A nursing decision-making system for breast cancer

Abstract

The present invention relates to the field of machine learning and discloses a nursing decision-making system for breast cancer, which aims to optimize nursing resource scheduling through multimodal data fusion and intelligent analysis. The system includes three modules: multimodal data acquisition, cluster analysis, and nursing decision-making. The cluster analysis module uses elastic diamond unit extraction and hierarchical mosaic structure to achieve spatial correlation modeling of cross-modal data and accurately mine the multi-scale distribution characteristics of nursing needs. The nursing decision-making module designs a spectrum graph fusion network, combines frequency domain modulation with spatial attention mechanism, dynamically analyzes nursing demand signals and generates graded predictions. The system breaks through the limitations of data fragmentation and feature decoupling of traditional methods, significantly improves the completeness of nursing demand representation and the accuracy of classification, realizes efficient decision-making in scenarios such as dynamic resource matching and emergency response, and provides adaptive and scalable technical support for complex nursing management.

Description

Technical Field

[0001] The present invention relates to the field of machine learning, and in particular to a nursing decision-making system for breast cancer. Background Art

[0002] With the development of the times, the nursing management of breast cancer patients faces increasingly complex multimodal data integration and precise decision-making needs, which places higher demands on existing technologies. Traditional nursing decision-making systems face several technical bottlenecks. First, traditional systems rely heavily on single-modal data, making it difficult to effectively integrate multi-source heterogeneous data. This leads to a lack of cross-modal feature correlation, which affects the comprehensiveness of nursing strategies. Second, medical data has complex spatial topological characteristics, and traditional clustering methods (such as K-means or hierarchical clustering) have difficulty capturing the multi-scale geometric structure and high-order adjacency relationships of lesion regions, limiting the granularity of feature expression. In addition, existing graph neural networks mostly focus on feature extraction in a single domain (spatial or frequency domain) and lack the collaborative modeling of the frequency domain response and spatial dependencies of the graph structure, which limits the accuracy of nursing grading predictions. These limitations pose severe challenges to existing systems in scenarios such as dynamic scheduling of nursing resources and personalized grading management. Technological innovation is urgently needed to achieve deep coupling of multimodal data and high-order feature mining to improve the scientific nature and operability of nursing decisions. Summary of the Invention

[0003] The present invention provides a nursing decision-making system for breast cancer, which aims to break through the bottlenecks of insufficient utilization of multimodal data and insufficient feature decoupling in traditional nursing decision-making through the collaborative innovation of high-order diamond mosaic clustering networks and spectral graph fusion network models, and realize the precise dynamic scheduling of nursing resources. Based on the high-order diamond mosaic clustering network, the system breaks through the limitations of rough data representation and multi-source information fragmentation in traditional nursing assessment: through elastic Delaunay topology construction and multi-scale diamond unit extraction, a spatial association network of cross-modal data is established, hierarchical features are mined from micro-nursing demand distribution to macro-resource allocation patterns, and low-dimensional clustering features are generated by combining density-sensitive optimization algorithms to accurately characterize the spatial heterogeneity of nursing needs. Further, a spectral graph fusion network model is designed, and the local fluctuation characteristics of nursing demand signals are analyzed using frequency domain modulation technology. The dynamic dependency relationship between nursing nodes is captured through the spatial attention mechanism, and the dual-path features are innovatively integrated to realize fine-grained prediction of nursing level. The system significantly improves the rationality of nursing resource allocation through topological association modeling and dynamic weight allocation, providing an adaptive and interpretable technical framework for hierarchical decision-making in complex nursing scenarios.

[0004] The present invention provides a nursing decision-making system for breast cancer, which includes a multimodal data acquisition module, a cluster analysis module, and a nursing decision-making module;

[0005] The multimodal data acquisition module, by connecting to the medical imaging system and the pathology information system, collects and structures breast cancer patients' ultrasound images, mammography X-rays, MRI sequence images, and pathology diagnosis reports to build a standardized multimodal dataset;

[0006] The cluster analysis module constructs a high-order diamond mosaic clustering network through multimodal spatial alignment, Delaunay topology construction, and diamond unit extraction methods. The standardized multimodal dataset is clustered using the high-order diamond mosaic clustering network to generate cluster feature data. The high-order diamond mosaic clustering network includes a multimodal alignment unit, a high-order structure generation unit, a diamond mosaic modeling unit, and a cluster feature generation unit. The multimodal alignment unit is electrically connected to the high-order structure generation unit, the high-order structure generation unit is electrically connected to the diamond mosaic modeling unit, and the diamond mosaic modeling unit is electrically connected to the cluster feature generation unit.

[0007] The nursing decision-making module constructs a spectrum-spectrum graph fusion network model through graph convolution, Chebyshev polynomials, multi-frequency modulation, graph attention mechanism and frequency-space fusion mechanism. The spectrum-spectrum graph fusion network model processes cluster feature data to generate nursing level predictions, thereby realizing reasonable scheduling and precise hierarchical management of nursing resources. The spectrum-spectrum graph fusion network model includes a spectrum structure construction unit, a frequency domain feature extraction unit, a spatial feature extraction unit, a feature fusion unit and a nursing prediction unit. The spectrum structure construction unit and the frequency domain feature extraction unit are electrically connected, and the spectrum structure construction unit and the spatial feature extraction unit are electrically connected. The frequency domain feature extraction unit, the spatial feature extraction unit and the feature fusion unit are electrically connected. The feature fusion unit and the nursing prediction unit are electrically connected.

[0008] Furthermore, the cluster analysis module generates cluster feature data, which specifically includes the following:

[0009] The multimodal alignment unit processes the image modality and text modality in the standardized multimodal dataset. For the image modality, it extracts the structural features of the lesion area, mass boundary, and mass boundary density, and constructs their corresponding spatial coordinate representation to generate image space point cloud data. For the text modality, it converts it into an embedded vector representation through text embedding methods. The image space point cloud data and the embedded vector representation are mapped to a unified spatial coordinate system through modality alignment, and then fused to form multimodal space point cloud data.

[0010] The high-order structure generation unit is based on multimodal spatial point cloud data, introduces Delaunay triangulation, and uses an error tolerance mechanism to relax the empty circularity constraint of Delaunay triangulation to construct a high-order Delaunay structure.

[0011] The diamond tessellation modeling unit extracts diamond units from high-order Delaunay structures of different orders, captures the high-order adjacency of multimodal spatial point cloud data, and introduces a diamond tessellation mechanism to construct structural mapping relationships between diamond units of different orders, generating a hierarchical diamond tessellation structure.

[0012] The cluster feature generation unit constructs the incident matrix based on the hierarchical diamond mosaic structure, calculates the cluster matrix and normalizes it to obtain the normalized cluster matrix, performs feature aggregation, and generates cluster feature data.

[0013] Furthermore, the nursing decision module generates a nursing level prediction process, which specifically includes the following:

[0014] The spectral structure construction unit combines clustering feature data to construct an adjacency matrix, calculates the degree matrix, and obtains a standard Laplace matrix. To achieve efficient and stable spectral graph convolution, scaling and shifting are performed on the basis of the standard Laplace matrix to construct a shifted Laplace matrix to achieve interval compression of the feature spectrum. The adjacency matrix and the shifted Laplace matrix are composed of nodes.

[0015] The frequency domain feature extraction unit performs spectral graph convolution on the shifted Laplace matrix using Chebyshev polynomials, extracts the local frequency response in the graph structure from a frequency domain perspective, generates initial graph convolution features, learns the frequency projection basis and phase offset for each node in the initial graph convolution features, and performs frequency modulation projection on adjacent nodes to generate a modulation response. Based on the modulation response, sine and cosine modulation encoding is performed to obtain a multi-frequency modulation vector, which is compressed into frequency domain side information using linear mapping, and then an aggregation function is used to summarize the frequency domain side information to obtain the node frequency domain features.

[0016] The spatial feature extraction unit is based on the adjacency matrix and adopts the graph attention mechanism and the frequency decoupled gated fusion method to model the spatial dependency between nodes and obtain the node spatial features.

[0017] The feature fusion unit concatenates the node frequency domain features and node spatial features, inputs them into the gated MLP, generates fusion coefficients, performs adaptive fusion, and generates node fusion representation;

[0018] The care prediction unit inputs the node fusion representation into the classification layer for prediction and generates a care level prediction.

[0019] By adopting the above scheme, the beneficial effects achieved by the present invention are as follows:

[0020] 1. Achieved efficient fusion of multimodal data and topological association modeling:

[0021] Through elastic topology construction technology and multi-scale diamond unit extraction mechanism, the limitations of traditional systems in isolated processing of multi-source data such as images and texts have been broken through, and a unified spatial association network for cross-modal data has been established. This technology solves the problem of information fragmentation caused by data heterogeneity, significantly improves the spatial matching accuracy of multimodal features, enhances the system's ability to express multi-level demand characteristics in complex nursing scenarios, and provides high-fidelity data support for dynamic resource scheduling.

[0022] 2. Improved the dynamic adaptability and grading accuracy of nursing decisions:

[0023] Based on the frequency-domain-space dual-path fusion architecture, the system achieves refined analysis of nursing demand signals through the synergy of frequency-domain feature modulation and dynamic attention mechanism. This technology overcomes the defect of traditional models in responding slowly to demand fluctuations, and can accurately capture the differentiated characteristics from normal care to emergencies. While ensuring the stability of the topological structure, it significantly improves the granularity and timeliness of hierarchical predictions, providing an intelligent decision-making basis for dynamic resource matching.

[0024] 3. Enhanced system robustness and scalability for complex scenarios:

[0025] Through hierarchical clustering optimization and dynamic feature fusion strategies, the system effectively reduces modeling deviations in high-density data areas while maintaining the integrity of topological associations. Combined with cross-scale feature transfer and adaptive weight distribution technology, it solves the overfitting risk of traditional methods in complex scenarios, enabling the system to flexibly adapt to nursing management needs of different scales and levels, and building a technical framework that is both stable and scalable for multi-dimensional resource scheduling. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a module diagram of a nursing decision-making system for breast cancer proposed by the present invention;

[0027] Figure 2 This is a schematic diagram of relaxing the Delaunay triangulation using the error tolerance mechanism provided in the second embodiment.

[0028] Figure 2 In the figure, the gray triangles represent the traditional Delaunay triangulation, the red points represent the connected nodes, and the blue dashed circles represent the error tolerance mechanism. DETAILED DESCRIPTION

[0029] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0030] Example 1, according to Figure 1 , the present invention provides a nursing decision-making system for breast cancer, which includes a multimodal data acquisition module, a cluster analysis module, and a nursing decision-making module;

[0031] The multimodal data acquisition module, by connecting to the medical imaging system and the pathology information system, collects and structures breast cancer patients' ultrasound images, mammography X-rays, MRI sequence images, and pathology diagnosis reports to build a standardized multimodal dataset;

[0032] The cluster analysis module constructs a high-order diamond mosaic clustering network through multimodal spatial alignment, Delaunay topology construction, and diamond unit extraction methods. The standardized multimodal dataset is clustered using the high-order diamond mosaic clustering network to generate cluster feature data. The high-order diamond mosaic clustering network includes a multimodal alignment unit, a high-order structure generation unit, a diamond mosaic modeling unit, and a cluster feature generation unit. The multimodal alignment unit is electrically connected to the high-order structure generation unit, the high-order structure generation unit is electrically connected to the diamond mosaic modeling unit, and the diamond mosaic modeling unit is electrically connected to the cluster feature generation unit.

[0033] The nursing decision-making module constructs a spectrum-spectrum graph fusion network model through graph convolution, Chebyshev polynomials, multi-frequency modulation, graph attention mechanism and frequency-space fusion mechanism. The spectrum-spectrum graph fusion network model processes cluster feature data to generate nursing level predictions, thereby realizing reasonable scheduling and precise hierarchical management of nursing resources. The spectrum-spectrum graph fusion network model includes a spectrum structure construction unit, a frequency domain feature extraction unit, a spatial feature extraction unit, a feature fusion unit and a nursing prediction unit. The spectrum structure construction unit and the frequency domain feature extraction unit are electrically connected, and the spectrum structure construction unit and the spatial feature extraction unit are electrically connected. The frequency domain feature extraction unit, the spatial feature extraction unit and the feature fusion unit are electrically connected. The feature fusion unit and the nursing prediction unit are electrically connected.

[0034] Example 2, according to Figure 2 This embodiment is based on the first embodiment. In this embodiment, the cluster analysis module generates cluster feature data in a process that specifically includes the following:

[0035] The multimodal alignment unit processes the image modality and text modality in the standardized multimodal dataset. For the image modality, it extracts the structural features of the lesion area, mass boundary, and mass boundary density, and constructs their corresponding spatial coordinate representation to generate image space point cloud data. For the text modality, it converts it into an embedded vector representation through text embedding methods. The image space point cloud data and the embedded vector representation are mapped to a unified spatial coordinate system through modality alignment, and then fused to form multimodal space point cloud data.

[0036] The high-order structure generation unit introduces Delaunay triangulation based on multimodal spatial point cloud data, and uses an error tolerance mechanism to relax the empty circularity constraint of Delaunay triangulation to construct a high-order Delaunay structure. The traditional Delaunay structure has a "hard boundary" that only accepts perfect geometric relationships. The high-order Delaunay structure has a "soft boundary" that allows a small error, making the network more robust and avoiding structural fracture or failure due to small disturbances. The high-order Delaunay structure can achieve robust partitioning of non-ideal point sets while maintaining local topological consistency. The Delaunay condition evaluation under a certain order tolerance allows moderate violation of the traditional Delaunay empty circularity constraint within the control order range, thereby generating a high-order triangulated network structure that is more suitable for noisy environments or density-varying areas.

[0037] The diamond tessellation modeling unit extracts diamond units from high-order Delaunay structures of different orders to capture the high-order adjacency of multimodal spatial point cloud data. The diamond tessellation mechanism is introduced to construct structural mapping relationships among diamond units of different orders, generating a hierarchical diamond tessellation structure. The diamond tessellation mechanism uses diamond units extracted from high-order Delaunay structures of different orders as basic units, and forms a hierarchical and robust spatial structure grid through connection, nesting, and aggregation.

[0038] High-order adjacency refers to the joint adjacency relationship formed by multiple points participating in a higher-order structural unit, including triples, quadruples, and higher-order simplexes. In a high-order Delaunay structure, a set of points forms a k-order Delaunay unit, and a high-order adjacency relationship exists between the point sets. This relationship explicitly characterizes the local collaborative structure between multiple points and can capture complex spatial geometric topology and multimodal semantic interaction features.

[0039] The cluster feature generation unit constructs the incident matrix based on the hierarchical diamond mosaic structure, calculates the cluster matrix and normalizes it to obtain the normalized cluster matrix, performs feature aggregation, and generates cluster feature data.

[0040] Example 3: This example is based on Example 1. In this example, the cluster analysis module generates cluster feature data by performing the following steps:

[0041] The multimodal alignment unit processes the image modality and text modality in the standardized multimodal dataset. For the image modality, it extracts the structural features of the lesion area, mass boundary, and mass boundary density, and constructs their corresponding spatial coordinate representation to generate image space point cloud data. For the text modality, it converts it into an embedded vector representation through text embedding methods. The image space point cloud data and the embedded vector representation are mapped to a unified spatial coordinate system through modality alignment, and then fused to form multimodal space point cloud data.

[0042] The high-order structure generation unit constructs a Delaunay structure based on multimodal spatial point cloud data;

[0043] The diamond tessellation modeling unit extracts diamond units from Delaunay structures of different orders, captures the high-order adjacency of multimodal spatial point cloud data, and introduces a diamond tessellation mechanism to construct a structural mapping relationship between diamond units of different orders to generate a hierarchical diamond tessellation structure.

[0044] The cluster feature generation unit constructs the incident matrix based on the hierarchical diamond mosaic structure, calculates the cluster matrix and normalizes it to obtain the normalized cluster matrix, performs feature aggregation, and generates cluster feature data.

[0045] Example 4: This example is based on Example 2. In this example, the nursing decision module generates a nursing level prediction process, which specifically includes the following:

[0046] The spectral structure construction unit combines clustering feature data to construct an adjacency matrix, calculates the degree matrix, and obtains a standard Laplace matrix. To achieve efficient and stable spectral graph convolution, scaling and shifting are performed on the basis of the standard Laplace matrix to construct a shifted Laplace matrix to achieve interval compression of the feature spectrum. The adjacency matrix and the shifted Laplace matrix are composed of nodes, and the formula used is as follows:

[0047] ;

[0048] in, represents the shifted Laplacian matrix, represents the standard Laplace matrix, Represents the maximum eigenvalue of the standard Laplace matrix, which is used to normalize the Laplace spectrum; Represents a zoom operation, represents the identity matrix, Indicates a shift operation;

[0049] Chebyshev polynomial convolution depends on the eigenvalue range of the Laplace matrix. The eigenvalue range of the standard Laplace matrix is ​​too large, which will cause numerical instability when calculating the polynomial. Therefore, a shift operation is performed to achieve interval compression of the characteristic spectrum and stabilize the range of the Chebyshev polynomial.

[0050] The frequency domain feature extraction unit performs spectral graph convolution on the shifted Laplace matrix using Chebyshev polynomials, extracts the local frequency response in the graph structure from a frequency domain perspective, generates the initial graph convolution feature, learns the frequency projection basis and phase offset for each node in the initial graph convolution feature, and performs frequency modulation projection on adjacent nodes to generate a modulation response; based on the modulation response, sine and cosine modulation encoding is performed to obtain a multi-frequency modulation vector, which is compressed into frequency domain side information using linear mapping, and then an aggregation function is used to summarize the frequency domain side information to obtain the node frequency domain feature. The formula used is as follows:

[0051] Multi-frequency modulation coding formula:

[0052] ;

[0053] in, represents the target node, represents neighbor nodes, Represents neighbor nodes At the target node The modulation response in represents the frequency scale index, represents the multi-frequency modulation vector, Indicates the frequency factors, and Indicates sine and cosine modulation; Represents vector concatenation;

[0054] Construct frequency domain side message formula:

[0055] ;

[0056] in, Indicates that the neighbor node To the target node Frequency domain side information, represents the linear mapping weight matrix, represents the modulation vector at the first frequency scale, represents the modulation vector at the second frequency scale, represents the modulation vector at the third frequency scale;

[0057] The spatial feature extraction unit is based on the adjacency matrix and adopts the graph attention mechanism and the frequency decoupled gated fusion method to model the spatial dependency between nodes and obtain the node spatial features.

[0058] The feature fusion unit concatenates the node frequency domain features and node spatial features, inputs them into the gated MLP, generates fusion coefficients, performs adaptive fusion, and generates node fusion representation;

[0059] The care prediction unit inputs the node fusion representation into the classification layer for prediction and generates a care level prediction.

[0060] Example 5: This example is based on Example 2. In this example, the nursing decision module generates a nursing level prediction process, which specifically includes the following:

[0061] The spectral structure construction unit combines the clustering feature data, constructs the adjacency matrix, calculates the degree matrix, and obtains the standard Laplace matrix;

[0062] The frequency domain feature extraction unit performs spectral convolution on the standard Laplace matrix through Chebyshev polynomials to obtain the node frequency domain features;

[0063] The spatial feature extraction unit is based on the adjacency matrix and adopts the graph attention mechanism and the frequency decoupled gated fusion method to model the spatial dependency between nodes and obtain the node spatial features.

[0064] The feature fusion unit concatenates the node frequency domain features and node spatial features, inputs them into the gated MLP, generates fusion coefficients, performs adaptive fusion, and generates node fusion representation;

[0065] The care prediction unit inputs the node fusion representation into the classification layer for prediction and generates a care level prediction.

[0066] Example 6. This example is based on Example 5. In this example, the nursing decision module constructs a spectrum graph fusion network model through graph convolution, Chebyshev polynomials, multi-frequency modulation, graph attention mechanism and frequency-space fusion mechanism. The spectrum graph fusion network model processes cluster feature data to generate nursing level predictions, thereby realizing reasonable scheduling and precise hierarchical management of nursing resources.

[0067] In this embodiment:

[0068] Image data acquisition:

[0069] Ultrasound images showed a mass in the upper quadrant of the left breast with unclear boundaries, approximately 2.4 cm × 2.0 cm;

[0070] Mammographic target X-ray showed uneven density of the left breast and local structural disorder;

[0071] MRI enhanced sequence revealed rapid enhancement of the mass, showing a typical "washout" curve;

[0072] Pathology text information:

[0073] Immunohistochemistry results: ER (+), PR (+), HER2 (-), Ki67 index: 30%;

[0074] Molecular typing: Luminal B;

[0075] High-order diamond tessellation clustering network:

[0076] The number of extracted diamond units: 438, forming a 5-layer hierarchical structure;

[0077] Nursing Level Prediction:

[0078]

[0079] Predicted level: Secondary care (moderate).

[0080] The present invention and its embodiments are described above. Such description is not restrictive. What is shown in the accompanying drawings is only one of the embodiments of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in this field are inspired by it and do not depart from the purpose of the invention, they can creatively design structural methods and embodiments similar to the technical solution, which should all fall within the scope of protection of the present invention.

Claims

1. A nursing decision-making system for breast cancer, comprising a multimodal data acquisition module, wherein the multimodal data acquisition module constructs a standardized multimodal dataset; characterized in that: The system also includes a cluster analysis module and a nursing decision module; The cluster analysis module constructs a high-order diamond mosaic clustering network, clusters the standardized multimodal data set through the high-order diamond mosaic clustering network, and generates cluster feature data; the high-order diamond mosaic clustering network includes a multimodal alignment unit, a high-order structure generation unit, a diamond mosaic modeling unit, and a cluster feature generation unit; The nursing decision module constructs a spectrum graph fusion network model, processes cluster feature data through the spectrum graph fusion network model, and generates a nursing level prediction; The spectrum graph fusion network model includes spectrum structure construction unit, frequency domain feature extraction unit, spatial feature extraction unit, feature fusion unit and nursing prediction unit; The high-order structure generation unit constructs multimodal spatial point cloud data based on a standardized multimodal dataset. It introduces Delaunay triangulation and uses an error tolerance mechanism to relax the empty circularity constraint of Delaunay triangulation. Combined with the multimodal spatial point cloud data, it constructs a high-order Delaunay structure. The diamond tessellation modeling unit extracts diamond units from high-order Delaunay structures of different orders, captures the high-order adjacency of multimodal spatial point cloud data, and introduces the diamond tessellation mechanism to construct structural mapping relationships and generate a hierarchical diamond tessellation structure. The cluster feature generation unit constructs the incident matrix based on the hierarchical diamond mosaic structure, calculates the cluster matrix and normalizes it to obtain the normalized cluster matrix, performs feature aggregation, and generates cluster feature data; The spectral structure construction unit combines cluster feature data to construct an adjacency matrix, calculates the degree matrix, obtains a standard Laplace matrix, and performs scaling and shifting on the basis of the standard Laplace matrix to construct a shifted Laplace matrix; the adjacency matrix and the shifted Laplace matrix are composed of nodes; The frequency domain feature extraction unit performs spectral graph convolution on the shifted Laplace matrix through Chebyshev polynomials to generate the initial graph convolution feature. It learns the frequency projection basis and phase offset for each node in the initial graph convolution feature, and performs frequency modulation projection on adjacent nodes to generate a modulation response. Based on the modulation response, sine and cosine modulation encoding is performed to obtain a multi-frequency modulation vector; Linear mapping is used to compress the multi-frequency modulation vector into frequency domain side information, and then an aggregation function is used to summarize the frequency domain side information to obtain the node frequency domain features.

2. A nursing decision system for breast cancer according to claim 1, characterized in that: The spatial feature extraction unit is based on the adjacency matrix and adopts the graph attention mechanism and the frequency decoupled gated fusion method to model the spatial dependency between nodes and obtain the node spatial features.

3. The nursing decision system for breast cancer according to claim 2, characterized in that: The feature fusion unit concatenates the node frequency domain features and node spatial features, uses gated MLP to generate fusion coefficients, performs adaptive fusion, and generates node fusion representation.

Images