Medical system for accurately identifying and positioning cancer focus based on deep learning framework

Through the subgraph extraction and weight allocation technology of the deep learning framework, the problem of low processing efficiency of the existing technology in small amounts of labeled data is solved, and more efficient cancer lesions identification and location are achieved, and the adaptability and generalization ability of the model are improved.

CN118942090AInactive Publication Date: 2024-11-12HARBIN MEDICAL UNIV DAQING BRANCH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410996178.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-24
Publication Date
2024-11-12
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing cancer lesion positioning system is inefficient when processing a small amount of labeled data, and cannot adjust the graph feature weight, which affects the analysis efficiency.

Method used

Adopt a medical system based on a deep learning framework, including data preprocessing, feature extraction, model training, lesion recognition and positioning modules, and adjust the subgraph feature weights and features through subgraph extraction, GraphSAGE model, weight allocation and graph convolution layer modules. Fusion to improve the model's adaptability to a small amount of labeled data.

Benefits of technology

It improves the classification and identification effect of cancer lesions, reduces data demand, simplifies model parameters, improves model efficiency and generalization capabilities, and can more accurately identify and locate cancer lesions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118942090A_ABST
    Figure CN118942090A_ABST
Patent Text Reader

Abstract

The invention discloses a medical system for accurately identifying and positioning a cancer focus based on a deep learning framework. In the invention, the graph convolution layer of the graph convolution layer module in the model training module updates the representation of the nodes by aggregating the features around the nodes so as to capture the local and global feature information, thereby considering the reduction of the data demand and the simplification of the model parameters, so that the model can better adapt to the condition of a small amount of marked data, and the accuracy of the model training is improved. And meanwhile, a weight distribution mechanism and SIN are introduced, so that the classification and identification effects of the diseases are improved. By using a sub-graph matching technology and similarity measurement, local sub-graphs can be extracted from a disease image as input data, so that the demand of data volume is reduced. Meanwhile, the weight of the sub-graph features can be adjusted by introducing a weight distribution mechanism, so that the model can better adapt to the condition of a small amount of marked data. Therefore, the demanded quantity of data is reduced, the parameters of the model are simplified, and the efficiency and generalization ability of the model are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of cancer lesion localization, and specifically is a medical system based on a deep learning framework for accurately identifying and localizing cancer lesions. Background Art

[0002] Cancer lesion localization is a medical technique used to determine the specific location of a tumor in the body. This technique is essential for the treatment of cancer and the evaluation of treatment effectiveness. In this introduction, we will explore the principles, methods, and applications of cancer lesion localization. The principle of cancer lesion localization is based on imaging examinations of the human body. Doctors use various imaging techniques, such as X-rays, CT scans, MRIs, and PET scans, to obtain images of the patient's body. These images show the structure and function of organs and tissues, and doctors can determine the location and size of tumors by analyzing these images. Therefore, the process of cancer lesion localization requires image recognition and analysis.

[0003] However, common positioning systems cannot adjust the weights of graph features during use, which makes the model unable to adapt to the situation of a small amount of labeled data, affecting the overall analysis efficiency. Summary of the invention

[0004] The purpose of the present invention is to provide a medical system based on a deep learning framework for accurately identifying and locating cancer lesions in order to solve the above-mentioned problems.

[0005] The technical solution adopted by the present invention is as follows: a medical system for accurately identifying and locating cancer lesions based on a deep learning framework, comprising a power supply module, a data preprocessing module, a feature extraction module, a model training module, a lesion identification and positioning module, a data visualization module, a subgraph extraction module, a subgraph feature representation module, a training module, a weight allocation module, a feature fusion module and a graph convolution layer module, characterized in that: the output end of 1) is connected to the input end of the data preprocessing module, the output end of the data preprocessing module is connected to the input end of the feature extraction module, the output end of the feature extraction module is connected to the input end of the model training module, the output end of the model training module is connected to the input end of the lesion identification and positioning module, and the output end of the lesion identification and positioning module is connected to the input end of the data visualization module.

[0006] In a preferred embodiment, the feature extraction module is internally provided with a sub-graph extraction module and a sub-graph feature representation module, and the model training module is internally provided with a training module, a weight allocation module, a feature fusion module and a graph convolution layer module.

[0007] In a preferred embodiment, the data preprocessing module performs image denoising preprocessing on the medical image data. The data preprocessing module performs online detection of the thickness of the cable insulation layer based on visual recognition, and estimates the range of noise using a least squares filter. This method has many advantages, such as a simple algorithm and good performance. When this method is used for valuation, 16 frames of images are collected continuously, and different data are collected for a single pixel. After filtering using the least squares method, an original image with less interference can be obtained. While restoring the clarity of the original image, some noise can be removed.

[0008] In a preferred embodiment, the sub-image extraction module uses the ROI detection algorithm to determine the position and size of the lesion area for each lesion image, and extracts the local disease sub-image according to a certain size or range; this can reduce the area scope processed by the model and improve the model's attention to disease features and recognition accuracy.

[0009] In a preferred embodiment, the subgraph feature representation module converts each subgraph into a feature vector representation, and integrates the information of the neighbor nodes into the feature representation of the current node by propagating and aggregating the features of the neighbor nodes to capture the feature relationship of the local field. CNN is used for feature extraction, that is, the subgraph is taken as input, and features are extracted through convolution and pooling layers to finally obtain a subgraph feature vector of fixed dimension; these subgraph feature vectors can be passed as input to the subsequent GraphSAGE model, and the subgraph feature vectors will transmit and aggregate information with the feature vectors of other subgraphs. Such a feature representation method can better capture the spatial information and structural characteristics of the disease.

[0010] In a preferred embodiment, during the training process, the training module does not use the pre-trained weights of ImageNet for initialization in the backbone network as a text encoder and the backbone network as an image encoder, but adopts a random initialization strategy, and fills the parameters of all linear layers with truncated normal distribution, and all initial bias parameters are set to 0;

[0011] The data preprocessing module uses linear projection on the feature vector outputs of the text encoder and the image encoder to project the feature vectors of text and images into a common embedding space.

[0012] The feature extraction module ensures that images and texts are paired one by one, and one image corresponds to only one sentence or a description text;

[0013] In the training phase, the model training module sets random image enhancement, including random cropping, random affine transformation, random horizontal flipping, random color jittering, and random rotation as possible image enhancement transformation operations. For each training image, two enhancement modes are randomly selected for application. In the test phase, no image enhancement is used.

[0014] The lesion recognition and positioning module sets a learnable temperature parameter T that affects the model output score. It is initialized to 0.07 and will be truncated if it exceeds 100.

[0015] In a preferred embodiment, the weight assignment module uses PageRank value as part of the weight assignment mechanism to adjust the weight of the subgraph feature; based on the similarity metric value, the contribution of the subgraph feature to the overall feature representation is increased or decreased.

[0016] In a preferred embodiment, the feature fusion module calibrates the input image features through a channel attention module, dynamically adjusts the contribution of different channel features to the overall features of the image, obtains weighted channel features, sums all channels, and obtains aggregated image features for participating in the loss function calculation;

[0017] The input text features will calculate the cosine similarity with the image channel features after channel attention weighting. The calculation result of the positive text features is represented as the channel positive score of the image, and the calculation result of the negative text features is represented as the channel negative score of the image.

[0018] The positive score of the image is obtained by summing the positive scores of all channels of the image, and the negative score of the image is obtained by summing the negative scores of all channels of the image. For a positive and negative text feature pair, the positive and negative scores of the image calculated for this image are mapped to positive probability and negative probability through the softmax function, where the sum of the positive probability and the negative probability is 1. The positive probability and the negative probability are compared. If the positive probability is greater than the negative probability, the pathological category corresponding to this positive and negative text feature pair in the prediction result is binary classified as 1, otherwise it is binary classified as 0.

[0019] In a preferred embodiment, the graph convolution layer module randomly selects m adjusted subgraph features as input data and passes them to the GraphSAGE model to reduce the parameters of the SIN module to reduce the computational complexity. The graph convolution layer module utilizes the transitivity of the graph structure to effectively transfer and learn the contextual information and associations between local subgraphs to learn a more global feature representation for disease identification and classification tasks. In the graph convolution network, each subgraph is used as a node, and the connection relationship between subgraphs is used as an edge. Features are extracted and classified through multiple layers of graph convolution layers.

[0020] In a preferred embodiment, the lesion identification and positioning module uses the trained model to classify and identify new disease images to obtain disease category labels; due to the extraction and feature representation of local sub-graphs, the model has a certain adaptability to new disease types or variants; at the same time, through the analysis and interpretation of sub-graph features, the interpretability of the model is improved, so that users can understand the decision-making process of the model;

[0021] The data visualization module visualizes the positioning and identification results to doctors or pathologists to help them analyze and diagnose the lesions.

[0022] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0023] 1. In the present invention, the graph convolution layer of the graph convolution layer module inside the model training module updates the representation of the node by aggregating the features around the node, thereby capturing local and global feature information, thereby taking into account the reduction of data demand and the simplification of model parameters, so that the model can better adapt to the situation of a small amount of labeled data. At the same time, a weight distribution mechanism and SIN are introduced on the basis of the GraphSAGE model, aiming to improve the classification and recognition effect of diseases. By using subgraph matching technology and similarity measurement, local subgraphs can be extracted from disease images as input data, thereby reducing the demand for data volume. At the same time, the introduction of a weight distribution mechanism can adjust the weights of subgraph features, so that the model can better adapt to the situation of a small amount of labeled data. This not only reduces the demand for data, but also simplifies the parameters of the model and improves the efficiency and generalization ability of the model.

[0024] 2. In the present invention, the feature fusion module uses the pre-training results on large-scale natural data sets to streamline the trainable network structure, and can adaptively focus on key local areas in the image, thereby improving the interpretability of the algorithm. The present invention can not only improve the accuracy of zero-sample classification of medical images, but also reduce the cost of application deployment, and conveniently, quickly and accurately identify chest lesion areas as imaging features of early diseases, provide a basis for clinical diagnosis of diseases, help quantitatively analyze the progression of disease development and the benign and malignant nature of lesion areas, and facilitate the development of objective diagnostic reference standards. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 is the overall system block diagram of the present invention;

[0026] Figure 2 This is a system block diagram of the feature extraction module in the present invention;

[0027] Figure 3 This is a system block diagram of the model training module in the present invention.

[0028] Markings in the figure: 1-power supply module, 2-data preprocessing module, 3-feature extraction module, 4-model training module, 5-lesion identification and positioning module, 6-data visualization module, 7-sub-graph extraction module, 8-sub-graph feature representation module, 9-training module, 10-weight allocation module, 11-feature fusion module, 12-graph convolutional layer module. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0030] Reference Figure 1-3 ,

[0031] Example:

[0032] A medical system for accurately identifying and locating cancer lesions based on a deep learning framework, comprising a power supply module 1, a data preprocessing module 2, a feature extraction module 3, a model training module 4, a lesion identification and positioning module 5, a data visualization module 6, a subgraph extraction module 7, a subgraph feature representation module 8, a training module 9, a weight distribution module 10, a feature fusion module 11 and a graph convolution layer module 12, wherein the output end of 1 is connected to the input end of the data preprocessing module 2, the output end of the data preprocessing module 2 is connected to the input end of the feature extraction module 3, the output end of the feature extraction module 3 is connected to the input end of the model training module 4, the output end of the model training module 4 is connected to the input end of the lesion identification and positioning module 5, and the output end of the lesion identification and positioning module 5 is connected to the input end of the data visualization module 6.

[0033] The feature extraction module 3 is internally provided with a sub-graph extraction module 7 and a sub-graph feature representation module 8, and the model training module 4 is internally provided with a training module 9, a weight allocation module 10, a feature fusion module 11 and a graph convolution layer module 12.

[0034] The data preprocessing module 2 performs image denoising preprocessing on the medical image data. The data preprocessing module 2 performs online detection of the thickness of the cable insulation layer based on visual recognition, and estimates the range of noise using a least squares filter. It has many advantages such as a simple algorithm and good performance. When using this method for valuation, 16 frames of images are collected continuously, and different data are collected for a single pixel. After filtering using the least squares method, an original image with less interference can be obtained. While restoring the clarity of the original image, some noise can be removed.

[0035] The sub-image extraction module 7 uses ROI detection algorithms such as Faster R-CNN to determine the location and size of the lesion area for each lesion image, and extracts local disease sub-images according to a certain size or range. This can reduce the area range processed by the model and improve the model's attention to disease features and recognition accuracy.

[0036] Subgraph feature representation module 8 converts each subgraph into a feature vector representation, and integrates the information of neighbor nodes into the feature representation of the current node by propagating and aggregating the features of neighbor nodes to capture the feature relationship of the local field. CNN is used for feature extraction, that is, the subgraph is taken as input, and features are extracted through convolution and pooling layers, and finally a subgraph feature vector of fixed dimension is obtained. These subgraph feature vectors can be passed as input to the subsequent GraphSAGE model, and the subgraph feature vectors will transmit and aggregate information with the feature vectors of other subgraphs. This feature representation method can better capture the spatial information and structural characteristics of the disease.

[0037] During the training process of training module 9, neither the backbone network used as the text encoder nor the backbone network used as the image encoder uses the ImageNet pre-trained weights for initialization. Instead, they all use a random initialization strategy and fill all the parameters of the linear layers with a truncated normal distribution. The initial bias parameters are all set to 0.

[0038] The data preprocessing module 2 uses linear projection for the feature vector outputs of the text encoder and the image encoder to project the feature vectors of the text and the image into a common embedding space.

[0039] The feature extraction module 3 ensures that the image and the text are matched one by one, and one image corresponds to only one sentence or a paragraph of descriptive text.

[0040] Model training module 4 sets random image enhancement in the training phase. Random cropping, random affine transformation, random horizontal flipping, random color jittering, and random rotation are possible image enhancement transformation operations. For each training image, two enhancement modes are randomly selected for application. In the testing phase, no image enhancement is used.

[0041] The lesion identification and positioning module 5 sets a learnable temperature parameter T that affects the model output score, which is initialized to 0.07 and will be truncated if it exceeds 100.

[0042] The weight assignment module 10 uses the PageRank value as part of the weight assignment mechanism to adjust the weight of the subgraph feature, and increases or decreases the contribution of the subgraph feature to the overall feature representation according to the similarity metric value.

[0043] The feature fusion module 11 calibrates the input image features through a channel attention module, dynamically adjusts the contribution of different channel features to the overall features of the image, obtains the weighted channel features, sums up all channels, and obtains the aggregated image features for participating in the loss function calculation.

[0044] The input text features will calculate the cosine similarity with the image channel features after channel attention weighting. The calculation result of the positive text features is represented as the channel positive score of the image, and the calculation result of the negative text features is represented as the channel negative score of the image.

[0045] The positive score of the image is obtained by summing the positive scores of all channels of the image, and the negative score of the image is obtained by summing the negative scores of all channels of the image. For a positive and negative text feature pair, the positive and negative scores of the image calculated for this image are mapped to positive probability and negative probability through the softmax function, where the sum of the positive probability and the negative probability is 1. Compare the positive probability and the negative probability. If the positive probability is greater than the negative probability, the pathological category corresponding to this positive and negative text feature pair in the prediction result is binary classified as 1, otherwise it is binary classified as 0.

[0046] The graph convolution layer module 12 randomly selects m adjusted subgraph features as input data and passes them to the GraphSAGE model to reduce the parameters of the SIN module to reduce the computational complexity. The graph convolution layer module 12 uses the transitivity of the graph structure to effectively transfer and learn the contextual information and associations between local subgraphs to learn a more global feature representation for disease identification and classification tasks. In the graph convolution network, each subgraph is used as a node, and the connection relationship between subgraphs is used as an edge. Features are extracted and classified through multiple layers of graph convolution layers.

[0047] The lesion identification and positioning module 5 uses the trained model to classify and identify new disease images and obtain the disease category label. Due to the extraction and feature representation of local sub-graphs, the model has a certain adaptability to new disease types or variants. At the same time, through the analysis and interpretation of sub-graph features, the interpretability of the model is improved, allowing users to understand the decision-making process of the model;

[0048] The data visualization module 6 visualizes the positioning and recognition results to doctors or pathologists to help them analyze and diagnose the lesions.

[0049] In the present invention, the graph convolution layer of the graph convolution layer module inside the model training module updates the representation of the node by aggregating the features around the node, thereby capturing local and global feature information, thereby taking into account the reduction of data demand and the simplification of model parameters, so that the model can better adapt to the situation of a small amount of labeled data, and at the same time introduces a weight distribution mechanism and SIN on the basis of the GraphSAGE model, aiming to improve the classification and recognition effect of diseases. By using subgraph matching technology and similarity measurement, local subgraphs can be extracted from disease images as input data, thereby reducing the demand for data volume. At the same time, the introduction of a weight distribution mechanism can adjust the weights of subgraph features, so that the model can better adapt to the situation of a small amount of labeled data. This not only reduces the demand for data, but also simplifies the parameters of the model, and improves the efficiency and generalization ability of the model.

[0050] In the present invention, the feature fusion module uses the pre-training results on large-scale natural data sets to streamline the trainable network structure, can adaptively focus on key local areas in the image, and improve the interpretability of the algorithm. The present invention can not only improve the accuracy of zero-sample classification of medical images, but also reduce the cost of application deployment, conveniently, quickly and accurately identify chest lesion areas as imaging features of early diseases, provide a basis for clinical diagnosis of diseases, help quantitatively analyze the progression of disease development and the benign and malignant nature of lesion areas, and facilitate the development of objective diagnostic reference standards.

[0051] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or still includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "including one..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0052] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A medical system for accurately identifying and locating cancer lesions based on a deep learning framework, comprising a power supply module (1), a data preprocessing module (2), a feature extraction module (3), a model training module (4), a lesion identification and positioning module (5), a data visualization module (6), a subgraph extraction module (7), a subgraph feature representation module (8), a training module (9), a weight distribution module (10), a feature fusion module (11) and a graph convolution layer module (12), Features: The output end of the aforementioned 1) is connected to the input end of the aforementioned data preprocessing module (2), the output end of the aforementioned data preprocessing module (2) is connected to the input end of the aforementioned feature extraction module (3), the output end of the aforementioned feature extraction module (3) is connected to the input end of the aforementioned model training module (4), the output end of the aforementioned model training module (4) is connected to the input end of the aforementioned lesion identification and positioning module (5), and the output end of the aforementioned lesion identification and positioning module (5) is connected to the input end of the aforementioned data visualization module (6).

2. A medical system for accurately identifying and locating cancer lesions based on a deep learning framework as claimed in claim 1, Features: The feature extraction module (3) is internally provided with a subgraph extraction module (7) and a subgraph feature representation module (8), and the model training module (4) is internally provided with a training module (9), a weight allocation module (10), a feature fusion module (11) and a graph convolution layer module (12).

3. A medical system for accurately identifying and locating cancer lesions based on a deep learning framework as claimed in claim 1, Features: The data preprocessing module (2) performs image denoising preprocessing on the medical image data. The data preprocessing module (2) performs online detection of the thickness of the cable insulation layer based on visual recognition and estimates the range of the noise by using a least squares filter. It has many advantages such as simple algorithm and good performance. When using this method for valuation, 16 frames of images are collected continuously, different data are collected for a single pixel, and after filtering using the least squares method, the original image with less interference can be obtained. While restoring the clarity of the original image, some noise can be removed.

4. A medical system for accurately identifying and locating cancer lesions based on a deep learning framework as claimed in claim 1, Features: The sub-image extraction module (7) uses a ROI detection algorithm (such as Faster R-CNN) to determine the location and size of the lesion area for each lesion image, and extracts a local disease sub-image according to a certain size or range; this can reduce the area range processed by the model and improve the model's attention to disease features and recognition accuracy.

5. A medical system for accurately identifying and locating cancer lesions based on a deep learning framework as claimed in claim 1, Features: The subgraph feature representation module (8) converts each subgraph into a feature vector representation, and integrates the information of the neighbor nodes into the feature representation of the current node by propagating and aggregating the features of neighbor nodes, so as to capture the feature relationship of the local field, and uses CNN for feature extraction, that is, taking the subgraph as input, extracting features through convolution and pooling layers, and finally obtaining a subgraph feature vector of fixed dimension; these subgraph feature vectors can be passed as input to the subsequent GraphSAGE model, and the subgraph feature vectors will transmit and aggregate information with the feature vectors of other subgraphs. Such a feature representation method can better capture the spatial information and structural characteristics of the disease.

6. A medical system for accurately identifying and locating cancer lesions based on a deep learning framework as claimed in claim 1, Features: During the training process, the training module (9) does not use the pre-trained weights of ImageNet for initialization in the backbone network used as the text encoder and the backbone network used as the image encoder. Instead, a random initialization strategy is adopted, and the parameters of all linear layers are filled with truncated normal distributions, and all initial bias parameters are set to 0. The data preprocessing module (2) uses linear projection to project the feature vectors of the text encoder and the image encoder into a common embedding space. The feature extraction module (3) ensures that images and texts are matched one by one, and one image corresponds to only one sentence or a paragraph of description text; The model training module (4) sets random image enhancement during the training phase, including random cropping, random affine transformation, random horizontal flipping, random color jittering, and random rotation as possible image enhancement transformation operations; For each training image, two enhancement modes are randomly selected and applied; in the test phase, no image enhancement is used; The lesion recognition and localization module (5) sets a learnable temperature parameter T that affects the model output score. It is initialized to 0.07 and will be truncated if it exceeds 100.

7. A medical system for accurately identifying and locating cancer lesions based on a deep learning framework as claimed in claim 1, Features: The weight distribution module (10) uses PageRank value as part of the weight distribution mechanism to adjust the weight of the subgraph feature; according to the similarity measurement value, the contribution of the subgraph feature to the overall feature representation is increased or decreased.

8. A medical system for accurately identifying and locating cancer lesions based on a deep learning framework as claimed in claim 1, Features: The feature fusion module (11) calibrates the input image features through a channel attention module, dynamically adjusts the contribution of different channel features to the overall features of the image, obtains weighted channel features, sums all channels, and obtains aggregated image features for participating in loss function calculation; The input text features will calculate the cosine similarity with the image channel features after channel attention weighting. The calculation result of the positive text features is represented as the channel positive score of the image, and the calculation result of the negative text features is represented as the channel negative score of the image. The positive score of the image is obtained by summing the positive scores of all channels of the image, and the negative score of the image is obtained by summing the negative scores of all channels of the image; For a positive and negative text feature pair, the positive and negative scores of the image calculated for this image are mapped to positive probability and negative probability through the softmax function, where the sum of the positive probability and the negative probability is 1; the positive probability and the negative probability are compared. If the positive probability is greater than the negative probability, the pathological category corresponding to this positive and negative text feature pair in the prediction result is binary classified as 1, otherwise it is binary classified as 0.

9. A medical system for accurately identifying and locating cancer lesions based on a deep learning framework as claimed in claim 1, Features: The graph convolution layer module (12) randomly selects m adjusted subgraph features as input data and passes them to the GraphSAGE model to reduce the parameters of the SIN module and reduce the computational complexity. The graph convolution layer module (12) utilizes the transitivity of the graph structure to effectively transfer and learn the context information and correlation between local subgraphs to learn a more global feature representation for disease identification and classification tasks. In the graph convolution network, each subgraph is used as a node and the connection relationship between subgraphs is used as an edge. Features are extracted and classified through multiple layers of graph convolution layers.

10. A medical system for accurately identifying and locating cancer lesions based on a deep learning framework as claimed in claim 1, Features: The lesion identification and positioning module (5) uses the trained model to classify and identify new disease images to obtain disease category labels; due to the extraction and feature representation of local sub-graphs, the model has a certain adaptability to new disease types or variants; at the same time, through the analysis and interpretation of sub-graph features, the interpretability of the model is improved, so that users can understand the decision-making process of the model; The data visualization module (6) visualizes the positioning and identification results to doctors or pathologists to help them analyze and diagnose the lesions.