A method and device for predicting the progression of myopathy based on spatiotemporal topology networks

Abstract

This invention provides a method and apparatus for predicting the progression of myopathy based on spatiotemporal topological networks. First, muscle biopsy sections from multiple time points of the patient are segmented and quantized to construct a static pathological topological network with individual muscle fibers as nodes and spatial adjacency relationships as edges. Nodes contain multidimensional features such as subtype, morphology, texture, and cell nucleus, and are arranged chronologically to form spatiotemporal graph data. This data is input into a pre-trained spatiotemporal graph neural network. Its spatial encoding layer aggregates neighborhood information through a graph convolutional network to capture spatial interactions, its temporal encoding layer learns temporal evolution patterns through a recurrent neural network, and its decoding layer predicts the network state at the next time point. Through iterative prediction, pathological network sequences at multiple future time points are generated. Key indicators are then extracted to plot quantitative prediction curves and generate a visualized pathological heatmap, achieving a leap from static pathology to dynamic progression, providing an objective and accurate tool for clinical prognostic assessment of myopathy.

Description

Technical Field

[0001] This invention relates to the field of medical data processing technology, and in particular to a method and apparatus for predicting the progression of myopathy based on spatiotemporal topology networks. Background Technology

[0002] Chronic progressive neuromuscular diseases, such as muscular dystrophy and autoimmune myopathy, typically exhibit highly heterogeneous pathological progression, with significant differences in disease progression rate, affected patterns, and prognostic outcomes among patients. Currently, in clinical practice, prognostic assessment and progression prediction for these diseases primarily rely on two types of information: first, longitudinal follow-up based on clinical functional scales, such as muscle strength scores and motor function assessments; and second, histopathological analysis based on muscle biopsies.

[0003] Muscle biopsy is considered the gold standard for diagnosis and assessment. Pathologists can identify key pathological changes such as muscle fiber atrophy, hypertrophy, necrosis, regeneration, and connective tissue hyperplasia by observing the sections under a microscope. However, this method is essentially a static morphological description of the disease at a single point in time. While it provides a snapshot of the diagnostic criteria and the current severity of the lesion, it cannot directly reveal the dynamic evolution of the disease over time. For example, it cannot answer key prognostic questions such as "What is the rate of muscle fiber atrophy in the patient?" or "How will the lesion spread in the future?" Physicians can only rely on personal experience to make rough and subjective judgments about the disease's trajectory, lacking objective and quantitative predictive tools.

[0004] In recent years, with the development of digital pathology and artificial intelligence technologies, several methods for quantitative analysis of pathological images have emerged. These methods typically employ computer vision or machine learning algorithms to segment and morphologically measure muscle fibers in a single biopsy slide, providing more precise quantitative indicators than manual observation. However, these existing technologies still have significant limitations: First, their analytical dimensions are limited to single, isolated pathological slides, failing to integrate longitudinal data from multiple time points of the same patient, thus making it impossible to construct time-series models of disease evolution. Second, their analytical focus is mostly on individual morphological parameters of muscle fibers, failing to model the tissue microenvironment as an interconnected network system, thereby neglecting the complex spatial interactions and neighborhood effects between muscle fibers and between muscle fibers and the stroma. These intercellular interactions are considered important mechanisms driving the spatial propagation and progression of disease.

[0005] Therefore, the existing technology system has a technological gap in achieving accurate and prospective prognostic prediction of progressive myopathy: that is, there is a lack of a technical means that can simultaneously integrate longitudinal information from multiple time points with the spatial topological relationship of pathological tissues, thereby moving from static description to dynamic evolution simulation. Summary of the Invention

[0006] In view of this, embodiments of the present invention provide a method and apparatus for predicting the progression of myopathy based on spatiotemporal topology networks, in order to eliminate or improve one or more defects existing in the prior art, and solve the problem that the current prognostic assessment of myopathy relies on static, single-time-point pathological analysis and cannot make dynamic and quantitative predictions of the future progression of the disease.

[0007] One aspect of the present invention provides a method for predicting the progression of myopathy based on spatiotemporal topology networks, the method comprising the following steps:

[0008] Muscle biopsy sections from a predetermined number of time points in the patient's past are obtained. After staining according to a preset scheme, single muscle fibers, cell nuclei, connective tissue, and microvessels in the muscle biopsy sections are identified based on a pre-trained image segmentation model. A node set is constructed using the single muscle fibers as nodes. Each node records subtype attributes, morphological parameter vectors, internal texture parameter vectors, and cell nucleus attribute vectors as node attributes. Adjacency relationships between nodes are constructed based on a preset spatial adjacency graph construction algorithm to obtain a static pathological topology network. The static case topology network corresponding to each time point is then constructed into time series graph data.

[0009] The time-series graph data is input into a pre-trained spatiotemporal graph neural network containing a spatial dependency encoding layer, a temporal evolution encoding layer, and a prediction decoding layer. The spatial dependency encoding layer aggregates information from each node and its neighboring nodes in the static pathological topology network using a graph convolutional network to extract feature representation vectors. The temporal evolution encoding layer performs graph pooling on the feature representation vectors to obtain a full graph representation vector, and extracts and encodes the hidden states of the entire pathological history from the global graph representation vector sequence using a recurrent neural network. This encoded state is then input into the prediction decoding layer to predict the static case topology network for the next time point based on a multilayer vector perceptron. The predicted static case topology network is then added to the tail of the time-series graph data and the head of the queue is removed. It is then re-input into the spatiotemporal graph neural network for continuous cyclic prediction to obtain the predicted static case topology network for the patient at multiple future time points.

[0010] Based on the predicted static case topology network of the patient at multiple future time points, a prediction report is generated. The prediction report includes: quantitative prediction curves for clinically relevant indicators, and a visualized pathological heatmap generated from the node attributes predicted for the target future time points.

[0011] In some embodiments, the pre-training step of the image segmentation model includes:

[0012] Obtain a first training sample set containing multiple samples, each sample containing a muscle biopsy digital slice sample stained with the preset scheme, and labeling the location of single muscle fibers, cell nuclei, connective tissue and microvessels as tags;

[0013] The initial image segmentation network is trained using the first training sample set. The initial image segmentation network takes the muscle biopsy digital slice sample as input and outputs the predicted location and predicted classification of single muscle fibers, cell nuclei, connective tissue and microvessels in the image. The initial image segmentation network is updated by constructing a loss based on the deviation between the predicted location, the predicted classification and the label to obtain the image segmentation network.

[0014] The initial image segmentation network includes a semantic segmentation network or an instance segmentation network based on a convolutional neural network architecture. The semantic segmentation network is a U-Net network or a DeepLab network, and the instance segmentation network is a Mask R-CNN network.

[0015] In some embodiments, the subtype attribute is a myofiber subtype classification based on metabolic enzyme staining results, including slow contraction oxidative type, fast contraction oxidative-glycolytic type and fast contraction glycolytic type, and the subtype attribute is represented in the form of one-heat coding;

[0016] The morphological parameter vector includes muscle fiber cross-sectional area, circularity, polygonality, and perimeter;

[0017] The internal texture parameter vector includes: a contrast texture that reflects grayscale changes and an energy texture that reflects the degree of texture uniformity.

[0018] The nuclear attribute vector includes the proportion of central nuclei, the total number of nuclei, and the number of peripheral nuclei;

[0019] The preset spatial adjacency graph construction algorithm is a Delaunay triangulation algorithm, a Voronoi graph algorithm, or a radius-based neighbor graph algorithm.

[0020] In some embodiments, the pre-training step of the spatiotemporal graph neural network includes:

[0021] Muscle biopsy sample slices from multiple patients at multiple time points are obtained. After staining according to the preset scheme, single muscle fibers, cell nuclei, connective tissue, and microvessels in the muscle biopsy sample slices are identified based on the image segmentation model. A node set is constructed using the single muscle fibers as nodes. Each node records subtype attributes, morphological parameter vectors, internal texture parameter vectors, and cell nucleus attribute vectors as node attributes. Adjacency relationships between nodes are constructed based on a preset spatial adjacency graph construction algorithm to obtain static pathological topology network samples at each time point, which are then used to construct time series graph data samples for each patient.

[0022] Continuous sampling is performed on the time series data samples to construct a second training sample set, each sample containing a second predetermined number of static pathological topology network samples at consecutive time points; the second predetermined number is the first predetermined number plus 1.

[0023] The second training sample set is used to train an initial spatiotemporal graph neural network containing an initial spatial dependency encoding layer, an initial temporal evolution law encoding layer, and an initial prediction decoding layer. The first set number of static pathological topology network samples in each sample are used as input. The initial spatial dependency encoding layer aggregates information of nodes and their neighboring nodes in each static pathological topology network sample based on a graph convolutional network to extract sample feature representation vectors. The initial temporal evolution law encoding layer performs graph pooling on the sample feature representation vectors to obtain a full graph representation vector of the sample, and extracts the hidden state of the sample encoding the entire pathological history from the global graph representation vector sequence of the sample based on a recurrent neural network. This hidden state is input to the initial prediction decoding layer to predict the last predicted static case topology network in each sample based on a multilayer vector perceptron. The parameters of the initial spatiotemporal graph neural network are updated based on the mean square error of node features between the predicted value and the true value to obtain the spatiotemporal graph neural network.

[0024] In some embodiments, a non-linear activation function layer is provided after the graph convolutional network, and the recurrent neural network adopts a long short-term memory network.

[0025] The Adam optimizer is used when updating the parameters of the initial spatiotemporal graph neural network;

[0026] The formula for calculating the mean square error is:

[0027] ;

[0028] Where L represents the mean square error. express The number of nodes in the static pathological topology network sample at the specified time point. express The predicted node attribute value of the i-th node in the static pathological topology network sample at the specified time point. express The true value of the node attribute of the i-th node in the static pathological topology network sample at the specified time point.

[0029] In some embodiments, the quantitative prediction curve for clinically significant indicators is a prediction curve for the average cross-sectional area of ​​muscle fibers, calculated as follows:

[0030] ;

[0031] in, Indicates the reference time point The expected average cross-sectional area of ​​the muscle fiber at the k-th time point thereafter. Indicates a point in time The predicted total number of muscle fiber nodes. Indicates a point in time The predicted node attribute value for the i-th node in the static pathological topology network sample. This is a projection operator used to extract the cross-sectional area of ​​muscle fibers from the predicted node attribute values.

[0032] In some embodiments, generating a visualized pathological heatmap for node attributes predicted for a target future time point includes: extracting the cross-sectional area of ​​the corresponding muscle fibers of each node from the predicted static case topology network for a specified future time point, measuring the degree of atrophy based on a preset evaluation standard, and generating a visualized pathological heatmap on the degree of muscle atrophy.

[0033] On the other hand, the present invention also provides a myopathy progression prediction device based on spatiotemporal topology network, including a processor, a memory, and a computer program or instructions stored in the memory, wherein the processor is used to execute the computer program or instructions, and when the computer program or instructions are executed, the device implements the steps of the above method.

[0034] On the other hand, the present invention also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0035] On the other hand, the present invention also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method.

[0036] The method and apparatus for predicting myopathy progression based on spatiotemporal topological networks described in this invention construct a static pathological topological network by segmenting and quantifying muscle biopsy sections from multiple time points longitudinally of the patient. This network is organized into nodes (single muscle fibers) and edges (spatial adjacency relationships between fibers). Each node contains multidimensional feature vectors representing subtype, morphology, texture, and cell nuclear attributes. The networks from each time point are then assembled sequentially to form spatiotemporal graph data. Next, a spatiotemporal graph neural network (ST-GNN) is used as the core analysis engine. Its spatial dependency encoding layer aggregates node and neighbor information through a graph convolutional network to capture the spatial interactions of the tissue microenvironment, while the temporal evolution encoding layer... A recurrent neural network is used to model the temporal full-map representation, learn the historical evolution pattern of pathological states, and output the prediction network for the next time point through a prediction decoding layer. Based on this, a sequence of predicted pathological networks for multiple future time points is generated through iterative forward extrapolation of a spatiotemporal graph neural network based on a single baseline pathological network. Finally, the numerical change trajectory of key pathological indicators is extracted from this prediction sequence and a quantitative prediction curve is plotted. At the same time, a future pathological heatmap that can intuitively display the spatial distribution and severity of lesions is generated based on node attributes. This completes the transition from static pathology to dynamic progression prediction, providing clinicians with an objective, accurate, and visualized tool for myopathy prognosis assessment.

[0037] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the description, or may be learned by practice of the invention. The objects and other advantages of the invention can be realized and obtained by means of the structures specifically pointed out in the specification and drawings.

[0038] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0039] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, are not intended to limit the scope of the invention. In the drawings:

[0040] Figure 1 This is a flowchart illustrating the method for predicting the progression of myopathy based on spatiotemporal topology networks according to an embodiment of the present invention.

[0041] Figure 2 This is a logical architecture diagram of the myopathy progression prediction system based on spatiotemporal topology network according to an embodiment of the present invention, which performs myopathy progression prediction. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0043] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0044] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0045] Current technologies for assessing the progression of myopathy mainly rely on pathological analysis of muscle biopsies at a single time point. While this method can provide diagnostic evidence and static lesion descriptions, it cannot capture the dynamic evolution of the disease over time. At the same time, existing quantitative image analysis methods are mostly limited to isolated morphological measurements of single slices. They fail to integrate data from multiple time points to construct temporal evolution models, nor do they model pathological tissues as a network system with spatial interactions. Therefore, they cannot quantitatively predict the future direction and speed of disease development, resulting in a lack of objective and accurate dynamic prediction tools for clinical prognosis.

[0046] One aspect of the present invention provides a method for predicting the progression of myopathy based on spatiotemporal topological networks, such as... Figure 1 As shown, the method includes the following steps S101~S103:

[0047] Step S101: Obtain muscle biopsy sections from the patient at a predetermined number of time points in the past. After staining with a preset scheme, identify single muscle fibers, cell nuclei, connective tissue, and microvessels in the muscle biopsy sections based on a pre-trained image segmentation model. Construct a node set using single muscle fibers as nodes. Each node records subtype attributes, morphological parameter vectors, internal texture parameter vectors, and cell nucleus attribute vectors as node attributes. Construct the adjacency relationships between nodes based on a preset spatial adjacency graph construction algorithm to obtain a static pathological topology network. Construct a time series graph data from the static case topology network corresponding to each time point.

[0048] Step S102: Input the time series graph data into a pre-trained spatiotemporal graph neural network containing a spatial dependency encoding layer, a temporal evolution pattern encoding layer, and a prediction decoding layer; the spatial dependency encoding layer aggregates information from nodes and their neighboring nodes in each static pathological topology network based on a graph convolutional network to extract feature representation vectors; the temporal evolution pattern encoding layer performs graph pooling on the feature representation vectors to obtain a full graph representation vector, and extracts and encodes the hidden states of the entire pathological history from the global graph representation vector sequence based on a recurrent neural network, which is then input into the prediction decoding layer to predict the static case topology network for the next time point based on a multilayer vector perceptron; the predicted static case topology network is added to the tail of the time series graph data and the head of the queue is removed, and then re-input into the spatiotemporal graph neural network for continuous cyclic prediction to obtain the predicted static case topology network for the patient at multiple future time points.

[0049] Step S103: Generate a prediction report based on the predicted static case topology network of the patient at multiple future time points. The prediction report includes: quantitative prediction curves for clinically important indicators, and a visualized pathological heatmap generated from the node attributes predicted for the target future time points.

[0050] Step S101 transforms traditional pathological images into a machine-learnable graph data structure that can characterize the spatiotemporal features of pathological tissue. First, a pre-trained image segmentation model segments the stained slices, identifying biological entities containing single muscle fibers, cell nuclei, connective tissue, and microvessels. This step achieves deconstruction from image to discrete objects. Next, each muscle fiber is used as a node, and node attributes are added, such as attribute vectors for single fiber subtype, morphology, texture, and cell nucleus feature parameters. These attribute vectors provide a comprehensive digital quantitative description of its pathological state, replacing the qualitative terms in traditional pathological descriptions. Then, a pre-defined spatial adjacency graph construction algorithm builds spatial adjacency relationships between nodes. This is based on the biological prior knowledge that adjacent cells in the tissue microenvironment interact, abstracting physical proximity relationships as edges in a graph structure. Thus, a slice at a single time point is transformed into a static pathological topological network, where nodes are muscle fibers carrying states, and edges represent potential interactions or influence paths between cells. Finally, multiple static networks from different time points of the same patient are arranged chronologically to form a time-series graph data. It encodes the historical information of the disease into a complete structured representation that includes both the micro-state at each moment and the spatial correlation and temporal order between states, which can be processed by deep learning models.

[0051] The pre-training steps of the image segmentation model used in this process include S1011~S1012:

[0052] Step S1011: Obtain a first training sample set containing multiple samples. Each sample contains a digital slice of muscle biopsy stained with a preset scheme, and the location of a single muscle fiber, cell nucleus, connective tissue and microvessel is marked as a label.

[0053] Step S1012: The initial image segmentation network is trained using the first training sample set. The initial image segmentation network takes the muscle biopsy digital slice sample as input and outputs the predicted location and predicted classification of single muscle fibers, cell nuclei, connective tissue and microvessels in the image. Based on the deviation between the predicted location, predicted classification and label, a loss is constructed to update the parameters of the initial image segmentation network to obtain the image segmentation network.

[0054] The initial image segmentation network includes a semantic segmentation network or an instance segmentation network based on a convolutional neural network architecture. The semantic segmentation network is either a U-Net network or a DeepLab network, and the instance segmentation network is a Mask R-CNN network.

[0055] In steps S1011-S1012, the sample can be stained with hematoxylin and eosin (H&E) to provide basic cell morphology and tissue structure for segmenting muscle fibers and cell nuclei and calculating morphological parameters. In other embodiments, Masson's trichrome staining can be used to distinguish muscle fibers from connective tissues such as collagen fibers with high contrast, facilitating precise segmentation and quantification of the degree of fibrosis. Alternatively, immunofluorescence staining can be used, employing specific antibodies to label different structural proteins, enabling multi-channel, highly specific, simultaneous segmentation, which is particularly suitable for distinguishing closely adjacent structures in complex microenvironments.

[0056] In some embodiments, the subtype attribute is based on the results of metabolic enzyme staining to classify myofibril subtypes, which can be divided into slow-contraction oxidative, fast-contraction oxidative-glycolytic, and fast-contraction glycolytic types. These three categories mainly correspond to type I, type IIA, and type IIB myofibrils, respectively. The subtype attribute is represented in the form of one-hot encoding. Specifically, using myosin ATPase staining, under different pH conditions, slow-contraction oxidative type I, fast-contraction oxidative-glycolytic type IIA, and fast-contraction glycolytic type IIB myofibrils will exhibit different patterns of deep staining, light staining, or no staining. Pathologists or trained algorithms interpret and classify each myofibril in the field of view according to the staining depth and pattern. Subsequently, when constructing the pathological topology network, each node representing a myofibril is assigned a three-dimensional one-hot encoding vector to characterize its subtype. For example, type I is encoded as [1, 0, 0], type IIA as [0, 1, 0], and type IIB as [0, 0, 1]. This encoding method transforms classification information into discrete, mutually exclusive binary vectors, which can be used as input features for subsequent neural network models.

[0057] The morphological parameter vector includes the cross-sectional area, circularity, polygonality, and perimeter of the muscle fiber. Specifically, after processing the stained muscle biopsy slices using an image segmentation model, a precise binary mask is generated for each individual muscle fiber. This mask defines the precise pixel region of the muscle fiber in the image. The cross-sectional area is directly obtained by multiplying the total number of pixels covered by the mask by the actual physical area of ​​each pixel; the perimeter is obtained by calculating the connection length of all pixels on the mask contour; further, the circularity is calculated using the following formula:

[0058] ;

[0059] Where C represents circularity, A represents cross-sectional area, and P represents perimeter.

[0060] The polygonal degree is calculated using the following formula:

[0061] ;

[0062] Where Poly represents polygonity and A represents cross-sectional area. This represents the convex hull area of ​​the mask corresponding to a single muscle fiber.

[0063] Furthermore, the morphological parameter vector may also include parameters such as aspect ratio, eccentricity, density, profile index, internal pore area or number, fiber spacing, and aggregation coefficient.

[0064] The internal texture parameter vector includes: a contrast texture reflecting grayscale changes and an energy texture reflecting texture uniformity. Specifically, the grayscale image within a single muscle fiber mask is extracted, and the grayscale co-occurrence matrix is ​​calculated; the contrast texture is calculated using the following formula: ;Calculate the energy texture, the formula is: ;

[0065] Where i and j represent the grayscale values ​​of pixel pairs with a preset spatial relationship. This represents the joint probability of a pixel pair with gray values ​​i and j appearing in the mask region.

[0066] Furthermore, the nuclear attribute vector includes the proportion of central nuclei, the total number of nuclei, and the number of peripheral nuclei. Depending on specific needs, nuclear area, nuclear perimeter, nuclear density, and nuclear membrane proximity can also be annotated.

[0067] The preset spatial adjacency graph construction algorithm is the Delaunay triangulation algorithm, the Voronoi diagram algorithm, or a radius-based nearest neighbor graph algorithm. During edge construction, based on the mask of a single muscle fiber identified by the image segmentation model, the centroid coordinates of the single muscle fiber in the image are determined, and edges are constructed based on these coordinates. For example, using the precise mask of each muscle fiber instance output by the trained image segmentation model, the geometric centroid coordinates of each muscle fiber on the two-dimensional image plane are obtained by calculating the mean of all pixel coordinates of the mask. This is used as the spatial position of the fiber node in the graph. Delaunay triangulation connects all centroids into triangles, ensuring no point lies inside the circumcircle of any triangle. This generates a neighborhood network that maximizes the minimum angle and avoids elongated triangles, naturally simulating nearest neighbor interactions. The Voronoi diagram algorithm divides the plane into regions based on centroids. If the Voronoi regions corresponding to the centroids of two regions share an edge, a connection is established between the two nodes, reflecting the adjacency of spheres of influence in space. The radius-based neighborhood graph algorithm requires a preset distance threshold R, calculating the Euclidean distance between all pairs of centroids. If the distance between two points is less than R, an edge is established between them. This method intuitively simulates local interactions within a fixed physical area. Ultimately, regardless of the algorithm used, the system outputs an adjacency matrix. This matrix clearly defines the edges between muscle fibers in the static pathological topology network, providing a topological foundation for subsequent spatial information aggregation in the graph neural network.

[0068] Step S102 utilizes a spatiotemporal graph neural network to automatically learn the potential dynamic patterns of disease progression from the time series graph data constructed in step S101, and performs iterative prospective inferences based on these patterns.

[0069] The model works through three core layers, corresponding to its encoding-decoding process. First, the spatial dependency encoding layer operates on each static pathological topology network. Its principle mimics the process of information propagation on a graph: the features of each muscle fiber node are aggregated and updated multiple times with the features of its neighboring nodes. This allows the model to learn the local spatial interaction patterns within the pathological microenvironment, such as how a healthy muscle fiber is affected by surrounding atrophied fibers, or how inflammatory foci spread spatially.

[0070] Secondly, the temporal evolution coding layer processes sequences composed of spatial snapshots at different time points. The principle is to utilize internal memory units, taking the full-map state summary vector obtained after GCN encoding and pooling at each time point as input for a time step. LSTM learns the temporal dependencies between these state summaries through its gating mechanism, thereby capturing dynamic evolution patterns across time, such as how the rate of myofibril atrophy changes over time, or how specific pathological events predict subsequent necrosis.

[0071] Finally, the prediction decoding layer receives the hidden states from the LSTM output, which encode complete historical information, and maps them to a prediction of the overall state at the next future time point. In the application phase, for a new patient, the system inputs their baseline network into the pre-trained model to obtain the prediction network for the next time point. Subsequently, an iterative autoregressive prediction principle is used: the output of this prediction is used as input again and fed into the model repeatedly, thus gradually generating a sequence of predicted pathological networks for multiple future time points.

[0072] In some embodiments, the pre-training step of the spatiotemporal graph neural network includes S1021~S1022:

[0073] Step S1021: Obtain muscle biopsy sample slices from multiple patients at multiple time points. After staining with a preset scheme, identify single muscle fibers, cell nuclei, connective tissue, and microvessels in the muscle biopsy sample slices based on an image segmentation model. Construct a node set using single muscle fibers as nodes. Each node records subtype attributes, morphological parameter vectors, internal texture parameter vectors, and cell nucleus attribute vectors as node attributes. Construct the adjacency relationship between nodes based on a preset spatial adjacency graph construction algorithm to obtain a static pathological topology network sample at each time point, and construct a time series graph data sample for each patient.

[0074] Step S1022: Continuously sample the time series data samples to construct a second training sample set, each sample containing a second set number of static pathological topology network samples at consecutive time points; the second set number is the first set number plus 1.

[0075] Step S1023: The initial spatiotemporal graph neural network, which includes an initial spatial dependency encoding layer, an initial temporal evolution law encoding layer, and an initial prediction decoding layer, is trained using the first set number of static pathological topology network samples in each sample as input. The initial spatial dependency encoding layer aggregates information from nodes and their neighboring nodes in each static pathological topology network sample based on a graph convolutional network to extract the sample feature representation vector. The initial temporal evolution law encoding layer performs graph pooling on the sample feature representation vector to obtain the full graph representation vector of the sample, and extracts the hidden state of the sample encoding the entire pathological history from the global graph representation vector sequence of the sample based on a recurrent neural network. This hidden state is input into the initial prediction decoding layer to predict the last predicted static case topology network in each sample based on a multilayer vector perceptron. The loss is constructed based on the mean square error of node features between the predicted value and the true value, and the parameters of the initial spatiotemporal graph neural network are updated to obtain the spatiotemporal graph neural network.

[0076] In some embodiments, a non-linear activation function layer is provided after the graph convolutional network, and a long short-term memory network is used for the recurrent neural network; the Adam optimizer is used when updating the parameters of the initial spatiotemporal graph neural network.

[0077] The formula for calculating the mean square error is:

[0078] ;

[0079] Where L represents the mean square error. express The number of nodes in a static pathological topology network sample at a given time point. express The predicted node attribute value of the i-th node in the static pathological topology network sample at a given time point. express The true value of the node attribute of the i-th node in the static pathological topology network sample at a given time point.

[0080] Step S103 performs information dimensionality reduction, refinement, and translation on the complex prediction data output from step S102, generating an intuitive report that aligns with clinical cognitive habits and can directly assist in decision-making. For the quantitative prediction curve, the principle is to extract specific indicator dimensions of clinical interest from the network node feature matrix at each prediction time point. By performing statistical calculations on the indicator values ​​of all nodes at that time point, the predicted value of that indicator at that future time can be obtained. Connecting the predicted values ​​at each time point forms a dynamic curve, thus summarizing the model's micro-predictions into macro-level, traceable trend indicators, such as the average CSA's downward trend curve over the next five years.

[0081] In some embodiments, the quantitative prediction curve for clinically significant indicators is a prediction curve for the average cross-sectional area of ​​muscle fibers, calculated as follows:

[0082] ;

[0083] in, Indicates the reference time point The expected average cross-sectional area of ​​the muscle fiber at the k-th time point. Indicates a point in time The predicted total number of muscle fiber nodes. Indicates a point in time The predicted node attribute value for the i-th node in a static pathological topology network sample. This is a projection operator used to extract the cross-sectional area of ​​muscle fibers from node attribute prediction values.

[0084] The principle behind visualized pathological heatmaps is to map abstract node attribute predictions back to the original image space coordinates. The system selects a target future time point and assigns a color intensity value to each node in the prediction network based on its specific attribute predictions; for example, a deeper red indicates more severe atrophy. Then, this color is filled into the outline of each muscle fiber at the corresponding anatomical location on the original biopsy slide, generating a pseudo-color image of the future pathology. This visualization method transforms numerical predictions into a visual form familiar to doctors, intuitively displaying the severity of future lesions and their spatial distribution in tissue, achieving a retrospective and presentation of prediction results from the data space to the visual space.

[0085] In some embodiments, generating a visualized pathological heatmap for node attributes predicted for a target future time point includes: extracting the cross-sectional area of ​​the corresponding muscle fibers of each node from the predicted static case topology network for a specified future time point, measuring the degree of atrophy based on a preset evaluation standard, and generating a visualized pathological heatmap on the degree of muscle atrophy.

[0086] On the other hand, the present invention also provides a myopathy progression prediction device based on spatiotemporal topology network, including a processor, a memory, and a computer program or instructions stored in the memory, wherein the processor is used to execute the computer program or instructions, and when the computer program or instructions are executed, the device implements the steps of the above method.

[0087] On the other hand, the present invention also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0088] On the other hand, the present invention also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method.

[0089] The present invention will now be described in conjunction with specific embodiments:

[0090] Example 1

[0091] This embodiment proposes a myopathy progression prediction system based on spatiotemporal topology networks, aiming to transform static pathological images from multiple time points into dynamic predictions of future disease progression through a modular collaborative workflow. The system mainly includes the following four core functional modules:

[0092] (1) Spatiotemporal data construction module

[0093] Function: This module is responsible for structuring and temporally sequencing multiple muscle biopsy slides from the same patient at different time points. It first constructs a static pathological topology network for each slide at each time point, with nodes representing muscle fibers and carrying rich quantitative attributes. Then, these static networks are organized chronologically to ultimately generate a time-series graph data for each patient containing their complete pathological evolution history.

[0094] (2) Spatiotemporal Evolution Law Learning Module

[0095] Function: The core of this module is a spatiotemporal graphical neural network (ST-GNN). It is responsible for learning and solidifying the deep-seated patterns of the evolution of muscle tissue pathological states over time from a large amount of longitudinal follow-up data. This model can simultaneously understand: Spatial dependence: how the state of a muscle fiber at any given moment is influenced by its surrounding neighbors. Temporal dependence: how the current state of the entire tissue microenvironment determines its state at future moments.

[0096] (3) Future disease course prediction module

[0097] Function: This module is the core application of this invention. It receives a baseline (single time point) pathological topology network of a new patient as input and uses a pre-trained spatiotemporal evolution learning module to perform iterative forward extrapolation. By predicting the topology network state at multiple future time points step by step, this module can generate a complete and personalized future pathological evolution trajectory.

[0098] (4) Predictive Report Generation Module

[0099] Function: This module is responsible for transforming the high-dimensional and complex prediction results output by the "Future Disease Course Prediction Module" into intuitive and clinically meaningful reports. It extracts and visualizes the future change curves of key pathological indicators (such as the rate of muscle fiber atrophy and the degree of fibrosis aggravation), and can further analyze and identify the dynamic biomarkers with the most predictive value for disease progression, providing strong data support for accurate prognostic assessment and intervention decisions.

[0100] Correspondingly, this embodiment adopts the following technical solution: a method for predicting the progression of myopathy based on spatiotemporal topology networks, the general implementation steps of which are as follows:

[0101] Step S11: Construction of spatiotemporal pathological map data.

[0102] First, a patient cohort with longitudinal follow-up data was collected, and muscle biopsy sections were obtained at baseline year, year 2, and year 5. For each time point, an automated image analysis method was used to transform the sections into a pathological topology network. In this network, each node It is a muscle fiber, and its feature vector It contains rich quantitative attributes, such as CSA, fiber type, number of central nuclei, and local fibrosis pressure. The edges of the network represent the spatial adjacency relationships between muscle fibers. Through this step, the pathological data of each patient is transformed from a series of isolated images into an ordered time series graph. This laid the foundation for subsequent dynamic modeling.

[0103] Step S12: Training the Spatiotemporal Graph Neural Network (ST-GNN) model.

[0104] The core of this embodiment is an ST-GNN model. It combines two powerful deep learning architectures:

[0105] 1. Graph Convolutional Network (GCN): At each time point t, the GCN is responsible for processing the topological network Gt. It updates the representation of a node by aggregating information about a node and its neighboring nodes, thereby effectively capturing the spatial dependencies in the pathological microenvironment.

[0106] 2. Recurrent Neural Networks (RNNs) or their variants (such as LSTM / GRU): RNNs handle the time dimension. They take snapshots of the entire graph extracted by the GCN at each time point as input to a sequence. The internal memory units of the RNN can learn the temporal evolution between these snapshots; for example, a continuous increase in the proportion of the central kernel often foreshadows increased myofibrillary necrosis 1-2 years later. Through end-to-end training on a longitudinal data queue, the ST-GNN model eventually learns a powerful mapping function.f It can predict the graph state at the next time point based on the graph state at one or more past time points, i.e. .

[0107] Step S13: Predict disease progression in new patients.

[0108] For a new patient with only baseline biopsy data, this system first converts their baseline slices into a baseline topology network. Then, When fed into a pre-trained ST-GNN model, the model will output the first future time point, such as the predictive topology network one year from now. This predicted network It not only includes the network structure, but more importantly, it predicts all the multidimensional properties of each muscle fiber node one year later.

[0109] Next, the system can perform iterative predictions: taking the newly generated... As new input, it is fed back into the model to predict the state two years later. By repeating this process, a complete trajectory of the disease's future evolution can be obtained. Finally, from this series of predictive topological networks, the future change curves of any pathological indicators of concern to clinicians can be easily extracted. For example, it can be predicted that the patient's average myofibrous CSA will decrease by 30% within the next 5 years, or that the connective tissue area will exceed the critical threshold of 20% in 3 years, and an intuitive predictive report can be generated.

[0110] Example 2

[0111] This embodiment uses a cohort containing longitudinal follow-up data of patients with muscular dystrophy to illustrate in detail the complete implementation steps of the method of the present invention. This cohort provides muscle biopsy samples for each patient at multiple time points, including, for example, baseline. , second year 5th year . Reference Figure 2 This embodiment includes the following steps:

[0112] Step S21: Construction of spatiotemporal pathological map data

[0113] The purpose of this step is to transform the raw, multi-timepoint pathological image sequences into a structured, machine-learnable time series graph data.

[0114] Construction of static pathological topology network ( ): For the patient at any point in time From the muscle biopsy sections, the system performs the following operations to construct a static pathological topology network. :

[0115] Image segmentation: Using a pre-trained deep learning segmentation model (such as Mask R-CNN), instance masks of single muscle fibers, cell nuclei, connective tissue, and microvessels are accurately segmented from digital slices.

[0116] Node definition ( ): The slice at that time point Root muscle fiber instances are defined as the set of nodes in a graph. .

[0117] Node attribute quantization ( ): For each node Calculate one 3D eigenvectors This vector comprehensively quantifies the pathological state of muscle fibers, and its composition is as follows: in:

[0118] Subtype attribute, represented by a one-hot encoded vector for myofiber subtype.

[0119] : A vector of morphological parameters, containing normalized cross-sectional area (CSA), circularity, polygonity, and other geometric measurements.

[0120] : Internal texture parameter vector, containing values ​​such as contrast, energy, and homogeneity extracted from gray-level co-occurrence matrix (GLCM), used to quantify internal pathological changes such as insect-eaten appearance.

[0121] : Cell nuclear attribute vector, including quantitative indicators such as the proportion of central nuclei and the total number of nuclei.

[0122] Network topology construction ( Based on the geometric centroids of all nodes, the Delaunay triangulation algorithm is used to construct the adjacency relationships between nodes, forming an edge set. Its corresponding adjacency matrix is ​​denoted as... .

[0123] Assembly of time series plots: The static plots constructed for each time point are arranged in chronological order to generate a time series plot for each patient. This constitutes the complete input for subsequent model training and prediction.

[0124] Step S22: Construction and training of the spatiotemporal graphical neural network (ST-GNN) model

[0125] This embodiment employs an encoder-decoder architecture that combines a graph convolutional network (GCN) and a long short-term memory network (LSTM).

[0126] Spatial dependency coding (GCN layer): at each time point The model uses a A layer of GCN is used to learn the spatial interactions between nodes. For the first layer... The updates to node features at each layer follow these propagation rules:

[0127] ;

[0128] in, It is the first The feature matrix of all nodes in the layer, and . It is the learnable weight matrix of this layer. It is a symmetric normalized adjacency matrix with added self-loops, where, yes The degree matrix. It is a non-linear activation function (such as ReLU). After... After layer propagation, each node Obtain an embedded representation that includes information about itself and its neighborhood. .

[0129] Time evolution pattern encoding (LSTM layer): Encoding each time point All node embeddings Graph pooling (e.g., averaging) yields a representation vector of the entire graph. Then, represent this graph as a sequence of vectors. The input is fed into an LSTM network to learn time dependencies. The LSTM update mechanism is as follows:

[0130] ;

[0131] ;

[0132] ;

[0133] ;

[0134] ;

[0135] in, These are the input gate, forget gate, and output gate, respectively. In cellular state, In hidden state, and These are learnable parameters.

[0136] Predictive decoder (MLP layer): The hidden state output by the LSTM at the last observation time point. As an encoding of the entire pathological history, it is input into a multilayer perceptron (MLP) to predict the next time point. Full graph representation vector Then, the predicted feature matrix for each node at that future time point is decoded through a reverse process (such as broadcasting or unpooling). .

[0137] Model training: Supervised learning is performed using longitudinal queue data. For a training sample... The goal of the model is to predict Loss function Defined as the mean squared error (MSE) between the predicted node features and the true features:

[0138] ;

[0139] in, It is a predictive feature. These are the true features. The loss function is minimized using optimizers such as Adam through the backpropagation algorithm.

[0140] Step S23: Iterative forward prediction for new patients

[0141] For a dataset with only baseline biopsy data For new patients, the system performs the following prediction process: [The system then analyzes the patient's baseline topology network.] As the initial input to the model, the model predicts the state of the pathological topology network at the first future time point, i.e., one year later. Note the topology here. It can be assumed to be the same as the baseline, or generated through an independent graph structure prediction model, but the core is to predict the node feature matrix. .

[0142] The prediction results generated in the previous step As new input, it is fed back into the ST-GNN model to predict the state at the second future time point, i.e., two years later. This process is automatically and repeatedly executed until the preset prediction period is reached, such as 5 years later.

[0143] Step S24: Predictive Report Generation and Insight Discovery

[0144] The system will generate the future topology network sequence in step S23. Transform it into a clinically friendly predictive report.

[0145] Quantitative prediction curve generation: For any clinically relevant indicator, such as average CSA, the system extracts values ​​for the corresponding dimensions from the predicted feature matrix sequence, calculates its expected value at each future time point, and finally plots a prediction curve of the indicator over time. For example, the prediction curve for average CSA. The calculation is as follows:

[0146] ,

[0147] in, Indicates the reference time point The expected average cross-sectional area of ​​the muscle fiber at the k-th time point. Indicates a point in time The predicted total number of muscle fiber nodes. Indicates a point in time The predicted node attribute value for the i-th node in a static pathological topology network sample. This is a projection operator used to extract the cross-sectional area of ​​muscle fibers from node attribute prediction values.

[0148] Future pathology heatmap visualization: By selecting a key future time point, such as year 5, the system can generate a pseudo-color image of the pathology. The color of each muscle fiber in the image is determined by its corresponding predicted attribute value. For example, it is colored according to the predicted CSA value, with red representing severe atrophy and blue representing normal, thus visually showing the area where the lesion is most severe.

[0149] Through the detailed implementation steps described above, this invention constructs a complete technical closed loop from data processing, model building, forward-looking prediction to result interpretation, achieving accurate, dynamic, and quantitative prediction of the course of myopathy.

[0150] Corresponding to the above method, the present invention also provides an apparatus / system including a computer device, the computer device including a processor and a memory, the memory storing computer instructions, the processor executing the computer instructions stored in the memory, and when the computer instructions are executed by the processor, the apparatus / system performs the steps of the method as described above.

[0151] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned edge computing server deployment method. The computer-readable storage medium can be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, floppy disks, hard disks, removable storage disks, CD-ROMs, or any other form of storage medium known in the art.

[0152] In summary, the method and apparatus for predicting myopathy progression based on spatiotemporal topological networks described in this invention constructs a static pathological topological network by segmenting and quantifying muscle biopsy slices from multiple time points longitudinally of the patient. This network consists of individual muscle fibers as nodes and spatial adjacency relationships between fibers as edges. Each node contains multidimensional feature vectors representing subtype, morphology, texture, and cell nuclear attributes. The networks at each time point are then assembled sequentially to form spatiotemporal graph data. Next, a spatiotemporal graph neural network (ST-GNN) is used as the core analysis engine. Its spatial dependency encoding layer aggregates node and neighbor information through a graph convolutional network to capture the spatial interactions of the tissue microenvironment and encodes the temporal evolution patterns. The layer then models the temporal full-map representation using a recurrent neural network, learns the historical evolution pattern of pathological states, and outputs a prediction network for the next time point through a prediction decoding layer. Based on this, a sequence of predicted pathological networks for multiple future time points is generated through iterative forward extrapolation of ST-GNN based on a single baseline pathological network. Finally, the numerical change trajectory of key pathological indicators is extracted from this prediction sequence and a quantitative prediction curve is plotted. At the same time, a future pathological heatmap that can intuitively display the spatial distribution and severity of lesions is generated based on node attributes. This completes the transition from static pathology to dynamic progression prediction, providing clinicians with an objective, accurate, and visualized tool for myopathy prognosis assessment.

[0153] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0154] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0155] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0156] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting myopathy progression based on a spatiotemporal topological network, characterized by, The method includes the following steps: Muscle biopsy sections from a predetermined number of time points in the patient's past are obtained. After staining according to a preset scheme, single muscle fibers, cell nuclei, connective tissue, and microvessels in the muscle biopsy sections are identified based on a pre-trained image segmentation model. A node set is constructed using the single muscle fibers as nodes. Each node records subtype attributes, morphological parameter vectors, internal texture parameter vectors, and cell nucleus attribute vectors as node attributes. Adjacency relationships between nodes are constructed based on a preset spatial adjacency graph construction algorithm to obtain a static pathological topology network. The static case topology network corresponding to each time point is then constructed into time series graph data. The time-series graph data is input into a pre-trained spatiotemporal graph neural network containing a spatial dependency encoding layer, a temporal evolution encoding layer, and a prediction decoding layer. The spatial dependency encoding layer aggregates information from each node and its neighboring nodes in the static pathological topology network using a graph convolutional network to extract feature representation vectors. The temporal evolution encoding layer performs graph pooling on the feature representation vectors to obtain a full graph representation vector, and extracts and encodes the hidden states of the entire pathological history from the global graph representation vector sequence using a recurrent neural network. This encoded state is then input into the prediction decoding layer to predict the static case topology network for the next time point based on a multilayer vector perceptron. The predicted static case topology network is then added to the tail of the time-series graph data and the head of the queue is removed. It is then re-input into the spatiotemporal graph neural network for continuous cyclic prediction to obtain the predicted static case topology network for the patient at multiple future time points. Based on the predicted static case topology network of the patient at multiple future time points, a prediction report is generated. The prediction report includes: quantitative prediction curves for clinically relevant indicators, and a visualized pathological heatmap generated from the node attributes predicted for the target future time points.

2. The method for predicting myopathy progression based on spatiotemporal topology networks according to claim 1, characterized in that, The pre-training steps of the image segmentation model include: Obtain a first training sample set containing multiple samples, each sample containing a muscle biopsy digital slice sample stained with the preset scheme, and labeling the location of single muscle fibers, cell nuclei, connective tissue and microvessels as tags; The initial image segmentation network is trained using the first training sample set. The initial image segmentation network takes the muscle biopsy digital slice sample as input and outputs the predicted location and predicted classification of single muscle fibers, cell nuclei, connective tissue and microvessels in the image. The initial image segmentation network is updated by constructing a loss based on the deviation between the predicted location, the predicted classification and the label to obtain the image segmentation network. The initial image segmentation network includes a semantic segmentation network or an instance segmentation network based on a convolutional neural network architecture. The semantic segmentation network is a U-Net network or a DeepLab network, and the instance segmentation network is a Mask R-CNN network.

3. The method for predicting myopathy progression based on spatiotemporal topological networks according to claim 1, characterized in that, The subtype attributes are based on the classification of myofibril subtypes according to the results of metabolic enzyme staining, including slow contraction oxidative type, fast contraction oxidative-glycolytic type and fast contraction glycolytic type. The subtype attributes are represented in the form of one-heat coding. The morphological parameter vector includes muscle fiber cross-sectional area, circularity, polygonality, and perimeter; The internal texture parameter vector includes: a contrast texture that reflects grayscale changes and an energy texture that reflects the degree of texture uniformity. The nuclear attribute vector includes the proportion of central nuclei, the total number of nuclei, and the number of peripheral nuclei; The preset spatial adjacency graph construction algorithm is a Delaunay triangulation algorithm, a Voronoi graph algorithm, or a radius-based neighbor graph algorithm.

4. The method for predicting myopathy progression based on spatiotemporal topology networks according to claim 1, characterized in that, The pre-training steps of the spatiotemporal graph neural network include: Muscle biopsy sample slices from multiple patients at multiple time points are obtained. After staining according to the preset scheme, single muscle fibers, cell nuclei, connective tissue, and microvessels in the muscle biopsy sample slices are identified based on the image segmentation model. A node set is constructed using the single muscle fibers as nodes. Each node records subtype attributes, morphological parameter vectors, internal texture parameter vectors, and cell nucleus attribute vectors as node attributes. Adjacency relationships between nodes are constructed based on a preset spatial adjacency graph construction algorithm to obtain static pathological topology network samples at each time point, which are then used to construct time series graph data samples for each patient. Continuous sampling is performed on the time series data samples to construct a second training sample set, each sample containing a second predetermined number of static pathological topology network samples at consecutive time points; the second predetermined number is the first predetermined number plus 1. The second training sample set is used to train an initial spatiotemporal graph neural network containing an initial spatial dependency encoding layer, an initial temporal evolution law encoding layer, and an initial prediction decoding layer. The first set number of static pathological topology network samples in each sample are used as input. The initial spatial dependency encoding layer aggregates information of nodes and their neighboring nodes in each static pathological topology network sample based on a graph convolutional network to extract sample feature representation vectors. The initial temporal evolution law encoding layer performs graph pooling on the sample feature representation vectors to obtain a full graph representation vector of the sample, and extracts the hidden state of the sample encoding the entire pathological history from the global graph representation vector sequence of the sample based on a recurrent neural network. This hidden state is input to the initial prediction decoding layer to predict the last predicted static case topology network in each sample based on a multilayer vector perceptron. The parameters of the initial spatiotemporal graph neural network are updated based on the mean square error of node features between the predicted value and the true value to obtain the spatiotemporal graph neural network.

5. The method for predicting myopathy progression based on spatiotemporal topology networks according to claim 4, characterized in that, The graph convolutional network is followed by a non-linear activation function layer, and the recurrent neural network uses a long short-term memory network. The Adam optimizer is used when updating the parameters of the initial spatiotemporal graph neural network; The formula for calculating the mean square error is: ； Where L represents the mean square error. express The number of nodes in the static pathological topology network sample at the specified time point. express The predicted node attribute value of the i-th node in the static pathological topology network sample at the specified time point. express The true value of the node attribute of the i-th node in the static pathological topology network sample at the specified time point.

6. The method for predicting myopathy progression based on spatiotemporal topology networks according to claim 1, characterized in that, The quantitative prediction curve for clinically relevant indicators is a prediction curve for the average cross-sectional area of ​​muscle fibers, calculated as follows: ； in, Indicates the reference time point The expected average cross-sectional area of ​​the muscle fiber at the k-th time point thereafter. Indicates at a point in time The predicted total number of muscle fiber nodes. Indicates a point in time The predicted node attribute value for the i-th node in the static pathological topology network sample. This is a projection operator used to extract the cross-sectional area of ​​muscle fibers from the predicted node attribute values.

7. The method for predicting myopathy progression based on spatiotemporal topology networks according to claim 1, characterized in that, Generate a visual pathological heatmap based on the node attributes predicted for a target future time point, including: a predicted static case topology network for a specified future time point, extracting the cross-sectional area of ​​the corresponding muscle fibers of each node, measuring the degree of atrophy based on preset evaluation criteria, and generating a visual pathological heatmap on the degree of muscle atrophy.

8. A device for predicting the progression of myopathy based on spatiotemporal topology networks, comprising a processor, a memory, and a computer program or instructions stored in the memory, characterized in that, The processor is configured to execute the computer program or instructions, and when the computer program or instructions are executed, the device implements the steps of the method as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method as described in any one of claims 1 to 7.

10. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 7.

Images