Stroke complication image analysis method based on prior guidance task decoupling graph learning

Abstract

The application discloses a stroke complication image analysis method based on prior guidance task decoupling graph learning, and belongs to the technical field of medical image analysis. The method acquires a brain region map formed by a preoperative non-enhanced CT image, and learns anatomical neighborhood context and left-right homology bias semantics respectively through a double-channel encoder; generates bias state soft probabilities of a hemorrhagic transformation task and a brain edema task according to a high-density voxel proportion difference and a low-density voxel proportion difference of left and right homologous brain regions, and performs decoupling contrast learning in an independent task embedding space; after gate fusion, task-specific attention convergence and stop gradient cross-task residual interaction, two complication prediction probabilities are generated. The application can effectively utilize brain region anatomical structure and left-right symmetry prior information, reduce task interference, and improve the accuracy of auxiliary image analysis of stroke complications.

Description

Technical Field

[0001] This invention relates to the field of medical image analysis technology, and in particular to a method for analyzing stroke complication images based on prior-guided task decoupling graph learning. Background Technology

[0002] Non-contrast computed tomography (CT) images are routinely acquired in the stroke emergency process. They are characterized by fast acquisition speed and high clinical accessibility, and therefore can serve as the data basis for preoperative image analysis.

[0003] Existing stroke risk analysis methods based on unenhanced computed tomography (CT) images typically employ artificial radiomics features, global image features, or tabular brain region statistical features. While artificial radiomics and tabular brain region features can describe changes in density, texture, and morphology, they usually treat the statistics of different brain regions as independent variables, failing to directly express the anatomical relationships between brain regions. Global image representation can reduce the need for artificial feature design, but it tends to weaken subtle abnormalities distributed across different brain regions and struggles to simultaneously preserve both local brain region information and inter-brain structural information.

[0004] Brain region mapping represents brain regions as nodes and their relationships as edges, thus explicitly preserving brain regions and their structural relationships. Existing brain region mapping methods typically aggregate neighborhood information along anatomical adjacency edges, primarily learning the local neighborhood context of each brain region. However, the brain's anatomical structure has approximately bilateral symmetry, and when observing non-contrast computed tomography (CT) images of stroke patients, physicians often identify subtle abnormalities by comparing left and right homologous brain regions. A single anatomical message passing pathway cannot adequately express the density shift of the affected brain region relative to its contralateral homologous region. If anatomical adjacency relationships and left-right homologous relationships are mixed in the same message passing path, the semantics of these two different types of brain regions may interfere with each other.

[0005] Hemorrhagic transformation and cerebral edema share a common ischemic pathological basis, thus joint analysis can utilize their shared imaging context. However, the relevant cues for both in non-contrast computed tomography (CT) images are not entirely consistent, potentially manifesting as different brain region distributions, density changes, and left-right asymmetry patterns. Training the two models completely independently will result in the loss of shared pathological information, while completely sharing a representation space will cause the discrimination cues for the two tasks to become entangled. Existing multi-task methods with shared encoders and independent output heads typically only distinguish tasks at the output end, failing to establish different representation spaces for different complications at brain region-level planes, and also making it difficult to control the backlash of one task's loss through shared structures affecting the other task.

[0006] Furthermore, training samples typically only provide object-level complication labels, without directly indicating which left-right homologous brain region pairs contain abnormalities associated with specific complications. Directly supervising all brain region pairs with object-level labels can easily lead to treating task-irrelevant brain region pairs as equally supervised objects, reducing the accuracy of task-specific brain region representation. Summary of the Invention

[0007] The purpose of this invention is to provide a method for analyzing stroke complication images based on prior-guided task decoupling graph learning, in order to solve the problems in the prior art, such as heterogeneous semantic interference caused by mixed modeling of anatomical adjacency and left-right homology, the inability of object-level labels to provide task-specific brain region pair supervision, and the easy occurrence of task entanglement and negative transfer when sharing representations of related complication tasks.

[0008] To achieve the above objectives, this invention provides a method for stroke complication image analysis based on prior-guided task decoupling graph learning, including a model training phase and an image analysis phase, the specific steps of which are as follows: Perform the following steps during the model training phase: S1. Obtain the set of training brain region maps and the corresponding hemorrhage transformation task labels and cerebral edema task labels; each training brain region map is formed based on the preoperative non-enhanced computed tomography images that have been obtained, including the set of brain region nodes, the set of node features, the set of anatomical adjacent edges, and the set of left and right homologous edges. S2. Input the node feature set into a dual-channel encoder, aggregate brain region neighborhood information along the anatomical adjacency edge set through the topological channel to obtain the topological channel representation of each brain region node; generate left and right homologous brain region pair representations based on the left and right homologous edge set through the symmetric channel, and generate the symmetric channel representation of each brain region node based on the left and right homologous brain region pair representations; S3. Determine the soft probability of the hemorrhage conversion task deviation state based on the hemorrhage conversion task correlation density deviation between the left and right homologous brain regions, and determine the soft probability of the cerebral edema task deviation state based on the cerebral edema task correlation density deviation between the left and right homologous brain regions; map the left and right homologous brain region pairs to the hemorrhage conversion task embedding space and the cerebral edema task embedding space respectively, and perform task-specific contrastive learning based on the corresponding task deviation state soft probabilities. S4. Generate a feature-dimensional gate vector based on the topological channel representation and the symmetric channel representation, and use the feature-dimensional gate vector to fuse the topological channel representation and the symmetric channel representation element by element to obtain the final representation of the brain region node; S5. Attention convergence is performed on the final representation of the brain region nodes using the hemorrhage transformation task query vector and the cerebral edema task query vector, respectively, to obtain the hemorrhage transformation graph-level task representation and the cerebral edema graph-level task representation. S6. After passing the stopping gradient operation and the hemorrhage transformation bottleneck mapping, the brain edema map-level task representation is injected into the hemorrhage transformation map-level task representation in residual form; after passing the stopping gradient operation and the brain edema bottleneck mapping, the hemorrhage transformation map-level task representation is injected into the brain edema map-level task representation in residual form to obtain the hemorrhage transformation interactive task representation and the brain edema interactive task representation; the model parameters are updated according to the hemorrhage transformation task label, the brain edema task label and the task-specific contrast loss to obtain the trained image analysis model; The following steps are performed during the image analysis phase: S7. Obtain the target brain region map formed based on the preoperative non-enhanced computed tomography image to be analyzed, input the target brain region map into the trained image analysis model, and sequentially perform dual-channel encoding, feature-dimensional gating fusion, task-specific attention convergence, and stopping gradient cross-task residual interaction, and generate the hemorrhage conversion prediction probability and cerebral edema prediction probability through the hemorrhage conversion output head and cerebral edema output head, respectively.

[0009] Preferably, in S1, the brain region map is represented as follows: ,in, i For object indexing, V A set of brain region nodes, Let be the set of edges. To dissect the set of adjacent edges, Let be the set of edges with the same origin on both sides. For the first i A set of brain region node features for each object; the brain region nodes form multiple pairs of left and right homologous brain regions according to their left-right homology relationship.

[0010] Preferably, in S1, the node features of each brain region node are formed by splicing together local brain region features, bilateral contrast features, and auxiliary pathological features; The local brain region features include one or more of the following: Henlein unit distribution statistics, high-density voxel ratio, low-density voxel ratio, texture description, and abnormal load statistics within the brain region; the high-density voxel ratio is the ratio of the number of effective voxels with computed tomography values ​​higher than a preset high-density threshold to the total number of effective voxels in the brain region; the low-density voxel ratio is the ratio of the number of effective voxels with computed tomography values ​​lower than a preset low-density threshold to the total number of effective voxels in the brain region; the threshold is determined before model training based on a preset Henlein unit range or training data, and remains consistent during the model training and image analysis phases. The bilateral contrast features include at least the feature difference and normalized density contrast between the current brain region and its left and right homologous brain regions; The auxiliary pathological features include at least contralateral density difference and low-density candidate features.

[0011] Preferably, in S2, for the first k Pairs of homologous brain regions were used to obtain the topological pathway representation of the left brain region. and topological pathways in the right brain region ,Will and sum, and The absolute difference and and The element-wise products are concatenated and input into a shared difference mapping network to obtain representations of left and right homologous brain regions. ; The representations of the left and right homologous brain regions are concatenated with the corresponding topological channel representations of the left and right brain regions, respectively, and shared projection is performed to obtain the symmetric channel representations of the nodes in the left and right brain regions, respectively.

[0012] Preferably, in S3, the hemorrhage conversion task-related density deviation is the absolute difference in the proportion of high-density voxels in the left and right homologous brain regions, and the cerebral edema task-related density deviation is the absolute difference in the proportion of low-density voxels in the left and right homologous brain regions. , ; in, and The first k The ratio of high-density voxels between the left and right homologous brain regions in the left and right brain regions. and These represent the corresponding low-density voxel ratios; For hemorrhage conversion task-related density bias, This is due to task-related density bias related to cerebral edema. set up t For task indexing, ,in, HT Indicates the bleeding transformation task. Ed Indicates cerebral edema task; will the first k Approximately 10 homologous brain regions are involved in the task. t The task-related density bias is denoted as ;when t = HT hour, ;when t = Ed hour, ; Task-related density bias According to the scale constant of the corresponding brain region Normalization truncates the denominator to a number not less than a preset positive number. e And through a learnable threshold S Shape functions generate task deviation state soft probabilities: , , ; in, Indicates the task-related density bias Normalized task-related density bias obtained after normalization processing The brain region scale constant estimated using only training fold data. As a preset positive number, This indicates taking the larger of the two values; This refers to the soft probability that the left and right homologous brain regions are in a biased state under the corresponding task; This is the Sigmoid activation function, used to map the input to the interval (0,1); To control the steepness of the soft state, a temperature parameter is assigned. The initial parameter is the learnable threshold, and its initial value is determined based on the task-related density bias distribution in the training trade-off. This is a positive value mapping function used to convert the original learnable threshold parameters into positive threshold values, and .

[0013] Preferably, in S3, the specific content of task-specific contrastive learning is as follows: Representing the left and right homologous brain regions Input the hemorrhage transformation task projection head and the cerebral edema task projection head respectively to obtain task-specific embeddings. ; For two brain region pairs within the same batch, the similarity is calculated using task-specific embedding temperature-scaled cosine similarity. When two brain region pairs correspond to the same left-right homologous brain region pair index and are both in a biased state or both in a normal state, the soft weight of the positive sample is determined based on the soft probability product of their task bias states. Specifically, for two brain region pairs within the same batch... a and b The cosine similarity of the task-specific embedding is divided by the contrast temperature. , obtain embedding similarity When the sample a and b When corresponding left and right homologous brain region pairs are indexed and both are in a deviated state or both are in a normal state, based on the sample a and b Task deviation state soft probability and The product determines the soft weights of positive samples. : ; in, This is an indicator function; it takes the value 1 when the condition within the square brackets is true, and 0 otherwise. This is a pair indexing function used to return the left and right homologous brain region pair indexes corresponding to the sample, where the two sample indices correspond to two brain region pairs within the batch, representing the sample. The first term in the table represents the soft joint weight when both samples are in a biased state, the second term represents the soft joint weight when both samples are in a normal state, and the sum of the two terms in the square brackets represents the soft weight of the positive sample when both samples are in the same state. Task t The contrast loss is expressed as: ; in, A valid set of anchor points consists of anchor points with at least one positive sample soft weight greater than zero.

[0014] Preferably, in S4, the feature dimension gate vector The final representation of brain region nodes They are represented as follows: , ; in, Indicates feature splicing, and These are brain region nodes. Topological channel representation and symmetric channel representation. and These are the learnable gated weight matrix and bias vector, respectively. Represents element-wise product. Indicates to Element-by-element complement.

[0015] Preferably, in S5, the query vector for the hemorrhage conversion task and the query vector for the cerebral edema task are initialized to be mutually orthogonal; For the task t Node attention weights and the corresponding graph-level task representation They are represented as follows: , ; in, For task query vectors, Dimensions are represented by nodes; In S6, the hemorrhage transformation interaction task representation and the cerebral edema interaction task representation are uniformly represented as follows: ; in, and These represent the current task and another task, respectively. This represents the task after the interaction. and This represents the graph-level task representation before interaction. For bottleneck mapping networks, To stop the gradient operation, This is the directional scaling factor. Injecting strength into learnable content This indicates the sum of residuals.

[0016] This invention also provides a stroke complication image analysis system based on prior-guided task decoupling graph learning, including a brain region map data acquisition module, a dual-channel encoding module, a task-specific bias generation module, a task decoupling comparison learning module, a gating fusion module, a task-specific convergence module, a stopping gradient cross-task interaction module, a model training module, and an image analysis module.

[0017] The brain region map data acquisition module is used to acquire brain region maps based on preoperative non-contrast computed tomography images; The dual-channel encoding module is used to learn the topological representation of brain regions along anatomical adjacent edges through the topological channel, and to generate brain region pair representations and symmetric channel representations based on left and right homologous brain regions through the symmetric channel. The task-specific bias generation module is used to generate soft probabilities of hemorrhage conversion task bias state and cerebral edema task bias state based on the difference in the proportion of high-density voxels and the difference in the proportion of low-density voxels in the left and right homologous brain regions, respectively. The task decoupling and contrastive learning module is used to perform bias-guided contrastive learning in the independent embedding spaces of hemorrhage transformation task and cerebral edema task; A gated fusion module is used to fuse the brain region topological representation and the symmetric channel representation; The task-specific aggregation module is used to generate graph-level task representations of hemorrhage transformation and cerebral edema using the hemorrhage transformation task query vector and the cerebral edema task query vector, respectively. The Stop Gradient Cross-Task Interaction module is used to inject the graph-level task representation of one task into another task branch in the form of residuals after stopping gradients and bottleneck mapping. The model training module is used to update the model parameters based on the hemorrhage conversion task label, the cerebral edema task label, and the bias-guided contrastive learning loss, so as to obtain the trained image analysis model. The image analysis module is used to input the target brain region map of the object to be analyzed into the trained image analysis model and output the hemorrhage conversion prediction probability and the cerebral edema prediction probability. The system can be deployed on a medical image processing workstation, server, or cloud computing platform.

[0018] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the stroke complication image analysis method based on prior-guided task decoupling graph learning.

[0019] Therefore, the stroke complication image analysis method based on prior guided task decoupling graph learning described above has the following beneficial effects: 1) Input the anatomical adjacency relationship and the left-right homology relationship into the topological channel and the symmetry channel respectively, so that the brain region neighborhood context and the left-right deviation semantics are learned separately before fusion, reducing the mutual coverage of the two types of heterogeneous relationships; 2) Construct the bias state soft probability of two tasks by using the high-density voxel ratio difference and low-density voxel ratio difference of the left and right homologous brain regions. Provide task-specific image soft supervision for brain region pair representation when the object-level label cannot point out the abnormal brain region pair. 3) Project the same brain region pairs onto two independent task embedding spaces so that the hemorrhage conversion task and the cerebral edema task can organize the brain region pairs according to their respective related density deviation patterns, thereby reducing task entanglement. 4) By using task-specific attention convergence and stopping gradient cross-task interaction, one task can read high-level image information of another task, while suppressing the back interference of the receiving task loss on the source task representation. 5) The key supervision signal uses the relative density deviation between the left and right homologous brain regions to reduce the model's dependence on the difference in absolute intensity scale between different scanning centers, which is beneficial to improving the cross-center generalization ability. 6) It generates two complication prediction probabilities simultaneously in a single forward calculation, improving the efficiency of combined image analysis and providing reference information for doctors to conduct auxiliary image analysis and risk stratification of stroke complications.

[0020] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall architecture of the stroke complication image analysis method based on prior guided task decoupling graph learning of the present invention; Figure 2 These are ablation experiment results for each component module of the present invention; Figure 3 This is a queue-averaged brain region attention distribution map for the two tasks of this invention; Figure 4 This is a statistical chart of the ten brain regions with the highest attention weights for the two tasks in this invention; Figure 5 This is a system implementation example diagram based on a browser, which is an example of the present invention. Detailed Implementation

[0022] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments. It should be understood that the following embodiments are used to illustrate the present invention and not to limit the scope of protection of the present invention. Equivalent substitutions made by those skilled in the art for graph convolution operators, mapping network structures, feature statistics, or task output formats without departing from the concept of the present invention can all be included within the scope of protection of the present invention.

[0023] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," etc., used in this invention are only used to distinguish different tasks, branches, outputs, or modules, and do not indicate order, quantity, or importance.

[0024] The object processed by this invention is preoperative non-enhanced computed tomography image data or brain region map data formed based on the image data. The hemorrhage conversion prediction probability and cerebral edema prediction probability generated by the model are computer-generated auxiliary image analysis information and do not directly constitute a disease diagnosis or final clinical conclusion.

[0025] Example 1 like Figure 1 As shown, this embodiment provides a method for analyzing stroke complication images based on prior-guided task decoupling graph learning, including a model training phase and an image analysis phase.

[0026] Perform the following steps during the model training phase: S1. Given a training brain region map dataset containing N objects and corresponding hemorrhage transformation task labels and cerebral edema task labels; the hemorrhage transformation task labels and cerebral edema task labels correspond to hemorrhage transformation-related image analysis tasks and cerebral edema-related image analysis tasks, respectively. The model performs a forward computation once for each brain region map, generating a hemorrhage transformation prediction probability. and the probability of predicting cerebral edema .

[0027] Each training brain region map includes a set of brain region nodes, a set of node features, a set of anatomical adjacent edges, and a set of left and right homologous edges. i The brain region map of each object is represented as follows: , V A set of brain region nodes, Let be the set of edges. To dissect the set of adjacent edges, Let be the set of edges with the same origin on both sides. It is a set of brain region node features; anatomical adjacent edges are used for topological channel message propagation, and left and right homologous edges are used for symmetric channel brain region pair comparison. The two types of relationships are encoded separately before entering gating fusion.

[0028] Figure 1 In the text, anatomical adjacent edges represent the connectivity between adjacent brain regions in space, and left and right homologous edges represent the pairing relationship between corresponding brain regions in the left and right hemispheres; the normal state and the biased state in the embedded space represent two types of soft states described by the soft probability of task biased states, respectively.

[0029] In this embodiment, each training brain region map is formed based on the preoperative non-enhanced computed tomography images. As a feasible data preparation method, the images can be registered to a preset standard brain space, and isovoxel resampling and skull removal can be performed. Then, a preset cerebral vascular supply area atlas is superimposed, and the target brain region in the atlas is used as a node, and formed according to the left-right anatomical correspondence. K Approximately 10 pairs of homologous brain regions. This data preparation method is only for illustrating how the brain region map input is obtained; standard brain space, resampling precision, cranial removal tools, and brain region maps can be replaced according to application conditions.

[0030] Each brain region node u The node features are composed of local brain region features bilateral contrast features and auxiliary pathological features It is formed by piecing together.

[0031] Local brain region features This is used to summarize evidence within the current brain region, including one or more of the following: Henlein unit distribution statistics, low-density voxel ratio, high-density voxel ratio, texture descriptor, and abnormal load statistics.

[0032] The high-density voxel ratio is the ratio of the number of valid voxels in the current brain region whose computed tomography (CT) values ​​exceed a preset high-density threshold to the total number of valid voxels in that brain region. The low-density voxel ratio is the ratio of the number of valid voxels in the current brain region whose CT values ​​fall below a preset low-density threshold to the total number of valid voxels in that brain region. Both thresholds can be determined based on a preset Henlein unit range or solely using the density distribution of the training fold data. Once determined, the thresholds remain unchanged throughout the corresponding training, validation, and image analysis processes using the trained model, and are not adjusted using the outcome labels of the validation or analysis subjects.

[0033] bilateral contrast features Used to express the differences between the current brain region and the left and right homologous brain regions, including one or more of the following: left-right feature difference, normalized density contrast, and water absorption-related normalized contrast.

[0034] Auxiliary pathological features Used to supplement descriptive information related to pathological changes, including one or more of refined contralateral density differences and low-density candidate features.

[0035] In addition to the standardized node features, this embodiment also preserves the original scale high-density voxel ratio for each brain region. and low-density voxel ratio It is used to calculate the left and right density deviations with physical meaning.

[0036] S2. Input the node feature set into the dual-channel encoder, and then analyze the dissected adjacent edge set through the topological channel. By aggregating neighborhood information of brain regions, topological channel representations of nodes in each brain region are obtained; these representations are then obtained using symmetric channels based on the set of left and right homologous edges. Generate left and right homologous brain region pairs representation And based on the left and right homologous brain regions, the representation Generate symmetrical channel representations for nodes in each brain region.

[0037] Topological pathways are used to learn the local context of brain regions within their anatomically adjacent structures. This is first achieved through input mapping. Map node features to the initial hidden state, then dissect the set of adjacent edges. The process is performed sequentially: layer normalization, graph convolution, nonlinear activation, random deactivation, and residual summation. This process is repeated. L The node topology channel representation is then obtained. Pre-normalization is used to stabilize the representation distribution in the early stages of training, and residual connections enable strong signal nodes to retain their original information. Topological channels propagate messages only along anatomically adjacent edges, not directly along left and right edges of the same origin, thus avoiding the mixing of anatomically neighboring semantics and cross-hemispherical bias semantics in the same propagation path.

[0038] like Figure 1 As shown, symmetric channels are used to explicitly learn density and representational biases between homologous left and right brain regions. For the first... k Approximately 10 pairs of homologous brain regions ( , ), respectively obtain the topological channel representation of the left brain region. and topological pathways in the right brain region ,Will and sum, and The absolute difference and and The element-wise products are concatenated and input into a shared difference mapping network to obtain representations of left and right homologous brain regions. : ; In the formula, Used to preserve the common context between the left and right hemispheres of the brain. The absolute value is used to express the intensity of the left and right deviations. This is used to represent joint activation of bilateral features. Using absolute difference enables commutativity between the left and right sides, eliminating the need to pre-fix the affected side in the model structure.

[0039] Subsequently, the left and right homologous brain regions were represented. It is concatenated with the corresponding topological channel representations of the left and right brain regions, respectively, and then shared projection is performed. The symmetrical channel representations of nodes in the left and right brain regions were obtained respectively: ; Thus, the left and right nodes in the same brain region pair obtain symmetrical channel representations containing information about the deviation between the brain regions.

[0040] S3. Determine the soft probability of hemorrhage conversion task deviation state based on the density correlation between hemorrhage conversion task and cerebral edema task deviation state based on the density correlation between cerebral edema task and cerebral edema task; represent the left and right homologous brain regions as pairs. They are respectively mapped to the embedding space of the hemorrhage transformation task and the embedding space of the cerebral edema task, and task-specific contrastive learning is performed based on the soft probabilities of the corresponding task bias states.

[0041] Object-level labels only indicate whether a corresponding complication has occurred, and do not directly indicate which left-right homologous brain regions contain task-related abnormalities. In this embodiment, left-right density deviation is extracted from the image itself as task-specific soft supervision at the brain region pair level.

[0042] The following deviation state calculations are performed independently for each object; object indices are omitted for simplicity. i For the first k The density biases related to hemorrhage conversion task and cerebral edema task, representing approximately 10 homologous brain region pairs, are respectively expressed as: , ; in, and The first k The ratio of high-density voxels between the left and right homologous brain regions in the left and right brain regions. and These represent the proportions of low-density voxels, respectively. The hemorrhage conversion task uses the high-density voxel proportion difference, while the cerebral edema task uses the low-density voxel proportion difference. These two indicators describe different aspects of the same ischemic density change, allowing the two tasks to represent brain regions using different density cues.

[0043] Furthermore, to unify the representation of the hemorrhage transformation task and the cerebral edema task, a set of... t For task indexing, ,in, HT Indicates the bleeding transformation task. Ed Indicates cerebral edema task; will the first k Approximately 10 homologous brain regions are involved in the task. t The task-related density bias is denoted as ;when t = HT hour, ;when t = Ed hour, ;therefore, yes and The representation under the unified task index is not a density bias calculated separately from either of them. Different brain region pairs have different normal left-right variation ranges; therefore, the brain region pair scale constant estimated in the training compromise is used. Task-related density bias Perform normalization and truncate the denominator to a minimum of a predetermined positive number. e To avoid a denominator of zero, a learnable threshold is used. S Shape functions generate task deviation state soft probabilities: , , ; in, Indicates the task-related density bias Normalized task-related density bias obtained after normalization processing The brain region scale constant estimated using only training fold data. As a preset positive number, This indicates taking the larger of the two values; This refers to the soft probability that the left and right homologous brain regions are in a biased state under the corresponding task; This is the Sigmoid activation function, used to map the input to the interval (0,1); To control the steepness of the soft state, a temperature parameter is assigned. The initial parameter for the learnable threshold is determined based on the task-related density bias distribution during training. From the learnable threshold parameter Obtained through positive value mapping; This is a positive value mapping function used to convert the original learnable threshold parameters into positive threshold values, and The soft probability of the task deviation state is obtained by normalizing and thresholding the corresponding task-related density deviation, and is used to determine subsequent tasks. t Comparative learning of positive sample soft weights.

[0044] The specific content of task-specific contrastive learning is as follows: Representing left and right homologous brain regions Input the bleeding conversion task projection head respectively and cerebral edema task projection head The projection results are then normalized using the L2 norm to obtain the task-specific embeddings for each of the two tasks. Independent projection heads are used for the two tasks, allowing the same brain region pairs to form different geometric structures in the hemorrhage transformation task space and the cerebral edema task space.

[0045] In containing M In the batch represented by each brain region pair, with a and b The batch represents the brain region pairs representing the sample index; the cosine similarity of the task-specific embedding is divided by the contrast temperature. , obtain embedding similarity When the sample a and b When corresponding left and right homologous brain region pairs are indexed and both are in a deviated state or both are in a normal state, based on the sample a and b Task deviation state soft probability and The product determines the soft weights of positive samples. : ; in, This is an indicator function; it takes the value 1 when the condition within the square brackets is true, and 0 otherwise. This is a pair indexing function used to return the left and right homologous brain region pair indexes corresponding to the sample, where the two sample indices correspond to two brain region pairs within the batch, representing the sample. The first term in the equation represents the soft joint weight when both samples are in a biased state, the second term represents the soft joint weight when both samples are in a normal state, and the sum of the two terms in square brackets represents the soft weight of the positive sample when both samples are in the same state. Using both states as positive sample conditions can avoid organizing the embedding space around only a single severity axis.

[0046] The set of effective anchor points consists of anchor points that have at least one positive sample soft weight greater than zero. ;Task t The contrast loss is expressed as: .

[0047] S4. Generate feature-dimensional gating vectors based on the topological channel representation and the symmetric channel representation. Use the feature-dimensional gating vectors to perform element-wise fusion of the topological channel representation and the symmetric channel representation to obtain the final representation of the brain region node.

[0048] Topological channel representations include anatomical neighborhood context, while symmetric channel representations include left-right homology bias; the two types of representations are complementary. For brain region nodes... u The feature dimension gate vector is generated based on the concatenation result of the two representations, and then fused element by element: , ; in, Represents the feature dimension gate vector, Indicates feature splicing, For the final representation of brain region nodes, This is a topology channel representation. For symmetrical channel representation, and These are the learnable gated weight matrix and bias vector, respectively. Represents element-wise product. The value range of each component is (0,1). Indicates to Element-wise complementation. The gating mechanism allows different brain regions and different feature dimensions to adaptively select more anatomical neighborhood information or more left-right deviation information, rather than using the same scalar weight for all nodes.

[0049] S5. Attention convergence is performed on the final representation of brain region nodes using the hemorrhage transformation task query vector and the cerebral edema task query vector, respectively, to obtain the hemorrhage transformation graph-level task representation and the cerebral edema graph-level task representation.

[0050] The fused brain region node representation serves as a shared key and value for both tasks, with the hemorrhage transformation task and the cerebral edema task each holding a query vector. and The two query vectors can be initialized to be mutually orthogonal, allowing the two tasks to focus on different brain region representations at the start of training.

[0051] For the task t Node attention weights and the corresponding graph-level task representation They are represented as follows: , ; in, For task query vectors, The nodes are represented by dimensions; thus, the hemorrhage transformation task and the cerebral edema task can form different attention distributions for the same set of brain region nodes, instead of using shared graph-level convergence results.

[0052] S6. After passing the stopping gradient operation and the hemorrhage transformation bottleneck mapping, the brain edema graph-level task representation is injected into the hemorrhage transformation graph-level task representation in the form of residuals; after passing the stopping gradient operation and the brain edema bottleneck mapping, the hemorrhage transformation graph-level task representation is injected into the brain edema graph-level task representation in the form of residuals, thus obtaining the hemorrhage transformation interactive task representation and the brain edema interactive task representation.

[0053] The hemorrhage transformation task and the cerebral edema task share a common ischemic pathological background, and completely isolating the two branches would result in the loss of shareable high-level information. To achieve information exchange without re-merging the representation spaces of the two tasks, this embodiment sets up bidirectional controlled cross-task interaction.

[0054] For the task t The tasks after interaction are uniformly represented as: ; in, and These represent the current task and another task, respectively. This represents the task after the interaction. and This represents the graph-level task representation before interaction. For bottleneck mapping networks, To stop the gradient operation, This is the directional scaling factor. Injecting strength into learnable content This indicates the addition of residuals. Stopping gradient operations allows the receiving task to read the graph-level representation of another task, but the receiving task's loss cannot modify the source task's representation through this interaction path. Bottleneck mapping and scaling factors restrict the interaction to controlled residuals, rather than full feature sharing.

[0055] Two independent output heads perform nonlinear mapping and Sigmoid activation function transformation on the hemorrhage transformation interactive task representation and the cerebral edema interactive task representation, respectively, to generate hemorrhage transformation prediction probability and cerebral edema prediction probability with values ​​ranging from 0 to 1.

[0056] Subsequently, the model parameters were updated based on the hemorrhage conversion task label, the cerebral edema task label, and the task-specific contrast loss to obtain the trained image analysis model. Specifically, the model was jointly trained using two object-level binary cross-entropy losses and a task decoupling contrast loss. The total loss was obtained by weighted summation of the hemorrhage conversion task cross-entropy loss, the cerebral edema task cross-entropy loss, and the task-specific contrast loss. When some brain region statistics were missing, a missing mask was used to skip the corresponding bias term; when the brain region pair scale constant was close to zero, the denominator was truncated to a positive number not less than a preset value; when a brain region pair in a batch lacked valid positive samples, the brain region pair was not included in the valid anchor set; when the valid anchor set was empty, the contrast loss for the corresponding task in this batch was skipped. Normalized statistics, brain region pair scale constants, and threshold initialization statistics were all calculated only based on the training fold.

[0057] The following steps are performed during the image analysis phase: S7. Obtain the target brain region map formed from the preoperative non-enhanced computed tomography images to be analyzed. Input the target brain region map into the trained image analysis model and sequentially perform dual-channel encoding, feature-dimensional gating fusion, task-specific attention convergence, and stopping gradient cross-task residual interaction. Generate the hemorrhage conversion prediction probability and the cerebral edema prediction probability through the hemorrhage conversion output head and the cerebral edema output head, respectively.

[0058] During the image analysis phase, only the acquired image data or the target brain region map derived from it is input; follow-up outcome labels for the subjects to be analyzed are not input. The system outputs computer-generated hemorrhage transformation prediction probabilities and cerebral edema prediction probabilities, used to assist in image analysis and risk stratification reference.

[0059] This embodiment uses data from a multicenter study of acute ischemic stroke patients who underwent endovascular treatment for validation, comprising 675 subjects. Site 1 included 102 subjects, of whom 55 had a positive label for the hemorrhage transformation task and 66 had a positive label for the cerebral edema task; Site 2 included 142 subjects, of whom 63 had a positive label for the hemorrhage transformation task and 71 had a positive label for the cerebral edema task; Site 3 included 431 subjects, of whom 195 had a positive label for the hemorrhage transformation task and 211 had a positive label for the cerebral edema task. Both complications can occur simultaneously, therefore the two labels are not mutually exclusive. All subjects had baseline non-contrast computed tomography images before treatment and follow-up images acquired approximately 24 hours after treatment to determine training and validation labels; follow-up labels were only used for model training or validation and were not required input for the model application phase.

[0060] Five-fold object hierarchical cross-validation was used to compare the proposed method with traditional machine learning methods, general graph neural network methods, and brain network learning methods. Evaluation metrics included hemorrhage-converted AUC, cerebral edema AUC, mean AUC, macro-mean precision, macro-mean F1, and macro-mean balanced precision. The complete results are shown in Table 1.

[0061] Table 1: Performance comparison of multicenter cohort combined image analysis (mean ± standard deviation, %) ; As shown in Table 1, the method of the present invention achieves the highest values ​​across all six indicators, with an average AUC 3.64 percentage points higher than the best comparison method.

[0062] To verify the function of each component module, five variations were set up: removing symmetric channels, removing task-specific contrastive learning, adopting shared contrastive learning, removing stopping gradient cross-task interaction, and adopting shared task convergence. Figure 2 The results for each variant are shown in terms of average AUC, macro-average precision, and macro-average balanced precision. Removing any module leads to a performance degradation, with removing the symmetry channel decreasing the average AUC by 2.33 percentage points, removing task-specific contrastive learning, and using shared contrastive learning decreasing it by 1.87 and 2.10 percentage points, respectively. This indicates that left-right homologous symmetry information, task decoupling supervision, task-specific convergence, and controlled cross-task interaction are all effective for joint image analysis.

[0063] The results of further comparing single-task training with the joint training of the present invention are shown in Table 2.

[0064] Table 2: Comparison of single-task training and joint training of the present invention (mean ± standard deviation, %) ; Table 2 shows that joint training did not cause significant negative transfer and achieved comparable or higher performance on both tasks and overall metrics.

[0065] To evaluate the cross-center generalization capability, a leave-one-site validation protocol was adopted. The system was trained at two sites and tested directly at a third site without fine-tuning. The results are shown in Table 3.

[0066] Table 3: External validation results of one of the three sites (%) ; The method of this invention achieved average AUCs of 70.58%, 60.71%, and 67.20% at the three leaveout sites, which were the highest values ​​for the corresponding sites. The support for relative density bias and task decoupling mechanisms helps to maintain cross-site analysis capabilities.

[0067] Figure 3The mean attention distribution of 675 subjects in the hemorrhage conversion task and cerebral edema task cohorts is shown. Figure 4 The ten brain regions with the highest attention weights for the two tasks are shown. These high-weighted regions are mainly distributed in the insular cortex of the middle cerebral artery supply area, the lenticulostriate artery region, and multiple cortical branches, and often appear in left-right hemisphere pairs. The attention results are used to illustrate the consistency between the model's attention location and the dual-channel structure, and should not be considered causal evidence or clinical diagnostic conclusions.

[0068] On a single NVIDIA RTX 3090 graphics processor, the method of this invention contains approximately 46,000 parameters, with a training time of approximately 0.89 seconds per round and a single-inference time of approximately 6.62 milliseconds. Figure 5 This document illustrates a browser-based implementation example, including image uploading, quality inspection, brain region segmentation, complication-assisted prediction, result interpretation, and report export processes. The complication-assisted prediction step generates hemorrhage conversion prediction probabilities and cerebral edema prediction probabilities. The result interpretation step can display the two prediction probabilities as high-risk or low-risk indicators according to preset judgment rules. These indicators are only auxiliary prediction results and do not directly constitute a clinical diagnostic conclusion. This system example is only for illustrating the engineering implementation method and does not limit the deployment interface or hardware environment of this invention.

[0069] Therefore, this invention employs the aforementioned method for stroke complication image analysis based on prior-guided task decoupling graph learning. It learns anatomical neighborhood context and left-right homology bias semantics through dual-channel encoding, avoiding interference from heterogeneous relationships. Task-specific soft supervision is constructed using left-right density bias, achieving effective supervision at the brain region pair level even when object-level labels cannot indicate abnormal brain region pairs. Independent task embedding spaces and stopping gradient cross-task interactions suppress task entanglement and negative transfer while sharing pathological information. In a multi-center stroke dataset, the average AUC of this invention reaches 72.47%, an improvement of 3.64 percentage points compared to the best comparison method. Furthermore, the average AUC of the three leave-one-out validation sites is the highest value for their respective sites, indicating that this invention can effectively assist in the imaging risk stratification analysis of hemorrhage transformation and cerebral edema.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for analyzing stroke complication images based on prior-guided task decoupling graph learning, characterized in that, The process includes a model training phase and an image analysis phase, with the following specific steps: Perform the following steps during the model training phase: S1. Obtain the set of training brain region maps and the corresponding hemorrhage transformation task labels and cerebral edema task labels; each training brain region map is formed based on the preoperative non-enhanced computed tomography images that have been obtained, including the set of brain region nodes, the set of node features, the set of anatomical adjacent edges, and the set of left and right homologous edges. S2. Input the node feature set into a dual-channel encoder, aggregate brain region neighborhood information along the anatomical adjacency edge set through the topological channel to obtain the topological channel representation of each brain region node; generate left and right homologous brain region pair representations based on the left and right homologous edge set through the symmetric channel, and generate the symmetric channel representation of each brain region node based on the left and right homologous brain region pair representations; S3. Determine the soft probability of the hemorrhage conversion task deviation state based on the hemorrhage conversion task correlation density deviation between the left and right homologous brain regions, and determine the soft probability of the cerebral edema task deviation state based on the cerebral edema task correlation density deviation between the left and right homologous brain regions; map the left and right homologous brain region pairs to the hemorrhage conversion task embedding space and the cerebral edema task embedding space respectively, and perform task-specific contrastive learning based on the corresponding task deviation state soft probabilities. S4. Generate a feature-dimensional gate vector based on the topological channel representation and the symmetric channel representation, and use the feature-dimensional gate vector to fuse the topological channel representation and the symmetric channel representation element by element to obtain the final representation of the brain region node; S5. Attention convergence is performed on the final representation of the brain region nodes using the hemorrhage transformation task query vector and the cerebral edema task query vector, respectively, to obtain the hemorrhage transformation graph-level task representation and the cerebral edema graph-level task representation. S6. After passing the stopping gradient operation and the hemorrhage transformation bottleneck mapping, the brain edema map-level task representation is injected into the hemorrhage transformation map-level task representation in residual form; after passing the stopping gradient operation and the brain edema bottleneck mapping, the hemorrhage transformation map-level task representation is injected into the brain edema map-level task representation in residual form to obtain the hemorrhage transformation interactive task representation and the brain edema interactive task representation; the model parameters are updated according to the hemorrhage transformation task label, the brain edema task label and the task-specific contrast loss to obtain the trained image analysis model; The following steps are performed during the image analysis phase: S7. Obtain the target brain region map formed based on the preoperative non-enhanced computed tomography image to be analyzed, input the target brain region map into the trained image analysis model, and sequentially perform dual-channel encoding, feature-dimensional gating fusion, task-specific attention convergence, and stopping gradient cross-task residual interaction, and generate the hemorrhage conversion prediction probability and cerebral edema prediction probability through the hemorrhage conversion output head and cerebral edema output head, respectively.

2. The method for stroke complication image analysis based on prior-guided task decoupling graph learning according to claim 1, characterized in that, In S1, the brain region map is represented as follows: ,in, i For object indexing, V A set of brain region nodes, Let be the set of edges. To dissect the set of adjacent edges, Let be the set of edges with the same origin on both sides. For the first i A set of brain region node features for each object; the brain region nodes form multiple pairs of left and right homologous brain regions according to their left-right homology relationship.

3. The method for stroke complication image analysis based on prior-guided task decoupling graph learning according to claim 2, characterized in that, In S1, the node features of each brain region node are formed by splicing together local brain region features, bilateral contrast features, and auxiliary pathological features; The local brain region features include one or more of the following: Henlein unit distribution statistics, high-density voxel ratio, low-density voxel ratio, texture description, and abnormal load statistics within the brain region; the high-density voxel ratio is the ratio of the number of effective voxels with computed tomography values ​​higher than a preset high-density threshold to the total number of effective voxels in the brain region; the low-density voxel ratio is the ratio of the number of effective voxels with computed tomography values ​​lower than a preset low-density threshold to the total number of effective voxels in the brain region; the threshold is determined before model training based on a preset Henlein unit range or training data, and remains consistent during the model training and image analysis phases. The bilateral contrast features include at least the feature difference and normalized density contrast between the current brain region and its left and right homologous brain regions; The auxiliary pathological features include at least contralateral density difference and low-density candidate features.

4. The method for stroke complication image analysis based on prior-guided task decoupling graph learning according to claim 3, characterized in that, In S2, for the first k Pairs of homologous brain regions were used to obtain the topological pathway representation of the left brain region. and topological pathways in the right brain region ,Will and sum, and The absolute difference and and The element-wise products are concatenated and input into a shared difference mapping network to obtain representations of left and right homologous brain regions. ; The representations of the left and right homologous brain regions are concatenated with the corresponding topological channel representations of the left and right brain regions, respectively, and shared projection is performed to obtain the symmetric channel representations of the nodes in the left and right brain regions, respectively.

5. The method for stroke complication image analysis based on prior-guided task decoupling graph learning according to claim 4, characterized in that, In S3, the density bias related to the hemorrhage conversion task is the absolute difference in the proportion of high-density voxels in the left and right homologous brain regions, and the density bias related to the cerebral edema task is the absolute difference in the proportion of low-density voxels in the left and right homologous brain regions. ， ； in, and The first k The ratio of high-density voxels between the left and right homologous brain regions in the left and right brain regions. and These represent the corresponding low-density voxel ratios; For hemorrhage conversion task-related density bias, This is due to task-related density bias related to cerebral edema. set up τ For task indexing, ,in, HT Indicates the bleeding transformation task. Ed Indicates cerebral edema task; will the first k Approximately 10 homologous brain regions are involved in the task. τ The task-related density bias is denoted as ;when τ = HT hour, ;when τ = Ed hour, ; Task-related density bias According to the scale constant of the corresponding brain region Normalization truncates the denominator to a number not less than a preset positive number. ε And through a learnable threshold S Shape functions generate task deviation state soft probabilities: ， ， ； in, Indicates the task-related density bias Normalized task-related density bias obtained after normalization processing The brain region scale constant estimated using only training fold data. As a preset positive number, This indicates taking the larger of the two values; This refers to the soft probability that the left and right homologous brain regions are in a biased state under the corresponding task; This is the Sigmoid activation function, used to map the input to the interval (0,1); To control the steepness of the soft state, a temperature parameter is assigned. The initial parameter is the learnable threshold, and its initial value is determined based on the task-related density bias distribution in the training trade-off. This is a positive value mapping function used to convert the original learnable threshold parameters into positive threshold values, and .

6. The method for stroke complication image analysis based on prior-guided task decoupling graph learning according to claim 5, characterized in that, In S3, the specific content of task-specific contrastive learning is as follows: Representing the left and right homologous brain regions Input the hemorrhage transformation task projection head and the cerebral edema task projection head respectively to obtain task-specific embeddings. ; For two brain region pairs within the same batch, the similarity is calculated using task-specific embedding temperature-scaled cosine similarity. When two brain region pairs correspond to the same left-right homologous brain region pair index and are both in a biased state or both in a normal state, the soft weight of the positive sample is determined based on the soft probability product of their task bias states. Specifically, for two brain region pairs within the same batch... a and b The cosine similarity of the task-specific embedding is divided by the contrast temperature. , obtain embedding similarity When the sample a and b When corresponding left and right homologous brain region pairs are indexed and both are in a deviated state or both are in a normal state, based on the sample a and b Task deviation state soft probability and The product determines the soft weights of positive samples. : ； in, This is an indicator function; it takes the value 1 when the condition within the square brackets is true, and 0 otherwise. This is a pair indexing function used to return the left and right homologous brain region pair indexes corresponding to the sample, where the two sample indices correspond to two brain region pairs within the batch, representing the sample. The first term in the table represents the soft joint weight when both samples are in a biased state, the second term represents the soft joint weight when both samples are in a normal state, and the sum of the two terms in the square brackets represents the soft weight of the positive sample when both samples are in the same state. Task τ The contrast loss is expressed as: ； in, A valid set of anchor points consists of anchor points with at least one positive sample soft weight greater than zero.

7. The method for stroke complication image analysis based on prior-guided task decoupling graph learning according to claim 6, characterized in that, In S4, the feature dimension gate vector The final representation of brain region nodes They are represented as follows: ， ； in, Indicates feature splicing, and These are brain region nodes. Topological channel representation and symmetric channel representation. and These are the learnable gated weight matrix and bias vector, respectively. Represents element-wise product. Indicates to Element-by-element complement.

8. The method for stroke complication image analysis based on prior-guided task decoupling graph learning according to claim 7, characterized in that, In S5, the query vectors for the hemorrhage transformation task and the cerebral edema task are initialized to be mutually orthogonal. For the task τ Node attention weights and the corresponding graph-level task representation They are represented as follows: ， ； in, For task query vectors, Dimensions are represented by nodes; In S6, the hemorrhage transformation interaction task representation and the cerebral edema interaction task representation are uniformly represented as follows: ； in, and These represent the current task and another task, respectively. This represents the task after the interaction. and This represents the graph-level task representation before interaction. For bottleneck mapping networks, To stop the gradient operation, This is the directional scaling factor. Injecting strength into learnable content This indicates the sum of residuals.

9. A stroke complication image analysis system based on prior-guided task decoupling graph learning, characterized in that, include: The brain region map data acquisition module is used to acquire brain region maps based on preoperative non-contrast computed tomography images; The dual-channel encoding module is used to learn the topological representation of brain regions along anatomical adjacent edges through the topological channel, and to generate brain region pair representations and symmetric channel representations based on left and right homologous brain regions through the symmetric channel. The task-specific bias generation module is used to generate soft probabilities of hemorrhage conversion task bias state and cerebral edema task bias state based on the difference in the proportion of high-density voxels and the difference in the proportion of low-density voxels in the left and right homologous brain regions, respectively. The task decoupling and contrastive learning module is used to perform bias-guided contrastive learning in the independent embedding spaces of hemorrhage transformation task and cerebral edema task; A gated fusion module is used to fuse the brain region topological representation and the symmetric channel representation; The task-specific aggregation module is used to generate graph-level task representations of hemorrhage transformation and cerebral edema using the hemorrhage transformation task query vector and the cerebral edema task query vector, respectively. The Stop Gradient Cross-Task Interaction module is used to inject the graph-level task representation of one task into another task branch in the form of residuals after stopping gradients and bottleneck mapping. The model training module is used to update the model parameters based on the hemorrhage conversion task label, the cerebral edema task label, and the bias-guided contrastive learning loss, so as to obtain the trained image analysis model. The image analysis module is used to input the target brain region map of the object to be analyzed into the trained image analysis model and output the hemorrhage conversion prediction probability and the cerebral edema prediction probability.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the stroke complication image analysis method based on prior guided task decoupling graph learning as described in any one of claims 1-8.

Images