Medical image reconstruction method based on deep learning

Abstract

The present invention belongs to the interdisciplinary technical field of medical image processing and artificial intelligence. Disclosed is a medical image reconstruction method based on deep learning. The method comprises: acquiring a medical image data set, and labeling corresponding key anatomical structure information; on the basis of a feature encoder and a graph structure processor, constructing a medical image reconstruction model; the feature encoder performing feature extraction on the medical image data set by means of a convolutional neural network, so as to obtain a multi-scale feature map; the graph structure processor constructing an initial landmark point set of the multi-scale feature map by means of a graph attention network, and analyzing topological relationships between landmark points, so as to obtain an optimized landmark point arrangement; on the basis of the optimized landmark point arrangement, constructing an anatomical structure graph; on the basis of the anatomical structure graph, designing an adaptive loss function to train the medical image reconstruction model; and on the basis of the trained medical image reconstruction model, obtaining a corresponding reconstructed image. The present invention achieves the objective of improving the detail representation of a reconstructed image while maintaining the accuracy of an anatomical structure.

Description

A Deep Learning-Based Medical Image Reconstruction Method Technical Field

[0001] This invention belongs to the field of medical image processing and artificial intelligence, and in particular relates to a medical image reconstruction method based on deep learning. Background Technology

[0002] Medical image reconstruction is a key technology in the field of medical imaging. With the rapid development of deep learning technology, deep learning-based medical image reconstruction methods have shown great potential and are expected to significantly improve the quality and accuracy of reconstructed images.

[0003] However, existing medical image reconstruction methods still have limitations when dealing with complex anatomical structures. Traditional methods often struggle to accurately capture the fine structures and topological relationships of human tissues and organs, leading to distortion or loss of detail in certain key areas of the reconstruction results. Furthermore, the arrangement of reference points significantly impacts reconstruction accuracy, but effective reference point optimization strategies are currently lacking. Particularly when dealing with highly complex human structures, existing methods struggle to simultaneously maintain overall structural integrity and local detail accuracy. These technical limitations result in reconstruction results that fail to meet the accuracy requirements of clinical applications, especially in scenarios demanding high-precision imaging.

[0004] Therefore, there is an urgent need to propose a deep learning-based medical image reconstruction method to improve the quality of medical image reconstruction. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a deep learning-based medical image reconstruction method to resolve the issues present in the prior art.

[0006] To achieve the above objectives, this invention provides a deep learning-based medical image reconstruction method, comprising the following steps:

[0007] Acquire a medical image dataset and annotate the key anatomical structure information corresponding to each image;

[0008] A medical image reconstruction model was constructed based on a feature encoder and a graph structure processor.

[0009] The feature encoder extracts features from the medical image dataset using a convolutional neural network to obtain multi-scale feature maps;

[0010] The graph structure processor constructs an initial set of reference points for a multi-scale feature map through a graph attention network, and dissects the topological relationships between the reference points to obtain an optimized arrangement of reference points.

[0011] An anatomical structure diagram is constructed based on the optimized arrangement of reference points, where nodes represent key anatomical structures and edges represent the topological relationships between structures.

[0012] An adaptive loss function was designed based on anatomical structure diagrams to train a medical image reconstruction model;

[0013] The medical image to be reconstructed is input into the trained medical image reconstruction model to obtain the corresponding reconstructed image.

[0014] Optionally, the process of acquiring a medical image dataset and annotating each image with corresponding key anatomical structure information includes:

[0015] The image content is extracted from the collected medical image data, and the anatomical structure category is determined by a pre-established classification model to obtain anatomical structure information;

[0016] Based on anatomical information, a convolutional neural network is used to identify key regions.

[0017] By using the feature data of key areas, the location coordinates of the areas are calculated to obtain key location information;

[0018] Based on key location information and anatomical structure information, key anatomical structure information is generated, and then medical images are annotated.

[0019] Optionally, the process by which the graph structure processor constructs an initial set of reference points for a multi-scale feature map using a graph attention network includes:

[0020] The distribution of high-response regions in the multi-scale feature map is obtained, and a clustering algorithm is used to group the high-response regions to determine the location of the reference point in each group.

[0021] Calculate the correspondence between the grouped reference points and anatomical structures to obtain the matching results;

[0022] If the matching result is lower than the preset threshold, the coordinates of the reference point are adjusted by weighted average method to obtain the optimized set;

[0023] Based on the optimized set, the distribution features of key points in medical images are extracted to determine the boundaries of anatomical structures;

[0024] By fusing boundary information to determine the distribution of high-response regions, the final mapping relationship between the reference point and the anatomical structure is determined.

[0025] After obtaining the mapping relationship, an initial set of reference points containing key anatomical structure information is generated.

[0026] Optionally, the process of obtaining the optimized reference point arrangement by defining the topological relationships between the anatomical reference points includes:

[0027] The initial set of reference points is processed by a graph attention network to obtain the topological relationships between the reference points and to obtain preliminary distribution characteristics.

[0028] Based on the preliminary distribution characteristics, network structure analysis is used to analyze the distribution of point sets and determine the correlation strength between reference points;

[0029] If the association strength is lower than a preset threshold, the distribution of the point set is adjusted by weighted calculation to obtain the enhanced topological relationship;

[0030] For the enhanced topological relationships, obtain the distribution characteristics of the optimized arrangement and determine the spatial consistency between point sets;

[0031] By analyzing the spatial consistency of the network processing results, the final optimized layout is determined, and thus the optimized baseline arrangement is obtained.

[0032] Optionally, the process of constructing the anatomical structure map based on the optimized reference point arrangement includes:

[0033] An anatomical diagram is generated by arranging optimized reference points.

[0034] The strength of topological relationships is calculated based on the node and edge data of the anatomical structure diagram. A weighted method is used to adjust the connection pattern to obtain the enhanced structural distribution characteristics.

[0035] Based on the enhanced structural distribution features, a convolutional neural network is used to process the spatial relationships between nodes, determine the stability of the connection patterns, and obtain the optimized initial representation.

[0036] Boundary features of key anatomical structures are extracted from the optimized initial representation. If the boundary features are consistent with the topological relationship, the current connection pattern is retained; otherwise, the node data is regrouped using a clustering algorithm.

[0037] Update the edge data based on the grouped node data, obtain the new structural distribution, and determine the global consistency of the anatomical structure diagram;

[0038] By analyzing the changing trends of the initial expression through global consistency analysis, and adjusting the distribution characteristics through iterative calculation, the final anatomical structure layout is obtained.

[0039] Optionally, the process of training the medical image reconstruction model by designing an adaptive loss function based on anatomical structure diagrams includes:

[0040] Topological data is extracted from anatomical structure diagrams, and global consistency features are calculated using a weighted method to obtain the initial loss function;

[0041] Based on the topological data analysis of the key area distribution, it is determined whether the reconstruction deviation exceeds the preset threshold. If it does, the loss weight is adjusted through feature enhancement to obtain the updated loss function.

[0042] For the updated loss function, obtain the overall integrity data of the network reconstruction, calculate the local feature distribution of key regions, and determine the bias correction parameters;

[0043] The initial representation is adjusted by bias correction parameters, and the spatial correlation of topological data is processed by convolutional neural network to obtain optimized reconstructed features.

[0044] Based on the optimized reconstruction features, determine whether the global consistency condition is met. If not, recalculate the loss weights using a weighted method to obtain a new loss function.

[0045] To address the new loss function, boundary features of key regions are extracted, and clustering algorithms are used to group topology data to obtain the final network reconstruction results, thereby completing the training of the medical image reconstruction model.

[0046] Optionally, after inputting the medical image to be reconstructed into the trained medical image reconstruction model to obtain the corresponding reconstructed image, the method further includes:

[0047] Reconstructed images are obtained through generative adversarial networks, and a discriminator evaluates the realism and anatomical integrity of the reconstructed images to obtain preliminary evaluation data.

[0048] Based on the preliminary evaluation data, a feedback mechanism was used to adjust the generator parameters, optimize the reconstructed image, and obtain improved image data.

[0049] For the improved image data, the discriminator evaluates the accuracy of the anatomical structure. If the accuracy of the anatomical structure is lower than a preset threshold, the medical image reconstruction model is further adjusted through network training until the accuracy of the anatomical structure meets the preset threshold, and the updated reconstruction result is obtained.

[0050] The present invention also proposes 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.

[0051] The present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method.

[0052] The present invention also proposes a computer program product, including a computer program that, when executed by a processor, implements the steps of the method.

[0053] Compared with the prior art, the present invention has the following advantages and technical effects:

[0054] This invention discloses a deep learning-based medical image reconstruction method. The method first acquires a dataset of medical images labeled with anatomical structure information and extracts multi-scale features using a convolutional neural network. Then, a graph attention network is used to process high-response regions in the feature map, learning the topological relationships between key anatomical structures to construct an anatomical structure map. Based on this structure map, an adaptive loss function is designed to guide the reconstruction process within a generative adversarial network framework, maintaining overall structural integrity and improving the accuracy of key regions. Through iterative training and parameter tuning, this invention achieves the goal of improving the detail representation of reconstructed images while maintaining the accuracy of anatomical structures, providing a new solution for high-quality reconstruction of complex human organ medical images. Attached Figure Description

[0055] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0056] Figure 1 is a schematic diagram of the method flow according to an embodiment of the present invention;

[0057] Figure 2 is a schematic diagram of the clustering and grouping of high-response regions according to an embodiment of the present invention;

[0058] Figure 3 is a flowchart illustrating the optimized anatomical structure according to an embodiment of the present invention. Detailed Implementation

[0059] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0060] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0061] Example 1

[0062] As shown in Figure 1, this embodiment provides a medical image reconstruction method based on deep learning, including the following steps:

[0063] Acquire a medical image dataset and annotate the key anatomical structure information corresponding to each image;

[0064] A medical image reconstruction model was constructed based on a feature encoder and a graph structure processor.

[0065] The feature encoder extracts features from the medical image dataset using a convolutional neural network to obtain multi-scale feature maps;

[0066] The graph structure processor constructs an initial set of reference points for a multi-scale feature map through a graph attention network, and dissects the topological relationships between the reference points to obtain an optimized arrangement of reference points.

[0067] An anatomical structure diagram is constructed based on the optimized arrangement of reference points, where nodes represent key anatomical structures and edges represent the topological relationships between structures.

[0068] An adaptive loss function was designed based on anatomical structure diagrams to train a medical image reconstruction model;

[0069] The medical image to be reconstructed is input into the trained medical image reconstruction model to obtain the corresponding reconstructed image.

[0070] The feasible process of acquiring a medical image dataset and annotating each image with corresponding key anatomical structure information includes:

[0071] The process involves extracting image content from collected medical image data, using a pre-established classification model to determine the anatomical structure category, and obtaining anatomical structure information. Based on the anatomical structure information, a convolutional neural network is used to determine key regions. Through the feature data of the key regions, the region location coordinates are calculated to obtain key location information. Based on the key location information and anatomical structure information, key anatomical structure information is generated, and then the medical images are labeled.

[0072] As a specific implementation method, acquiring medical image datasets is fundamental to the entire process, typically achieved through public databases or interfaces with medical institutions. For example, datasets containing lung CT images can be obtained from public databases such as TCIA, or cardiac MRI images can be collected using the PACS system interface of a hospital in collaboration with them. This data is usually stored in DICOM format, containing rich metadata and image content, providing raw material for subsequent analysis.

[0073] When extracting image content from image data, image processing techniques can be used to separate the foreground and background. For example, in a lung CT image, thresholding can be used to extract the lung tissue region, removing bone and external air. One possible approach is to use OpenCV to preprocess the grayscale image to obtain clear organ outlines, laying the foundation for subsequent classification. This approach improves data quality and reduces irrelevant interference.

[0074] Determining the category of anatomical structures using a pre-established classification model is a crucial step. For example, a trained ResNet model can be used to classify extracted lung images, determining whether they belong to a lobe, pulmonary artery, or bronchus. The model, trained on a large amount of labeled data, can recognize the texture and morphological features of anatomical structures. Specifically, given a 512x512 pixel CT slice as input, it outputs the category probability, such as 0.9 for a lung lobe. The efficiency of this method lies in its rapid identification of structural categories, providing direction for subsequent analysis. When analyzing complex structural features using convolutional neural networks based on the extracted structural information, key regions can be further focused.

[0075] In one embodiment, a U-Net network is applied to pulmonary artery images to extract key regional features of vascular branches, such as edge sharpness and curvature variations.

[0076] Understandably, the accuracy of coordinate calculation directly affects the quality of subsequent annotation. When annotating image content based on location and structural information, annotation data containing category and coordinates can be generated. For example, a lung CT image can be annotated as "pulmonary artery, (256, 300), 20x20".

[0077] In one embodiment, this information is stored in JSON format to facilitate data integration.

[0078] It's important to note that the intuitiveness of the annotations helps doctors quickly understand the image content. When comparing the annotated data with the original image content, if the difference exceeds a preset threshold, such as a key region overlap rate of less than 80%, structural information needs to be re-extracted. For example, if the annotation box deviates from the actual location of the pulmonary artery, the U-Net parameters are adjusted, the heatmap is regenerated, and the coordinates are updated. This iterative optimization improves annotation consistency and ensures data reliability. When integrating the adjusted annotated data into a medical image dataset, a complete dataset containing anatomical structure and location information can be output.

[0079] The feasible process of using a convolutional neural network to extract features from an input medical image to obtain a feature map containing multi-scale feature information includes:

[0080] First, the medical images are preprocessed by normalization and data augmentation. Then, a convolutional neural network model is constructed, which includes multi-scale feature extraction modules (such as different convolutional kernels, dilated convolution, ASPP, etc.) and an attention mechanism. The medical images are then input into the model, and multi-scale features are extracted and fused through convolutional layers to generate a feature map containing multi-scale feature information for subsequent medical image analysis tasks.

[0081] As a specific implementation method, preprocessing of medical images is a crucial step in ensuring data quality. For example, when normalizing CT images, the pixel value range can be adjusted from the original 0-4096 to the 0-1 range, unifying the grayscale range through linear transformation. This method can reduce the differences in pixel values ​​between different devices.

[0082] In one possible implementation, data augmentation can be achieved through rotation, flipping, or adding noise. For example, applying a random 5-10 degree rotation to lung MRI images or adding Gaussian noise to simulate scanning interference can expand the dataset size and improve the model's generalization ability. It is important to note that augmentation must preserve the realism of the anatomical structures to avoid excessive deformation that could lead to feature distortion.

[0083] When building a convolutional neural network model, the multi-scale feature extraction module is the core. Specifically, different sizes of convolutional kernels, such as 3x3 and 5x5, can be used to process lung CT images in parallel, capturing features from small blood vessels to large lung lobes. After inputting a lung CT image into the model, early convolutional layers can extract edge details such as blood vessel contours, middle convolutional layers capture texture features such as lung lobe boundaries, and high-level convolutional layers fuse global information to determine the overall structural category.

[0084] As shown in Figure 2, the feasible process of the graph structure processor constructing an initial set of reference points for a multi-scale feature map through a graph attention network includes:

[0085] The distribution of high-response regions in multi-scale feature maps is obtained, and clustering algorithms are used to group these regions, determining the location of reference points within each group. The correspondence between the grouped reference points and anatomical structures is calculated to obtain matching results. If the matching results are lower than a preset threshold, the reference point coordinates are adjusted using a weighted average method to obtain an optimized set. Based on the optimized set, the distribution features of key points in medical images are extracted to determine the boundaries of anatomical structures. The distribution of high-response regions is fused using boundary information to determine the final mapping relationship between reference points and anatomical structures. After obtaining the mapping relationship, an initial set of reference points containing key anatomical structure information is generated.

[0086] As a specific implementation, when analyzing the distribution of high-response regions through feature map analysis to obtain an initial set of reference points, one can start with the feature map of a lung CT image. For example, high-response regions typically correspond to key structures such as pulmonary vessels or lung lobe boundaries, and areas with higher pixel values ​​may reflect significant anatomical features. In one possible implementation, assuming the feature map resolution is 256x256 and the pixel value range is between 0 and 1, a threshold such as 0.8 can be set to filter out high-response points as initial reference points. For instance, the pixel values ​​in the pulmonary artery region may be concentrated above 0.9, while the background region is close to 0.2; this distribution difference helps in initial localization.

[0087] When using clustering algorithms to group high-response regions and determine the location of reference points within each group, the K-means algorithm is a common choice. Specifically, high-response points can be divided into 3-5 clusters based on pixel values ​​and spatial coordinates, representing different anatomical regions. For example, for lung images, clustering may result in three groups: pulmonary artery, pulmonary vein, and lung lobe edge, with the reference point location within each group represented by the cluster center.

[0088] It should be noted that distance weights can be introduced during clustering to ensure that spatially close points are more likely to be grouped together, avoiding grouping distortion caused by scattered distribution. When calculating the correspondence between the grouped baseline points and anatomical structures, matching can be performed using pre-annotated lung structure templates.

[0089] In one embodiment, assuming the actual location of the pulmonary artery branch is known, the Euclidean distance between the reference point and the template is calculated. If the average distance is less than 5 pixels, the matching degree is considered to be high.

[0090] Preferably, directional features, such as gradient information of blood vessel direction, can be introduced to further verify the correspondence. For example, the reference point in the pulmonary artery region should be consistent with the direction of blood vessel extension. If the deviation angle exceeds 15 degrees, the matching result may be lower than expected. If the matching result is lower than a preset threshold, when adjusting the reference point coordinates using a weighted average method, weights can be assigned based on pixel values ​​and neighborhood information.

[0091] In one embodiment, the weight of a pixel with a value of 0.9 is set to 0.7, and the weight of neighboring pixels is set to 0.3. After adjustment, the reference point is closer to the high-response core region. For example, the initial coordinates of a reference point are (100, 120), which may become (102, 118) after adjustment, closer to the center of the pulmonary vessels.

[0092] Understandably, this method can smooth out the effects of noise and make the dataset more stable. Based on the optimized dataset, the distribution features of key points in the image are extracted. When determining the boundaries of anatomical structures, the changes in distance and density between points can be analyzed. For example, the point distribution at the lung lobe boundary is relatively continuous with a spacing of less than 3 pixels, while the point distribution in the vascular region is branched, and the spacing may increase to 10 pixels.

[0093] In one possible implementation, the boundary outline, such as the dividing line of a lung lobe region, is delineated by connecting adjacent key points.

[0094] It should be noted that boundary determination needs to incorporate global information from the feature map to avoid interference from local noise. When determining the mapping relationship between the final reference point and the anatomical structure by fusing the distribution of high-response areas through boundary information, overlay analysis can be used. Specifically, the boundary contour is overlaid with the feature map to reassess the attribution of each reference point. For example, points near the pulmonary artery boundary can be identified as vascular-related points, while points near the edge of the lung lobe are classified as boundary points.

[0095] Preferably, this fusion improves mapping accuracy, especially for complex structures such as bronchial branches. After obtaining the mapping relationship, when generating aggregate data containing the locations of key anatomical structures, a structured coordinate set can be output.

[0096] In one embodiment, the aggregate data includes coordinates of the pulmonary artery center point, such as (150, 200), and lung lobe boundary points, such as (180, 220), in a tabular or list format. For example, for a 512x512 CT slice, the final aggregate may contain 50 key points, clearly labeled with their anatomical affiliation.

[0097] Understandably, this data can directly support subsequent analysis, such as vessel tracking or region segmentation, significantly improving processing efficiency.

[0098] The feasible process of obtaining an optimized arrangement of reference points by establishing the topological relationships between the anatomical reference points includes:

[0099] The initial set of reference points is processed using a graph attention network to obtain the topological relationships between the reference points and obtain preliminary distribution characteristics. Based on the preliminary distribution characteristics, the network structure analysis is used to analyze the distribution of the point set and determine the correlation strength between the reference points. If the correlation strength is lower than a preset threshold, the distribution of the point set is adjusted through weighted calculation to obtain the enhanced topological relationship. For the enhanced topological relationship, the distribution characteristics of the optimized arrangement are obtained to determine the spatial consistency between the point sets. The spatial consistency analysis of the network processing results determines the final optimized layout, thereby obtaining the optimized arrangement of reference points.

[0100] As a specific implementation method, when processing the initial set of reference points using a graph attention network, the reference points can be regarded as nodes in a graph, and the attention mechanism can be used to capture the dependencies between points. For example, in a lung CT image, assuming that the initial set of reference points includes points of pulmonary vessels and lung lobe boundaries, the graph attention network can assign attention weights based on the spatial distance and feature similarity between points.

[0101] In one embodiment, points closer to the pulmonary artery may receive higher weights due to similar feature values, reflecting strong topological associations. After obtaining the topological relationships between reference points, preliminary distribution characteristics can be represented by the connection strength between points. Specifically, points in the pulmonary vascular region may form dense connections, while points at the lung lobe boundaries exhibit a linear distribution. In one possible implementation, the connection strength is determined by attention weights; for example, point pairs with weights higher than 0.7 are considered strongly associated.

[0102] It should be noted that this distribution characteristic can initially reveal the geometric properties of anatomical structures. When analyzing point set distribution using network structure analysis, topological features can be extracted through graph convolutional networks. For example, the point set of pulmonary artery branches may exhibit a tree-like structure, with the association strength quantified by the edge weights.

[0103] Preferably, if the weight of a point pair is below a threshold, such as 0.5, its association is considered weak. For weak associations, spatial proximity can be introduced as an auxiliary factor when adjusting the distribution through weighted calculation. For example, if a point initially has a weight of 0.4, but its neighbors are all strongly associated points, its weight can be increased to 0.6, forming an enhanced topological relationship. This adjustment enhances the coherence of the point set. For the enhanced topological relationship, the spatial consistency of the optimized arrangement's distribution characteristics can be determined by the distance and angle information between points.

[0104] In one embodiment, the spacing between lung lobe boundary points should be uniform, such as an average of 3 pixels, while the spacing between points in the vascular region may be larger. The spatial consistency analysis network can identify outliers by comparing the ideal distribution with the actual distribution. For example, if a point deviates from the expected trajectory by more than 10 pixels, it can be marked as an object to be optimized.

[0105] Understandably, this method improves the overall regularity of the point set. Clustering algorithms can be used for grouping when determining the final optimized layout. For example, the K-means algorithm can divide lung reference points into vascular and boundary groups.

[0106] In one embodiment, after clustering, the vascular group points are concentrated in areas with feature values ​​above 0.8, and the distance between boundary group points is less than 5 pixels. When fusing the topological relationships of the classified distribution features, the globally optimized arrangement can be confirmed through overlay analysis. Specifically, vascular group points match the tree-like topology, and boundary group points correspond to the linear topology. This fusion ensures that the point set grouping is consistent with the anatomical structure. For example, the optimization of the point set for the pulmonary artery region can be considered from multiple perspectives. In one possible implementation, the initial point set contains 20 points, the attention network identifies 15 strongly correlated points, and after clustering, they are divided into two groups, corresponding to the trunk and branches, respectively. After adjusting the weakly correlated points, the average distance between points is reduced from 8 pixels to 4 pixels, and the spatial consistency is significantly improved.

[0107] Preferably, after integrating the topological relationships, the deviation between the main trunk point and the blood vessel direction is less than 10 degrees, and the optimized layout is closer to the real anatomical structure.

[0108] Understandably, this multi-level processing can effectively improve the accuracy of subsequent analyses.

[0109] As shown in Figure 3, the feasible process of constructing the anatomical structure map based on the optimized reference point arrangement includes:

[0110] An anatomical structure diagram is generated by arranging optimized reference points. The strength of topological relationships is calculated based on the node and edge data of the diagram, and a weighted method is used to adjust the connection patterns, resulting in enhanced structural distribution features. For these enhanced features, a convolutional neural network is used to process the spatial relationships between nodes, determining the stability of the connection patterns and obtaining an optimized initial representation. Boundary features of key anatomical structures are extracted from the optimized initial representation. If the boundary features are consistent with the topological relationships, the current connection patterns are retained; otherwise, node data is regrouped using a clustering algorithm. Edge data is updated based on the grouped node data to obtain a new structural distribution and determine the global consistency of the anatomical structure diagram. The changing trend of the initial representation is analyzed through global consistency analysis, and iterative calculations are used to adjust the distribution features, resulting in the final anatomical structure diagram layout.

[0111] As a specific implementation method, extracting node data and edge data from the anatomical structure diagram is crucial when generating the diagram through an optimized arrangement of reference points. For example, in a lung CT image, node data represents key points of the pulmonary vessels, while edge data reflects the connections between these points. In one embodiment, assuming there are 10 nodes on a pulmonary artery branch, the edge data records the distances and directions between them, and a preliminary structural diagram can be drawn based on this.

[0112] It should be noted that the accuracy of node data directly affects the determination of the distribution location of anatomical structures. When determining the distribution location of key anatomical structures, the strength of the topological relationship can be calculated based on the node data and edge data. For example, the topological strength of pulmonary vascular nodes can be calculated by weighting the number and length of edges; nodes closer to the trunk have higher strength due to more connections.

[0113] When adjusting the connection pattern using a weighted method, it can be understood that if an edge is less than 5 pixels long and connects two high-intensity nodes, its weight can be increased to 0.8, forming an enhanced structural distribution feature. This adjustment can more clearly reflect the hierarchy of the anatomical structure. For the enhanced structural distribution feature, convolutional neural networks can be used to process the spatial relationships between nodes.

[0114] Specifically, the network can analyze the spatial proximity of nodes, such as the distance between nodes at the lung lobe boundary, which is usually maintained at about 3 pixels, and determine the stability of the connection pattern through convolutional layers.

[0115] In one possible implementation, if the node spacing in a certain connection pattern fluctuates by more than 20%, it is considered unstable and the initial expression needs to be optimized.

[0116] Preferably, a stable connection pattern provides a reliable foundation for subsequent boundary feature extraction. When extracting boundary features of key anatomical structures from the optimized initial representation, it is necessary to verify their consistency with the topological relationships. For example, the linear distribution of lung lobe boundaries should match the edge data between nodes. If a boundary point deviates from the expected trajectory by more than 8 pixels, the node data can be regrouped using the K-means clustering algorithm.

[0117] In one embodiment, after regrouping, the spacing between boundary group points is reduced to within 4 pixels, and the updated edge data better reflects the structural distribution. This method helps improve the local accuracy of anatomical structure maps. After obtaining the new structural distribution, global consistency analysis can be adjusted based on the changing trend of the initial expression. For example, if the node distribution in the pulmonary vascular region changes from scattered to concentrated, iterative calculations can gradually adjust the features, reducing the average distance between points from 10 pixels to 6 pixels.

[0118] It should be noted that the iterative process can gradually approximate the actual anatomical layout. After the final anatomical structure diagram is formed, it is crucial to verify the integrity of key anatomical structures through topological relationships. Specifically, the tree-like topology of the pulmonary artery trunk should contain at least 8 strongly connected nodes; if any are missing, the edge data weights can be adjusted and rematched.

[0119] In one possible implementation, for a lung anatomy diagram, the initial node data contains 15 points, and the edge data initially generates 20 connections. After optimization, the number of strong connections increases to 25, and integrity verification shows that the distribution of trunks and branches deviates from the actual anatomical structure by less than 5 degrees.

[0120] Understandably, this multi-level validation can significantly improve the practicality of structural diagrams and provide solid support for subsequent medical analysis.

[0121] The feasible process of training the medical image reconstruction model using an adaptive loss function based on anatomical structure diagrams includes:

[0122] Topological data is extracted from anatomical structures, and a weighted method is used to calculate global consistency features to obtain an initial loss function. Based on the topological data, the distribution of key regions is analyzed to determine if the reconstruction bias exceeds a preset threshold. If it does, feature enhancement is used to adjust the loss weights, resulting in an updated loss function. For the updated loss function, overall network reconstruction integrity data is obtained, and the local feature distribution of key regions is calculated to determine bias correction parameters. The initial representation is adjusted using these bias correction parameters, and a convolutional neural network is used to process the spatial correlation of the topological data to obtain optimized reconstruction features. Based on the optimized reconstruction features, it is determined whether global consistency is satisfied. If not, the weighted method is used to recalculate the loss weights, resulting in a new loss function. For the new loss function, boundary features of key regions are extracted, and a clustering algorithm is used to group the topological data to obtain the final network reconstruction result, thus completing the training of the medical image reconstruction model.

[0123] As a specific implementation method, the extraction of topological data is crucial in anatomical structure reconstruction. For example, for lung CT images, topological data can be obtained by analyzing the spatial distribution of vascular branch points. Specifically, the tree-like structure formed by the main pulmonary artery and its branches can be represented by nodes and edges, and the connections between nodes constitute the topological data. When calculating global consistency features using weighted methods, vessel diameter and branch angles can be considered.

[0124] In one possible implementation, the main pulmonary artery nodes have a higher weight, while the distal small branches have a lower weight. For example, the weight of the main pulmonary artery initiation node could be set to 0.9, while the weight of the terminal capillary node could be only 0.1.

[0125] The construction of the initial loss function needs to consider multiple factors. In one embodiment, node position deviation, edge length error, and branch angle error can be taken into account.

[0126] Preferably, each error can be assigned a different weight according to its impact on the overall structure. Key area distribution analysis is crucial for reconstruction quality.

[0127] It should be noted that structurally complex regions such as the interlobular spaces and hilar region are often more prone to reconstruction errors. For example, if the vascular node density in the hilar region is less than 20% of the expected density, it may be necessary to adjust the feature extraction strategy for that region. Feature enhancement is an effective means to improve reconstruction accuracy.

[0128] Specifically, key region features can be enhanced by increasing the number of convolutional layers or using attention mechanisms. In one embodiment, for the interlobar region, multi-scale feature fusion can be used to improve boundary recognition accuracy.

[0129] Obtaining overall integrity data involves multiple levels. Understandably, it requires considering not only the connectivity of the vascular system but also its relative positional relationship with surrounding tissues. For example, the parallelism between the pulmonary artery trunk and the bronchi can serve as one indicator of integrity.

[0130] Calculating the distribution of local features helps in refining the reconstruction. For example, in the lung apex region, the density and orientation of small vessel branches can be analyzed in detail to ensure consistency with the actual anatomical structure.

[0131] Preferably, a standard feature template for the region can be established by comparing multiple case samples. The determination of the deviation correction parameters requires comprehensive consideration of both global and local features.

[0132] In one possible implementation, different weights can be assigned based on the importance of different anatomical regions. For example, the weight of the correction parameters for the hilar region might be higher than that for the peripheral lung region. Spatial correlation processing of topological data is a key step in optimizing reconstructed features. Specifically, the continuity of vascular orientation can be inferred by analyzing the positional relationships between adjacent nodes.

[0133] In one embodiment, if the angle change between two adjacent nodes exceeds 45 degrees, a transition node may need to be inserted to ensure smoothness. The criteria for judging global consistency may include multiple aspects.

[0134] It is important to note that not only should the accuracy of the topological structure be considered, but also its consistency with clinical reality should be assessed. For example, the number of reconstructed pulmonary artery branches should correspond to anatomical knowledge, and the angular deviation of the main branches should not exceed 10 degrees. Boundary feature extraction is crucial for accurately delineating anatomical structures. For instance, in the interlobar region, the location of the interlobar fissure can be precisely determined by analyzing changes in CT value gradients.

[0135] In one possible implementation, combining morphological operations and edge detection algorithms can improve the robustness of boundary localization. Clustering algorithms play a crucial role in grouping topological data. Specifically, clustering can be performed based on the spatial location and connectivity of nodes to form different levels of vascular branch groups. For example, hierarchical clustering can divide the pulmonary artery system into multiple levels, such as the trunk, interlobar branches, and intersegmental branches. The final reconstruction results require multi-faceted evaluation for verification.

[0136] Understandably, in addition to visual morphological contrast, functional rationality should also be considered. In one embodiment, the physiological feasibility of reconstructing the vascular network can be verified by simulating hemodynamics, ensuring that the reconstruction result is not only morphologically accurate but also reflects the true physiological function.

[0137] Implementable, after inputting the medical image to be reconstructed into the trained medical image reconstruction model to obtain the corresponding reconstructed image, the method further includes:

[0138] Reconstructed images are obtained through a generative adversarial network (GAN). A discriminator evaluates the realism and anatomical integrity of the reconstructed images to obtain preliminary evaluation data. Based on the preliminary evaluation data, a feedback mechanism is used to adjust the generator parameters and optimize the reconstructed images to obtain improved image data. For the improved image data, the discriminator evaluates the accuracy of the anatomical structures. If the accuracy of the anatomical structures is lower than a preset threshold, the medical image reconstruction model is further adjusted through network training until the accuracy of the anatomical structures meets the preset threshold, and an updated reconstruction result is obtained.

[0139] As a specific implementation method, generative adversarial networks (GANs) play a crucial role in medical image reconstruction. For example, in the reconstruction of lung CT images, the generator can be constructed using a convolutional neural network, taking a low-dose CT image as input and outputting a high-quality reconstruction result. The discriminator then employs multi-scale feature extraction to evaluate the realism and anatomical integrity of the reconstructed image.

[0140] In one possible implementation, the initial evaluation data may include indicators such as structural similarity index and anatomical landmark deviation. Specifically, if the structural similarity index is below 0.85, or the deviation of key anatomical landmarks exceeds 2 mm, a feedback mechanism is triggered. The feedback mechanism enhances the reconstruction accuracy of key structures by adjusting the weights of the generator's loss function.

[0141] It should be noted that when the discriminator assesses the accuracy of anatomical structure evaluation, it can focus on complex areas such as interlobar spaces and the hilar region. For example, if the angle deviation of the vascular branches in the hilar region exceeds 15 degrees, it is considered insufficient in accuracy, triggering network training adjustments.

[0142] Preferably, a progressive training strategy can be employed to gradually improve the discriminator's discrimination ability. Anatomical feature extraction is fundamental to assessing global consistency. In one embodiment, global consistency can be determined by analyzing the topology of the vascular tree and the continuity of the lung lobe boundaries. Specifically, if the number of major vascular branches differs from the standard anatomical structure by more than 10%, it is considered insufficiently consistent.

[0143] Understandably, realism measurement and accuracy judgment need to be considered comprehensively. For example, in addition to structural similarity, the consistency of texture features can also be evaluated. In one possible implementation, if the CT value distribution of a local region differs from the real image by more than 50 HU, it is considered to have a significant deviation. Generator optimization is a key step in improving reconstruction quality.

[0144] Specifically, the accuracy of reconstructing key anatomical structures can be enhanced by introducing attention mechanisms. For example, for the interlobar region, multi-scale feature fusion can be used to improve the accuracy of boundary recognition.

[0145] Preferably, morphological manipulation can be combined to further optimize the representation of details.

[0146] As an additional implementation, this embodiment also includes comparing the reconstructed image with the original image, calculating evaluation metrics such as structural similarity index and peak signal-to-noise ratio, and quantitatively evaluating the performance of the reconstruction algorithm.

[0147] Specifically, by comparing the reconstructed image with the original image, a convolutional neural network is used to obtain structural similarity data. Features are extracted from the structural similarity data to determine the peak signal-to-noise ratio (PSNR) calculation range. For the PNR calculation range, mean squared error (MSE) is used to calculate the quantitative evaluation result. If the quantitative evaluation result is lower than a preset threshold, parameters are adjusted to update the data and obtain optimized evaluation data. Based on the optimized evaluation data, performance index features are extracted to determine the algorithm's stability. Consistency verification is performed using the algorithm stability data to determine the final performance index. Key region data is obtained from the final performance index to determine reconstruction consistency.

[0148] Example 2

[0149] This embodiment also proposes 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.

[0150] Example 3

[0151] This embodiment also proposes a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method.

[0152] Example 4

[0153] This embodiment also proposes a computer program product, including a computer program that, when executed by a processor, implements the steps of the method.

[0154] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A deep learning-based medical image reconstruction method, characterized by, Includes the following steps: Acquire a medical image dataset and annotate the key anatomical structure information corresponding to each image; A medical image reconstruction model was constructed based on a feature encoder and a graph structure processor. The feature encoder extracts features from the medical image dataset using a convolutional neural network to obtain multi-scale feature maps; The graph structure processor constructs an initial set of reference points for a multi-scale feature map through a graph attention network, and dissects the topological relationships between the reference points to obtain an optimized arrangement of reference points. An anatomical structure diagram is constructed based on the optimized arrangement of reference points, where nodes represent key anatomical structures and edges represent the topological relationships between structures. An adaptive loss function was designed based on anatomical structure diagrams to train a medical image reconstruction model; The medical image to be reconstructed is input into the trained medical image reconstruction model to obtain the corresponding reconstructed image.

2. The method of claim 1, wherein, The process of acquiring a medical image dataset and annotating each image with corresponding key anatomical structure information includes: The image content is extracted from the collected medical image data, and the anatomical structure category is determined by a pre-established classification model to obtain anatomical structure information; Based on anatomical information, a convolutional neural network is used to identify key regions. By using the feature data of key areas, the location coordinates of the areas are calculated to obtain key location information; Based on key location information and anatomical structure information, key anatomical structure information is generated, and then medical images are annotated.

3. The method according to claim 1, characterized in that, The process by which the graph structure processor constructs an initial set of reference points for a multi-scale feature map using a graph attention network includes: The distribution of high-response regions in the multi-scale feature map is obtained, and a clustering algorithm is used to group the high-response regions to determine the location of the reference point in each group. Calculate the correspondence between the grouped reference points and anatomical structures to obtain the matching results; If the matching result is lower than the preset threshold, the coordinates of the reference point are adjusted by weighted average method to obtain the optimized set; Based on the optimized set, the distribution features of key points in medical images are extracted to determine the boundaries of anatomical structures; By fusing boundary information to determine the distribution of high-response regions, the final mapping relationship between the reference point and the anatomical structure is determined. After obtaining the mapping relationship, an initial set of reference points containing key anatomical structure information is generated.

4. The method according to claim 3, characterized in that, The process of obtaining the optimized reference point arrangement by establishing the topological relationships between the anatomical reference points includes: The initial set of reference points is processed by a graph attention network to obtain the topological relationships between the reference points and to obtain preliminary distribution characteristics. Based on the preliminary distribution characteristics, network structure analysis is used to analyze the distribution of point sets and determine the correlation strength between reference points; If the association strength is lower than a preset threshold, the distribution of the point set is adjusted by weighted calculation to obtain the enhanced topological relationship; For the enhanced topological relationships, obtain the distribution characteristics of the optimized arrangement and determine the spatial consistency between point sets; By analyzing the spatial consistency of the network processing results, the final optimized layout is determined, and thus the optimized baseline arrangement is obtained.

5. The method according to claim 1, characterized in that, The process of constructing the anatomical structure diagram based on the optimized reference point arrangement includes: An anatomical diagram is generated by arranging optimized reference points. The strength of topological relationships is calculated based on the node and edge data of the anatomical structure diagram. A weighted method is used to adjust the connection pattern to obtain the enhanced structural distribution characteristics. Based on the enhanced structural distribution features, a convolutional neural network is used to process the spatial relationships between nodes, determine the stability of the connection patterns, and obtain the optimized initial representation. Boundary features of key anatomical structures are extracted from the optimized initial representation. If the boundary features are consistent with the topological relationship, the current connection pattern is retained; otherwise, the node data is regrouped using a clustering algorithm. Update the edge data based on the grouped node data, obtain the new structural distribution, and determine the global consistency of the anatomical structure diagram; By analyzing the changing trends of the initial expression through global consistency analysis, and adjusting the distribution characteristics through iterative calculation, the final anatomical structure layout is obtained.

6. The method according to claim 1, characterized in that, The process of training the medical image reconstruction model by designing an adaptive loss function based on anatomical structure diagrams includes: Topological data is extracted from anatomical structure diagrams, and global consistency features are calculated using a weighted method to obtain the initial loss function; Based on the topological data analysis of the key area distribution, it is determined whether the reconstruction deviation exceeds the preset threshold. If it does, the loss weight is adjusted through feature enhancement to obtain the updated loss function. For the updated loss function, obtain the overall integrity data of the network reconstruction, calculate the local feature distribution of key regions, and determine the bias correction parameters; The initial representation is adjusted by bias correction parameters, and the spatial correlation of topological data is processed by convolutional neural network to obtain optimized reconstructed features. Based on the optimized reconstruction features, determine whether the global consistency condition is met. If not, recalculate the loss weights using a weighted method to obtain a new loss function. To address the new loss function, boundary features of key regions are extracted, and clustering algorithms are used to group topology data to obtain the final network reconstruction results, thereby completing the training of the medical image reconstruction model.

7. The method according to claim 1, characterized in that, The process of inputting the medical image to be reconstructed into the trained medical image reconstruction model to obtain the corresponding reconstructed image further includes: Reconstructed images are obtained through generative adversarial networks, and a discriminator evaluates the realism and anatomical integrity of the reconstructed images to obtain preliminary evaluation data. Based on the preliminary evaluation data, a feedback mechanism was used to adjust the generator parameters, optimize the reconstructed image, and obtain improved image data. For the improved image data, the discriminator evaluates the accuracy of the anatomical structure. If the accuracy of the anatomical structure is lower than a preset threshold, the medical image reconstruction model is further adjusted through network training until the accuracy of the anatomical structure meets the preset threshold, and the updated reconstruction result is obtained.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1-7.

Images