component analysis method to obtain a dimensionality-reduced feature set;

[0020] Use a pre-trained convolutional neural network to perform classification prediction on the dimensionality-reduced feature set to judge the quality level of the image data and obtain a classification result;

[0021] According to the classification results, the weighted fusion method is used to integrate the texture and structure information of the reduced dimension feature set to generate the target feature set.

[0022] Preferably, the process of performing gradient calculation processing on the first feature set to generate a second feature set including edge distribution features further includes:

[0023] Based on the second feature set, a feature extraction method is used to separate edge distribution features from non-edge distribution features to obtain an edge feature subset;

[0024] If the distribution characteristics of the edge feature subset meet the preset threshold, the feature conversion method is determined through data analysis to obtain the converted feature set; if not, the first feature set is returned to re-execute the gradient calculation;

[0025] According to the converted feature set, the random forest algorithm is applied to sort the features according to their importance to obtain the sorted feature set;

[0026] For the sorted feature set, a data processing method is used to remove features whose importance is lower than a preset threshold to obtain a streamlined feature set;

[0027] A feature generation operation is performed based on the simplified feature set to construct an enhanced feature set including distribution characteristics. A clustering algorithm is used on the enhanced feature set to determine the distribution correlation between features to obtain a target feature set.

[0028] Preferably, according to the comparison result of the edge detail intensity of the second feature set and a preset threshold, a channel attention mechanism is used for processing to generate a third feature set, which includes:

[0029] Obtain edge detail information of the second feature set, extract edge strength using an edge detection algorithm, and obtain edge strength distribution;

[0030] If the intensity of at least one area in the edge intensity distribution is lower than a preset threshold, the Sobel operator is used to enhance the edge details and generate enhanced edge features;

[0031] According to the enhanced edge features, a channel attention mechanism is used to assign weights to obtain a weighted feature set;

[0032] Using the weighted feature set, a convolutional neural network is used to extract deep features to generate a third feature set;

[0033] If the feature dimension of the third feature set exceeds a preset range, principal component analysis is used to reduce the dimension to obtain a reduced-dimensional optimized feature set;

[0034] According to the dimension reduction optimization feature set, mean pooling is used to determine a third feature set of final feature representation.

[0035] Preferably, the process of optimizing the generation of the first reconstructed image after evaluating the feature distribution difference through a discriminator network includes:

[0036] Normalize the third feature set to obtain a normalized feature set;

[0037] Process the normalized feature set through an adversarial generation network to generate a preliminary reconstructed image;

[0038] Use a discriminator network to evaluate the feature distribution difference between the preliminary reconstructed image and the target image to obtain a distribution difference value;

[0039] If the distribution difference value is greater than a preset threshold, adjust the parameters of the adversarial generation network through an optimization algorithm to generate an optimized reconstructed image;

[0040] Evaluate and analyze the clarity of the optimized reconstructed image to obtain a quality evaluation result;

[0041] According to the quality evaluation result, post-process the optimized reconstructed image to obtain the first reconstructed image.

[0042] Preferably, the process of performing artifact correction processing on the first reconstructed image to generate a second reconstructed image that satisfies physical constraints includes:

[0043] Obtain the pixel intensity data and artifact distribution characteristics of the first reconstructed image, and use a feature extraction algorithm to determine the artifact area to obtain an artifact distribution feature map;

[0044] According to the artifact distribution feature map, extract imaging model parameters from a pre-established physical imaging model to determine a parameter set for correction processing;

[0045] Use a constrained optimization method, combined with physical constraint conditions and imaging model parameters, to perform optimization calculations on the artifact distribution feature map to obtain a corrected pixel intensity distribution;

[0046] Reconstruct the spatial consistency of the image through the corrected pixel intensity distribution to generate a preliminary corrected image;

[0047] If the image reconstruction quality of the preliminary corrected image is lower than a preset threshold, use an iterative optimization algorithm to adjust the pixel intensity to obtain an optimized corrected image, and verify whether the optimized corrected image satisfies the physical constraint conditions to obtain a second reconstructed image;

[0048] Calculate the spatial consistency and artifact residue degree of the second reconstructed image to judge the accuracy of the correction algorithm to obtain a finally corrected second reconstructed image.

[0049] Preferably, the process of performing intensity correction on the second reconstructed image according to a pre-trained device deviation pattern to generate a third reconstructed image includes:

[0050] Obtain pixel intensity data from the second reconstructed image, and perform matching through a pre-trained device deviation pattern to obtain a deviation distribution feature;

[0051] If the deviation distribution feature exceeds a preset threshold, an intensity correction algorithm is used to adjust the pixel intensity to generate corrected intensity data;

[0052] According to the corrected intensity data, combined with an image reconstruction algorithm, a preliminary third reconstructed image is generated;

[0053] Obtain the edge feature of the preliminary third reconstructed image, and verify the deviation compensation effect through pattern matching to obtain a verification result;

[0054] If the verification result shows insufficient deviation compensation, the edge feature is intensity-adjusted to generate an optimized third reconstructed image;

[0055] Through an image generation algorithm, pixel smoothing processing is performed on the optimized third reconstructed image to obtain a final third reconstructed image;

[0056] Extract quality parameters from the final third reconstructed image, and evaluate through a preset quality standard to obtain a third reconstructed image that completes image correction.

[0057] Preferably, the process of using a multi-scale convolutional network to perform feature extraction and fusion processing on the third reconstructed image to generate a fourth reconstructed image includes:

[0058] Perform feature extraction on the third reconstructed image through a preset multi-scale convolutional network to obtain a multi-scale feature set;

[0059] Use a multi-scale convolutional network to perform fusion processing on the multi-scale feature set to generate a fused feature representation;

[0060] If the resolution of the fused feature representation is lower than a preset threshold, the feature resolution is adjusted through an upsampling operation to obtain an adjusted feature representation;

[0061] Perform reconstruction processing on the adjusted feature representation through a decoding network to generate a fourth reconstructed image;

[0062] Obtain the feature space distribution of the fourth reconstructed image, and determine whether it meets a preset feature consistency condition. If the feature consistency condition is not met, adjust the feature extraction and fusion process by optimizing network parameters to obtain an optimized fourth reconstructed image;

[0063] According to the optimized fourth reconstructed image, use a preset evaluation function to calculate the image quality and determine the finally output fourth reconstructed image.

[0064] Preferably, the process of extracting feature data from the final high-quality reconstructed image and generating a heat map and locating a suspicious lesion area using a dual-path network includes:

[0065] Obtain pixel-level feature data from the final high-quality reconstructed image, and use a preprocessing algorithm to remove noise to obtain a denoised feature set;

[0066] Use a convolutional neural network to perform preliminary feature extraction on the denoised feature set to generate a feature set after feature extraction;

[0067] Process the feature set after feature extraction through a dual-path network to generate first heat map data;

[0068] If the significance value of the first heat map data is greater than a preset threshold, use a clustering algorithm to segment the heat map data to determine the suspicious lesion area;

[0069] Extract regional boundary features based on the segmented suspicious lesion area to generate second heat map data;

[0070] Perform coordinate mapping on the second heat map data through a geometric analysis algorithm to obtain the precise position coordinates of the lesion area;

[0071] If the deviation of the precise position coordinates is less than a preset threshold, output the final localization result to determine the lesion area.

[0072] The present invention also provides a medical image reconstruction system based on deep learning, including:

[0073] An initial feature extraction module, configured to obtain low-quality image data based on a medical imaging device, and use a pre-trained convolutional neural network to extract initial features to generate a first feature set including texture and structure information;

[0074] A gradient feature generation module, configured to perform gradient calculation processing on the first feature set to generate a second feature set including edge distribution features;

[0075] An attention processing module, configured to perform processing using a channel attention mechanism according to the comparison result between the edge detail intensity of the second feature set and a preset threshold to generate a third feature set;

[0076] An adversarial generation module, configured to input the third feature set into an adversarial generation network to generate a preliminary reconstructed image, and optimize and generate a first reconstructed image after evaluating the feature distribution difference through a discriminator network;

[0077] An artifact correction module, configured to perform artifact correction processing on the first reconstructed image to generate a second reconstructed image that satisfies physical constraints;

[0078] An intensity correction module for intensity-correcting the second reconstructed image according to a pre-trained device deviation pattern to generate a third reconstructed image;

[0079] A multi-scale fusion module for performing feature extraction and fusion processing on the third reconstructed image by using a multi-scale convolutional network to generate a fourth reconstructed image;

[0080] A residual refinement module for inputting the fourth reconstructed image into a residual refinement network for detail enhancement processing to generate a final high-quality reconstructed image;

[0081] A lesion localization module for extracting feature data from the final high-quality reconstructed image, and using a dual-path network to generate a heat map and locate a suspicious lesion area;

[0082] A diagnostic annotation module for performing type annotation processing on the suspicious lesion area according to pathological prior knowledge to generate a diagnostic auxiliary image.

[0083] Compared with the prior art, the present invention has the following advantages and technical effects:

[0084] For the low-quality images obtained by medical imaging devices, the present invention first extracts initial features by using a pre-trained convolutional neural network, enhances edge details through gradient calculation and channel attention mechanism, and then uses a generative adversarial network for image reconstruction. Subsequently, the present invention corrects the reconstructed image by combining a physical imaging model and a device deviation pattern, and further improves the image quality through a multi-scale convolutional network and a residual refinement network. Finally, the present invention uses a dual-path network to generate a heat map to locate the suspicious lesion area and performs annotation according to pathological prior knowledge to generate a diagnostic auxiliary image. This method can effectively improve the quality of medical images, enhance the visualization effect of the lesion area, and provide strong support for subsequent clinical diagnosis by doctors. Description of the Drawings

[0085] The drawings constituting a part of this application are used to provide a further understanding of this application. The schematic embodiments of this application and their descriptions are used to explain this application and do not constitute an improper limitation to this application. In the drawings:

[0086] Figure 1 is a schematic flowchart of the method according to an embodiment of the present invention;

[0087] Figure 2 is a schematic structural diagram of the system according to an embodiment of the present invention. Detailed Embodiments

[0088] It should be noted that, without conflict, the embodiments in this application and the features in the embodiments can be combined with each other. The following will refer to the drawings and combine the embodiments to detail this application.

[0089] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. And although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.

[0090] Embodiment 1

[0091] As Figure 1 shown, in this embodiment, a medical image reconstruction method based on deep learning is provided, including:

[0092] Obtain low-quality image data based on a medical imaging device, and use a pre-trained convolutional neural network to extract initial features to generate a first feature set containing texture and structure information;

[0093] Perform gradient calculation processing on the first feature set to generate a second feature set containing edge distribution features;

[0094] According to the comparison result between the edge detail intensity of the second feature set and a preset threshold, use a channel attention mechanism for processing to generate a third feature set;

[0095] Input the third feature set into a generative adversarial network to generate a preliminary reconstructed image, and optimize to generate a first reconstructed image after evaluating the feature distribution difference through a discriminator network;

[0096] Perform artifact correction processing on the first reconstructed image to generate a second reconstructed image that meets physical constraints;

[0097] Perform intensity correction on the second reconstructed image according to a pre-trained device deviation pattern to generate a third reconstructed image;

[0098] Use a multi-scale convolutional network to perform feature extraction and fusion processing on the third reconstructed image to generate a fourth reconstructed image;

[0099] Input the fourth reconstructed image into a residual refinement network for detail enhancement processing to generate a final high-quality reconstructed image;

[0100] Extract feature data from the final high-quality reconstructed image, and use a dual-path network to generate a heat map and locate suspicious lesion areas;

[0101] Perform type annotation processing on the suspicious lesion areas according to pathological prior knowledge to generate a diagnostic auxiliary image.

[0102] Furthermore, the process of generating the first feature set containing texture and structure information includes:

[0103] Obtain low-quality image data output by a medical imaging device, and use a pre-trained convolutional neural network to extract initial features to generate a first feature set containing texture and structure information;

[0104] Process the low-quality image data corresponding to the first feature set through adaptive histogram equalization to enhance the image contrast and obtain an image data set;

[0105] Use a pre-trained convolutional neural network to perform secondary feature extraction on the image data set to generate an information feature set containing enhanced texture and structure information;

[0106] If the texture information entropy value of the information feature set is lower than a preset threshold, optimize the texture of the information feature set through a generative adversarial network to obtain an optimized feature set; otherwise, directly output the information feature set as the optimized feature set;

[0107] According to the optimized feature set, use the principal component analysis method to perform dimensionality reduction processing on the texture and structure information to obtain a dimensionality-reduced feature set;

[0108] Use a pre-trained convolutional neural network to perform classification prediction on the dimensionality-reduced feature set to judge the quality level of the image data and obtain a classification result;

[0109] According to the classification result, use a weighted fusion method to integrate the texture and structure information of the dimensionality-reduced feature set to generate a target feature set.

[0110] Exemplarily, during the process of a medical imaging device acquiring low-quality image data, first, a CT scanner scans at a voltage of 120 kV and a current of 200 mA to obtain DICOM format images with a resolution of 512×512 pixels. Due to factors such as equipment limitations and patient movement, these images have noise and blur problems.

[0111] To extract initial features, a pre-trained ResNet-50 convolutional neural network is used. This network is trained on the ImageNet data set and contains 50 convolutional layers and fully connected layers. Input the low-quality images into ResNet-50, and extract the output of the 4th convolutional layer through forward propagation to obtain 256-dimensional feature vectors. These feature vectors are processed by L2 normalization to generate a first feature set containing texture and structure information.

[0112] To further optimize the features, the principal component analysis (PCA) algorithm is used to reduce the 256-dimensional features to 128 dimensions, retaining 95% of the variance information. Through the K-means clustering algorithm, the 128-dimensional feature set is divided into 10 clusters, and each cluster represents different texture and structure patterns. Finally, the t-SNE algorithm is used to map the high-dimensional features to a two-dimensional space to generate a visualization image for subsequent analysis.

[0113] Furthermore, the process of generating a second feature set containing edge distribution features by performing gradient calculation processing on the first feature set further includes:

[0114] Based on the second feature set, a feature extraction method is used to separate edge distribution features from non-edge distribution features to obtain an edge feature subset;

[0115] If the distribution characteristics of the edge feature subset meet the preset threshold, the feature conversion method is determined through data analysis to obtain the converted feature set; if not, the first feature set is returned to re-execute the gradient calculation;

[0116] According to the converted feature set, the random forest algorithm is applied to sort the features according to their importance to obtain the sorted feature set;

[0117] For the sorted feature set, a data processing method is used to remove features whose importance is lower than a preset threshold to obtain a streamlined feature set;

[0118] Based on the simplified feature set, the feature generation operation is performed to construct an enhanced feature set containing distribution characteristics. A clustering algorithm is used on the enhanced feature set to determine the distribution correlation between features and obtain the target feature set.

[0119] Exemplarily, after obtaining the first feature set, this embodiment generates a second feature set including edge distribution features through gradient calculation. The core of gradient calculation is to capture the change of pixel intensity in the image, especially in the edge area.

[0120] In a possible implementation, the image data in the first feature set is processed using a Sobel operator, and gradients in the horizontal and vertical directions are calculated respectively to generate an edge intensity map.

[0121] For example, for a 512×512 pixel CT image, the Sobel operator can generate edge distribution features and output a gradient feature map with a dimension of 512×512, emphasizing the texture changes in the edge area.

[0122] Specifically, for the second feature set, it is necessary to separate the marginal distribution features and the non-marginal distribution features to obtain the marginal feature subset.

[0123] In one embodiment, a threshold segmentation method may be used to mark areas in the gradient feature map where the gradient value is higher than a certain threshold as edge features, and the rest as non-edge features.

[0124] For example, by setting the threshold to the 75% quantile of the gradient value and extracting the edge feature subset, the dimension may be reduced to about 10,000 feature points.

[0125] It should be noted that the distribution characteristics of the edge feature subset must meet the preset threshold, such as the feature point coverage must reach 20% of the image area. If not, return to the first feature set and readjust the Sobel operator parameters, such as increasing the convolution kernel size to 5×5 and recalculating the gradient.

[0126] In one embodiment, if the edge feature subset meets the threshold, the feature transformation method can be determined through data analysis.

[0127] Preferably, the linear Sexually Transmitted Infections (STI) algorithm is used to standardize the features, mapping the numerical range of the edge feature subset to 0 to 1 to generate the transformed feature set.

[0128] For example, for 10,000 feature points, a 128-dimensional feature vector with a unified dimension is generated after standardization. Based on the transformed feature set, the random forest algorithm is applied to generate the feature importance ranking. The random forest evaluates the contribution of each feature to the classification task by constructing multiple decision trees.

[0129] For example, the edge strength feature may have an importance score of 0.85, while the smooth area feature is only 0.2, generating the sorted feature set.

[0130] It can be understood that for the sorted feature set, features with an importance lower than the threshold, such as 0.3, are removed to obtain the refined feature set. In one possible implementation, the first 50 high-importance features are retained, and the dimension is reduced from 128 to 50 dimensions. Based on the refined feature set, a feature generation operation is performed to construct the enhanced feature set.

[0131] For example, new features are generated through feature combination, such as the product of edge strength and direction, generating an enhanced feature set containing distribution characteristics, and the dimension may increase to 60 dimensions. For the enhanced feature set, the DBSCAN clustering algorithm is used to judge the distribution correlation between features.

[0132] For example, setting the radius parameter to 0.5, 3 feature clusters are identified, generating the final feature set, which contains a feature subset with high correlation for subsequent analysis.

[0133] Furthermore, according to the comparison result between the edge detail intensity of the second feature set and the preset threshold, the process of generating the third feature set by using the channel attention mechanism includes:

[0134] Obtain the edge detail information of the second feature set, and use the edge detection algorithm to extract the edge strength to obtain the edge strength distribution;

[0135] If the intensity of at least one region in the edge strength distribution is lower than the preset threshold, the Sobel operator is used to enhance the edge details to generate the enhanced edge features;

[0136] According to the enhanced edge features, weights are assigned using the channel attention mechanism to obtain the weighted feature set;

[0137] Through the weighted feature set, deep features are extracted using a convolutional neural network to generate the third feature set;

[0138] If the feature dimension of the third feature set exceeds the preset range, principal component analysis is used for dimensionality reduction to obtain a dimension-reduced and optimized feature set;

[0139] According to the dimension-reduced and optimized feature set, average pooling is used for processing to determine the third feature set of the final feature representation.

[0140] Exemplarily, when obtaining the edge detail information of the second feature set, it can be achieved by analyzing the edge characteristics of the image data. Edge details usually refer to the areas where pixel values change significantly in the image, reflecting the boundaries or textures of objects. Taking the medical image data being processed as a CT scan image as an example, the edge details may correspond to tissue boundaries. Edge information can be extracted through gradient magnitude analysis to generate an edge distribution map.

[0141] For example, in a 512x512 pixel CT image, the edge distribution map shows the pixel changes at the junction of bone and soft tissue.

[0142] In a possible implementation, an edge detection algorithm is used to extract the edge intensity, and the Canny algorithm can be used. The Canny algorithm smooths the image through Gaussian filtering, then calculates the gradient magnitude and direction to generate an edge intensity distribution. Suppose the edge intensity distribution shows that the intensity values of bone edges are in the range of 100 - 200, while the edge intensity of some soft tissues is below 50. If the intensity of at least one region is below the preset threshold of 50, the edge details need to be enhanced.

[0143] For example, the edge intensity of the soft tissue region is 30, triggering the enhancement operation.

[0144] Specifically, when using the Sobel operator to enhance the edge details, the Sobel operator calculates the gradient through convolution kernels in the horizontal and vertical directions to highlight the edges. Continuing with the CT image as an example, after applying the Sobel operator, the edge intensity of the soft tissue is increased from 30 to 80, generating enhanced edge features. This enhancement helps to retain more details during subsequent feature extraction.

[0145] Preferably, based on the enhanced edge features, a channel attention mechanism is used to assign weights. The channel attention mechanism dynamically adjusts the weights by learning the importance of different feature channels.

[0146] For example, in the enhanced edge features of the CT image, a weight of 0.7 is assigned to the bone edges and a weight of 0.3 is assigned to the soft tissue edges, generating a weighted feature set. This weight assignment makes the model pay more attention to key regions.

[0147] In one embodiment, through the weighted feature set, a convolutional neural network is used to extract deep features to generate a third feature set. The convolutional neural network captures the spatial relationships of the edges through multiple layers of convolution and pooling.

[0148] For example, in CT images, the network can extract the geometric shape features of bone edges and generate a third feature set with a dimension of 1024. If the dimension exceeds the preset range, such as 512, dimensionality reduction is required.

[0149] It can be understood that when using principal component analysis for dimensionality reduction, principal component analysis retains the main features through linear transformation and reduces the dimension.

[0150] For example, reducing the 1024-dimensional feature set to 256 dimensions generates an optimized feature set. This dimensionality reduction retains the edge information of bones and soft tissues while reducing the computational complexity.

[0151] For example, when using average pooling to process the optimized feature set, average pooling generates the final feature representation by taking the average of the feature map regions. Assume the optimized feature set contains 256 feature maps, each with 16x16 pixels. After average pooling, a 256-dimensional vector is generated, representing the edge characteristics of the CT image. This representation can be used for subsequent classification or segmentation tasks.

[0152] It should be noted that each step of this embodiment focuses on the extraction and optimization of edge features, and the logic progresses layer by layer. Edge detection, enhancement, weight assignment, deep feature extraction, dimensionality reduction, and pooling together construct an efficient feature processing flow. This flow is particularly important in medical image analysis and can provide effective diagnostic basis for doctors.

[0153] Furthermore, the process of optimizing and generating the first reconstructed image by evaluating the feature distribution difference through a discriminator network includes:

[0154] Performing standardization processing on the third feature set to obtain a normalized feature set;

[0155] Processing the normalized feature set through an adversarial generation network to generate a preliminary reconstructed image;

[0156] Using a discriminator network to evaluate the feature distribution difference between the preliminary reconstructed image and the target image to obtain a distribution difference value;

[0157] If the distribution difference value is greater than the preset threshold, adjust the parameters of the adversarial generation network through an optimization algorithm to generate an optimized reconstructed image;

[0158] Evaluating and analyzing the clarity of the optimized reconstructed image to obtain a quality evaluation result;

[0159] According to the quality evaluation result, perform post-processing on the optimized reconstructed image to obtain the first reconstructed image.

[0160] Exemplarily, when the third feature set is input into the generative adversarial network, the generator network structure is first a U-Net, which includes 5 layers of downsampling and 5 layers of upsampling. The number of convolutional kernels in each layer is 64, 128, 256, 512, and 1024 respectively. The LeakyReLU activation function (negative slope of 0.2) and batch normalization layer are used. The input feature dimension is 256×256×64, and it is gradually upsampled to a 512×512×3 RGB image through transposed convolution to generate a preliminary reconstructed image. The discriminator adopts a PatchGAN structure, which consists of 4 layers of convolution (convolutional kernels 64→128→256→512). The last layer outputs a 30×30 feature map, and each pixel corresponds to a 70×70 receptive field of the input image. By calculating the difference in the feature distributions between the generated image and the real image, the Wasserstein distance is used as the loss function (gradient penalty coefficient λ = 10), the initial learning rate is set to 0.0001, and the Adam optimizer is used (β1 = 0.5, β2 = 0.999). During the optimization process, the generator loss function includes content loss (L1 loss weight 100) and adversarial loss (weight 1). After 50,000 iterations of training, the difference in feature distributions drops from the initial 1.83 to 0.12. Finally, the PSNR of the first reconstructed image reaches 28.6 dB, and the SSIM is 0.91. During the training process, the learning rate is dynamically adjusted every 1000 iterations (decay rate 0.95), and spectral normalization is used to constrain the discriminator weights to ensure training stability.

[0161] Exemplarily, after obtaining the third feature set, standardization processing needs to be performed. The purpose of standardization processing is to unify the dimension and distribution of feature data for subsequent network processing.

[0162] For example, when processing the third feature set of CT images, assume that the feature set contains a 1024-dimensional vector, and the data range is between 0 - 255. The Z-score method can be used for standardization to convert the feature values into a distribution with a mean of 0 and a standard deviation of 1.

[0163] Specifically, for each feature dimension, its mean and standard deviation are calculated and then normalized to generate a normalized feature set. This processing ensures that the data distributions of different feature channels are consistent and improves the stability of the model.

[0164] In a possible implementation, the normalized feature set is input into the generative adversarial network to generate a preliminary reconstructed image. The generative adversarial network consists of a generator and a discriminator. The generator is responsible for reconstructing the image from the feature set, and the discriminator evaluates the difference between the generated image and the real image.

[0165] For example, in CT image reconstruction, the generator receives a 256-dimensional normalized feature set and outputs a preliminary reconstructed image of 512x512 pixels. The preliminary image may contain bone and soft tissue regions, but the details are blurred. The discriminator network calculates the difference in feature distribution by comparing the pixel distribution of the preliminary reconstructed image with that of the target CT image.

[0166] For example, assume that the discriminator outputs a distribution difference value of 0.8, while the preset threshold is 0.5, indicating a large difference.

[0167] It should be noted that if the distribution difference value is greater than the threshold, the parameters of the adversarial generation network need to be adjusted through an optimization algorithm.

[0168] Preferably, the Adam optimization algorithm is adopted to update the weights of the generator through backpropagation to reduce the distribution difference.

[0169] For example, after adjustment, the difference value drops to 0.3, generating an optimized reconstructed image with clearer bone edges and more natural soft tissue textures. This optimization improves the authenticity of the image.

[0170] Specifically, the clarity of the optimized reconstructed image needs to be analyzed by an image quality assessment module. The peak signal-to-noise ratio method can be used for quality assessment to compare the pixel differences between the optimized image and the target image.

[0171] For example, the evaluation result shows that the peak signal-to-noise ratio is 30 dB, indicating that the image quality is high but there is still noise. The denoising module then performs post-processing on the optimized reconstructed image.

[0172] In one embodiment, the non-local means denoising algorithm is adopted to smooth the noise region by analyzing the pixel values in similar regions of the image.

[0173] For example, the noise points in the soft tissue region are reduced, generating a first reconstructed image with smoother details.

[0174] It can be understood that the first reconstructed image is saved by the storage module to generate the final output data.

[0175] For example, the image is stored in DICOM format, including a grayscale image of 512x512 pixels and metadata such as patient information. This storage method facilitates subsequent diagnostic analysis and ensures data integrity.

[0176] Furthermore, the process of performing artifact correction on the first reconstructed image to generate a second reconstructed image that meets physical constraints includes:

[0177] Obtain the pixel intensity data and artifact distribution characteristics of the first reconstructed image, and use a feature extraction algorithm to determine the artifact region to obtain an artifact distribution feature map;

[0178] Extract imaging model parameters from a pre-established physical imaging model according to the artifact distribution feature map, and determine a parameter set for correction processing;

[0179] Adopt a constrained optimization method, combine physical constraint conditions and imaging model parameters, perform optimization calculation on the artifact distribution feature map, and obtain the corrected pixel intensity distribution;

[0180] Reconstruct the image spatial consistency through the corrected pixel intensity distribution to generate a preliminary corrected image;

[0181] If the image reconstruction quality of the preliminary corrected image is lower than a preset threshold, use an iterative optimization algorithm to adjust the pixel intensity to obtain an optimized corrected image, and verify whether the optimized corrected image meets the physical constraint conditions to obtain a second reconstructed image;

[0182] Calculate the image spatial consistency and the degree of artifact residue of the second reconstructed image to judge the accuracy of the correction algorithm, and obtain the finally corrected second reconstructed image.

[0183] Exemplarily, when performing artifact correction in this embodiment, first establish a linear attenuation coefficient matrix based on the attenuation characteristics of ray propagation. For example, for CT imaging, assume that the original projection data received by the detector is 1200 groups, and each group contains 512 sampling points. Calculate the initial reconstructed image through an iterative reconstruction algorithm (such as the SART algorithm). Physical constraint conditions are introduced during the iteration. For example, set the reasonable range of the attenuation coefficient to be 0.005 / mm to 0.03 / mm, and the pixel values outside this range are corrected by non-negative least squares. For the metal artifact problem, adopt a correction method based on energy spectrum decomposition. Assume that the scanning voltage is 140 kVp, and the basis material images of iron (Fe) and titanium (Ti) are separated from the dual-energy spectrum data, and their K-edge energies are 7.11 keV and 4.96 keV respectively. Use polynomial fitting to correct the energy spectrum drift error. For scatter artifacts, use Monte Carlo simulation to calculate the scatter distribution, set the scatter contribution ratio in the water phantom to be 15%, and construct a scatter correction model through a convolutional neural network (CNN). The network contains 5 convolutional layers, and each layer uses the ReLU activation function, and finally outputs the scatter-corrected projection data. Input the processed data into the FDK reconstruction algorithm, set the filtering function to the Hamming window, and the cut-off frequency to 0.8 Nyquist, and reconstruct a second reconstructed image that meets the physical constraints, and its SSIM index is improved from 0.72 to 0.89.

[0184] Further, the process of generating a third reconstructed image by performing intensity correction on the second reconstructed image according to a pre-trained device deviation pattern includes:

[0185] Obtain pixel intensity data from the second reconstructed image, perform matching through the pre-trained device deviation pattern, and obtain the deviation distribution feature;

[0186] If the deviation distribution characteristics exceed the preset threshold, an intensity correction algorithm is used to adjust the pixel intensity to generate corrected intensity data;

[0187] According to the corrected intensity data, combined with the image reconstruction algorithm, a preliminary third reconstructed image is generated;

[0188] Obtain the edge features of the preliminary third reconstructed image, verify the deviation compensation effect through pattern matching, and obtain the verification result;

[0189] If the verification result shows insufficient deviation compensation, the intensity of the edge features is adjusted to generate an optimized third reconstructed image;

[0190] Through the image generation algorithm, pixel smoothing processing is performed on the optimized third reconstructed image to obtain the final third reconstructed image;

[0191] Extract quality parameters according to the final third reconstructed image, and evaluate through the preset quality standard to obtain the third reconstructed image that completes image correction.

[0192] Exemplarily, the device deviation pattern pre-trained in this embodiment is implemented based on a deep learning convolutional neural network (CNN). For example, the ResNet50 architecture is adopted, and image features are extracted by loading pre-trained weights (such as a model trained on the ImageNet dataset). For the second reconstructed image (size 512×512 pixels, grayscale value range 0-255), first input it into the model, extract the feature map output by the 4th layer convolution (dimension 256×64×64), and calculate the global average pooling value of this feature map as the quantization index of the device deviation. Assuming the deviation value is 0.15 (range -1 to 1 after standardization), a linear correction algorithm is used to adjust the image intensity, and the correction formula is I_corrected = I_original×(1 + 0.15×0.8), where 0.8 is an empirical attenuation coefficient used to prevent overcorrection. When generating the third reconstructed image after correction, the grayscale value needs to be restricted within the range of 0-255, and the exceeded part is truncated. For example, if the original value of a certain pixel is 200, after correction it is 200×1.12 = 224, which does not exceed the threshold and is retained; if the corrected value is 260, it is forced to be set to 255. The entire process is implemented through the PyTorch framework, and CUDA is used to accelerate the calculation. The processing time for a single image is about 23 milliseconds (NVIDIA V100 graphics card). To verify the correction effect, the peak signal-to-noise ratio (PSNR) of the image before and after correction can be calculated. If the PSNR improvement exceeds 3dB, the correction is considered effective.

[0193] Furthermore, the process of using a multi-scale convolutional network to perform feature extraction and fusion processing on the third reconstructed image to generate the fourth reconstructed image includes:

[0194] Extract features from the third reconstructed image through a preset multi-scale convolutional network to obtain a multi-scale feature set;

[0195] Use a multi-scale convolutional network to fuse the multi-scale feature set to generate a fused feature representation;

[0196] If the resolution of the fused feature representation is lower than the preset threshold, adjust the feature resolution through an upsampling operation to obtain an adjusted feature representation;

[0197] Reconstruct the adjusted feature representation through a decoding network to generate a fourth reconstructed image;

[0198] Obtain the feature space distribution of the fourth reconstructed image, and determine whether it meets the preset feature consistency condition. If the feature consistency condition is not met, adjust the feature extraction and fusion process by optimizing network parameters to obtain an optimized fourth reconstructed image;

[0199] According to the optimized fourth reconstructed image, use a preset evaluation function to calculate the image quality and determine the finally output fourth reconstructed image.

[0200] Exemplarily, when using a multi-scale convolutional network to extract features and fuse the third reconstructed image, first perform convolution operations on the image through convolutional kernels of different scales (such as 3×3, 5×5, 7×7) to capture detail information at different levels.

[0201] For example, use a 3×3 convolutional kernel to extract local features, a 5×5 convolutional kernel to extract medium-range features, and a 7×7 convolutional kernel to extract global features. After each convolutional layer, connect a ReLU activation function to enhance the non-linear expression ability. Then, fuse the feature maps of different scales through a feature fusion module, using the weighted average method, and the weight values are dynamically adjusted according to the importance of the feature maps. For example, the local feature weight is 0.4, the medium-range feature weight is 0.3, and the global feature weight is 0.3. The fused feature map undergoes dimensionality reduction through a 1×1 convolutional layer to reduce the computational amount. Finally, restore the feature map to the original image size through an upsampling operation to generate a fourth reconstructed image. Throughout the process, the Adam optimizer is used for parameter update, the learning rate is set to 0.001, and the number of iterations is 100 times to ensure that the model converges and generates a high-quality fourth reconstructed image.

[0202] Furthermore, the process of inputting the fourth reconstructed image into a residual refinement network for detail enhancement to generate a final high-quality reconstructed image includes:

[0203] Extract features from the fourth reconstructed image through a residual refinement network to obtain detail enhancement features;

[0204] If the resolution of the detail enhancement feature is lower than the preset threshold, the resolution is enhanced through a super-resolution network to obtain an optimized feature image;

[0205] An image fusion algorithm is used to fuse the optimized feature image with the fourth reconstructed image to generate a preliminary high-quality image;

[0206] The clarity of the preliminary high-quality image is detected by a quality assessment module to determine whether the clarity meets the preset standard;

[0207] If the clarity does not meet the preset standard, secondary enhancement is performed through a residual refinement network to obtain a final high-quality image;

[0208] The final high-quality image is obtained, and the image format is adjusted through a format conversion module to generate a target output image;

[0209] Specifically, the feature extraction of the fourth reconstructed image by the residual refinement network can be realized through a multi-layer residual connection structure. The residual network enhances the ability to express details by introducing "shortcut connections" to learn the differences between features.

[0210] In one embodiment, the network includes 5 residual blocks, each composed of a 3×3 convolutional kernel and a batch normalization layer, with 64 convolutional kernels. The extracted detail enhancement features contain texture and edge information. If the resolution of the detail enhancement feature is lower than the preset threshold, such as 512×512 pixels, it needs to be processed by a super-resolution network.

[0211] In a possible implementation, the super-resolution network can adopt a generative adversarial network structure. The generator enhances the resolution by 4 times through upsampling, and the discriminator determines the authenticity of the generated image.

[0212] For example, the generator includes 8 convolutional layers with a convolutional kernel size of 3×3. After upsampling, the resolution reaches 2048×2048 to obtain an optimized feature image.

[0213] Preferably, the image fusion algorithm can adopt a fusion method based on an attention mechanism. The principle is to highlight the detail regions of the optimized feature image through attention weights and fuse them with the global information of the fourth reconstructed image.

[0214] In one embodiment, the attention weights are calculated based on gradient information, and the weight of the detail region is set to 0.6 to generate a preliminary high-quality image.

[0215] It should be noted that the clarity detection of the quality assessment module can be realized by calculating the gradient magnitude and entropy value of the image.

[0216] For example, when the gradient magnitude is lower than 50 or the entropy value is lower than 7, it is determined that the clarity does not meet the standard and secondary enhancement is required. The secondary enhancement still passes through the residual refinement network, adjusts the number of convolutional kernels to 128, increases the feature expression ability, and generates the final high-quality image.

[0217] It can be understood that the format conversion module can adjust the image to a specific format, such as JPEG or PNG.

[0218] For example, for the JPEG format, the compression quality is set to 90 to ensure the balance between visual quality and storage efficiency.

[0219] In one embodiment, a strict logical chain is formed through image processing, feature extraction, resolution improvement, image fusion, quality assessment, and format conversion, and it unfolds around the business goal of generating high-quality images.

[0220] Details of residual refinement and super-resolution improvement, the super-resolution network supplements when the resolution is insufficient, the fusion algorithm integrates multi-source information, quality assessment ensures output standards, and format conversion adapts to application requirements. Each link supports each other, jointly achieving the goal of generating high-quality images, with a strict and efficient logic.

[0221] Furthermore, the process of extracting feature data from the final high-quality reconstructed image and generating a heat map and locating the suspicious lesion area using a dual-path network includes:

[0222] Obtain pixel-level feature data from the final high-quality reconstructed image, use a preprocessing algorithm to remove noise, and obtain a denoised feature set;

[0223] Use a convolutional neural network to perform preliminary feature extraction on the denoised feature set to generate a feature set after feature extraction;

[0224] Process the feature set after feature extraction through a dual-path network to generate the first heat map data;

[0225] If the significance value of the first heat map data is greater than a preset threshold, then use a clustering algorithm to segment the heat map data to determine the suspicious lesion area;

[0226] According to the segmented suspicious lesion area, extract the regional boundary features to generate the second heat map data;

[0227] Perform coordinate mapping on the second heat map data through a geometric analysis algorithm to obtain the precise position coordinates of the lesion area;

[0228] If the deviation of the precise position coordinates is less than a preset threshold, then output the final localization result to determine the lesion area.

[0229] Exemplarily, feature data is extracted from the final high-quality reconstructed image. First, a convolutional neural network (CNN) is used for feature extraction, with ResNet-50 as the base model. The input image size is 224×224. Multilevel features are extracted through 5 convolutional layers and pooling layers, and the output feature map size is 7×7×2048. Then, a dual-path network (DPN) is used to generate a heatmap. The DPN consists of two parallel paths. One path uses a 3×3 convolutional kernel to extract local features, and the other path uses a 1×1 convolutional kernel to extract global features. The features of the two paths are fused in the last layer to generate a heatmap with a size of 14×14×1. After the heatmap is generated, a threshold segmentation algorithm is used to locate the suspicious lesion area. The threshold is set to 0.7, and the areas in the heatmap greater than this threshold are marked as suspicious lesion areas. The connected component analysis algorithm is used to cluster the marked areas to obtain the final position of the lesion area. Finally, a support vector machine (SVM) is used to classify the extracted features. The radial basis function (RBF) is used as the kernel function, and the optimal parameters C = 1.0 and gamma = 0.01 are determined through cross-validation, and the classification accuracy reaches 95.6%. The whole process is realized through an automated process without manual intervention, ensuring processing efficiency and accuracy.

[0230] Embodiment 2

[0231] As Figure 2 shown, based on the same inventive concept, this embodiment also provides a medical image reconstruction system based on deep learning, including:

[0232] An initial feature extraction module, configured to obtain low-quality image data based on a medical imaging device, and extract initial features using a pre-trained convolutional neural network to generate a first feature set containing texture and structure information;

[0233] A gradient feature generation module, configured to perform gradient calculation processing on the first feature set to generate a second feature set containing edge distribution features;

[0234] An attention processing module, configured to perform processing using a channel attention mechanism according to the comparison result between the edge detail intensity of the second feature set and a preset threshold to generate a third feature set;

[0235] An adversarial generation module, configured to input the third feature set into an adversarial generation network to generate a preliminary reconstructed image, and optimize and generate a first reconstructed image after evaluating the feature distribution difference through a discriminator network;

[0236] An artifact correction module, configured to perform artifact correction processing on the first reconstructed image to generate a second reconstructed image that meets physical constraints;

[0237] An intensity correction module, configured to perform intensity correction on the second reconstructed image according to a pre-trained device deviation pattern to generate a third reconstructed image;

[0238] A multi-scale fusion module, configured to perform feature extraction and fusion processing on the third reconstructed image by using a multi-scale convolutional network to generate a fourth reconstructed image;

[0239] A residual refinement module, configured to input the fourth reconstructed image into a residual refinement network for detail enhancement processing to generate a final high-quality reconstructed image;

[0240] A lesion localization module, configured to extract feature data from the final high-quality reconstructed image, and use a dual-path network to generate a heat map and localize a suspicious lesion area;

[0241] A diagnostic annotation module, configured to perform type annotation processing on the suspicious lesion area according to pathological prior knowledge to generate a diagnostic auxiliary image.

[0242] A medical image reconstruction system based on deep learning provided in this embodiment has all the advantages of the medical image reconstruction method based on deep learning provided in Embodiment 1.

[0243] Embodiment 3

[0244] This embodiment also discloses a computer device, including a memory, a processor, and a computer program stored on the memory, where the processor executes the computer program to implement the steps of the method described in Embodiment 1.

[0245] Embodiment 4

[0246] This embodiment also discloses a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the method described in Embodiment 1 are implemented.

[0247] Embodiment 5

[0248] This embodiment also discloses a computer program product, including a computer program, and when the computer program is executed by a processor, the steps of the method described in Embodiment 1 are implemented.

[0249] The above is only a preferred specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present application should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A medical image reconstruction method based on deep learning, characterized in that, Including: Obtain low-quality image data based on a medical imaging device, and use a pre-trained convolutional neural network to extract initial features, generating a first feature set containing texture and structural information; Perform gradient calculation processing on the first feature set to generate a second feature set containing edge distribution features; According to the comparison result between the edge detail intensity of the second feature set and a preset threshold, perform processing using a channel attention mechanism to generate a third feature set; Input the third feature set into an adversarial generation network to generate a preliminary reconstructed image, and optimize to generate a first reconstructed image after evaluating the feature distribution difference through a discriminator network; Perform artifact correction processing on the first reconstructed image to generate a second reconstructed image that meets physical constraints; Perform intensity correction on the second reconstructed image according to a pre-trained device deviation pattern to generate a third reconstructed image; Use a multi-scale convolutional network to perform feature extraction and fusion processing on the third reconstructed image to generate a fourth reconstructed image; Input the fourth reconstructed image into a residual refinement network for detail enhancement processing to generate a final high-quality reconstructed image; Extract feature data from the final high-quality reconstructed image, and use a dual-path network to generate a heat map and locate suspicious lesion areas; Perform type annotation processing on the suspicious lesion areas according to pathological prior knowledge to generate a diagnostic assistance image.

2. The method according to claim 1, wherein: The process of generating a first feature set containing texture and structural information includes: Obtain low-quality image data output by a medical imaging device, use a pre-trained convolutional neural network to extract initial features, generating a first feature set containing texture and structural information; Enhance the image contrast by performing adaptive histogram equalization on the low-quality image data corresponding to the first feature set to obtain an image data set; Use a pre-trained convolutional neural network to perform secondary feature extraction on the image data set to generate an information feature set containing enhanced texture and structural information; If the texture information entropy value of the information feature set is lower than a preset threshold, perform texture optimization on the information feature set through a generative adversarial network to obtain an optimized feature set; otherwise, directly output the information feature set as the optimized feature set; According to the optimized feature set, perform dimensionality reduction processing on the texture and structural information using the principal component analysis method to obtain a dimensionality-reduced feature set; Perform classification prediction on the dimensionality-reduced feature set through a pre-trained convolutional neural network to judge the quality level of the image data and obtain a classification result; According to the classification result, use a weighted fusion method to integrate the texture and structural information of the dimensionality-reduced feature set to generate a target feature set.

3. The method according to claim 1, wherein: The process of performing gradient calculation processing on the first feature set to generate a second feature set containing edge distribution features further includes: Based on the second feature set, use a feature extraction method to separate the edge distribution features and non-edge distribution features to obtain an edge feature subset; If the distribution characteristics of the edge feature subset meet a preset threshold, determine a feature conversion method through data analysis to obtain a converted feature set; if not, return the first feature set and re-execute the gradient calculation; According to the converted feature set, the random forest algorithm is applied to sort the features according to their importance to obtain the sorted feature set; For the sorted feature set, a data processing method is used to remove features whose importance is lower than a preset threshold to obtain a streamlined feature set; A feature generation operation is performed based on the simplified feature set to construct an enhanced feature set including distribution characteristics. A clustering algorithm is used on the enhanced feature set to determine the distribution correlation between features to obtain a target feature set.

4. The method according to claim 1, characterized in that: According to the comparison result of the edge detail intensity of the second feature set and the preset threshold, the channel attention mechanism is used for processing, and the process of generating the third feature set includes: Obtain edge detail information of the second feature set, extract edge strength using an edge detection algorithm, and obtain edge strength distribution; If the intensity of at least one area in the edge intensity distribution is lower than a preset threshold, the Sobel operator is used to enhance the edge details and generate enhanced edge features; According to the enhanced edge features, a channel attention mechanism is used to assign weights to obtain a weighted feature set; Using the weighted feature set, a convolutional neural network is used to extract deep features to generate a third feature set; If the feature dimension of the third feature set exceeds a preset range, principal component analysis is used to reduce the dimension to obtain a reduced-dimensional optimized feature set; According to the dimension reduction optimization feature set, mean pooling is used to determine a third feature set of final feature representation.

5. The method according to claim 1, characterized in that The process of optimizing and generating the first reconstructed image after evaluating the feature distribution difference through the discriminator network includes: Performing standardization processing on the third feature set to obtain a normalized feature set; Processing the normalized feature set through a generative adversarial network to generate a preliminary reconstructed image; Using a discriminator network to evaluate the difference in feature distribution between the preliminary reconstructed image and the target image, and obtain a distribution difference value; If the distribution difference value is greater than a preset threshold, the adversarial generative network parameters are adjusted through an optimization algorithm to generate an optimized reconstructed image; Evaluating and analyzing the clarity of the optimized reconstructed image to obtain a quality evaluation result; According to the quality assessment result, the optimized reconstructed image is post-processed to obtain a first reconstructed image.

6. The method according to claim 1, characterized in that The process of performing artifact correction processing on the first reconstructed image to generate a second reconstructed image that satisfies physical constraints includes: Acquire pixel intensity data and artifact distribution characteristics of the first reconstructed image, determine the artifact area using a feature extraction algorithm, and obtain an artifact distribution feature map; Extracting imaging model parameters from a pre-established physical imaging model according to the artifact distribution characteristic map, and determining a parameter set for correction processing; Using a constrained optimization method, combined with physical constraints and imaging model parameters, an optimization calculation is performed on the artifact distribution characteristic map to obtain a corrected pixel intensity distribution; Reconstruct the image spatial consistency through the corrected pixel intensity distribution and generate a preliminary corrected image; If the image reconstruction quality of the preliminary corrected image is lower than a preset threshold, an iterative optimization algorithm is used to adjust the pixel intensity to obtain an optimized corrected image, and it is verified whether the optimized corrected image meets the physical constraint conditions to obtain a second reconstructed image; Calculate the image spatial consistency and the degree of artifact residue of the second reconstructed image to judge the accuracy of the correction algorithm, and obtain the finally corrected second reconstructed image.

7. The method according to claim 1, wherein The process of generating a third reconstructed image by performing intensity correction on the second reconstructed image according to a pre-trained device deviation pattern includes: Obtain pixel intensity data from the second reconstructed image, and perform matching through a pre-trained device deviation pattern to obtain a deviation distribution feature; If the deviation distribution feature exceeds a preset threshold, an intensity correction algorithm is used to adjust the pixel intensity to generate corrected intensity data; According to the corrected intensity data, combined with an image reconstruction algorithm, generate a preliminary third reconstructed image; Obtain the edge feature of the preliminary third reconstructed image, and verify the deviation compensation effect through pattern matching to obtain a verification result; If the verification result shows insufficient deviation compensation, perform intensity adjustment on the edge feature to generate an optimized third reconstructed image; Through an image generation algorithm, perform pixel smoothing processing on the optimized third reconstructed image to obtain a final third reconstructed image; Extract quality parameters from the final third reconstructed image, and evaluate through a preset quality standard to obtain a third reconstructed image that completes image correction.

8. The method according to claim 1, wherein The process of generating a fourth reconstructed image by using a multi-scale convolutional network to perform feature extraction and fusion processing on the third reconstructed image includes: Perform feature extraction on the third reconstructed image through a preset multi-scale convolutional network to obtain a multi-scale feature set; Use a multi-scale convolutional network to perform fusion processing on the multi-scale feature set to generate a fused feature representation; If the resolution of the fused feature representation is lower than a preset threshold, adjust the feature resolution through an upsampling operation to obtain an adjusted feature representation; Perform reconstruction processing on the adjusted feature representation through a decoding network to generate a fourth reconstructed image; Obtain the feature space distribution of the fourth reconstructed image, and judge whether it meets the preset feature consistency condition. If the feature consistency condition is not met, adjust the feature extraction and fusion process by optimizing network parameters to obtain an optimized fourth reconstructed image; According to the optimized fourth reconstructed image, use a preset evaluation function to calculate the image quality and determine the finally output fourth reconstructed image.

9. The method according to claim 1, wherein The process of extracting feature data from the finally high-quality reconstructed image and using a dual-path network to generate a heat map and locate suspicious lesion areas includes: Obtain pixel-level feature data from the finally high-quality reconstructed image, and use a preprocessing algorithm to remove noise to obtain a denoised feature set; Use a convolutional neural network to perform preliminary feature extraction on the denoised feature set to generate a feature set after feature extraction; Process the feature set after feature extraction through a dual-path network to generate first heatmap data; If the significance value of the first heatmap data is greater than a preset threshold, use a clustering algorithm to segment the heatmap data to determine a suspicious lesion area; Extract regional boundary features based on the segmented suspicious lesion area to generate second heatmap data; Perform coordinate mapping on the second heatmap data through a geometric analysis algorithm to obtain the exact position coordinates of the lesion area; If the deviation of the exact position coordinates is less than a preset threshold, output the final localization result to determine the lesion area.

10. A medical image reconstruction system based on deep learning, characterized in that, Including: An initial feature extraction module for acquiring low-quality image data based on a medical imaging device and using a pre-trained convolutional neural network to extract initial features to generate a first feature set containing texture and structural information; A gradient feature generation module for performing gradient calculation processing on the first feature set to generate a second feature set containing edge distribution features; An attention processing module for processing according to the comparison result between the edge detail intensity of the second feature set and a preset threshold and using a channel attention mechanism to generate a third feature set; An adversarial generation module for inputting the third feature set into an adversarial generation network to generate a preliminary reconstructed image and optimizing to generate a first reconstructed image after evaluating the feature distribution difference through a discriminator network; An artifact correction module for performing artifact correction processing on the first reconstructed image to generate a second reconstructed image that meets physical constraints; An intensity correction module for performing intensity correction on the second reconstructed image according to a pre-trained device deviation pattern to generate a third reconstructed image; A multi-scale fusion module for performing feature extraction and fusion processing on the third reconstructed image using a multi-scale convolutional network to generate a fourth reconstructed image; A residual refinement module for inputting the fourth reconstructed image into a residual refinement network for detail enhancement processing to generate a final high-quality reconstructed image; A lesion localization module for extracting feature data from the final high-quality reconstructed image and using a dual-path network to generate a heatmap and locate the suspicious lesion area; A diagnostic annotation module for performing type annotation processing on the suspicious lesion area according to pathological prior knowledge to generate a diagnostic auxiliary image.

Images