Image system for accurate segmentation and function protection of endangered organs in radiotherapy plan
By combining graph neural networks and quantum graph convolution techniques with self-supervised learning, the problems of insufficient segmentation accuracy and functional protection of organs at risk in radiotherapy planning have been solved, achieving high-precision segmentation of organs at risk and functional avoidance.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- AFFILIATED HOSPITAL OF JINING MEDICAL UNIV
- Filing Date
- 2026-01-27
- Publication Date
- 2026-05-08
AI Technical Summary
Existing technologies in radiotherapy planning suffer from insufficient organ segmentation accuracy, are unable to effectively capture non-local correlations of complex anatomical structures, and fail to take into account internal organ functional differences during segmentation, thus failing to guide functional avoidance radiotherapy.
A non-Euclidean graph structure is constructed using graph neural networks. By combining quantum graph convolution and self-supervised learning, high-precision segmentation and functional protection of organs at risk are achieved through graph attention segmentation and morphological function mapping.
It enables high-precision segmentation of organs at risk in complex anatomical structures, automatically identifies and protects high-functioning regions inside organs, and improves the intelligent decision support capability for radiotherapy planning.
Smart Images
Figure CN121999221A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of technology, and more specifically to an imaging system for precise segmentation and functional protection of organs at risk in radiotherapy planning. Background Technology
[0002] In radiotherapy planning, accurately delineating the extent of organs at risk is crucial for protecting normal tissues and reducing side effects. However, traditional methods mainly rely on manual delineation or threshold / template-based automatic segmentation, which is time-consuming, labor-intensive, and has limited accuracy. In recent years, deep learning techniques such as convolutional neural networks (CNNs) have been introduced into the field of automatic medical image segmentation. For example, the U-Net model has achieved significant results in tasks including lung CT segmentation. However, conventional CNNs, based on regular grid convolutions, have fixed receptive fields, making it difficult to capture non-local correlations in organ morphology and unable to fully handle the global complex relationships of anatomical structures. To enhance global features, some studies have introduced attention mechanisms or multi-scale feature fusion, but these still have limitations in handling complex structures and long-distance dependencies.
[0003] Graph Neural Networks (GNNs), as a deep learning method for processing non-Euclidean structural data, have gained attention in medical image analysis in recent years. Modeling medical images or anatomical structures as graphs can more effectively represent the spatial relationships and topological structures between different regions. For example, some studies have combined GNNs with traditional segmentation networks to characterize the complex heterogeneity of lung tumors and tissues, improving segmentation accuracy and boundary details. These methods demonstrate that introducing graph structures helps capture complex spatial relationships, improving segmentation performance while maintaining computational efficiency. However, most current applications of GNNs are limited to Euclidean spatial nearest neighbor relationships, and there is still room for improvement in handling more complex non-Euclidean geometric relationships (such as curved surface adjacencies of organs) and fusing information from different modalities.
[0004] Furthermore, existing automated delineation methods rarely consider the differences in functional importance among different regions within an organ. In radiotherapy, incorporating functional imaging information can aid in the development of "functional avoidance" radiotherapy plans, that is, minimizing irradiation of high-functioning areas to reduce treatment side effects. For example, ventilation / perfusion functional imaging of the lungs can guide the avoidance of areas with good healthy lung function. However, obtaining such functional images (such as SPECT ventilation / perfusion scans) in clinical practice is often costly and complex. Therefore, some studies have explored using deep learning to synthesize functional distribution maps from conventional CT images, such as using 3D CT features to predict lung perfusion maps as an alternative to SPECT scans. This indicates that inferring organ function based on anatomical morphology is feasible, but currently these functional predictions are usually independent of the segmentation process, and a complete system integrating segmentation and functional assessment is still lacking.
[0005] On the other hand, with the development of computing technology, new algorithmic paradigms are bringing potential breakthroughs to medical image analysis. For example, quantum computing is being explored for optimization problems such as image segmentation. Some studies have modeled image pixel segmentation as a graph partitioning problem and solved it using quantum annealing. The results show that in some unsupervised segmentation tasks, quantum methods have faster solution speeds and better results than classical methods. This suggests that introducing quantum computing into medical image processing can overcome the bottleneck of classical computing and achieve globally optimal graph partitioning results. Furthermore, self-supervised learning, as a new trend in machine learning, can utilize unlabeled data for model pre-training, improving feature representation capabilities. In the field of medical imaging, spatiotemporal structure networks have been used to improve segmentation performance by utilizing the temporal information of CT sequences. Learning the spatiotemporal features of images through self-supervised learning can improve the model's adaptability to organ motion and morphological changes even with limited labeled data.
[0006] In summary, existing technologies have the following shortcomings in terms of automatic segmentation and functional protection of organs at risk: conventional deep learning segmentation methods lack the ability to capture complex non-local anatomical relationships, thus limiting segmentation accuracy; and the segmentation process does not take into account the functional differences within the organ, making it impossible to directly guide subsequent functional avoidance radiotherapy. Summary of the Invention
[0007] Technical Objective: To address the shortcomings of existing technologies, this invention discloses an imaging system for precise segmentation and functional protection of organs at risk in radiotherapy planning. This system can perform high-precision segmentation of organs at risk in medical images and assess the functional heterogeneity within the organs, thereby providing intelligent decision support for radiotherapy planning.
[0008] Technical solution: To achieve the above technical objectives, the present invention adopts the following technical solution:
[0009] An imaging system for precise segmentation and function preservation of organs at risk in radiotherapy planning, comprising:
[0010] An image acquisition module is used to acquire and preprocess the patient's medical image data, which includes at least CT images.
[0011] The graph construction module is used to represent target organs and related anatomical regions in medical image data as non-Euclidean graph structures, including dividing the target region into multiple nodes and establishing graph edges based on the spatial adjacency and anatomical similarity between nodes.
[0012] The graph attention segmentation module is used to apply a non-Euclidean graph attention mechanism to organ segmentation on graph structures. It extracts the features of each node and calculates the attention weight between the node and its neighboring nodes to aggregate domain information in a weighted manner, thereby identifying the set of nodes belonging to the target endangered organ and obtaining the initial organ segmentation result.
[0013] The morphological function mapping module is used to predict or obtain the functional indicators of each part inside the target organ based on the anatomical morphological features of medical images. It assigns functional importance weights to each location within the target organ region in the initial organ segmentation result and generates the corresponding organ function distribution map.
[0014] The quantum graph convolution prediction module is used to map the initial organ segmentation results and functional distribution map into a graph partitioning optimization problem and apply the quantum graph convolution algorithm for global optimization. It corrects the boundary of the initial organ segmentation results to protect high functional importance regions. The quantum graph convolution algorithm is based on the principle of quantum computing to perform convolution propagation on graph node features or uses quantum annealing to solve the optimal segmentation state.
[0015] The spatiotemporal self-supervised reconstruction module is used to improve feature representation capabilities through self-supervised learning during the model training phase. This includes partially occluding medical image data in the spatial or temporal dimensions and having the model reconstruct the occluded content to learn the spatiotemporal feature representation of organs, thereby improving the robustness and accuracy of segmentation during the inference phase.
[0016] The output interface is used to output optimized target organ segmentation images and their functional distribution information for reference when formulating radiotherapy plans, enabling automatic delineation of organs at risk and functional protection prompts.
[0017] Preferably, the graph construction module divides the target organ region to be segmented into several sub-regions as graph nodes, and establishes an undirected weighted graph structure based on the spatial distance and image similarity between the sub-regions; the edge weights between adjacent nodes in the graph structure are jointly determined by the distance between them in three-dimensional space and the difference in grayscale distribution, so as to reflect the proximity and similarity of organ morphology.
[0018] Preferably, when the graph attention segmentation module calculates the attention coefficient between graph node i and its neighboring node j, it considers the similarity of node feature vectors and the non-Euclidean distance between nodes, assigning higher weights to nodes that are closer or have similar shapes, in order to capture the non-local correlation of anatomical structures, thereby improving the coherence and accuracy of the segmentation results.
[0019] Preferably, the formula for calculating the attention coefficient is as follows:
[0020] ,
[0021] Where i is the index of the current center node in the graph, j is the index of a neighboring node adjacent to node i, and k is the traversal variable for all the indices of the neighboring nodes of node i. and Let i and j represent the feature vectors of node i and node j in the graph, respectively. The weight matrix for the feature transformation. For training vectors used to compute attention; symbol This represents the vector concatenation operation; This represents the non-Euclidean distance or morphological difference measure between nodes i and j. This represents the non-Euclidean distance or morphological difference measure between node i and node j. To convert distance into a function of influence factor, To adjust the parameters that affect the weight of distance, Let i be the set of adjacent nodes of node i. It is a linear rectified activation function with leakage.
[0022] Preferably, the morphological function mapping module extracts morphological features from medical images and predicts the functional values of corresponding parts through a pre-trained deep learning model, generating a functional distribution map of the target organ; or when a 4D temporal image of the target organ is acquired, the morphological function mapping module calculates functional indicators by analyzing the intensity or volume changes of voxels at different time phases, so as to obtain the relative functional information of each part of the organ without the need for additional functional imaging.
[0023] Preferably, the quantum graph convolution prediction module optimizes the initial organ segmentation result of the target organ as an energy minimization problem of graph partitioning, and uses quantum annealing or quantum superposition search mechanism to solve it to obtain the globally optimal segmentation label configuration; wherein quantum graph convolution achieves efficient aggregation of graph node features by performing convolution operations in parallel on qubits, which can accelerate the finding of the segmentation scheme with the global minimum energy, thereby improving the accuracy of the segmentation boundary and ensuring that high functional areas are preserved and protected.
[0024] Preferably, the spatiotemporal self-supervised reconstruction module adopts an occlusion reconstruction strategy during model training: for unlabeled medical image data, a portion of the spatial region or certain time series frames are randomly occluded, and the occluded image content is reconstructed through the network; the features learned through this self-supervised task are used for model initialization or as regularization constraints, thereby improving the generalization performance of the segmentation model under limited labeled data, as well as its adaptability to organ morphology and temporal changes.
[0025] Preferably, the system is deployed on the embedded computing platform of the radiotherapy equipment or integrated with the medical imaging workstation, and adopts model compression and hardware acceleration technologies to meet the requirements of real-time processing, so as to realize the automatic delineation of organs at risk and functional area prompts in real time, thereby improving the efficiency of clinical radiotherapy planning.
[0026] A method for precise segmentation and function preservation imaging of organs at risk in radiotherapy planning, applied to the aforementioned system for precise segmentation and function preservation imaging of organs at risk in radiotherapy planning, specifically includes the following steps:
[0027] S1. Acquire the patient's medical imaging data and preprocess the medical imaging data, including CT images;
[0028] S2. In the preprocessed medical image data, determine the region of interest of the target organ at risk, divide the region of interest into multiple sub-regions, use each sub-region as a graph node, establish graph edges based on the spatial adjacency and anatomical similarity between sub-regions, and construct a non-Euclidean graph structure representing the target organ at risk and its surrounding anatomical structure.
[0029] S3. Input the non-Euclidean graph structure into the graph attention segmentation model, perform non-Euclidean graph attention calculation on the graph nodes, calculate the attention weight based on the node feature similarity and the non-Euclidean distance between nodes, perform weighted aggregation of the features of neighboring nodes, identify the set of nodes belonging to the target endangered organ, and generate the initial segmentation result of the target endangered organ.
[0030] S4. Perform morphological-functional mapping on the target organ at risk region in the initial segmentation results. Based on the anatomical morphological features of medical images, predict or calculate the functional indicators of each location inside the target organ at risk. Assign functional importance weights to each location within the target organ at risk region and generate a functional distribution map of the target organ at risk that corresponds one-to-one with the initial segmentation results.
[0031] S5. The initial segmentation results and functional distribution map are mapped to a graph partitioning optimization problem. The quantum graph convolution prediction algorithm is used to solve the graph partitioning optimization problem, and the segmentation results of the target endangered organs after quantum graph convolution optimization are obtained. The quantum graph convolution prediction algorithm is based on the principle of quantum computing to perform convolution propagation on the graph node features or to use quantum annealing to solve the segmentation state with the minimum energy, so as to protect the regions with high functional importance weights during the optimization process.
[0032] S6. Output the segmentation results of the target organs at risk optimized by quantum graph convolution and the corresponding functional distribution map, which can be used to automatically delineate the target organs at risk and provide functional protection prompts during the radiotherapy planning process.
[0033] S7. Before performing the graph attention segmentation step, the graph attention segmentation model is pre-trained using unlabeled medical image data through a spatiotemporal self-supervised reconstruction task. This includes partially occluding the medical image data in the spatial or temporal dimensions and reconstructing the occluded content by the graph attention segmentation model. This enables the graph attention segmentation model to learn the spatiotemporal feature representation of the target endangered organ, thereby improving the robustness and accuracy of the target endangered organ segmentation when performing the graph attention segmentation step and the quantum graph convolution prediction step.
[0034] Beneficial Effects: The imaging system for precise segmentation and function protection of organs at risk in radiotherapy planning provided by this invention has the following beneficial effects:
[0035] 1. This invention solves the technical problems of discontinuous segmentation boundaries and poor robustness to organ deformation and low contrast in existing convolutional neural networks by using a collaborative technique of graph construction, graph attention segmentation, and spatiotemporal self-supervised reconstruction training. It achieves the effect of stable output of high-precision organ segmentation results under various patient anatomical differences and different respiratory phases. Specifically, in the graph construction step, this invention first subdivides the region of interest (ROI) containing the target endangered organ into several sub-region nodes and establishes a non-Euclidean weighted graph by integrating 3D spatial distance and image similarity, explicitly characterizing the real spatial topological relationship of the organ from the data structure level. Then, in the graph attention segmentation step, a graph attention mechanism with a non-Euclidean distance correction term is adopted to jointly weight the feature similarity and spatial relationship between each node and its neighboring nodes, considering not only local texture but also long-range structural linkage, thereby automatically suppressing noisy regions and isolated artifacts, and significantly improving the continuity and smoothness of the overall organ contour. At the same time, in the spatiotemporal self-supervised reconstruction training step, a large number of unlabeled static or 4D temporal medical images are used to pre-train the segmentation network through occlusion reconstruction tasks, so that the model has learned the organ shape integrity constraints and the movement patterns that change over time before formal segmentation, fundamentally improving the model's adaptability to abnormal anatomical morphology and respiratory movements.
[0036] 2. This invention solves the technical problems of existing radiotherapy planning, which, through the combination of morphological function mapping and quantum graph convolution prediction, only reflects the anatomical boundaries of the segmentation results of organs at risk and cannot automatically identify and protect high-functioning areas inside the organs. It also solves the problems of traditional optimization based on local heuristics being prone to getting trapped in local optima and failing to take into account both functional protection and geometric accuracy on a global scale. This invention achieves the effect of functional avoidance radiotherapy without the need for additional complex functional imaging equipment and automatically corrects the segmentation boundaries at the global level to protect functional critical areas. Specifically, the morphological-functional mapping step utilizes a pre-trained deep model or 4D temporal image analysis to map morphological features such as density distribution and motion amplitude in conventional anatomical images like CT scans into a functional importance weight map within the target organ at risk. This gives each segmented voxel or graph node dual attributes of anatomical label and functional weight, thereby reconstructing the functional distribution at the image level without the need for additional equipment. Building upon this, the quantum graph convolution prediction step constructs a graph partitioning energy function from the initial segmentation results and functional distribution. Regions with high functional weights are embedded into the energy model as strong protection terms, and quantum annealing or quantum graph convolution is used to search in parallel across the entire graph's state space for the segmentation configuration with the minimum energy, optimizing the organ boundary for global consistency. Because quantum graph convolution prediction considers node features, graph topology, and functional weights simultaneously, and efficiently escapes local minima through the parallel evolution of quantum states, in the optimized segmentation results, high functional weight regions are systematically preserved within the organ at risk contour or marked as highly avoidable regions in dose planning. Low-functional or diseased regions can be appropriately contracted or adjusted while maintaining overall morphological rationality, achieving an integrated goal of anatomical accuracy and functional priority. Attached Figure Description
[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.
[0038] Figure 1 This is a schematic diagram of the overall system structure of the present invention;
[0039] Figure 2 Construct a schematic diagram for non-Euclidean graphs;
[0040] Figure 3 This is a schematic diagram of the graph attention segmentation process;
[0041] Figure 4 This is a schematic diagram of morphological function mapping and functional distribution;
[0042] Figure 5 A schematic diagram illustrating the optimization of quantum graph convolution prediction;
[0043] Figure 6 A schematic diagram of the spatiotemporal self-supervised reconstruction training process. Detailed Implementation
[0044] The present invention will now be described more clearly and completely by way of a preferred embodiment in conjunction with the accompanying drawings, but this does not limit the invention to the scope of the described embodiment.
[0045] like Figure 1 As shown, an imaging system for precise segmentation and function preservation of organs at risk in radiotherapy planning includes:
[0046] The image acquisition module is used to acquire and preprocess the patient's medical image data, which includes at least CT images.
[0047] This module is used to acquire and preprocess the patient's medical imaging data, which can include multimodal images such as CT, MRI, and PET. Taking lung cancer radiotherapy as an example, the main source of imaging data is the treatment planning CT scan, and lung function-related imaging (such as 4DCT sequences for extracting ventilation information, or PET for metabolic function) can be optionally added. The image acquisition module performs routine preprocessing on the input images, such as denoising, resampling, and registration, to ensure the accuracy of subsequent analysis.
[0048] The graph construction module is used to represent target organs and related anatomical regions in medical image data as non-Euclidean graph structures, including dividing the target region into multiple nodes and establishing graph edges based on the spatial adjacency and anatomical similarity between nodes.
[0049] After obtaining the preprocessed image, this invention represents the target organ and related anatomical structures in the image as a non-Euclidean graph data structure through a graph construction module. Specifically, traditional algorithms or depth models can be used to initially locate or coarsely segment the target organ at risk to determine the region of interest (ROI). Subsequently, the set of voxels / pixels within the ROI is discretized into graph nodes, for example, clustered into several sub-regions according to anatomical regions or density similarity, with each sub-region serving as a node. Edges are established between nodes based on spatial proximity and anatomical connectivity, and the edge weights can be defined as a function of the distance between nodes and image similarity. For example, adjacent nodes can have high edge weights, while nodes that are far apart or have significantly different tissue properties have low edge weights, thus forming a weighted undirected graph structure.
[0050] like Figure 2As shown, the process of constructing a non-Euclidean graph in this embodiment of the invention includes extracting the region of interest (ROI) of the organ at risk from CT image slices and subdividing the ROI into sub-regions. The left side shows the original CT image slice, where the ROI of the organ at risk is marked with a dashed box; the middle section shows the sub-region division result after magnifying the ROI, where the magnified area is regularly divided into multiple sub-regions, each corresponding to a subsequent graph node; the right side shows the graph structure composed of the graph nodes corresponding to each sub-region, with multiple graph nodes interconnected by edges representing spatial proximity and anatomical similarity. In this way, the three-dimensional anatomical structure surrounding the organ at risk is explicitly modeled as a non-Euclidean graph, enabling the subsequent graph attention segmentation module to comprehensively consider node features and structural topological information in the graph space, thereby improving the completeness and continuity of the organ at risk boundary identification.
[0051] The graph attention segmentation module is used to apply a non-Euclidean graph attention mechanism to organ segmentation on graph structures. It extracts the features of each node and calculates the attention weight between the node and its neighboring nodes to aggregate domain information in a weighted manner, thereby identifying the set of nodes belonging to the target endangered organ and obtaining the initial organ segmentation result.
[0052] This module performs segmentation operations on the constructed graph data. An improved graph attention network is employed, which weights and aggregates the feature vector of each node with the features of its neighboring nodes to classify the set of nodes belonging to the target organ. The improvement in the graph attention mechanism is reflected in the weight allocation, introducing the combined effect of node feature similarity and topological distance. Specifically, the attention coefficient is calculated as follows:
[0053] ,
[0054] Where i is the index of the current center node in the graph, j is the index of a neighboring node adjacent to node i, and k is the traversal variable for all the indices of the neighboring nodes of node i. and Let i and j represent the feature vectors of node i and node j in the graph, respectively. The weight matrix for the feature transformation. For training vectors used to compute attention; symbol This represents the vector concatenation operation; This represents the non-Euclidean distance or morphological difference measure between nodes i and j. This represents the non-Euclidean distance or morphological difference measure between node i and node j. To convert distance into a function of influence factor, To adjust the parameters that affect the weight of distance, Let i be the set of adjacent nodes of node i. This is a linear rectified activation function with leakage. Compared to existing graph attention formulas, the formula in this invention adds an additional term to the summation term. This improved attention mechanism assigns greater attention values to nodes that are spatially close or morphologically similar. It assigns explicit quantitative weights to the "neighbor-distant, different-near" relationships of anatomical structures within the graph network. Through training, this module can automatically learn to aggregate node features related to the target organ on the graph while suppressing irrelevant regions, thus obtaining accurate initial organ segmentation results. For example, in spinal cord segmentation tasks, even if a segment of the spinal cord's ROI is significantly deformed due to tumor compression, this module can still effectively integrate information from adjacent normal spinal cord segments through graph attention, outputting continuous and smooth segmentation boundaries.
[0055] Unlike traditional CNNs based on regular grid convolutions, graph attention mechanisms adaptively allocate weights based on node feature similarity and topological distance, enabling the characterization of non-local relationships within anatomical structures. For example, spatially distant but physiologically related structures can influence each other under the attention mechanism, thereby improving the completeness and accuracy of segmentation.
[0056] like Figure 3 As shown, in this embodiment of the invention, the graph attention segmentation process uses the local graph structure as the basic unit and performs weighted aggregation on the relationship between the central node and its neighboring nodes. Figure 3 The left side shows a local graph structure containing node i and its multiple neighboring nodes j, k, l, and m, with non-Euclidean adjacency relationships between nodes represented by edge connections. The middle section is a graph attention segmentation module, whose input includes the feature vectors of each node and auxiliary information reflecting spatial distance and anatomical similarity between nodes. Within this module, attention weights are calculated based on the correlation between nodes, and the features of neighboring nodes are weighted and aggregated to output an updated representation of the center node. The right side shows the category labels of different nodes in the segmentation results, where nodes labeled "OAR" represent nodes identified as organs at risk, while nodes labeled "BG" represent background nodes that are not organs at risk. Through this graph attention mechanism, this invention can suppress the interference of noisy nodes in complex anatomical structures and highlight nodes consistent with the morphology of organs at risk, thereby obtaining more accurate and continuous initial segmentation results for organs at risk.
[0057] The morphological function mapping module is used to predict or obtain the functional indicators of each part inside the target organ based on the anatomical morphological features of medical images. It assigns functional importance weights to each location within the target organ region in the initial organ segmentation result and generates the corresponding organ function distribution map.
[0058] After obtaining the initial segmentation results of the organ, this invention assesses the functional importance of different parts within the organ through a morphological-functional mapping module. This module combines anatomical morphological features and functional imaging / prediction models to obtain a spatial functional distribution map. There are two specific implementation methods: First, deep learning-based functional prediction. A deep neural network is trained using prior data to map anatomical images to functional indicators. For example, for the lungs, the network inputs a patient's CT image block and outputs the relative perfusion value of the corresponding region, thereby approximating a perfusion map of the entire lung. The network can use multimodal inputs (such as CT+PET) to improve prediction accuracy, or introduce an attention mechanism to focus on image features affecting function. Second, direct calculation based on time-series images. When time-series data such as 4D CT is available, ventilation function can be calculated by the intensity changes of voxels during the respiratory cycle. For example, taking inspiratory and expiratory CT scans, the difference in CT values or the proportion of volume change for each voxel is calculated to obtain indicators reflecting local ventilation. Similarly, for other organs requiring protection (such as the filtration function of the kidneys, the neural functional areas of the brain, etc.), imaging physiological models from their respective fields can be used to convert morphological changes in images into functional parameters. Using this method, the morphological-functional mapping module generates a functional importance matrix or atlas corresponding to the anatomical regions of the organs, where each voxel / node is assigned a relative functional weight (e.g., a normalized value between 0 and 1, with 1 representing the highest functional importance). In the lung perfusion example, the functional value of the large vessel region near the hilum may be higher, while the functional value of the fibrotic lesion region tends to be close to 0. This functional distribution map provides a basis for subsequent dose optimization: high-functional-value areas should be avoided from high-dose irradiation as much as possible.
[0059] like Figure 4 As shown, in this embodiment of the invention, the morphological function mapping process takes the segmented region of the endangered organ obtained in the previous step as input and converts the anatomical morphology information into a corresponding functional distribution map. Figure 4 The left side shows the outline of the segmented region of the organ at risk, representing the organ range identified by the graph attention segmentation module. The middle section is the morphological function mapping module, which predicts the functional indicators of various locations within the organ based on image morphological information such as the shape, density distribution, and possible motion characteristics of the segmented region. The right side shows a functional distribution map, in which the organ at risk is divided into high-function, medium-function, and low-function regions, each with different importance weights. By superimposing the functional distribution onto the segmentation results, this invention not only provides the geometric outline of the organ at risk but also distinguishes between high-function regions requiring priority protection and low-function regions where dose constraints can be relatively relaxed, providing a basis for subsequent function-prioritized segmentation optimization by the quantum graph convolution prediction module.
[0060] The quantum graph convolution prediction module is used to map the initial organ segmentation results and functional distribution map into a graph partitioning optimization problem and apply the quantum graph convolution algorithm for global optimization. It corrects the boundaries of the initial organ segmentation results to protect high functional importance regions. The quantum graph convolution algorithm is based on the principle of quantum computing to perform convolution propagation on graph node features or uses quantum annealing to solve the optimal segmentation state.
[0061] This module aims to globally optimize and correct the initial segmentation results, with a particular focus on protecting high-functionality regions from missegmentation or over-irradiation. The core idea is to transform the organ segmentation optimization problem into an energy minimization problem on a graph, and then use quantum algorithms to solve it, obtaining the globally optimal segmentation configuration. The specific steps are as follows: First, a segmentation energy function is constructed based on the initial segmentation results and the functional distribution map. For example, an energy function can be defined using a Markov random field (MRF) model. ,in The segmentation label of node i is represented (e.g., foreground = organ, background = non-organ). It is a single-node energy term, which can be determined by the initial segmentation confidence and functional weight (if a region with a high functional value is marked as foreground, its energy will be reduced, indicating that it should be retained as an organ protection zone). This refers to the interaction energy between adjacent nodes, encouraging them to adopt the same label to ensure a smooth partition boundary. This module then uses a quantum graph convolution / quantum annealing algorithm to solve this energy optimization problem. In the quantum annealing implementation, the energy function can be mapped to the Hamiltonian of the Ising model, each node label can be considered as the spin state of a qubit, and the interaction between nodes can be considered as quantum coupling. The lowest energy state is then determined on a quantum annealer (such as a D-Wave quantum computer). This method can utilize the parallel tunneling effect of quantum mechanics to quickly escape local minima, thus efficiently finding a partitioning scheme close to the global optimum. For environments without quantum hardware, simulated quantum annealing can be used to approximate the solution on a classical computer. Another implementation is an end-to-end optimization based on quantum graph convolutional neural networks (QGCN): a quantum circuit layer is designed to perform graph convolution operations, allowing node features to evolve in quantum states through parametric quantum gates, achieving feature propagation similar to classical convolution. The output probability distribution is then obtained by measuring the quantum states. This quantum convolutional layer, capable of handling a large number of configurations under superposition simultaneously, possesses potential exponential parallelism, which can be used to improve the inference speed or optimization quality of segmentation models. Regardless of the approach used, the output of the quantum graph convolutional prediction module is an optimized organ segmentation result, with smoother boundaries and greater consistency with the functional graph. If a high-functioning, important region belongs to an organ after optimization, the surrounding organ labels are preserved by quantum optimization, avoiding misclassification due to local algorithmic errors. Conversely, if a high-functioning region does not belong to the target organ (e.g., a healthy tissue area adjacent to a tumor), the system will, guided by functional weights, try to remove it from the target region or reduce its importance, thus providing a clear boundary reference for dose allocation in subsequent radiotherapy.
[0062] like Figure 5 As shown, in this embodiment of the invention, the quantum graph convolution prediction module jointly optimizes the initial segmentation result and functional distribution to obtain the final segmentation result of the endangered organ that balances boundary accuracy and functional protection. Figure 5The upper left shows the initial segmentation result, where the boundaries are not smooth enough in some areas or deviate from the actual anatomical boundaries. The lower left shows the functional distribution map corresponding to the segmented region, clearly identifying high-functional areas that require priority protection. The middle section is the quantum graph convolution prediction module, which maps the initial segmentation and functional weights to a graph partitioning energy function. Through quantum graph convolution or quantum annealing mechanisms, it searches for the segmentation configuration with the minimum energy in the graph's state space, giving high-functional areas a higher protection weight in the energy function. The right side shows the optimized segmentation result, showing smoother organ boundaries and better consistency with anatomical structures. High-functional areas are preserved entirely within the outline of organs at risk or explicitly marked in dosage planning as areas requiring increased protection. Thus, this invention achieves priority protection of critical functional areas while maintaining geometric segmentation accuracy.
[0063] The spatiotemporal self-supervised reconstruction module is used to improve feature representation capabilities through self-supervised learning during the model training phase. This includes partially occluding medical image data in the spatial or temporal dimensions and having the model reconstruct the occluded content to learn the spatiotemporal feature representation of organs, thereby improving the robustness and accuracy of segmentation during the inference phase.
[0064] The system of this invention utilizes spatiotemporal self-supervised learning during training to improve the model's feature generalization ability and sensitivity to organ changes. This module does not directly act on the input during the inference stage, but is integrated as a model training strategy. Specifically, it performs self-supervised pre-training on a large amount of unlabeled medical image data (which may include historical CT and MRI sequences of the same location). Taking 4D CT sequences as an example, the sequences are first arranged chronologically, then the entire organ region or a portion of a slice is randomly occluded at a certain time frame. The model then attempts to reconstruct the occluded portion based on the unoccluded spatiotemporal proximity information. The model learns inherent spatiotemporal correlation features by minimizing the reconstruction error. For example, for CT images within a call cycle, the model will learn the patterns of lung tissue density changes and understand which structural changes are normal respiratory deformations. Similarly, by training on occlusion reconstruction of 3D anatomical structures, the model will master the integrity constraints of organ shape. After several rounds of such self-supervised training, the model's encoder can extract robust morphological and temporal features. Then, supervised fine-tuning on a small amount of labeled data significantly improves the accuracy and robustness of the final segmentation model. Experiments demonstrate that the self-supervised pre-trained model can still effectively distinguish organ boundaries and is less susceptible to noise interference when encountering rare anatomical variations or low-contrast images. This is particularly important in medical settings, as patients exhibit significant individual differences in anatomical structures, and tumors and organs may be severely deformed. Self-supervised pre-training endows the model with analogical learning capabilities, making it more adaptable and reliable in clinical applications.
[0065] See Figure 6 In this embodiment of the invention, the spatiotemporal self-supervised reconstruction training process utilizes unlabeled 4D medical image sequences to pre-train the graph attention segmentation model, thereby improving the model's ability to model organ morphological changes and respiratory movements. Figure 6 The left side shows an unlabeled 4D medical image sequence composed of multiple temporal phases t1, t2, and t3, in which local regions of certain phases are artificially occluded as prediction targets for self-supervised learning. The middle section shows a spatiotemporal self-supervised reconstruction module, which takes the occluded temporal images as input and learns to reconstruct the occluded image content simultaneously in both spatial and temporal dimensions, thereby capturing the spatiotemporal relevant features of organs at risk and their surrounding tissues. The right side shows the reconstruction results and the pre-trained segmentation model, where the reconstruction results show that the occluded regions were successfully filled in, and the pre-trained segmentation model transfers the learned spatiotemporal features to the subsequent graph attention segmentation task. Through this self-supervised reconstruction training, this invention can acquire rich prior knowledge of organ structure and motion without additional manual annotation, enabling the formal segmentation stage to maintain high segmentation accuracy and robustness even with small samples or abnormal anatomy.
[0066] The output interface is used to output optimized target organ segmentation images and their functional distribution information for reference when formulating radiotherapy plans, enabling automatic delineation of organs at risk and functional protection prompts.
[0067] The aforementioned modules can be implemented through a combination of software and hardware. The software portion of the system consists of a deep learning model and optimization algorithms, which can be programmed using a mix of Python and C++. The deep learning model is built using the TensorFlow or PyTorch framework, and quantum annealing calls the D-Wave Ocean interface or corresponding simulation libraries. On the hardware side, the system can be deployed on a radiotherapy planning workstation or in the image guidance device integrated into a linear accelerator. Considering the real-time requirements of the clinical environment, this invention optimizes the algorithms and models in its implementation: on the one hand, the size of the deep learning model is compressed through model pruning and quantization, accelerating inference speed and facilitating operation on embedded GPUs / FPGAs; on the other hand, for the quantum optimization part, optimization strategies for common organs can be pre-calculated offline or a lightweight alternative model can be trained to approximate real-time quantum computing during actual operation, thereby meeting clinical timeliness requirements. In practical applications, after obtaining patient CT images, doctors input them into the system; after processing by the aforementioned modules, the system outputs outlines and functional distribution suggestions for each organ at risk within approximately a few seconds to a minute. Based on this, doctors can adjust the radiotherapy plan, such as replanning the irradiation field to avoid marked high-functioning areas. The entire process of the system of this invention can be embedded into existing radiotherapy planning software to achieve human-machine collaboration: automated segmentation and functional analysis provide preliminary plans, while manual review and adjustment ensure safety and reliability, thereby significantly improving the efficiency and quality of radiotherapy planning.
[0068] A method for precise segmentation and function preservation imaging of organs at risk in radiotherapy planning, applied to the aforementioned system for precise segmentation and function preservation imaging of organs at risk in radiotherapy planning, specifically includes the following steps:
[0069] S1. Acquire the patient's medical imaging data and preprocess the medical imaging data, including CT images;
[0070] S2. In the preprocessed medical image data, determine the region of interest of the target organ at risk, divide the region of interest into multiple sub-regions, use each sub-region as a graph node, establish graph edges based on the spatial adjacency and anatomical similarity between sub-regions, and construct a non-Euclidean graph structure representing the target organ at risk and its surrounding anatomical structure.
[0071] S3. Input the non-Euclidean graph structure into the graph attention segmentation model, perform non-Euclidean graph attention calculation on the graph nodes, calculate the attention weight based on the node feature similarity and the non-Euclidean distance between nodes, perform weighted aggregation of the features of neighboring nodes, identify the set of nodes belonging to the target endangered organ, and generate the initial segmentation result of the target endangered organ.
[0072] S4. Perform morphological-functional mapping on the target organ at risk region in the initial segmentation results. Based on the anatomical morphological features of medical images, predict or calculate the functional indicators of each location inside the target organ at risk. Assign functional importance weights to each location within the target organ at risk region and generate a functional distribution map of the target organ at risk that corresponds one-to-one with the initial segmentation results.
[0073] S5. The initial segmentation results and functional distribution map are mapped to a graph partitioning optimization problem. The quantum graph convolution prediction algorithm is used to solve the graph partitioning optimization problem, and the segmentation results of the target endangered organs after quantum graph convolution optimization are obtained. The quantum graph convolution prediction algorithm is based on the principle of quantum computing to perform convolution propagation on the graph node features or to use quantum annealing to solve the segmentation state with the minimum energy, so as to protect the regions with high functional importance weights during the optimization process.
[0074] S6. Output the segmentation results of the target organs at risk optimized by quantum graph convolution and the corresponding functional distribution map, which can be used to automatically delineate the target organs at risk and provide functional protection prompts during the radiotherapy planning process.
[0075] S7. Before performing the graph attention segmentation step, the graph attention segmentation model is pre-trained using unlabeled medical image data through a spatiotemporal self-supervised reconstruction task. This includes partially occluding the medical image data in the spatial or temporal dimensions and reconstructing the occluded content by the graph attention segmentation model. This enables the graph attention segmentation model to learn the spatiotemporal feature representation of the target endangered organ, thereby improving the robustness and accuracy of the target endangered organ segmentation when performing the graph attention segmentation step and the quantum graph convolution prediction step.
[0076] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. An imaging system for precise segmentation and function preservation of organs at risk in radiotherapy planning, characterized in that, include: An image acquisition module is used to acquire and preprocess the patient's medical image data, which includes at least CT images. The graph construction module is used to represent target organs and related anatomical regions in medical image data as non-Euclidean graph structures, including dividing the target region into multiple nodes and establishing graph edges based on the spatial adjacency and anatomical similarity between nodes. The graph attention segmentation module is used to apply a non-Euclidean graph attention mechanism to organ segmentation on graph structures. It extracts the features of each node and calculates the attention weight between the node and its neighboring nodes to aggregate domain information in a weighted manner, thereby identifying the set of nodes belonging to the target endangered organ and obtaining the initial organ segmentation result. The morphological function mapping module is used to predict or obtain the functional indicators of each part inside the target organ based on the anatomical morphological features of medical images. It assigns functional importance weights to each location within the target organ region in the initial organ segmentation result and generates the corresponding organ function distribution map. The quantum graph convolution prediction module is used to map the initial organ segmentation results and functional distribution map into a graph partitioning optimization problem and apply the quantum graph convolution algorithm for global optimization. It corrects the boundary of the initial organ segmentation results to protect high functional importance regions. The quantum graph convolution algorithm is based on the principle of quantum computing to perform convolution propagation on graph node features or uses quantum annealing to solve the optimal segmentation state. The spatiotemporal self-supervised reconstruction module is used to improve feature representation capabilities through self-supervised learning during the model training phase. This includes partially occluding medical image data in the spatial or temporal dimensions and having the model reconstruct the occluded content to learn the spatiotemporal feature representation of organs, thereby improving the robustness and accuracy of segmentation during the inference phase. The output interface is used to output optimized target organ segmentation images and their functional distribution information for reference when formulating radiotherapy plans, enabling automatic delineation of organs at risk and functional protection prompts.
2. The imaging system for precise segmentation and function protection of organs at risk in radiotherapy planning according to claim 1, characterized in that, The graph construction module divides the target organ region to be segmented into several sub-regions as graph nodes, and establishes an undirected weighted graph structure based on the spatial distance and image similarity between the sub-regions. The edge weights between adjacent nodes in the graph structure are determined by the distance between them in three-dimensional space and the difference in grayscale distribution, so as to reflect the proximity and similarity of organ morphology.
3. The imaging system for precise segmentation and function preservation of organs at risk in radiotherapy planning according to claim 1, characterized in that, When calculating the attention coefficient between graph node i and its neighboring node j, the graph attention segmentation module considers the similarity of node feature vectors and the non-Euclidean distance between nodes. It assigns higher weights to nodes that are closer or have similar shapes in order to capture the non-local correlation of anatomical structures, thereby improving the coherence and accuracy of the segmentation results.
4. The imaging system for precise segmentation and function preservation of organs at risk in radiotherapy planning according to claim 3, characterized in that, The formula for calculating the attention coefficient is as follows: ， Where i is the index of the current center node in the graph, j is the index of a neighboring node adjacent to node i, and k is the traversal variable for all the indices of the neighboring nodes of node i. and Let i and j represent the feature vectors of node i and node j in the graph, respectively. The weight matrix for the feature transformation. For training vectors used to compute attention; symbol This represents the vector concatenation operation; This represents the non-Euclidean distance or morphological difference measure between nodes i and j. This represents the non-Euclidean distance or morphological difference measure between node i and node j. To convert distance into a function of influence factor, To adjust the parameters that affect the weight of distance, Let i be the set of adjacent nodes of node i. It is a linear rectified activation function with leakage.
5. The imaging system for precise segmentation and function preservation of organs at risk in radiotherapy planning according to claim 1, characterized in that, The morphological function mapping module extracts morphological features from medical images and predicts the functional values of corresponding parts through a pre-trained deep learning model, generating a functional distribution map of the target organ; or when a 4D temporal image of the target organ is acquired, the morphological function mapping module calculates functional indicators by analyzing the intensity or volume changes of voxels at different time phases, so as to obtain the relative functional information of each part of the organ without the need for additional functional imaging.
6. The imaging system for precise segmentation and function preservation of organs at risk in radiotherapy planning according to claim 1, characterized in that, The quantum graph convolution prediction module optimizes the initial organ segmentation results of the target organ as an energy minimization problem of graph partitioning, and uses quantum annealing or quantum superposition search mechanism to solve it to obtain the globally optimal segmentation label configuration; Quantum graph convolution achieves efficient aggregation of graph node features by performing convolution operations in parallel on qubits, which can accelerate the finding of the segmentation scheme with the global minimum energy, thereby improving the accuracy of the segmentation boundary and ensuring that high-functionality regions are preserved and protected.
7. The imaging system for precise segmentation and function preservation of organs at risk in radiotherapy planning according to claim 1, characterized in that, The spatiotemporal self-supervised reconstruction module employs an occlusion reconstruction strategy during model training: for unlabeled medical image data, a portion of the spatial region or certain time-series frames are randomly occluded, and the occluded image content is reconstructed through the network; the features learned through this self-supervised task are used for model initialization or as regularization constraints, thereby improving the generalization performance of the segmentation model under limited labeled data, as well as its adaptability to organ morphology and temporal changes.
8. The imaging system for precise segmentation and function preservation of organs at risk in radiotherapy planning according to claim 1, characterized in that, The system is deployed on the embedded computing platform of radiotherapy equipment or integrated with medical imaging workstations. It uses model compression and hardware acceleration technologies to meet the requirements of real-time processing, enabling immediate automatic delineation of organs at risk and functional area prompts, thereby improving the efficiency of clinical radiotherapy planning.
9. A method for precise segmentation and function preservation of organs at risk in radiotherapy planning, characterized in that, An imaging system for precise segmentation and function preservation of organs at risk in radiotherapy planning, as described in any one of claims 1-8, specifically includes the following steps: S1. Acquire the patient's medical imaging data and preprocess the medical imaging data, including CT images; S2. In the preprocessed medical image data, determine the region of interest of the target organ at risk, divide the region of interest into multiple sub-regions, use each sub-region as a graph node, establish graph edges based on the spatial adjacency and anatomical similarity between sub-regions, and construct a non-Euclidean graph structure representing the target organ at risk and its surrounding anatomical structure. S3. Input the non-Euclidean graph structure into the graph attention segmentation model, perform non-Euclidean graph attention calculation on the graph nodes, calculate the attention weight based on the node feature similarity and the non-Euclidean distance between nodes, perform weighted aggregation of the features of neighboring nodes, identify the set of nodes belonging to the target endangered organ, and generate the initial segmentation result of the target endangered organ. S4. Perform morphological-functional mapping on the target organ at risk region in the initial segmentation results. Based on the anatomical morphological features of medical images, predict or calculate the functional indicators of each location inside the target organ at risk. Assign functional importance weights to each location within the target organ at risk region and generate a functional distribution map of the target organ at risk that corresponds one-to-one with the initial segmentation results. S5. The initial segmentation results and functional distribution map are mapped to a graph partitioning optimization problem. The quantum graph convolution prediction algorithm is used to solve the graph partitioning optimization problem, and the segmentation results of the target endangered organs after quantum graph convolution optimization are obtained. The quantum graph convolution prediction algorithm is based on the principle of quantum computing to perform convolution propagation on the graph node features or to use quantum annealing to solve the segmentation state with the minimum energy, so as to protect the regions with high functional importance weights during the optimization process. S6. Output the segmentation results of the target organs at risk optimized by quantum graph convolution and the corresponding functional distribution map, which can be used to automatically delineate the target organs at risk and provide functional protection prompts during the radiotherapy planning process. S7. Before performing the graph attention segmentation step, the graph attention segmentation model is pre-trained using unlabeled medical image data through a spatiotemporal self-supervised reconstruction task. This includes partially occluding the medical image data in the spatial or temporal dimensions and reconstructing the occluded content by the graph attention segmentation model. This enables the graph attention segmentation model to learn the spatiotemporal feature representation of the target endangered organ, thereby improving the robustness and accuracy of the target endangered organ segmentation when performing the graph attention segmentation step and the quantum graph convolution prediction step.