CN-121999221-A - Precise segmentation and functional protection image system for organs at risk in radiotherapy plan

CN121999221ACN 121999221 ACN121999221 ACN 121999221ACN-121999221-A

Abstract

The invention discloses an organ-at-risk accurate segmentation and function protection image system in radiotherapy planning. The system utilizes a non-Euclidean graph attention mechanism to automatically divide the organ of the medical image, introduces a quantum drawing convolution algorithm to realize global optimization of the division result, and obtains the functional distribution information inside the organ through a morphological function mapping technology. The system also adopts space-time self-supervision learning to improve the characteristic expression capability of the model. The system can be embedded into radiotherapy equipment or an image platform, realizes high-precision automatic sketching of organs at risk and protection of key functional areas, and provides reliable decision support for radiotherapy planning.

Inventors

ZHANG XIAO
LI PENG
WANG GUIJUAN
ZHANG XINHUA
WANG HUILI

Assignees

济宁医学院附属医院

Dates

Publication Date: 20260508
Application Date: 20260127

Claims (9)

1. An organ-at-risk accurate segmentation and functional protection image system in a radiation therapy plan, comprising: the image acquisition module is used for acquiring medical image data of a patient and preprocessing the medical image data, wherein the medical image data at least comprises CT images; The image construction module is used for representing a target organ and a related anatomical region in the medical image data as a non-Euclidean image structure and comprises the steps of dividing the target region into a plurality of nodes and establishing image edges according to the spatial adjacent relation and the anatomical similarity between the nodes; The diagram attention segmentation module is used for carrying out organ segmentation on the diagram structure by applying a non-Euclidean diagram attention mechanism, extracting the characteristic of each node and calculating the attention weight between the node and the neighborhood node thereof so as to weight and aggregate the domain information, thereby identifying the node set belonging to the target organ at risk and obtaining an initial organ segmentation result; The morphological function mapping module is used for predicting or acquiring the functional index of each part in the target organ based on the anatomical morphological characteristics of the medical image, giving functional importance weight to each position in the target organ area in the initial organ segmentation result and generating a corresponding organ functional distribution diagram; the quantum graph convolution prediction module is used for mapping an initial organ segmentation result and a functional distribution diagram into a diagram division optimization problem, performing global optimization by using a quantum diagram volume integration algorithm, and correcting the boundary of the initial organ segmentation result to protect a high functional importance region, wherein the quantum graph convolution algorithm performs convolution propagation on the diagram node characteristics based on a quantum calculation principle or solves an optimal segmentation state by using quantum annealing; The space-time self-supervision reconstruction module is used for improving the characteristic representation capability through self-supervision learning in the model training stage, and comprises the steps of partially masking medical image data in the space or time dimension and reconstructing the masked content by the model to learn the space-time characteristic representation of organs, so that the robustness and accuracy of segmentation are improved in the reasoning stage; The output interface is used for outputting the optimized target organ segmentation image and the function distribution information thereof for reference during the radiotherapy planning, and realizing automatic sketching and function protection prompt of the organs at risk.
2. The system for accurately segmenting and functionally protecting the organs at risk in the radiotherapy plan according to claim 1, wherein the graph construction module divides the target organ region to be segmented into a plurality of subareas as graph nodes, establishes an undirected weighted graph structure according to the space distance and the image similarity between the subareas, and the edge weight between the adjacent nodes in the graph structure is jointly determined by the distance and the gray distribution difference of the adjacent nodes in the three-dimensional space so as to reflect the adjacency and the similarity of the organ morphology.
3. The system of claim 1, wherein the attention coefficient between the nodes i and the adjacent nodes j is calculated by the attention segmentation module, and the similarity of the feature vectors of the nodes and the non-euclidean distance factor between the nodes are considered, and the nodes with a closer distance or similar morphology are given higher weight to capture the non-local correlation of the anatomical structure, so as to improve the consistency and the accuracy of the segmentation result.
4. A system for accurately segmenting and functionally protecting an image of a organs at risk in a radiation therapy planning according to claim 3, wherein the attention factor is calculated as follows: , where i is the index of the current center node in the graph, j is the index of a certain neighboring node in the graph that is adjacent to node i, k is the traversal variables of all neighboring node indexes of node i, And Feature vectors of node i and node j in the diagram are represented respectively, For the weight matrix of the feature transformation, For trainable vectors for calculating attention, symbols Representing a vector tandem operation; Representing a non-euclidean distance or morphology difference metric between nodes i and j, Representing a non-euclidean distance or morphology difference metric between node i and node j, To convert the distance into a function of the impact factor, In order to adjust the parameters of the distance-influencing weights, Is a set of contiguous nodes of node i, The function is activated for linear rectification with leakage.
5. The system of claim 1, wherein the morphology function mapping module extracts morphology features of the medical image and predicts functional values of the corresponding parts to generate a functional distribution map of the target organ through a pre-trained deep learning model, or when a 4D time-series image of the target organ is acquired, the morphology function mapping module calculates functional indexes by analyzing intensity or volume changes of voxels at different time phases to obtain relative functional information of each part of the organ without additional functional imaging.
6. The precise organ-at-risk segmentation and functional protection image system in radiotherapy planning according to claim 1, wherein the quantum graph convolution prediction module optimally models the initial organ segmentation result of the target organ as an energy minimization problem of graph division and solves the problem by utilizing a quantum annealing or quantum superposition search mechanism to obtain a globally optimal segmentation label configuration, wherein the quantum graph convolution realizes efficient aggregation of graph node characteristics by performing convolution operation on quantum bits in parallel, and can accelerate finding of a segmentation scheme with globally minimum energy, thereby improving the accuracy of segmentation boundaries and ensuring that high-functional areas are preserved and protected.
7. The system for accurately segmenting and functionally protecting the image of the organs at risk in the radiotherapy plan according to claim 1, wherein the space-time self-supervision reconstruction module adopts a masking reconstruction strategy in the model training process, wherein a part of spatial areas or certain time sequence frames are randomly masked for unlabeled medical image data, the masked image content is reconstructed through a network, and the features learned by the self-supervision task are used for model initialization or used as regular constraints, so that the generalization performance of the segmented model under the condition of limited labeling data and the adaptability to organ morphology and time sequence changes are improved.
8. The precise organ-at-risk segmentation and functional protection image system in radiotherapy planning according to claim 1, wherein the system is deployed on an embedded computing platform of radiotherapy equipment or integrated with a medical image workstation, and model compression and hardware acceleration technologies are adopted to meet the requirements of real-time processing, so that immediate automatic organ-at-risk sketching and functional area prompt are realized, and the clinical radiotherapy planning efficiency is improved.
9. A method for accurately segmenting and functionally protecting an image of a organs at risk in a radiotherapy plan, which is applied to an image system for accurately segmenting and functionally protecting an organs at risk in a radiotherapy plan according to any one of claims 1 to 8, and specifically comprises the following steps: s1, acquiring medical image data of a patient and preprocessing the medical image data, wherein the medical image data comprises CT images; S2, determining an interested region of a target endangered organ in the preprocessed medical image data, dividing the interested region into a plurality of subareas, taking each subarea as a graph node, establishing a graph edge according to a spatial adjacent relation and an anatomical similarity between the subareas, and constructing a non-Euclidean graph structure representing the target endangered organ and surrounding anatomical structures of the target endangered organ; S3, inputting a non-Euclidean graph structure into a graph attention segmentation model, performing non-Euclidean graph attention calculation on graph nodes, calculating attention weights based on node feature similarity and non-Euclidean distance between nodes, performing weighted aggregation on neighborhood node features, identifying a node set belonging to a target organ at risk, and generating an initial segmentation result of the target organ at risk; S4, performing morphological function mapping on the target organ-at-risk area in the initial segmentation result, predicting or calculating functional indexes of all positions in the target organ-at-risk based on anatomical morphological characteristics of the medical image, giving functional importance weights to all positions in the target organ-at-risk area, and generating a target organ-at-risk functional distribution map which corresponds to the initial segmentation result one by one; S5, mapping an initial segmentation result and a functional distribution diagram into a diagram division optimization problem, solving the diagram division optimization problem by adopting a quantum diagram convolution prediction algorithm to obtain a target organ-at-risk segmentation result optimized by quantum diagram convolution, and carrying out convolution propagation on diagram node characteristics by the quantum diagram convolution prediction algorithm or solving a segmentation state with minimum energy by utilizing quantum annealing based on a quantum computing principle so as to protect a region with higher functional importance weight in the optimization process; S6, outputting a target organ-at-risk segmentation result and a corresponding functional distribution diagram which are subjected to quantum graph convolution optimization, and performing automatic sketching and functional protection prompting on the target organ-at-risk in the radiation treatment planning process; S7, before the graph attention segmentation step is executed, pre-training the graph attention segmentation model through a space-time self-supervision reconstruction task by utilizing unlabeled medical image data, wherein the pre-training comprises the steps of partially masking the medical image data in a space dimension or a time dimension and reconstructing masked contents by the graph attention segmentation model so that the graph attention segmentation model learns the space-time characteristic representation of the target organ at risk, and therefore the robustness and the accuracy of the target organ at risk segmentation are improved when the graph attention segmentation step and the quantum graph convolution prediction step are executed.

Description

Precise segmentation and functional protection image system for organs at risk in radiotherapy plan Technical Field The invention relates to the technical field, in particular to an organ-at-risk accurate segmentation and function protection image system in radiotherapy planning. Background In the radiation treatment planning process, the precise delineation of the scope of the organs at risk is important for protecting normal tissues and reducing side effects. However, the conventional method mainly relies on manual sketching or automatic segmentation based on threshold/templates, which is time-consuming, laborious and limited in accuracy. In recent years, deep learning techniques such as Convolutional Neural Networks (CNNs) have been introduced into the field of automatic segmentation of medical images, such as U-Net models, to achieve significant effects in tasks including lung CT segmentation. However, the conventional CNN is based on convolution of a regular grid, has a fixed receptive field, is difficult to capture non-local association on the morphology of an organ, and cannot fully process the global complex relationship of an anatomical structure. To enhance global features, some studies have introduced attention mechanisms or multi-scale feature fusion, but have limitations in dealing with complex structures and long-range dependencies. The graph neural network (Graph Neural Network, GNN) has recently received attention in medical image analysis as a deep learning method to process non-euclidean structural data. Modeling a medical image or anatomical structure as a map may more effectively represent spatial relationships and topologies between different regions. For example, there are studies to combine the graph neural network with the conventional segmentation network to characterize the complexity and heterogeneity of lung tumor and tissue, and to improve segmentation accuracy and boundary details. These methods prove that introducing graph structures helps to capture complex spatial relationships, improving segmentation performance while maintaining computational efficiency. However, the application of the graph neural network in the prior art is mostly limited to the Euclidean spatial neighbor relation, and there is still room for improvement in terms of more complex non-Euclidean geometric relation (such as curved surface adjacency of organ surfaces) and different modality information fusion. In addition, existing automatic delineation methods rarely take into account differences in the importance of the function of regions within an organ. In radiotherapy, functional image information is introduced to assist in making a 'functional avoidance' radiotherapy plan, namely, a high-functional area is prevented from being irradiated as much as possible, so that the side effect of treatment is reduced. For example, ventilation/perfusion functional imaging of the lungs may guide avoidance of areas of better healthy lung function. However, clinically acquiring such functional images (e.g., SPECT ventilation/perfusion scans) tends to be costly and complex. Thus, some studies explore the use of deep learning to synthesize functional profiles from conventional CT images, such as predicting lung perfusion maps using 3D CT features, instead of SPECT scans. This suggests that it is feasible to infer organ function based on anatomical morphology, but these functional predictions are usually independent of the segmentation process and there is no complete system for integrating segmentation and functional assessment. On the other hand, with the development of computing technology, new algorithm paradigms bring about potential breakthroughs for medical image analysis. For example, quantum computing is attempted for optimization problems such as image segmentation. There are studies modeling image pixel segmentation as a graph partitioning problem and solving using quantum annealing, and the results show that quantum methods have faster solving speeds and better results than classical methods in some unsupervised segmentation tasks. This means that introducing quantum computation into medical image processing can break through classical computation bottlenecks, achieving globally optimal graph partitioning results. In addition, self-supervision learning is used as a new trend of machine learning, model pre-training can be performed by utilizing unlabeled data, and feature representation capability is improved. In the medical imaging field, a spatio-temporal structural network has been used to improve segmentation performance using the temporal information of CT sequences. The space-time characteristics of the images are learned in a self-supervision mode, so that the adaptability of the model to organ movement and morphological changes can be improved under the condition of limited annotation data. In summary, the prior art has the defects that the conventional deep learning segmentation method lacks capture o