Spinal disease recognition system based on relation perception and condition classification

Abstract

The embodiment of the invention discloses a spinal disease recognition system based on relation perception and condition classification. According to one specific embodiment, the system comprises a data processing server, a feature extraction server, a disease recognition server, a report generation server and a user interaction terminal, and the data processing server is used for carrying out preset standardization processing on data of patients with spinal diseases; the feature extraction server is used for performing feature extraction and segmentation on the standard image data; the disease recognition server is used for performing spinal disease recognition to obtain a spinal disease recognition result; the report generation server is used for carrying out multi-modal fusion and generating a spinal disease structured recognition report; and the user interaction terminal is used for displaying the image and the report and controlling at least one of the robot and the chemical allocation equipment to perform preset operation according to the electric signal. According to the embodiment, the display waiting time of the system is shortened, and the redundant calculation cycle is reduced.

Description

Technical Field

[0001] Embodiments of this disclosure relate to the field of computer technology, and more specifically to a spinal disease identification system based on relation awareness and conditional classification. Background Technology

[0002] With the aging global population and the increasing demand for precision medicine, the market for spinal disease diagnosis is huge and growing. Against the backdrop of a projected significant expansion in the AI ​​medical imaging system market, there is a clear need for improved diagnostic speed and consistency. Currently, the common methods for diagnosing spinal diseases include: radiologists using AI medical imaging systems (such as U-Net, Vision Transformer ViT, etc.) to assist in interpreting medical images (such as CT and MRI); and the use of computer-aided diagnostic (CADx) systems for assisted diagnosis.

[0003] However, when using the above system to identify spinal diseases, the following technical problems often arise: The system experiences long display times. Specifically, systems using convolutional neural networks (CNNs, such as U-Net) are limited by their inherent receptive field. Their performance significantly degrades when identifying and accurately segmenting lesions of varying shapes and blurred boundaries, requiring multiple calls to the diagnostic model for verification. This results in long display times for disease identification results.

[0004] The system suffers from redundant computational loops. Specifically, Transformer-based systems (such as the Vision Transformer ViT) lack a detailed understanding of the three-dimensional anatomical relationships of the spine and fail to fully utilize the topological information between vertebrae. Existing computer-aided diagnostic (CADx) systems lack the ability to integrate different information for comprehensive differential diagnosis and cannot automatically and seamlessly convert the quantitative analysis results of AI into high-quality diagnostic text. Spinal disease identification reports displayed on user terminals are inconsistent with the facts, requiring the system to re-execute, resulting in redundant computational loops.

[0005] The information disclosed in this background section is only intended to enhance the understanding of the background of the inventive concept, and therefore may contain information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The summary portion of this disclosure is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description portion. This summary portion is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0007] Some embodiments of this disclosure propose a spinal disease identification system based on relation awareness and condition classification to address one or more of the technical problems mentioned in the background section above.

[0008] Some embodiments of this disclosure provide a spinal disease identification system based on relation awareness and conditional classification. The system includes: a data processing server, a feature extraction server, a disease identification server, a report generation server, and a user interaction terminal, all interconnected. The data processing server performs pre-defined standardization processing on spinal disease patient data to generate standard image data and standard clinical data. The feature extraction server performs feature extraction and segmentation on the standard image data to generate a set of vertebral node differential feature vectors and a set of medical image segmentation masks for spinal diseases. The disease identification server generates a set of reconstructed images and spinal disease identification results based on the set of spinal disease symptoms, the standard image data, and the set of vertebral node differential feature vectors. The report generation server performs multimodal fusion of the standard clinical data, the set of vertebral node differential feature vectors, and the spinal disease identification results received from the disease identification server to obtain multimodal vectors, and generates a structured spinal disease identification report based on these multimodal vectors.

[0009] The various embodiments of this disclosure have the following beneficial effects: The spinal disease recognition system based on relational awareness and conditional classification, through some embodiments of this disclosure, models the biomechanical dependence between vertebrae and generates anatomical-aware reconstructions conditionally using high-dimensional graph embedding, providing an interpretable decision-making mechanism for large language models, reducing system display waiting time, and reducing redundant computational loops. Specifically, the system needs to call the diagnostic model multiple times for verification, resulting in long display waiting times for disease recognition results, and the spinal disease recognition report displayed on the user's interactive terminal is inconsistent with the facts, requiring the system to re-execute, causing redundant computational loops. The reason for this is the long display waiting time. Specifically, systems using convolutional neural networks (CNNs, such as U-Net) are limited by their inherent receptive field, and their performance significantly degrades when recognizing and accurately segmenting lesions with varying shapes and blurred boundaries, requiring multiple calls to the diagnostic model for verification. This leads to long display waiting times for disease recognition results. The system has redundant computational loops. Specifically, systems based on Transformers (such as the Vision Transformer ViT) lack a detailed understanding of the three-dimensional anatomical relationships of the spine and fail to fully utilize the topological information between vertebrae. Existing computer-aided diagnostic (CADx) systems lack the ability to integrate different information for comprehensive differential diagnosis and cannot automatically and seamlessly convert the quantitative analysis results of AI into high-quality diagnostic text. Spinal disease identification reports displayed on user terminals often do not match the facts, requiring the system to re-execute, resulting in redundant computational loops. Based on this, some embodiments of the spinal disease identification system based on relational awareness and conditional classification disclosed herein include a data processing server, a feature extraction server, a disease identification server, a report generation server, and a user terminal that are interconnected. First, the data processing server performs pre-defined standardization processing on spinal disease patient data to generate standard image data and standard clinical data. This standardization of raw medical images and clinical text generates image and clinical data in a unified format, providing high-quality input for downstream analysis. Second, the feature extraction server performs feature extraction and segmentation on the standard image data to generate a set of vertebral node differential feature vectors and a set of spinal disease medical image segmentation masks. Then, the vertebral node differential features and segmentation masks are extracted through a graph transformer differential information interpolator (GT-DII) to quantify spinal anatomical structural abnormalities. Then, the aforementioned disease identification server generates a set of reconstructed images and spinal disease identification results based on the spinal lesion set, the aforementioned standard image data, and the aforementioned vertebral node differential feature vector set. Thus, reconstructed images are generated based on the Conditional Diffusion Diagnostic Network (CD-DN), and reconstruction errors are calculated, achieving lesion type identification and uncertainty quantification.Next, the aforementioned report generation server performs multimodal fusion on the aforementioned standard clinical data, the aforementioned set of differential feature vectors for vertebral nodes, and the aforementioned spinal disease identification results received from the aforementioned disease identification server to obtain multimodal vectors. Based on these multimodal vectors, a structured spinal disease identification report is generated. Thus, a structured diagnostic report is generated by fusing multimodal data, and fine-tuned using the LoRA method based on a manually constructed instruction-based fine-tuning dataset, using frameworks such as llama-factory. Then, an actionable clinical insight is output using the fine-tuned dedicated large language model. Finally, combined with the aforementioned user interaction terminal, the aforementioned spinal disease medical images, the aforementioned spinal disease medical image segmentation mask set, and the aforementioned structured spinal disease identification report are displayed. At least one of a preset robot and a preset chemical dispensing device is controlled to perform preset operations based on electrical signals. Thus, the images, masks, and reports are visualized, and the robot and dispensing equipment are controlled via interactive buttons to perform precise treatment operations. This implementation reduces system display waiting time and redundant computational loops. Attached Figure Description

[0010] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and elements are not necessarily drawn to scale.

[0011] Figure 1 This is a schematic diagram of the structure of some embodiments of the spinal disease identification system based on relation awareness and condition classification according to the present disclosure; Figure 2 These are schematic diagrams of structures of other embodiments of the spinal disease identification system based on relation awareness and condition classification according to this disclosure; Figure 3 This is a front-end user interface diagram of a spinal disease identification system based on relation awareness and condition classification according to this disclosure; Figure 4 This is a sample training data structure diagram of a spinal disease identification system based on relation awareness and conditional classification according to this disclosure. Detailed Implementation

[0012] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0013] It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other.

[0014] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0015] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0016] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0017] This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0018] Please see Figure 1 , Figure 1 This is a schematic diagram of the structure of some embodiments of the spinal disease recognition system based on relation awareness and conditional classification disclosed herein. The spinal disease recognition system based on relation awareness and conditional classification includes: a data processing server 101, a feature extraction server 102, a disease recognition server 103, a report generation server 104, and a user interaction terminal 105.

[0019] Data processing server 101 is a server used to perform pre-defined standardization processing on spinal disease patient data to generate standard imaging data and standard clinical data.

[0020] In some embodiments, the data processing server 101 can perform pre-defined standardization processing on spinal disease patient data, which includes medical images and clinical text data of spinal diseases, to generate standard image data and standard clinical data. The data processing server 101 may include, but is not limited to, at least one of the following: a processor, memory, a database, and a dedicated data integration module (ETL engine, API gateway, etc.). The spinal disease medical images may include: DICOM format computed tomography (CT) images or magnetic resonance imaging (MRI) images of the spine. The clinical text data may include: structured text data (e.g., JSON format data, including: patient age, gender, chief complaint, past medical history, laboratory test values) and unstructured text data (e.g., free text clinical notes, pathology reports). The specific structure of spinal disease patient data can be found in [reference needed]. Figure 4 Example training data structure diagram of a spinal disease identification system based on relation awareness and conditional classification according to this disclosure.

[0021] In some optional implementations of certain embodiments, the data processing server 101 described above can be configured to perform the following steps: The first step is to resample the aforementioned computed tomography (CT) images and magnetic resonance imaging (MRI) images to generate a spatially standardized medical image set for spinal diseases.

[0022] In practice, firstly, the aforementioned implementing entity can resample the computed tomography (CT) images using trilinear interpolation to obtain interpolated CT images. Then, it can resample the magnetic resonance imaging (MRI) images using B-spline interpolation to obtain interpolated MRI images. Finally, the interpolated CT images and the interpolated MRI images are combined to form a spatially normalized medical image set for spinal diseases. The interpolated MRI images may include T1-weighted images (T1WI) and T2-weighted images (T2WI).

[0023] The second step is to normalize the aforementioned spatially standardized medical image set of spinal diseases to generate standard image data.

[0024] In practice, firstly, the aforementioned implementing entity can use Z-score normalization to normalize the interpolated computed tomography (CT) images in the spatially standardized spinal disease medical image set, thereby standardizing grayscale values ​​and obtaining normalized CT images. Then, histogram matching can be used to normalize the T1-weighted images in the interpolated MRI images, resulting in normalized T1-weighted MRI images. Secondly, quantile normalization (QN) is used to normalize the T2-weighted images in the interpolated MRI images, resulting in normalized T2-weighted MRI images. Finally, the normalized CT images, the normalized T1-weighted MRI images, and the normalized T2-weighted MRI images are identified as standard image data.

[0025] In some alternative implementations of certain embodiments, the data processing server 101 described above may be further configured to perform the following steps: The first step involves parsing and encoding the structured text data to generate a structured clinical feature vector. In practice, the data processing server 101 can extract key fields (e.g., patient age, gender, chief complaint, past medical history, laboratory test values) from the structured text data to obtain a set of key fields, and then perform one-hot encoding on the set of key fields to obtain the structured clinical feature vector.

[0026] The second step involves cleaning the unstructured text data to generate cleaned unstructured data. In practice, the data processing server 101 can first remove special characters (e.g., !, @, #) from the unstructured text data using regular expressions. The data processing server 101 can then use a predefined term mapping table to replace the original terms in the unstructured text data without special characters with standardized terms, resulting in cleaned unstructured data.

[0027] The third step involves extracting key entities from the cleaned unstructured data to generate clinical entity vectors. In practice, firstly, the data processing server 101 can use a pre-trained model in the biomedical field (a general medical language representation model, such as BioBERT) to generate vectors from the cleaned unstructured data, producing context-sensitive vector representations (such as 768-dimensional vectors). Secondly, a pre-trained Named Entity Recognition (NER) model is used to extract key clinical entities (such as disease names, anatomical locations, and symptom descriptions) from the context-sensitive vector representations, resulting in clinical entity vectors. The pre-trained Named Entity Recognition (NER) model can be a recognition model pre-trained based on de-identified medical corpora such as PubMed or MIMIC-III. For example, the pre-trained Named Entity Recognition (NER) model could be BioBERT-Base.

[0028] The fourth step is to determine the above-mentioned structured clinical feature vectors and clinical entity vectors as standard clinical data.

[0029] Feature extraction server 102 is used to extract and segment features from standard image data to generate a set of vertebral node differential feature vectors and a set of medical image segmentation masks for spinal diseases. For specific procedures, please refer to [reference needed]. Figure 2 The following is a schematic diagram of the structure of some other embodiments of the spinal disease identification system based on relation awareness and condition classification according to the present disclosure.

[0030] In some embodiments, the feature extraction server 102 can perform feature extraction and segmentation on the standard image data received from the data processing server based on a preset three-dimensional convolutional neural network to generate a set of vertebral node differential feature vectors and a set of medical image segmentation masks for spinal diseases. The feature extraction server 102 may include, but is not limited to, at least one of the following: a graphics processing unit (GPU) and a feature extraction engine. The preset three-dimensional convolutional neural network may be a three-dimensional convolutional neural network that performs three-dimensional feature extraction on all vertebrae in CT / MRI images and outputs a high-precision three-dimensional segmentation mask. For example, the preset three-dimensional convolutional neural network may be a lightweight 3D U-Net. Here, the preset three-dimensional convolutional neural network includes an encoder and a decoder.

[0031] In some alternative implementations of certain embodiments, the feature extraction server 102 described above can be configured to perform the following steps: The first step is to generate a high-dimensional spinal disease feature map, a set of vertebral nodes, a set of vertebral node feature vectors, and a set of vertebral edge relationships based on standard image data.

[0032] In practice, the aforementioned feature extraction server 102 can input the standard image data into a preset three-dimensional convolutional neural network, and obtain a high-dimensional spinal disease feature map after the second downsampling layer in the encoder path of the preset three-dimensional convolutional neural network. The encoder of the preset three-dimensional convolutional neural network can include convolution block 1 (3D convolution, batch normalization, and ReLU activation function), downsampling 1 (max pooling), convolution block 2 (3D convolution, batch normalization, and ReLU activation function), and downsampling 2 (max pooling). The high-dimensional spinal disease feature map contains the location of the vertebral bodies, and each spatial location corresponds to a feature vector.

[0033] The aforementioned feature extraction server 102 can determine the position of each vertebra in the spine based on the location of all vertebrae in the high-dimensional spinal disease feature map. The corresponding center point of the vertebral body is used as the vertebral node. This yields the set of vertebral nodes.

[0034] The aforementioned feature extraction server 102 can generate a set of vertebral node feature vectors and a set of vertebral edge relationships based on the aforementioned high-dimensional spinal disease feature map and the aforementioned set of vertebral nodes. First, the feature extraction server 102 can extract a 3×3×3 neighborhood voxel feature cube based on the position of each vertebral node in the aforementioned set of vertebral nodes, and compress the 3×3×3 neighborhood voxel feature cube into a vector using max pooling, which serves as the vertebral node feature vector, thus obtaining the set of vertebral node feature vectors. Second, based on the spinal structure, the feature extraction server 102 can establish undirected edges for each vertebral node in the aforementioned set of vertebral nodes, connecting only adjacent vertebrae, and determine the coordinate difference vector between two vertebral nodes in each undirected edge as the vertebral edge relationship, thus obtaining the set of vertebral edge relationships.

[0035] The second step involves aggregating the mean values ​​of the vertebral node feature vector set based on the aforementioned set of vertebral edge relationships to generate a neighborhood mean feature set. In practice, the feature extraction server 102 can perform the following formula on the vertebral nodes... neighborhood The mean of the feature vectors of at least one corresponding vertebral node is aggregated: .

[0036] in, , Indicates different vertebrae, . This indicates the total number of vertebrae. Represents the vertebral body The neighborhood (e.g., the neighborhood of vertebral body L3 is vertebral body L2 and vertebral body L4). Represents the vertebral body The vertebral node feature vector. Represents the vertebral body The neighborhood mean feature. As an example, the feature vector of the cone node. The neighborhood mean characteristic is vertebral node feature vector The neighborhood mean characteristic is ( + ) / 2.

[0037] The third step involves constructing an undirected graph structure based on the aforementioned set of vertebral nodes and edge relationships, and generating a set of vertebral attention weight vectors based on this undirected graph structure. In practice, due to the bidirectional nature of biomechanical relationships, the feature extraction server 102 can construct an undirected graph based on the aforementioned set of vertebral nodes and edge relationships. (Undirected graph structure) The model is constructed based on "vertebral body center point" and "anatomical connectivity" to ensure that it understands long-range dependencies (such as spinal curves). For example, the set of vertebral body nodes. vertebral body side relationship Undirected graph structure .

[0038] The aforementioned feature extraction server 102 can determine the attention weight of each vertebral node feature vector to its adjacent vertebral node feature vectors in the aforementioned vertebral node feature vector set based on the Graph Attention Network (GAT) mechanism and the aforementioned undirected graph structure, using the following formula, to obtain the vertebral attention weight vector set: .

[0039] .

[0040] in, Represents the vertebral body The vertebral node feature vector. Represents the vertebral body The vertebral node feature vector. This represents the shared linear transformation matrix for pre-training. This represents a pre-trained learnable attention vector. This represents the transpose of the pre-trained learnable attention vector. Represents the vertebral body and vertebral body The relationship between the vertebral body and the lateral side. Indicates it belongs to the vertebral body Neighborhood The vertebral body. Represents the vertebral body Its neighborhood The cone attention weight vector. This represents the Leaky ReLU (Leaky Rectified Linear Unit) nonlinear activation function.

[0041] The fourth step involves generating a set of differentially embedded vectors based on the preset healthy vertebral prototype embedding vector, the aforementioned set of vertebral node feature vectors, and the aforementioned set of neighborhood mean feature vectors. In practice, the feature extraction server 102 can generate the set of differentially embedded vectors using the following formula. : .

[0042] .

[0043] in, Represents the vertebral body The difference embedding vector. Represents the vertebral body The vertebral node feature vector. This represents the preset healthy vertebral prototype embedding vector. This represents the set of differential embedding vectors. The aforementioned preset healthy vertebral prototype embedding vector is a global "healthy vertebral prototype" embedding vector obtained by the feature extraction server 102 during the training process by aggregating the features of healthy vertebrae in all training samples. This represents the connection string. This represents a first preset multilayer perceptron. The aforementioned first preset multilayer perceptron can be a small multilayer perceptron with an input dimension of 768 (…). The first preset multilayer perceptron can include a fully connected layer 1 and a fully connected layer 2. The output dimension of the fully connected layer 1 is 512, and it is activated by the ReLU function. The output dimension of the fully connected layer 2 is 256, and it has no activation function.

[0044] Fifth, based on the aforementioned set of differential embedding vectors, a differential embedding gating function is constructed. In practice, the feature extraction server 102 can be based on a second preset multilayer perceptron and the aforementioned set of differential embedding vectors. Construct differential embedding gate function The second preset multilayer perceptron can be a multilayer perceptron with the same structure as the first preset multilayer perceptron, except that its input dimension can be 256 and its output dimension can be 1. The fully connected layer 1 of the second preset multilayer perceptron has an output dimension of 64 and is activated by the ReLU function. The fully connected layer 2 of the second preset multilayer perceptron has an output dimension of 1 and is activated by the Sigmoid function. Difference embedding gate function. It is a vector of difference embeddings The control gating function. The above-mentioned differential embedding gating function can be: .

[0045] in, Represents the vertebral body The difference embedding vector. This represents the differential embedding gate function. This refers to the aforementioned second preset multilayer perceptron. This represents the Sigmoid activation function.

[0046] Step 6: Based on the aforementioned set of vertebral node feature vectors, the aforementioned set of vertebral attention weight vectors, and the aforementioned differential embedding gating function, determine the set of vertebral node differential feature vectors. In practice, the aforementioned feature extraction server 102 can determine the set of vertebral node differential feature vectors based on the aforementioned set of vertebral node feature vectors, the aforementioned set of vertebral attention weight vectors, and the aforementioned differential embedding gating function using the following formula: .

[0047] in, Represents the vertebral body The vertebral node difference feature vector. Represents the vertebral body and its neighboring areas The attention weight vector between the cones. Represents the vertebral body The vertebral node feature vector. Represents the vertebral body The vertebral node feature vector. Represents the vertebral body Difference embedding gate function. This represents element-wise multiplication. It is a linear transformation matrix used for neighbor information paths and self-difference information paths. This represents the neighborhood information transformation matrix during pre-training. This represents the self-feature transformation matrix of the pre-trained model. This represents the activation function (e.g., the ELU activation function).

[0048] Step 7: Based on the decoder of the preset 3D convolutional neural network, the aforementioned set of differential feature vectors of the vertebral nodes is decoded to generate a medical image segmentation mask set for spinal diseases. In practice, the aforementioned feature extraction server 102 can input the aforementioned differential feature vectors of the vertebral nodes into the decoder of the preset 3D convolutional neural network, and obtain the medical image segmentation mask set for spinal diseases after the output layer. The decoder of the aforementioned preset 3D convolutional neural network may include a feature remapping layer (1×1×1 convolution), an upsampling layer 1 (3D deconvolution), a skip connection layer (feature concatenation), an upsampling layer 2 (3D deconvolution), and an output layer (1×1×1 convolution and Sigmoid activation function).

[0049] The disease identification server 103 is a server used to generate a set of reconstructed images and spinal disease identification results based on a set of spinal disease symptoms, standard image data, and a set of differential feature vectors of vertebral nodes.

[0050] In some embodiments, the disease identification server 103 can generate a reconstructed image set based on a set of spinal lesions, a preset classifier, the standard image data received from the data processing server, and the set of vertebral node differential feature vectors received from the feature extraction server. It then determines a set of reconstruction errors in the reconstructed image set to perform spinal disease identification and obtain a spinal disease identification result. The disease identification server 103 may include, but is not limited to, at least one of the following: a processor, a memory, and a disease identification engine. The set of spinal lesions may include: degenerative lesions, neoplastic lesions, fractures, and infections. The preset classifier may be a classifier that classifies the standard image data. For example, the preset classifier may be a conditional diffusion model based on 3D U-Net. The aforementioned pre-defined classifier may include: an encoder (composed of multiple layers of downsampling blocks, each downsampling block containing a 3D ResNet block, a self-attention block, and a downsampling operation), a bottleneck layer (located between the encoder and the decoder, typically composed of one or more 3D ResNet blocks), and a decoder (basically symmetrical to the encoder, composed of multiple upsampling blocks, each upsampling block containing an upsampling operation, a feature concatenation operation, a 3D ResNet block, and a self-attention block).

[0051] In addressing the aforementioned technical problems, the following technical problem often arises: Traditional systems use simple category labels (such as "tumor" or "infection") as input conditions, failing to encode complex anatomical relationships (such as the topological connections between intervertebral discs and vertebral bodies). The system cannot distinguish between lesions with similar images but different anatomical mechanisms, requiring multiple calls to the diagnostic model for verification, resulting in long waiting times for disease identification results. Conventional solutions to this second technical problem typically involve traditional machine learning methods, medical image classification models such as U-Net and ResNet, and recognition methods based on VerFormer (ViT with vertebral body-aware attention). However, considering the shortcomings of these methods—lack of anatomical relationship modeling (local receptive fields struggle to capture global context and long-range dependencies), flattened anatomical structures (compressing the 3D spine into a 2D patch sequence, losing intervertebral disc-vertebral body topological connection information), and low sensitivity to anomalies (uniform attention mechanisms cannot focus on lesion areas, such as early tumor signals)—and considering the current state of available technology, the following solution can be adopted: In some alternative implementations of certain embodiments, the disease identification server 103 described above can be configured to perform the following steps: The first step involves adding Gaussian noise to the standard image data based on a preset time step to generate a noisy image set at that time step. In practice, the disease identification server 103 can add Gaussian noise to the standard image data based on a preset time step using the following formula to generate a noisy image set at that time step: .

[0052] in, Indicates time step Noisy images. Indicates the preset time step. The aforementioned preset time step can be a pre-defined unit duration. This indicates the maximum duration of the preset time step. This indicates the preset noise control parameters (to control noise intensity). Follow (It decreases as it increases). This indicates a standard image. Standard Gaussian noise ( , Represents a multivariate standard normal distribution. (Indicates unit covariance).

[0053] The second step involves creating a semantic embedding vector for each spinal symptom in the aforementioned set of spinal symptom conditions, thereby generating a spinal symptom vector set. In practice, the disease identification server 103 can use a learnable embedding layer to map spinal symptom conditions to fixed-length vectors, creating a semantic embedding vector for each spinal symptom in the aforementioned set of spinal symptom conditions. The learnable embedding layer can be a pre-trained embedding layer that maps discrete category identifiers (such as text, ID) to continuous low-dimensional dense vectors.

[0054] The third step is to pool the aforementioned set of differential feature vectors of the vertebral nodes to generate a high-dimensional feature set. In practice, the disease recognition server 103 can use global max pooling to pool the aforementioned set of differential feature vectors of the vertebral nodes to generate a high-dimensional feature set.

[0055] The fourth step involves generating a joint conditional vector based on the aforementioned high-dimensional feature set and the aforementioned set of spinal disease vectors. In practice, the disease identification server 103 can concatenate the aforementioned high-dimensional feature set and the aforementioned set of spinal disease vectors (e.g., using the concat function) to generate the joint conditional vector.

[0056] The fifth step involves generating a reconstructed image set based on a preset classifier, the aforementioned noisy image set at the aforementioned time step, the aforementioned joint conditional vector, and the aforementioned preset time step. In practice, the aforementioned disease recognition server 103 can input the aforementioned noisy image set at the aforementioned time step, the aforementioned joint conditional vector, and the aforementioned preset time step into the preset classifier for denoising and obtain the reconstructed image set.

[0057] The sixth step is to determine the reconstruction error of each reconstructed image in the above reconstructed image set, thus obtaining a reconstruction error set. In practice, the disease recognition server 103 can quantify the reconstruction error of each reconstructed image in the above reconstructed image set and use the error to reflect the diagnostic matching degree. .

[0058] in, Indicates noisy images Vector of spinal diseases The reconstructed image corresponding to the joint conditional vector. This indicates a standard image. This indicates the reconstruction error.

[0059] Step 7: Based on the aforementioned set of reconstruction errors, generate spinal disease identification results. In practice, the disease identification server 103 can select the reconstruction errors from the aforementioned set of reconstruction errors. Spinal disorders corresponding to the minimum value : .

[0060] in, This indicates the results of spinal disease identification. The above spinal disease identification results... It belongs to the aforementioned set of vectors for spinal diseases. This indicates finding the factor that reduces the reconstruction error. Minimize spinal diseases .

[0061] The first to seventh steps described above are an inventive point of this disclosure, solving technical problem two: "Traditional systems use simple category labels (such as 'tumor' or 'infection') as conditional inputs, which cannot encode complex anatomical relationships (such as the topological connection between intervertebral discs and vertebral bodies). The system cannot distinguish lesions with similar images but different anatomical mechanisms, leading to multiple calls to the diagnostic model for verification and resulting in long waiting times for displaying disease identification results." The reasons for this are as follows: Traditional models use simple category labels (such as 'tumor' or 'infection') as conditional inputs, which cannot encode complex anatomical relationships (such as the topological connection between intervertebral discs and vertebral bodies), and the model cannot distinguish lesions with similar images but different anatomical mechanisms. Solving these factors can reduce the waiting time for displaying disease identification results. To achieve this, firstly, based on a preset time step, Gaussian noise is added to the aforementioned standard image data to generate a time-step noisy image set. Thus, through time-step-driven controllable noise injection, the dynamic evolution of lesions in the images is simulated, providing standardized input for the diffusion model. Second, a semantic embedding vector is created for each spinal symptom in the aforementioned spinal symptom set to generate a spinal symptom vector, resulting in a spinal symptom vector set. This transforms medical symptom names into machine-understandable semantic vectors, constructing a mathematical representation space for the symptom. Third, the aforementioned vertebral node differential feature vector set is pooled to generate a high-dimensional feature set. This aggregates vertebral-level local features into a global spinal representation, extracting key anatomical information for decision-making. Fourth, a joint condition vector is generated based on the aforementioned high-dimensional feature set and the aforementioned spinal symptom vector set. This fuses anatomical features and symptom semantics to generate a joint condition signal guiding the diffusion model reconstruction. Fifth, a reconstructed image set is generated based on a preset classifier, the aforementioned noisy image set at the aforementioned time step, the aforementioned joint condition vector, and the aforementioned preset time step. This performs inverse diffusion under the anatomical-symptom joint condition constraint to generate specific reconstructed images for the disease hypothesis. Sixth, the reconstruction error of each reconstructed image in the aforementioned reconstructed image set is determined, resulting in a reconstruction error set. This quantifies the matching differences between different disease hypotheses and the original images, generating an objective decision-making basis. Seventh, based on the aforementioned reconstruction error set, spinal disease identification results are generated. Therefore, based on the principle of minimizing reconstruction error, the lesion type is determined, and a quantifiable diagnostic confidence level is provided. Finally, combining the aforementioned report generation server 104 and the aforementioned user interaction terminal 105, the aforementioned spinal disease identification results are fused with the aforementioned standard clinical data and the aforementioned vertebral node differential feature vector set in a multimodal manner to obtain a structured spinal disease identification report, which is then displayed, thereby reducing the waiting time for displaying disease identification results.

[0062] The report generation server 104 is used to perform multimodal fusion of standard clinical data, vertebral node differential feature vector set and spinal disease identification results to obtain multimodal vectors, and generate a structured identification report of spinal diseases based on the multimodal vectors.

[0063] In some embodiments, the report generation server 104 can perform multimodal fusion on the standard clinical data received from the data processing server, the set of differential feature vectors of vertebral nodes received from the feature extraction server, and the spinal disease identification results received from the disease identification server to obtain a multimodal vector. Based on the multimodal vector, a parameter fine-tuning method, and a preset large language model, a structured identification report of spinal diseases can be generated. The report generation server 104 may include, but is not limited to, at least one of the following: a report generation engine.

[0064] In some alternative implementations of certain embodiments, the report generation server 104 described above can be configured to perform the following steps: The first step is to concatenate the above set of differential feature vectors of vertebral nodes, the above results of spinal disease identification, and the above standard clinical data to obtain a multimodal vector.

[0065] In practice, firstly, the report generation server 104 can perform pooling operations on the aforementioned set of differential feature vectors of the vertebral nodes to obtain the global features of the vertebral nodes: .

[0066] in, This represents the global features of the vertebral nodes. Indicates the vertebral body number. This indicates the total number of vertebrae. Represents the vertebral body The vertebral node difference feature vector.

[0067] Then, the global features of the vertebral nodes, the spinal disease identification results, and the standard clinical data are concatenated to obtain a multimodal vector.

[0068] The second step involves obtaining a large language model for spinal diseases based on the aforementioned low-rank adaptive method and the pre-defined large language model. In practice, firstly, the aforementioned report generation server 104 freezes all original parameters. Secondly, while maintaining the original weights from pre-training... Without changing the parameters, introduce two small, trainable low-rank matrices. and This reduces the number of parameters that need to be trained. Here, the model's weight updates are approximated as... The forward propagation of the model becomes: .

[0069] in, This represents the output of a structured identification report for spinal diseases. This indicates input. This represents the original weights from the pre-training process. This indicates the weight update of the model. Rank much smaller and .

[0070] Then, the training objective of the above-mentioned pre-defined large language model is set as follows: .

[0071] The aforementioned structured identification report for spinal diseases includes Each token. This indicates the lexical sequence number of the above-mentioned structured identification report of spinal diseases. This indicates the word length of the structured identification report for the aforementioned spinal disease. This indicates the first [item] in the above structured identification report of spinal diseases. Each word element. express All previous morphemes. Represents a low-rank matrix and Trainable parameters. This represents the aforementioned multimodal vector. It represents the log-likelihood probability. This represents the cross-entropy loss function.

[0072] Finally, the aforementioned multimodal vector and low-rank adaptive method is used to train the pre-defined large language model to obtain a large language model for spinal diseases. This pre-defined large language model can be used to output a medical recognition report based on input structured feature data and structured text data; for example, it could be Meta-Llama or Llama-3-70B-Instruct.

[0073] The third step involves inputting the aforementioned multimodal vectors into the aforementioned large-scale language model for spinal diseases to generate a structured identification report for spinal diseases. This structured identification report may include: a patient summary (e.g., 65-year-old male, with worsening chronic low back pain and radiating pain in the left lower extremity), detailed findings (e.g., 5mm posterior disc protrusion at L4-L5, compressing the left L5 nerve root (confidence 94%)), and differential diagnosis notes (based on uncertainty). Values ​​generate comments. For example, When the value is >0.15, add "spinal canal tumor cannot be ruled out" and a recommended treatment plan. (Uncertainty exists.) The smaller the value, the more uncertain the model is in distinguishing between the two options; that is, the higher the uncertainty. The aforementioned uncertainty... The values ​​are as follows: .

[0074] in, Indicates noisy images Results of spinal disease identification The reconstructed image corresponding to the joint conditional vector. This indicates a standard image. This represents the suboptimal identification result of spinal diseases (compared to the reconstruction errors in the above reconstruction error set). The second smallest value corresponds to the spinal condition. ). Indicates noisy images Suboptimal identification results of spinal diseases The reconstructed image corresponding to the joint conditional vector.

[0075] In some optional implementations of certain embodiments, the system further includes: User interaction terminal 105 is a terminal used to display medical images of spinal diseases, a segmentation mask set of medical images of spinal diseases, and a structured recognition report of spinal diseases, and to control at least one of a preset robot and a preset chemical dispensing device to perform preset operations based on electrical signals. For a specific user interaction terminal, please refer to [reference needed]. Figure 3 The diagram shows the front-end user interface of the spinal disease recognition system based on relation awareness and condition classification according to this disclosure.

[0076] In some embodiments, the user interaction terminal 105 can display the spinal disease medical images, the spinal disease medical image segmentation mask set received from the feature extraction server, and the spinal disease structured recognition report received from the report generation server, and control at least one of a preset robot and a preset chemical dispensing device to perform preset operations based on the received electrical signals. The report generation server 104 may include, but is not limited to, at least one of the following: a display, a memory, a user input unit, and a control interface.

[0077] In addressing the aforementioned technical problems using technical solutions, the following third technical issue often arises: Under conditions requiring high-precision operation (sub-millimeter tolerance scenarios) and timely emergency intervention (infection control, intraoperative massive bleeding, etc.), the user interface terminal lacks a rule engine to translate diagnostic conclusions into device commands, leading to delayed system emergency responses and decreased patient safety. A conventional solution to this third technical problem is that the user interface terminal only displays images of spinal disease lesions. Doctors must repeatedly switch between the imaging terminal (displaying the lesion) and the operating console (controlling the device). However, considering the drawback of this method causing a break in the diagnostic and treatment process, and taking into account the current technological capabilities, the following solution can be adopted: In some optional implementations of some embodiments, the user interaction terminal 105 described above is configured as follows: The first step involves fusing the aforementioned spinal disease medical images and the aforementioned spinal disease medical image segmentation mask set based on the alpha blending algorithm to obtain a spinal disease lesion tissue image. In practice, the aforementioned user interaction terminal 105 can use the following alpha blending formula to overlay the original image with the segmentation mask to highlight the lesion area: .

[0078] .

[0079] in, Images representing lesions in spinal diseases. This refers to medical images representing spinal diseases. This indicates the transparency of medical images related to spinal diseases. This represents a medical image segmentation mask set for spinal diseases. This indicates a lesion highlighted in red. This indicates element-wise multiplication.

[0080] The second step involves projecting the spinal disease lesion tissue images onto a three-dimensional coordinate system using WebGL for 3D display. In practice, the user interaction terminal 105 can establish a three-dimensional spatial coordinate system with the sacrum (S1 vertebra) as the origin using the medical imaging standard (RAS: Right, Anterior, Superior). Based on the geometric center coordinates of the vertebral segmentation mask set, each vertebra is mapped to three-dimensional space to preserve the true anatomical positional relationships. Then, each voxel of the spinal disease lesion tissue image (including healthy tissue and diseased areas) is converted into a three-dimensional mesh vertex to construct a triangular facet structure and generate vertex normal vectors to simulate realistic lighting reflection effects, thereby enhancing the three-dimensionality of the anatomical structure.

[0081] The third step is to display the structured identification report of the spinal disease as text. In practice, the user interaction terminal 105 can display the structured identification report of the spinal disease as text on the display screen.

[0082] The fourth step involves generating interactive operation buttons based on the aforementioned structured spinal disease identification report and predefined treatment rule base. In practice, the user interaction terminal 105 can generate interactive operation buttons based on the aforementioned structured spinal disease identification report and predefined treatment rule base. The user interaction terminal 105 can then display these interactive operation buttons on a monitor. These interactive operation buttons include: treatment plans for degenerative diseases, fracture treatment plans, tumor treatment plans, and infection treatment plans. The predefined treatment rule base can be a rule base that includes a set of spinal diseases and their corresponding treatment plans.

[0083] Fifth step: In response to receiving the trigger electrical signal of the aforementioned interactive operation button, control at least one of the preset robot and preset chemical dispensing device to perform a preset operation according to the type of the aforementioned interactive operation button, including the following sub-steps, wherein the aforementioned preset robot includes a first preset robot and a second preset robot, and the aforementioned preset chemical dispensing device includes a first preset chemical dispensing device and a second preset chemical dispensing device: In the first sub-step, in response to the interactive operation button representing the treatment plan for degenerative lesions, the lesion location is determined based on the aforementioned spinal disease medical image segmentation mask set, and a first preset robot is controlled to perform vertebral stabilization operations based on the lesion location. The first preset robot can be a robot capable of performing vertebral stabilization operations (such as pedicle screw implantation or percutaneous resection). For example, the first preset robot could be the Mazor X Stealth Edition (Medtronic).

[0084] Sub-step two involves responding to the interactive operation button representing the fracture treatment plan, determining the vertebroplasty puncture path based on the aforementioned spinal disease medical image segmentation mask set, and controlling the second preset robot to perform bone cement injection based on the aforementioned vertebroplasty puncture path. In practice, the aforementioned user interaction terminal 105 can determine the vertebroplasty puncture path according to the following formula: .

[0085] in, This indicates the puncture path for vertebroplasty. The aforementioned second-preset robot can be a robot capable of performing bone cement injection according to the vertebroplasty puncture path. For example, the aforementioned second-preset robot can be a ROSA (Robot of Stereotactic Assistant) surgical robot.

[0086] Sub-step three involves responding to the interactive operation button representing the treatment plan for the tumorous lesion, determining the target drug concentration based on the aforementioned structured identification report of spinal diseases, and then invoking the first preset chemical dispensing device to synthesize a sustained-release formulation based on the target drug concentration. In practice, the aforementioned user interaction terminal 105 can determine the target drug concentration according to the following formula: .

[0087] in, This indicates the concentration of the targeted drug, and the unit can be... The aforementioned first pre-set chemical dispensing device can be a device capable of synthesizing sustained-release formulations based on the target drug concentration. For example, the aforementioned first pre-set chemical dispensing device can be the YuMi robot from ABB, Switzerland.

[0088] Sub-step four: In response to the interactive operation button representing the infection treatment plan, based on the aforementioned structured identification report of spinal diseases and the acquired pathogen identification report, a second preset chemical dispensing device is invoked to prepare antibiotic sustained-release microparticles. The acquired pathogen identification report may include the pathogen type (e.g., pathogen = "MRSA"). The aforementioned second preset chemical dispensing device may be a device capable of preparing antibiotic sustained-release microparticles according to the pathogen type (e.g., synthesizing vancomycin-PLGA microparticles based on MRSA resistance). For example, the aforementioned second preset chemical dispensing device may be an MIT sustained-release microparticle robot.

[0089] The first to fifth steps and related content described above, as an inventive point of this disclosure, solve technical problem three: "In situations requiring high-precision operation (sub-millimeter tolerance scenarios) and timely emergency intervention (infection control, intraoperative massive bleeding, etc.), the user interaction terminal lacks a rule engine to convert diagnostic conclusions into device instructions, leading to delayed system emergency response and decreased patient safety." Factors causing delayed system emergency response and decreased patient safety often include: the interaction terminal lacks a rule engine to convert diagnostic conclusions into device instructions, making it unable to generate executable operations. Solving these factors can achieve timely system emergency response and increase patient safety. To achieve this effect, firstly, based on the alpha fusion algorithm, the aforementioned spinal disease medical images and the aforementioned spinal disease medical image segmentation mask set are fused to obtain a spinal disease lesion tissue image. The lesion tissue area is then highlighted to provide input for subsequent visualization. Secondly, based on WebGL, the aforementioned spinal disease lesion tissue image is projected onto a three-dimensional coordinate system for three-dimensional display. This enables an interactive and intuitive display of the lesion's spatial location and morphology. Thirdly, the aforementioned structured spinal disease identification report is displayed as text. Therefore, the aforementioned structured identification report of spinal diseases is clearly presented in structured text form, providing doctors with clear diagnostic basis and treatment suggestions. Fourth, based on the aforementioned structured identification report of spinal diseases and a predefined treatment rule base, interactive operation buttons are generated. This allows diagnostic conclusions to be transformed into executable treatment plan options with a single click. Fifth, in response to the trigger electrical signal received from the aforementioned interactive operation buttons, at least one of the preset robot and preset chemical dispensing equipment is controlled to perform a preset operation, depending on the type of the interactive operation button. This allows the system to receive button instructions and drive the corresponding robot and dispensing equipment to perform precise treatment operations, completing the steps from decision-making to execution. Finally, this achieves timely emergency response and increases patient safety.

[0090] The various embodiments of this disclosure have the following beneficial effects: The spinal disease recognition system based on relational awareness and conditional classification, through some embodiments of this disclosure, models the biomechanical dependence between vertebrae and generates anatomical-aware reconstructions conditionally using high-dimensional graph embedding, providing an interpretable decision-making mechanism for large language models, reducing system display waiting time, and reducing redundant computational loops. Specifically, the system needs to call the diagnostic model multiple times for verification, resulting in long display waiting times for disease recognition results, and the spinal disease recognition report displayed on the user's interactive terminal is inconsistent with the facts, requiring the system to re-execute, causing redundant computational loops. The reason for this is the long display waiting time. Specifically, systems using convolutional neural networks (CNNs, such as U-Net) are limited by their inherent receptive field, and their performance significantly degrades when recognizing and accurately segmenting lesions with varying shapes and blurred boundaries, requiring multiple calls to the diagnostic model for verification. This leads to long display waiting times for disease recognition results. The system has redundant computational loops. Specifically, systems based on Transformers (such as the Vision Transformer ViT) lack a detailed understanding of the three-dimensional anatomical relationships of the spine and fail to fully utilize the topological information between vertebrae. Existing computer-aided diagnostic (CADx) systems lack the ability to integrate different information for comprehensive differential diagnosis and cannot automatically and seamlessly convert the quantitative analysis results of AI into high-quality diagnostic text. Spinal disease identification reports displayed on user terminals often do not match the facts, requiring the system to re-execute, resulting in redundant computational loops. Based on this, some embodiments of the spinal disease identification system based on relational awareness and conditional classification disclosed herein include a data processing server, a feature extraction server, a disease identification server, a report generation server, and a user terminal that are interconnected. First, the data processing server performs pre-defined standardization processing on spinal disease patient data to generate standard image data and standard clinical data. This standardization of raw medical images and clinical text generates image and clinical data in a unified format, providing high-quality input for downstream analysis. Second, the feature extraction server performs feature extraction and segmentation on the standard image data to generate a set of vertebral node differential feature vectors and a set of spinal disease medical image segmentation masks. Then, the vertebral node differential features and segmentation masks are extracted through a graph transformer differential information interpolator (GT-DII) to quantify spinal anatomical structural abnormalities. Then, the aforementioned disease identification server generates a set of reconstructed images and spinal disease identification results based on the spinal lesion set, the aforementioned standard image data, and the aforementioned vertebral node differential feature vector set. Thus, reconstructed images are generated based on the Conditional Diffusion Diagnostic Network (CD-DN), and reconstruction errors are calculated, achieving lesion type identification and uncertainty quantification.Next, the aforementioned report generation server performs multimodal fusion on the aforementioned standard clinical data, the aforementioned set of differential feature vectors for vertebral nodes, and the aforementioned spinal disease identification results received from the aforementioned disease identification server to obtain multimodal vectors. Based on these multimodal vectors, a structured spinal disease identification report is generated. Thus, a structured diagnostic report is generated by fusing multimodal data, and fine-tuned using the LoRA method based on a manually constructed instruction-based fine-tuning dataset, using frameworks such as llama-factory. Then, an actionable clinical insight is output using the fine-tuned dedicated large language model. Finally, combined with the aforementioned user interaction terminal, the aforementioned spinal disease medical images, the aforementioned spinal disease medical image segmentation mask set, and the aforementioned structured spinal disease identification report are displayed. At least one of a preset robot and a preset chemical dispensing device is controlled to perform preset operations based on electrical signals. Thus, the images, masks, and reports are visualized, and the robot and dispensing equipment are controlled via interactive buttons to perform precise treatment operations. This implementation reduces system display waiting time and redundant computational loops.

[0091] The above description is merely a selection of preferred embodiments of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A spinal disease identification system based on relation perception and conditional classification, wherein, The spinal disease identification system includes: a data processing server, a feature extraction server, a disease identification server, and a report generation server that are interconnected. The data processing server is used to perform pre-defined standardization processing on spinal disease patient data to generate standard imaging data and standard clinical data. The feature extraction server is used to extract and segment features from the standard image data to generate a set of vertebral node differential feature vectors and a set of medical image segmentation masks for spinal diseases. The disease identification server is used to generate a set of reconstructed images and spinal disease identification results based on the set of spinal diseases, the standard image data, and the set of differential feature vectors of vertebral nodes. The report generation server is used to perform multimodal fusion of the standard clinical data, the set of differential feature vectors of vertebral nodes and the spinal disease identification results to obtain multimodal vectors, and generate a structured identification report of spinal diseases based on the multimodal vectors.

2. The method according to claim 1, wherein, The system also includes: The user interaction terminal is used to display the spinal disease medical images, the spinal disease medical image segmentation mask set, and the spinal disease structured identification report, and to control at least one of the preset robot and preset chemical dispensing equipment to perform preset operations according to electrical signals.

3. The system according to claim 1, wherein, The spinal disease patient data includes: spinal disease medical imaging and clinical text data.

4. The system according to claim 3, wherein, The medical imaging of the spinal disease includes: computed tomography (CT) images and magnetic resonance imaging (MRI) images, and The data processing server is configured to: The computed tomography images and the magnetic resonance imaging images are resampled to generate a spatially standardized medical image set for spinal diseases; The spatially standardized medical image set of spinal diseases is normalized to generate standard image data.

5. The system according to claim 4, wherein, The clinical text data includes: structured text data and unstructured text data, and The data processing server is further configured to: The structured text data is parsed and encoded to generate structured clinical feature vectors; The unstructured text data is cleaned to generate cleaned unstructured data. Key entities are extracted from the cleaned unstructured data to generate clinical entity vectors; The structured clinical feature vector and the clinical entity vector are determined as standard clinical data.

6. The system according to claim 1, wherein, The report generation server is configured to: The vertebral node differential feature vector set, the spinal disease identification results, and the standard clinical data are concatenated to obtain a multimodal vector; Based on the low-rank adaptive method and the preset large language model, a large language model for spinal diseases is generated. The multimodal vectors are input into the large language model of spinal diseases to generate a structured recognition report of spinal diseases.

7. The system according to claim 1, wherein, The feature extraction server is configured to: Based on standard image data, a high-dimensional spinal disease feature map, a set of vertebral nodes, a set of vertebral node feature vectors, and a set of vertebral edge relationships are generated. Based on the set of vertebral edge relationships, the set of vertebral node feature vectors is aggregated by mean to generate a neighborhood mean feature set. Based on the set of vertebral nodes and the set of vertebral edge relationships, an undirected graph structure is constructed, and a set of vertebral attention weight vectors is generated according to the undirected graph structure. Based on the preset healthy vertebral body prototype embedding vector, the vertebral body node feature vector set, and the neighborhood mean feature set, a differential embedding vector set is generated; Based on the set of differential embedding vectors, a differential embedding gate function is constructed; Based on the set of vertebral node feature vectors, the set of vertebral attention weight vectors, and the differential embedding gating function, the set of vertebral node differential feature vectors is determined; A decoder based on a pre-defined three-dimensional convolutional neural network decodes the set of differential feature vectors of the vertebral nodes to generate a medical image segmentation mask set for spinal diseases.

Images