A medical image reconstruction method and system based on deep learning

Abstract

The application discloses a kind of medical image reconstruction method and system based on deep learning, including: for the low-quality image obtained by medical imaging device, first, pre-training convolutional neural network is used to extract initial feature, edge details are enhanced by gradient calculation and channel attention mechanism, then image reconstruction is carried out using adversarial generative network.Subsequently, the reconstructed image is corrected in combination with physical imaging model and device deviation mode, and the image quality is further improved by multi-scale convolution network and residual refining network.Finally, the application generates a heat map using a dual-path network to locate suspicious lesion areas and labels them based on pathological prior knowledge to generate diagnostic auxiliary images.The application can effectively improve the quality of medical images and enhance the visualization of lesion areas.

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 medical imaging, crucial for improving diagnostic accuracy and treatment outcomes. High-quality medical images clearly depict human anatomy and pathological features, providing a reliable basis for clinical decision-making. However, existing medical image reconstruction methods often struggle to balance the preservation of image detail with the suppression of artifacts when processing incomplete, low-quality data, 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 diverse imaging devices and complex pathological scenarios. The reconstructed results are often compromised by noise, bias, and distortion, reducing their value. Deep learning-based medical image reconstruction faces several core challenges in ensuring the fidelity of reconstructed images. Image reconstruction requires recovering 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 between image features and the complete image. This directly results in the loss of critical anatomical details or the introduction of artifacts that are inconsistent with physical laws in the reconstructed images. Furthermore, due to systematic variations introduced by different imaging devices and parameter settings, reconstructed images can exhibit inconsistencies across different devices, impacting the reliability of multicenter studies and remote diagnosis. Furthermore, medical images often contain complex textures and lesions. Models must emphasize details in key areas while maintaining overall image quality, which increases the complexity of the reconstruction task. Therefore, designing a deep learning method that can accurately restore realistic and consistent medical images from incomplete data while adaptively correcting for device bias and emphasizing lesions has become a key issue in the field of medical image reconstruction. Summary of the Invention

[0003] 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 a pre-trained convolutional neural network is used to extract initial features to generate a first feature set containing texture and structure information;

[0005] Performing gradient calculation 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 the result to generate a third feature set;

[0007] Inputting the third feature set into a generative adversarial network to generate a preliminary reconstructed image, and optimizing and generating a first reconstructed image after evaluating feature distribution differences through a discriminator network;

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

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

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

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

[0012] Extracting feature data from the final high-quality reconstructed image, generating a heat map using a dual-pathway network, and locating suspicious lesion areas;

[0013] The suspicious lesion area is labeled according to prior knowledge of pathology to generate a diagnosis-assisted image.

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

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

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

[0017] Performing secondary feature extraction on the image dataset using a pre-trained convolutional neural network to generate an information feature set containing enhanced texture and structural information;

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

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

[0020] Performing classification prediction on the dimensionality reduction feature set through a pre-trained convolutional neural network, judging the quality level of the image data, and obtaining a classification result;

[0021] According to the classification results, the weighted fusion method is used to integrate the texture and structure information of the dimensionality reduction 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 gradient calculation is performed again by returning to the first feature set;

[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, data processing methods are 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 applied to the enhanced feature set to determine the distribution correlation between features and obtain a target feature set.

[0028] Preferably, the process of generating the third feature set by using a channel attention mechanism to process the comparison result of the edge detail intensity of the second feature set with a preset threshold includes:

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

[0030] If the intensity of at least one region 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 performed to determine a third feature set for final feature representation.

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

[0036] performing standardization processing on the third feature set to obtain a normalized feature set;

[0037] Processing the normalized feature set through a generative adversarial network to generate a preliminary reconstructed image;

[0038] 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;

[0039] 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;

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

[0041] According to the quality assessment result, the optimized reconstructed image is post-processed to obtain a 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] Obtaining pixel intensity data and artifact distribution characteristics of the first reconstructed image, determining the artifact area using a feature extraction algorithm, and obtaining an artifact distribution feature map;

[0044] 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;

[0045] 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;

[0046] The spatial consistency of the image is reconstructed 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, 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;

[0048] The image spatial consistency and the degree of residual artifacts are calculated for the second reconstructed image, the accuracy of the correction algorithm is determined, and the final corrected second reconstructed image is obtained.

[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 comprises:

[0050] Obtain pixel intensity data based on the second reconstructed image, and match it with 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] generating a preliminary third reconstructed image based on the corrected intensity data in combination with an image reconstruction algorithm;

[0053] Acquiring edge features of the preliminary third reconstructed image, verifying the deviation compensation effect through pattern matching, and obtaining a verification result;

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

[0055] Performing pixel smoothing processing on the optimized third reconstructed image using an image generation algorithm to obtain a final third reconstructed image;

[0056] Quality parameters are extracted from the final third reconstructed image and evaluated using a preset quality standard to obtain a third reconstructed image that has completed image correction.

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

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

[0059] Using a multi-scale convolutional network to fuse 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, adjusting the feature resolution by upsampling to obtain an adjusted feature representation;

[0061] Reconstructing the adjusted feature representation through a decoding network to generate a fourth reconstructed image;

[0062] Obtaining a feature spatial distribution of the fourth reconstructed image, determining whether a preset feature consistency condition is satisfied, and if the feature consistency condition is not satisfied, adjusting 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, the image quality is calculated using a preset evaluation function to determine the fourth reconstructed image to be finally output.

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

[0065] Obtaining pixel-level feature data from the final high-quality reconstructed image, removing noise by using a preprocessing algorithm to obtain a denoised feature set;

[0066] Performing preliminary feature extraction on the denoised feature set by using a convolutional neural network to generate a feature set after feature extraction;

[0067] Generating first heat map data by processing the feature set after feature extraction through a double-path network;

[0068] If the saliency value of the first heat map data is greater than a preset threshold, segmenting the heat map data by using a clustering algorithm to determine a suspicious lesion area;

[0069] According to the segmented suspicious lesion area, extracting regional boundary features to generate second heat map data;

[0070] Performing coordinate mapping on the second heat map data by using a geometric analysis algorithm to obtain accurate position coordinates of the lesion area;

[0071] If the deviation of the accurate position coordinates is less than a preset threshold, outputting a final positioning result to determine the lesion area.

[0072] The application also provides a medical image reconstruction system based on deep learning, comprising:

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

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

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

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

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

[0078] an intensity correction module, configured to perform intensity correction on the second reconstructed image according to a pre-trained device bias pattern to generate a third reconstructed image;

[0079] 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;

[0080] a residual refining module, configured to input the fourth reconstructed image into a residual refining network to perform detail enhancement processing to generate a final high-quality reconstructed image;

[0081] a lesion positioning module, configured to extract feature data from the final high-quality reconstructed image, and generate a heat map and locate a suspicious lesion area by using a dual-path network;

[0082] a diagnosis labeling module, configured to perform type labeling processing on the suspicious lesion area according to pathological prior knowledge to generate a diagnosis auxiliary image.

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

[0084] The present application is aimed at low-quality images obtained by a medical imaging device, and first extracts initial features by using a pre-trained convolutional neural network, enhances edge details through gradient calculation and channel attention mechanism, and then performs image reconstruction by using a generative adversarial network. Subsequently, the present application corrects the reconstructed image by combining a physical imaging model and a device bias pattern, and further improves the image quality by using a multi-scale convolutional network and a residual refining network. Finally, the present application generates a heat map by using a dual-path network to locate a suspicious lesion area, and performs labeling according to pathological prior knowledge to generate a diagnosis auxiliary image. This method can effectively improve the quality of medical images and enhance the visualization effect of lesion areas, thereby providing strong support for subsequent clinical diagnosis by doctors. BRIEF DESCRIPTION OF DRAWINGS

[0085] The accompanying drawings, which form a part of this application, are intended to provide further understanding of the application and are incorporated herein in their entirety, and the illustrative embodiments thereof and their description serve to explain the present application. In the drawings:

[0086] Figure 1 is a method flowchart of an embodiment of the present application;

[0087] Figure 2 is a system structure schematic diagram of an embodiment of the present application. DETAILED DESCRIPTION

[0088] It should be noted that the embodiments and features in the present application can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

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

[0090] Example 1

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

[0092] Low-quality image data is acquired based on medical imaging equipment, and a pre-trained convolutional neural network is used to extract initial features to generate a first feature set containing texture and structure information;

[0093] Performing gradient calculation on the first feature set to generate a second feature set including edge distribution features;

[0094] 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 to process it to generate the third feature set;

[0095] The third feature set is input into the adversarial generative network to generate a preliminary reconstructed image, and the discriminator network evaluates the feature distribution difference and then optimizes to generate the first reconstructed image;

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

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

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

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

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

[0101] Based on prior knowledge of pathology, suspicious lesion areas are labeled and processed to generate diagnostic auxiliary images.

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

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

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

[0105] A pre-trained convolutional neural network is used to perform secondary feature extraction on the image dataset to generate an informative feature set containing enhanced texture and structural information;

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

[0107] According to the optimized feature set, the principal component analysis method is used to reduce the dimension of texture and structure information to obtain the reduced dimension feature set;

[0108] The pre-trained convolutional neural network is used to classify and predict the reduced dimension feature set, determine the quality level of the image data, and obtain the classification result;

[0109] 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.

[0110] For example, when medical imaging equipment acquires low-quality image data, a CT scanner is first used to scan at 120 kV and 200 mA, acquiring DICOM format images with a resolution of 512 × 512 pixels. These images are subject to noise and blurring due to factors such as equipment limitations and patient movement.

[0111] To extract initial features, we used a pretrained ResNet-50 convolutional neural network, trained on the ImageNet dataset, consisting of 50 convolutional and fully connected layers. Low-quality images were fed into the ResNet-50 network, and the output of the fourth convolutional layer was extracted through forward propagation, resulting in a 256-dimensional feature vector. These feature vectors were then L2-normalized to generate the first feature set containing texture and structural information.

[0112] To further optimize the features, principal component analysis (PCA) was used to reduce the 256-dimensional features to 128 dimensions, preserving 95% of the variance. K-means clustering was used to partition the 128-dimensional feature set into 10 clusters, each representing a distinct texture and structural pattern. Finally, the t-SNE algorithm was used to map the high-dimensional features into a two-dimensional space, generating visualizations for subsequent analysis.

[0113] Furthermore, the process of performing gradient calculation processing on the first feature set to generate a second feature set including edge distribution features 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 gradient calculation is performed again by returning to the first feature set;

[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, data processing methods are 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, 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] For example, 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 changes in pixel intensity in the image, especially in the edge area.

[0120] In one possible implementation, the Sobel operator is used to process the image data in the first feature set, and the 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 marginal distribution features from non-marginal distribution features to obtain a 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 a preset threshold, such as feature point coverage of 20% of the image area. If not, return to the first feature set, readjust the Sobel operator parameters, such as increasing the convolution kernel size to 5×5, and recalculate the gradient.

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

[0127] Preferably, a linear Sexually Transmitted Infections (STI) algorithm is used to normalize the features, and the numerical range of the edge feature subset is mapped to 0 to 1 to generate a converted feature set.

[0128] For example, for 10,000 feature points, normalization generates a 128-dimensional feature vector with a uniform dimension. Based on this transformed feature set, the random forest algorithm is applied to generate a feature importance ranking. Random forests construct multiple decision trees to evaluate the contribution of each feature to the classification task.

[0129] For example, the edge strength feature may have an importance score of 0.85, while the smooth region feature may only have an importance score of 0.2, resulting in a ranked feature set.

[0130] It is understood that, for the sorted feature set, features with importance below a threshold, such as 0.3, are removed to obtain a reduced feature set. In one possible implementation, the top 50 highly important features are retained, and the dimensionality is reduced from 128 to 50. Based on the reduced feature set, feature generation operations are performed to construct an enhanced feature set.

[0131] For example, by combining features to generate new features, such as the product of edge strength and direction, we can generate an enhanced feature set that includes distribution characteristics, with the dimension potentially increasing to 60. For this enhanced feature set, we use the DBSCAN clustering algorithm to determine the distribution correlation between features.

[0132] For example, by setting the radius parameter to 0.5, three feature clusters are identified and the final feature set is generated, which contains a subset of features with high correlation, which is convenient for subsequent analysis.

[0133] Furthermore, based on the comparison result of the edge detail intensity of the second feature set with a preset threshold, a channel attention mechanism is used for processing, and the process of generating the third feature set includes:

[0134] Obtain edge detail information of the second feature set, extract edge intensity using an edge detection algorithm, and obtain edge intensity distribution;

[0135] If the intensity of at least one region 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;

[0136] According to the enhanced edge features, the channel attention mechanism is used to assign weights to obtain a weighted feature set;

[0137] By using the weighted feature set, a convolutional neural network is used to extract deep features and generate the third feature set;

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

[0139] According to the dimensionality reduction optimization feature set, mean pooling is used to determine the third feature set of the final feature representation.

[0140] For example, edge detail information from the second feature set can be obtained by analyzing the edge characteristics of the image data. Edge detail typically refers to areas in an image where pixel values ​​change significantly, reflecting the boundaries or texture of an object. For example, in the case of CT scan images, where the medical imaging data being processed is a CT scan, edge detail may correspond to tissue boundaries. Edge information can be extracted through gradient amplitude 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 one possible implementation, edge detection algorithms, such as the Canny algorithm, are used to extract edge intensity. The Canny algorithm smoothes the image using a Gaussian filter and then calculates the gradient magnitude and direction to generate an edge intensity distribution. Suppose the edge intensity distribution shows that the intensity of bone edges is between 100 and 200, while the intensity of some soft tissue edges is below 50. If the intensity of at least one region falls below a preset threshold of 50, the edge details need to be enhanced.

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

[0144] Specifically, when using the Sobel operator to enhance edge detail, it calculates gradients using horizontal and vertical convolution kernels to highlight edges. Continuing with the example of a CT image, applying the Sobel operator increases the intensity of soft tissue edges from 30 to 80, generating enhanced edge features. This enhancement helps preserve more detail 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 CT images, a weight of 0.7 is assigned to the bone edge and a weight of 0.3 is assigned to the soft tissue edge to generate a weighted feature set. This weight distribution makes the model focus more on key areas.

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

[0148] For example, in the CT image, the network can extract the geometric features of the bone edge, generating a third feature set with a dimension of 1024. If the dimension exceeds a preset range, such as 512, dimension reduction is required.

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

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

[0151] For example, when mean pooling is used to process the optimized feature set, mean pooling generates the final feature representation by taking the average value of the feature map region. Assuming that the optimized feature set contains 256 feature maps, each with 16x16 pixels, after mean 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 the embodiment is centered around the extraction and optimization of edge features, and logically progresses layer by layer. Edge detection, enhancement, weight assignment, deep feature extraction, dimension reduction, and pooling together form an efficient feature processing flow. This flow is particularly important in medical image analysis, as it can provide effective diagnostic evidence for doctors.

[0153] Further, the process of optimizing the generation of the first reconstructed image by evaluating the feature distribution difference through the discriminator network includes:

[0154] The third feature set is standardized to obtain a normalized feature set;

[0155] The normalized feature set is processed by the generative adversarial network to generate a preliminary reconstructed image;

[0156] The discriminator network is used to evaluate the feature distribution difference between the preliminary reconstructed image and the target image, and a distribution difference value is obtained;

[0157] If the distribution difference value is greater than a preset threshold, the generative adversarial network parameters are adjusted through an optimization algorithm to generate an optimized reconstructed image;

[0158] The clarity of the optimized reconstructed image is evaluated to obtain a quality evaluation result;

[0159] According to the quality evaluation result, the optimized reconstructed image is post-processed to obtain the first reconstructed image.

[0160] For example, when the third feature set is fed into a generative adversarial network, a U-Net generator is first employed, consisting of five layers of downsampling and five layers of upsampling. The number of convolution kernels in each layer is 64, 128, 256, 512, and 1024, respectively. LeakyReLU activation functions (with a negative slope of 0.2) and batch normalization layers are used. The input feature dimensions are 256×256×64, which are progressively upsampled to 512×512×3 RGB images through transposed convolutions to generate a preliminary reconstructed image. The discriminator adopts a PatchGAN architecture, consisting of four convolution layers (kernels 64→128→256→512). The final layer outputs a 30×30 feature map, with each pixel corresponding to a 70×70 receptive field of the input image. The Wasserstein distance loss function (with a gradient penalty coefficient of λ=10) was used as the loss function. The initial learning rate was set to 0.0001, and the Adam optimizer (β1=0.5, β2=0.999) was employed. During the optimization process, the generator loss function included a content loss (L1 loss with weight 100) and an adversarial loss (weight 1). After 50,000 training iterations, the feature distribution difference decreased from an initial 1.83 to 0.12, and the final output of the first reconstructed image achieved a PSNR of 28.6dB and an SSIM of 0.91. During training, the learning rate was dynamically adjusted every 1,000 iterations (with a decay rate of 0.95), and spectral normalization was used to constrain the discriminator weights to ensure training stability.

[0161] For example, after obtaining the third feature set, a normalization process is required. The normalization process aims to unify the dimension and distribution of the feature data to facilitate subsequent network processing.

[0162] For example, when processing the third feature set of CT images, assuming that the feature set contains a 1024-dimensional vector and the data range is between 0 and 255, the standardization can use the Z-score method 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, the mean and standard deviation are calculated and then normalized to generate a standardized feature set. This process ensures consistent data distribution across different feature channels, improving model stability.

[0164] In one possible implementation, the normalized feature set is fed into a generative adversarial network (GAN) to generate a preliminary reconstructed image. The GAN consists of a generator (which reconstructs the image from the feature set) and a discriminator (which 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. This preliminary image may contain bone and soft tissue areas, but details are blurred. The discriminator network compares the pixel distribution of the preliminary reconstructed image with that of the target CT image and calculates the difference in feature distribution.

[0166] For example, suppose 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 adversarial generative network parameters need to be adjusted through the optimization algorithm.

[0168] Preferably, the Adam optimization algorithm is used to update the generator weights through back propagation to reduce the distribution difference.

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

[0170] Specifically, the optimized reconstructed image is analyzed for clarity through an image quality assessment module. This quality assessment can be performed using the peak signal-to-noise ratio method, comparing the pixel differences between the optimized image and the target image.

[0171] For example, the evaluation results show a peak signal-to-noise ratio of 30dB, indicating that the image quality is high but still noisy. The denoising module then post-processes the optimized reconstructed image.

[0172] In one embodiment, a non-local means denoising algorithm is used to smooth noisy areas by analyzing pixel values ​​of similar areas in the image.

[0173] For example, the noise points in the soft tissue area are reduced, and the first reconstructed image is generated with smoother details.

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

[0175] For example, images are stored in the DICOM format, which contains a 512x512 pixel grayscale image 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 processing on the first reconstructed image to generate a second reconstructed image that satisfies the physical constraints includes:

[0177] Obtaining pixel intensity data and artifact distribution characteristics of the first reconstructed image, determining the artifact area using a feature extraction algorithm, and obtaining an artifact distribution feature map;

[0178] 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;

[0179] By using the constrained optimization method, combined with physical constraints and imaging model parameters, the artifact distribution feature map is optimized to obtain the corrected pixel intensity distribution.

[0180] The spatial consistency of the image is reconstructed 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, an iterative optimization algorithm is used to adjust the pixel intensity to obtain an optimized corrected image, and the optimized corrected image is verified to see whether it meets the physical constraints to obtain a second reconstructed image;

[0182] The image spatial consistency and the degree of artifact residue are calculated for the second reconstructed image, the accuracy of the correction algorithm is determined, and the final corrected second reconstructed image is obtained.

[0183] Exemplarily, when performing artifact correction in this embodiment, a linear attenuation coefficient matrix is ​​first established based on the attenuation characteristics of ray propagation. For example, for CT imaging, assuming the detector receives 1200 sets of raw projection data, each containing 512 sampling points, an iterative reconstruction algorithm (such as the SART algorithm) is used to calculate the initial reconstructed image. Physical constraints are introduced during the iterative process, such as setting a reasonable range for the attenuation coefficient to 0.005 / mm to 0.03 / mm. Pixel values ​​outside this range are corrected using non-negative least squares. To address metal artifacts, a correction method based on energy spectrum decomposition is employed. Assuming a scanning voltage of 140 kVp, dual-energy spectrum data is used to separate the base material images of iron (Fe) and titanium (Ti), with K-edge energies of 7.11 keV and 4.96 keV, respectively. Polynomial fitting is then used to correct energy spectrum drift errors. To address scattering artifacts, Monte Carlo simulation was used to calculate the scattering distribution, assuming a 15% scattering contribution from the water phantom. A scattering correction model was constructed using a convolutional neural network (CNN). The network consisted of five convolutional layers, each using a Reinforced Luminance (ReLU) activation function. The resulting projection data were then fed into the FDK reconstruction algorithm, using a Hamming window filter with a Nyquist cutoff frequency of 0.8. This resulted in a second reconstruction image that met physical constraints, improving the SSIM performance from 0.72 to 0.89.

[0184] Furthermore, intensity correction is performed on the second reconstructed image according to the pre-trained device bias pattern to generate a third reconstructed image, including:

[0185] Obtain pixel intensity data based on the second reconstructed image, match it with the pre-trained device deviation pattern, and obtain a deviation distribution feature;

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

[0187] Generate a preliminary third reconstructed image based on the corrected intensity data in combination with an image reconstruction algorithm;

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

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

[0190] Performing pixel smoothing processing on the optimized third reconstructed image through an image generation algorithm to obtain a final third reconstructed image;

[0191] Quality parameters are extracted from the final third reconstructed image, and the image is evaluated using a preset quality standard to obtain a third reconstructed image that has completed image correction.

[0192] Exemplarily, the pre-trained device bias model in this embodiment is implemented using a deep learning convolutional neural network (CNN), for example, the ResNet50 architecture. Image features are extracted by loading pre-trained weights (e.g., a model trained on the ImageNet dataset). The second reconstructed image (512×512 pixels, grayscale range 0-255) is first input into the model. The feature map output by the fourth convolution layer (256×64×64 dimensions) is extracted. The global average pooling value of this feature map is calculated as a quantitative indicator of device bias. Assuming a bias value of 0.15 (normalized to the range -1 to 1), a linear correction algorithm is used to adjust the image intensity. The correction formula is I_corrected = I_original × (1 + 0.15 × 0.8), where 0.8 is an empirical decay factor to prevent overcorrection. When generating the third reconstructed image after correction, the grayscale values ​​are constrained to the range 0-255, and any excess is truncated. For example, if a pixel's original value is 200 and its corrected value is 200 × 1.12 = 224, it will be retained if it does not exceed the threshold. If the corrected value is 260, it will be forced to 255. The entire process is implemented in the PyTorch framework, using CUDA acceleration, and processing a single image takes approximately 23 milliseconds (on an NVIDIA V100 graphics card). To verify the effectiveness of the correction, the peak signal-to-noise ratio (PSNR) of the images before and after correction is calculated. If the PSNR improves by more than 3dB, the correction is considered effective.

[0193] Furthermore, a multi-scale convolutional network is used to perform feature extraction and fusion processing on the third reconstructed image to generate a fourth reconstructed image, including:

[0194] Performing feature extraction on the third reconstructed image through a preset multi-scale convolutional network to obtain a multi-scale feature set;

[0195] A multi-scale convolutional network is used to fuse multi-scale feature sets to generate fused feature representations;

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

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

[0198] Obtaining a feature space distribution of the fourth reconstructed image and determining whether a preset feature consistency condition is satisfied; if the feature consistency condition is not satisfied, adjusting 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, the image quality is calculated using a preset evaluation function to determine the fourth reconstructed image to be output finally.

[0200] For example, when a multi-scale convolutional network is used to perform feature extraction and fusion processing on the third reconstructed image, the image is first convolved with convolution kernels of different scales (such as 3×3, 5×5, 7×7) to capture detail information at different levels.

[0201] For example, a 3×3 convolution kernel is used to extract local features, a 5×5 convolution kernel is used to extract mid-range features, and a 7×7 convolution kernel is used to extract global features. Each convolution layer is followed by a ReLU activation function to enhance nonlinear expression capabilities. Next, the feature maps of different scales are fused through a feature fusion module. Using a weighted average method, the weight values ​​are dynamically adjusted based on the importance of the feature map, such as a weight of 0.4 for local features, 0.3 for mid-range features, and 0.3 for global features. The fused feature maps are then processed through a 1×1 convolution layer for dimensionality reduction to reduce computational complexity. Finally, an upsampling operation is used to restore the feature map to its original image size to generate the fourth reconstructed image. Throughout the entire process, the Adam optimizer is used for parameter updates, with a learning rate set to 0.001 and 100 iterations to ensure model convergence and generate a high-quality fourth reconstructed image.

[0202] Furthermore, the fourth reconstructed image is input into a residual refinement network for detail enhancement processing to generate a final high-quality reconstructed image, which includes:

[0203] Perform feature extraction on the fourth reconstructed image through the 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 improved through the 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 quality assessment module performs a clarity test on the preliminary high-quality image to determine whether the clarity meets the preset standard;

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

[0208] Obtain the final high-quality image, adjust the image format through the format conversion module, and generate the target output image;

[0209] Specifically, the residual refinement network extracts features from the fourth reconstructed image through a multi-layer residual connection structure. The residual network introduces "shortcut connections" to learn the differences between features, enhancing the ability to express details.

[0210] In one embodiment, the network includes five residual blocks, each consisting of a 3×3 convolution kernel and a batch normalization layer, with a total of 64 convolution kernels. The extracted detail-enhanced features contain texture and edge information. If the resolution of the detail-enhanced features is below a preset threshold, such as 512×512 pixels, super-resolution network processing is required.

[0211] In one possible implementation, the super-resolution network can adopt a generative adversarial network structure, where the generator improves the resolution by 4x upsampling, and the adversary determines the authenticity of the generated image.

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

[0213] Preferably, the image fusion algorithm may adopt a fusion method based on an attention mechanism, the principle of which is to highlight and optimize the detail area of ​​the feature image through the attention weight, and fuse it with the global information of the fourth reconstructed image.

[0214] In one embodiment, the attention weight is calculated based on gradient information, and the weight of the detail area 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 achieved by calculating the gradient amplitude and entropy value of the image.

[0216] For example, if the gradient amplitude is less than 50 or the entropy value is less than 7, the image clarity is considered substandard and requires secondary enhancement. This secondary enhancement also uses the residual refinement network, adjusting the number of convolution kernels to 128 to increase feature expression and generate a high-quality final image.

[0217] It is understandable 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 a balance between visual quality and storage efficiency.

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

[0220] Residual refinement and super-resolution enhance details, super-resolution networks supplement when resolution is insufficient, fusion algorithms integrate multi-source information, quality assessment ensures output standards, and format conversion adapts to application requirements. Each link supports each other to achieve the goal of high-quality image generation, with rigorous logic and high efficiency.

[0221] Furthermore, the process of extracting feature data from the final high-quality reconstructed image, generating a heat map using a dual-pathway network, and locating suspicious lesion areas 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] A convolutional neural network is used to perform preliminary feature extraction on the denoising feature set to generate a feature set after feature extraction;

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

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

[0226] Extracting regional boundary features based on the segmented suspicious lesion area to generate second heat map data;

[0227] Performing coordinate mapping on the second thermal 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 the preset threshold, the final positioning result is output and the lesion area is determined.

[0229] For example, feature data is extracted from the final high-quality reconstructed image. A convolutional neural network (CNN) is first used for feature extraction, using ResNet-50 as the base model. The input image size is 224×224. Five convolutional and pooling layers are used to extract multi-level features, resulting in an output feature map of 7×7×2048. Next, a dual-path network (DPN) is used to generate a heat map. The DPN consists of two parallel paths: one path extracts local features using a 3×3 convolution kernel, and the other path extracts global features using a 1×1 convolution kernel. The features from both paths are fused in the final layer, resulting in a heat map of 14×14×1. After the heat map is generated, a threshold segmentation algorithm is used to locate suspicious lesions. A threshold of 0.7 is set, and regions in the heat map with a value greater than this threshold are marked as suspicious lesions. Connected component analysis is then used to cluster these marked regions to determine the final lesion locations. Finally, a support vector machine (SVM) was used to classify the extracted features, using the radial basis function (RBF) as the kernel function. Cross-validation determined the optimal parameters C = 1.0 and gamma = 0.01, achieving a classification accuracy of 95.6%. This entire process was automated, requiring no human intervention, ensuring both efficiency and accuracy.

[0230] Example 2

[0231] like Figure 2 As 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 is used to obtain low-quality image data based on medical imaging equipment and use a pre-trained convolutional neural network to extract initial features 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 including edge distribution features;

[0234] an attention processing module, configured to generate a third feature set by using a channel attention mechanism based on a comparison result of the edge detail intensity of the second feature set with a preset threshold;

[0235] An adversarial generation module is configured to input the third feature set into the adversarial generation network to generate a preliminary reconstructed image, and optimize and generate a first reconstructed image after evaluating the feature distribution difference through the 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 satisfies 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 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] The lesion localization module is used to extract feature data from the final high-quality reconstructed image, generate a heat map using a dual-pathway network, and locate suspicious lesion areas;

[0241] The diagnostic labeling module is used to label the suspicious lesion areas according to the prior knowledge of pathology and generate diagnostic auxiliary images.

[0242] The deep learning-based medical image reconstruction system provided in this embodiment has all the advantages of the deep learning-based medical image reconstruction method provided in Example 1.

[0243] Example 3

[0244] This embodiment further discloses a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method described in the first embodiment.

[0245] Example 4

[0246] This embodiment further discloses a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the method described in the first embodiment are implemented.

[0247] Example 5

[0248] This embodiment further discloses a computer program product, including a computer program, which implements the steps of the method described in the first embodiment when executed by a processor.

[0249] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A medical image reconstruction method based on deep learning, characterized in that: include: Low-quality image data is acquired based on medical imaging equipment, and a pre-trained convolutional neural network is used to extract initial features to generate a first feature set containing texture and structure information; Performing gradient calculation on the first feature set to generate a second feature set including edge distribution features; 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 the result to generate a third feature set; Inputting the third feature set into a generative adversarial network to generate a preliminary reconstructed image, and optimizing and generating a first reconstructed image after evaluating feature distribution differences through a discriminator network; performing artifact correction processing on the first reconstructed image to generate a second reconstructed image that satisfies physical constraints; performing intensity correction on the second reconstructed image according to a pre-trained device bias pattern to generate a third reconstructed image; Performing feature extraction and fusion processing on the third reconstructed image using a multi-scale convolutional network to generate a fourth reconstructed image; Inputting the fourth reconstructed image into a residual refinement network for detail enhancement processing to generate a final high-quality reconstructed image; Extracting feature data from the final high-quality reconstructed image, generating a heat map using a dual-pathway network, and locating suspicious lesion areas; Performing type labeling processing on the suspicious lesion area based on prior pathological knowledge to generate a diagnosis-assisted image; The process of performing artifact correction processing on the first reconstructed image to generate a second reconstructed image that satisfies physical constraints includes: Obtaining pixel intensity data and artifact distribution characteristics of the first reconstructed image, determining the artifact area using a feature extraction algorithm, and obtaining 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; The spatial consistency of the image is reconstructed through the corrected pixel intensity distribution to 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; The image spatial consistency and the degree of residual artifacts are calculated for the second reconstructed image, the accuracy of the correction algorithm is determined, and the final corrected second reconstructed image is obtained.

2. The method according to claim 1, characterized in that The process of generating the first feature set containing texture and structure information includes: Obtain low-quality image data output by medical imaging equipment, use a pre-trained convolutional neural network to extract initial features, and generate a first feature set containing texture and structure information; Processing low-quality image data corresponding to the first feature set by adaptive histogram equalization to enhance image contrast and obtain an image data set; Performing secondary feature extraction on the image dataset using a pre-trained convolutional neural network 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, the texture of the information feature set is optimized by generating an adversarial network to obtain an optimized feature set; otherwise, the information feature set is directly output as the optimized feature set; According to the optimized feature set, a principal component analysis method is used to perform dimensionality reduction processing on texture and structure information to obtain a dimensionality reduction feature set; Performing classification prediction on the dimensionality reduction feature set through a pre-trained convolutional neural network, judging the quality level of the image data, and obtaining a classification result; According to the classification results, the weighted fusion method is used to integrate the texture and structure information of the dimensionality reduction feature set to generate the target feature set.

3. The method according to claim 1, characterized in that The process of performing gradient calculation on the first feature set to generate a second feature set including edge distribution features further includes: 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; 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 gradient calculation is performed again by returning to the first feature set; 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, data processing methods are 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 applied to the enhanced feature set to determine the distribution correlation between features and obtain a target feature set.

4. The method according to claim 1, wherein Based on the comparison result of the edge detail intensity of the second feature set with a preset threshold, a channel attention mechanism is used to process the result to generate a third feature set, which includes: Obtain edge detail information of the second feature set, extract edge intensity using an edge detection algorithm, and obtain edge intensity distribution; If the intensity of at least one region 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 performed to determine a third feature set for final feature representation.

5. The method according to claim 1, wherein 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 intensity correction on the second reconstructed image according to a pre-trained device bias pattern to generate a third reconstructed image includes: Obtain pixel intensity data based on the second reconstructed image, and match it with 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; generating a preliminary third reconstructed image based on the corrected intensity data in combination with an image reconstruction algorithm; Acquiring edge features of the preliminary third reconstructed image, verifying the deviation compensation effect through pattern matching, and obtaining a verification result; If the verification result shows that the deviation compensation is insufficient, intensity adjustment is performed on the edge feature to generate an optimized third reconstructed image; Performing pixel smoothing processing on the optimized third reconstructed image using an image generation algorithm to obtain a final third reconstructed image; Quality parameters are extracted from the final third reconstructed image and evaluated using a preset quality standard to obtain a third reconstructed image that has completed image correction.

7. The method according to claim 1, characterized in that The process of performing feature extraction and fusion processing on the third reconstructed image using a multi-scale convolutional network to generate a fourth reconstructed image includes: Performing feature extraction on the third reconstructed image through a preset multi-scale convolutional network to obtain a multi-scale feature set; Using a multi-scale convolutional network to fuse 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, adjusting the feature resolution by upsampling to obtain an adjusted feature representation; Reconstructing the adjusted feature representation through a decoding network to generate a fourth reconstructed image; Obtaining a feature spatial distribution of the fourth reconstructed image, determining whether a preset feature consistency condition is satisfied, and if the feature consistency condition is not satisfied, adjusting the feature extraction and fusion process by optimizing network parameters to obtain an optimized fourth reconstructed image; According to the optimized fourth reconstructed image, the image quality is calculated using a preset evaluation function to determine the fourth reconstructed image to be output finally.

8. The method according to claim 1, characterized in that The process of extracting feature data from the final high-quality reconstructed image, generating a heat map using a dual-pathway network, and locating suspicious lesion areas includes: Acquire pixel-level feature data from the final high-quality reconstructed image, remove noise using a preprocessing algorithm, and obtain a denoised feature set; Performing preliminary feature extraction on the denoising feature set using a convolutional neural network to generate a feature set after feature extraction; Processing the feature set after feature extraction through a dual-path network to generate first heat map data; If the significance value of the first heat map data is greater than a preset threshold, a clustering algorithm is used to segment the heat map data to determine a suspicious lesion area; Extracting regional boundary features based on the segmented suspicious lesion area to generate second heat map data; Performing coordinate mapping on the second thermal map data by using a geometric analysis algorithm to obtain the precise position coordinates of the lesion area; If the deviation of the precise position coordinates is less than a preset threshold, the final positioning result is output to determine the lesion area.

9. A medical image reconstruction system based on deep learning, characterized in that: include: An initial feature extraction module is used to obtain low-quality image data based on medical imaging equipment and use a pre-trained convolutional neural network to extract initial features to generate a first feature set containing texture and structure information; 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; an attention processing module, configured to generate a third feature set by using a channel attention mechanism based on a comparison result of the edge detail intensity of the second feature set with a preset threshold; An adversarial generation module is configured to input the third feature set into a generative adversarial network to generate a preliminary reconstructed image, and to optimize and generate a first reconstructed image after evaluating feature distribution differences through a discriminator network; 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; 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; a multi-scale fusion module, configured to perform 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, configured to input 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, configured to extract feature data from the final high-quality reconstructed image, generate a heat map using a dual-pathway network, and locate suspicious lesion areas; A diagnostic labeling module is used to label the suspicious lesion area according to prior knowledge of pathology and generate a diagnostic auxiliary image; The process of performing artifact correction processing on the first reconstructed image to generate a second reconstructed image that satisfies physical constraints includes: Obtaining pixel intensity data and artifact distribution characteristics of the first reconstructed image, determining the artifact area using a feature extraction algorithm, and obtaining 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; The spatial consistency of the image is reconstructed through the corrected pixel intensity distribution to 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; The image spatial consistency and the degree of residual artifacts are calculated for the second reconstructed image, the accuracy of the correction algorithm is determined, and the final corrected second reconstructed image is obtained.

Images