Ovarian cancer database construction and intelligent analysis system
By constructing an ovarian cancer database, combining imaging and pathological images, using neural network models to extract features and generate high-precision heat maps, the technical difficulties of mapping pathological information on imaging images are solved, and the accuracy of ovarian cancer diagnosis is improved sex.
Patent Information
- Application Number
- CN202411183098.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-27
- Publication Date
- 2024-12-17
AI Technical Summary
In medical imaging diagnosis, especially in the diagnosis of ovarian cancer, how to accurately map key information in pathological images to imaging images and generate intuitive heat maps to show the location and severity of the lesion is still an unfinished solution. technical difficulties.
By constructing an ovarian cancer database, combining imaging images and pathological images, multi-source data acquisition module, data processing module, model training module and deployment application module are used to extract imaging and pathological features using neural network models to generate high-precision thermal picture.
Improve the accuracy of image analysis and form high-precision heat maps for doctors to analyze and diagnose, solving the challenge of converting pathological information into forms that clinicians can intuitively understand.
Smart Images
Figure CN119153067A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of computing software technology, and more particularly to an ovarian cancer database construction and intelligent analysis system. Background Art
[0002] Currently, in medical imaging diagnosis, especially for diseases such as ovarian cancer, researchers have been able to effectively identify key areas of suspected tumors and accurately assess the extent of cell changes through computer vision technology; however, how to map this key information from pathological images to imaging images and generate heat maps to intuitively display the location and severity of lesions is still a technical problem that has not yet been fully solved;
[0003] Although modern medical imaging technology can provide rich pathological information, it is still challenging to convert this information into a form that clinicians can intuitively understand. For example, pathological images provide detailed information about cell structure and tissue morphology, but for radiologists, this information usually needs to be combined with imaging images to better understand the specific location of the lesion and its scope of influence.
[0004] Therefore, how to accurately map the key information in the pathological image to the imaging image is an urgent problem to be solved by those skilled in the art. Summary of the invention
[0005] In view of this, the present invention provides an ovarian cancer database construction and intelligent analysis system, which constructs an ovarian cancer database based on imaging images and pathological images, and jointly trains macroscopic and microscopic features to improve the accuracy of image analysis, thereby forming a high-precision thermal map for doctors to analyze and diagnose.
[0006] In order to achieve the above object, the present invention adopts the following technical solution:
[0007] In a first aspect, the present invention provides an ovarian cancer database construction and intelligent analysis system, comprising:
[0008] The multi-source data acquisition module is used to acquire ovarian cancer data, wherein the ovarian cancer data includes imaging images and pathological images.
[0009] The data processing module is used to pair the imaging images with the corresponding set of pathological images, and to label the imaging images and the pathological images respectively to construct a database.
[0010] The model training module is used to sample the database, extract tumor areas based on imaging image samples, and extract pathological features from corresponding pathological image samples to calculate the degree of tumor deterioration; it supervises tumor extraction and pathological feature extraction based on the labels during annotation and optimizes model parameters.
[0011] Deploy an application module to use the trained model to identify tumor areas and tumor progression and generate heat maps.
[0012] Preferably, the data processing module includes a sample marking submodule and a label annotation submodule; the sample marking submodule is used to add sample marks according to the sources of the imaging image and the pathological image, and perform pairing according to the added sample marks;
[0013] The sample annotation module is used to annotate key feature areas in the imaging image and the pathological image.
[0014] Preferably, the sample mark includes a primary sample mark and a secondary sample mark, and the primary sample mark and the secondary sample mark are used to mark the imaging image and the pathological image, respectively.
[0015] Preferably, the sample marking submodule confirms the imaging image to which each pathological image belongs and the position in the corresponding imaging image according to the first sample marking and the second sample marking, and performs pairing processing.
[0016] Preferably, the neural network model trained by the model training module includes: a radiological image processing layer and a pathological image processing layer;
[0017] The imaging image processing layer is used to extract imaging features and identify tumor areas;
[0018] The pathological image processing layer is used to extract the pathological features of the tumor area and predict the degree of deterioration.
[0019] Preferably, the total loss function used by the model training module is:
[0020] L total =α1L cls +α2L reg +α3L score
[0021] Among them, α1, α2, and α3 are the weight factors of the corresponding losses; L total is the total loss, L cls is the tumor region classification loss, L reg is the tumor bounding box regression loss, L score Predict the loss for the degree of deterioration; the specific loss functions are as follows:
[0022] L cls =-α t (1-p t ) γ log(p t )
[0023] In the above formula, p t Represents the predicted probability, is the balance factor of positive and negative samples, and is a hyperparameter for controlling the balance of difficult and easy samples;
[0024]
[0025] In the above formula, x represents the deviation between the predicted boundary and the true bounding box; i represents the index of the predicted bounding box;
[0026]
[0027] In the above formula, L score Predict the loss for the deterioration degree, y score is the true value of the deterioration degree, is the predicted value of the degree of deterioration.
[0028] Preferably, the neural network model also includes a multimodal fusion layer, which fuses the imaging features and pathological features through a fusion network, and uses the labeled tumor location and deterioration degree information to adjust the parameters of the entire network through back propagation.
[0029] Preferably, the deployment and application module includes a model storage submodule, a data interface submodule, a visual configuration submodule and a visual output submodule;
[0030] The storage submodule is used to store the trained model;
[0031] The data interface submodule is used to configure the data source and input the data received by the data source into the trained model for intelligent analysis;
[0032] The visual submodule is used to configure corresponding visual expression effects for different deterioration levels;
[0033] The visualization output submodule is used to output a heat map according to the intelligent analysis results and the configuration parameters of the vision submodule.
[0034] In a second aspect, the present invention provides a model training method, which is used for the above-mentioned ovarian cancer database construction and intelligent analysis system, comprising the following steps:
[0035] S1: inputting the imaging images in the database into the pre-built first model, extracting imaging features through the first model, identifying the tumor area, calculating the loss according to the tumor area label and optimizing the first model, and fixing the model parameters of the first model when the first model meets the preset conditions;
[0036] S2: Output the tumor region through the trained first model, and take out the corresponding pathological image from the database according to the tumor region and output it to the pre-built second model;
[0037] S3: extracting pathological features through the second model and calculating the degree of tumor deterioration; calculating the loss and optimizing the second model according to the annotated deterioration degree label, and fixing the model parameters of the second model when the second model meets the preset conditions.
[0038] In a third aspect, the present invention provides a model training method, which is used for the above-mentioned ovarian cancer database construction and intelligent analysis system, comprising the following steps:
[0039] S1: inputting the imaging images in the database into the pre-built first model, and extracting imaging features through the first model to distinguish the tumor area from the background;
[0040] S2: extracting corresponding pathological images from the database according to the tumor area and outputting them to the pre-built second model;
[0041] S3: extracting pathological features through the second model and calculating the degree of tumor progression;
[0042] S4: Calculate the classification loss and bounding box regression loss of tumor region recognition based on the true label of the tumor region in the imaging image; calculate the deterioration degree prediction loss based on the tumor deterioration degree label;
[0043] S5: According to the total loss, adjust the parameters of the first network model and the second network model through back propagation.
[0044] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses an ovarian cancer database construction and intelligent analysis system, which constructs an ovarian cancer database based on imaging images and pathological images, and jointly trains macroscopic and microscopic features to improve the accuracy of image analysis, thereby forming a high-precision thermal map for doctors to analyze and diagnose. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.
[0046] Figure 1 A schematic diagram of the structure of an ovarian cancer database construction and intelligent analysis system provided by the present invention;
[0047] Figure 2 A schematic diagram of a model training method provided in Example 2 of the present invention;
[0048] Figure 3 Schematic diagram of the model training method provided in Example 3 of the present invention. DETAILED DESCRIPTION
[0049] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0050] Example 1
[0051] like Figure 1 The embodiment of the present invention discloses an ovarian cancer database construction and intelligent analysis system, including:
[0052] A multi-source data acquisition module, used to acquire ovarian cancer data, including imaging images and pathological images;
[0053] A data processing module is used to pair the imaging images with the corresponding set of pathological images, and to label the imaging images and pathological images respectively to build a database;
[0054] The model training module is used to sample the database, extract the tumor area based on the imaging image samples, and extract the pathological features of the corresponding pathological image samples to calculate the degree of tumor deterioration; supervise the tumor extraction and pathological feature extraction according to the labels during annotation, and optimize the model parameters;
[0055] Deploy an application module to use the trained model to identify tumor areas and tumor progression and generate heat maps.
[0056] In order to further implement the above technical solution, the multi-source data acquisition module includes a sample marking submodule and a label annotation submodule;
[0057] The sample labeling submodule is used to add sample labels according to the sources of the imaging images and pathological images, and to pair them according to the added sample labels;
[0058] The sample annotation module is used to annotate key feature areas in imaging images and pathological images.
[0059] Furthermore, sample labels need to be divided according to different user sources. At the same time, in order to describe the sources of tissue sections in pathological images, different identifiers are set respectively. Specifically, sample labels include primary sample labels and secondary sample labels. The primary sample labels and the secondary sample labels are used to mark the imaging images and the pathological images, respectively. The sample labeling submodule confirms the imaging images to which each pathological image belongs and its position in the corresponding imaging image according to the first sample label and the second sample label, and performs pairing processing.
[0060] The specific marking form can be achieved through custom coding. After completing the above marking, when there are multiple tumor areas in the imaging image with corresponding pathological images, the second identifier can be used to confirm the imaging image to which the pathological image itself belongs, and its own position in the imaging image can be further clarified.
[0061] Furthermore, the sample annotation process includes annotation of imaging images and annotation of pathological images.
[0062] In order to further implement the above technical solution, the neural network model trained by the type training module includes: a radiological image processing layer and a pathological image processing layer.
[0063] The imaging image processing layer is used to extract imaging features and identify tumor areas; the pathological image processing layer is used to extract pathological features of tumor areas and predict the degree of deterioration.
[0064] In this implementation, the specific structure of the imaging image processing layer includes: a first feature extractor, a region proposal network, a region interest pooling network, a classifier and a bounding box regressor; the feature extractor can extract features from the imaging image, the region proposal network generates a candidate tumor region based on the extracted features, and then the region interest pooling network extracts a fixed-size feature map from the candidate and inputs it into the classifier; finally, the classified tumor region is regressed through the bounding box regressor to achieve more accurate positioning.
[0065] The specific structure of the pathology image processing layer includes: a second feature extractor, a pathology feature extractor and a deterioration degree predictor; the second feature extractor can extract features from the pathology image, and then identify and extract pathology-related features through the pathology feature extractor; the deterioration degree predictor calculates the deterioration degree through a linear layer.
[0066] Among them, the first feature extractor and the second feature extractor can use CNN network, ResNet, Inception network and other networks.
[0067] In order to further implement the above technical solution, the total loss function used in the model training module is:
[0068] L total =α1L cls +α2L reg +α3L score
[0069] Among them, α1, α2, and α3 are the weight factors of the corresponding losses; L total is the total loss, L cls is the tumor region classification loss, L reg is the tumor bounding box regression loss, L score Predict the loss for the degree of deterioration; the specific loss functions are as follows:
[0070] L cls =-α t (1-p t ) γ log(p t )
[0071] In the above formula, p t Represents the predicted probability, is the balance factor of positive and negative samples, and is a hyperparameter for controlling the balance of difficult and easy samples;
[0072]
[0073] In the above formula, x represents the deviation between the predicted boundary and the true bounding box; i represents the index of the predicted bounding box;
[0074]
[0075] In the above formula, L score Predict the loss for the deterioration degree, y score is the true value of the deterioration degree, is the predicted value of the degree of deterioration.
[0076] In order to further implement the above technical solution, the neural network model also includes a multimodal fusion layer, which fuses the imaging features and pathological features through a fusion network, and uses the annotated tumor location and deterioration degree information to adjust the parameters of the entire network through back propagation.
[0077] In order to further implement the above technical solution, the deployment and application module includes a model storage submodule, a data interface submodule, a visual configuration submodule and a visual output submodule.
[0078] The storage submodule is used to store the trained model; the data interface submodule is used to configure the data source and input the data received by the data source into the trained model for intelligent analysis; the visual submodule is used to configure the corresponding visual expression effects for different degrees of deterioration; the visualization output submodule is used to output the heat map according to the intelligent analysis results and the configuration parameters of the visual submodule.
[0079] Example 2
[0080] like Figure 2 Based on the same inventive concept, an embodiment of the present invention discloses a model training method, comprising the following steps:
[0081] S1-1: Input the imaging images in the database into the pre-built first model, extract imaging features through the first model, identify the tumor area, calculate the loss according to the tumor area label and optimize the first model, and when the first model meets the preset conditions, fix the model parameters of the first model;
[0082] S1-2: Output the tumor region through the trained first model, and take out the corresponding pathological image from the database according to the tumor region and output it to the pre-built second model;
[0083] S1-3: Pathological features are extracted through the second model to calculate the degree of tumor deterioration; the loss is calculated and the second model is optimized according to the annotated deterioration degree labels, and when the second model meets the preset conditions, the model parameters of the second model are fixed.
[0084] In this embodiment, the first model and the second model can be constructed as the imaging image processing layer and the pathological image processing layer in Example 1, respectively. First, the first model for processing imaging images is trained, and then the corresponding pathological images can be accurately retrieved through the trained first model for training. In this way, the losses of the two stages in the training task can be considered separately, thereby making it easier to debug and optimize the model; when training the second model, you can focus on learning pathological features without worrying about the performance of the first model, reducing the risk of overfitting.
[0085] Example 3
[0086] like Figure 3 Based on the same inventive concept, an embodiment of the present invention discloses a model training method, comprising the following steps:
[0087] S2-1: inputting the imaging images in the database into the pre-built first model, and extracting imaging features through the first model to distinguish the tumor area from the background;
[0088] S2-2: extracting corresponding pathological images from the database according to the tumor area and outputting them to the pre-built second model;
[0089] S2-3: extracting pathological features through the second model and calculating the degree of tumor progression;
[0090] S2-4: Calculate the classification loss and bounding box regression loss of tumor region recognition based on the true label of the tumor region in the imaging image; calculate the deterioration degree prediction loss based on the tumor deterioration degree label;
[0091] S2-5: According to the total loss, adjust the parameters of the first network model and the second network model through back propagation.
[0092] In this embodiment, the parameters of the first network model and the second network model are simultaneously affected by back propagation. Therefore, the two network models can influence each other during training and share information, which helps to improve the overall performance of the model and better capture the relationship between different modalities.
[0093] It should be emphasized that in order to avoid the first network model missing the tumor area and causing the second network model to be unable to detect it, the first network model performs binary classification on the entire image area after extracting the imaging features, introduces background classification, and adds an additional loss to punish this situation, as follows:
[0094]
[0095] Among them, L bg represents the background area penalty. When a pixel is marked as background but is actually a tumor, the corresponding loss value is calculated. Otherwise, it is 0. bg is the predicted probability of the background category, α bg is the balancing factor for the background category.
[0096] The present invention constructs an ovarian cancer database based on imaging images and pathological images, and jointly trains the macroscopic features reflected in the imaging images and the microscopic features reflected in the pathological images, thereby improving the accuracy of image analysis, thereby forming a high-precision thermal map for doctors to analyze and diagnose.
[0097] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.
[0098] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. An ovarian cancer database construction and intelligent analysis system, characterized in that: include: A multi-source data acquisition module, used to acquire ovarian cancer data, wherein the ovarian cancer data includes imaging images and pathological images; A data processing module, used to pair the imaging images with the corresponding set of pathological images, and label the imaging images and the pathological images respectively to construct a database; The model training module is used to sample the database, extract the tumor area based on the imaging image samples, and extract the pathological features of the corresponding pathological image samples to calculate the degree of tumor deterioration; supervise the tumor extraction and pathological feature extraction according to the labels during annotation, and optimize the model parameters; Deploy an application module to use the trained model to identify tumor areas and tumor progression and generate heat maps.
2. The ovarian cancer database construction and intelligent analysis system according to claim 1, characterized in that: The data processing module includes a sample marking submodule and a label annotation submodule; The sample labeling submodule is used to add sample labels according to the sources of the imaging image and the pathological image, and perform pairing according to the added sample labels; The sample annotation module is used to annotate key feature areas in the imaging image and the pathological image.
3. The ovarian cancer database construction and intelligent analysis system according to claim 2, characterized in that: The sample mark includes a primary sample mark and a secondary sample mark, and the primary sample mark and the secondary sample mark are used to mark the imaging image and the pathological image respectively.
4. The ovarian cancer database construction and intelligent analysis system according to claim 3, characterized in that: The sample marking submodule confirms the imaging image to which each pathological image belongs and the position in the corresponding imaging image according to the first sample marking and the second sample marking, and performs pairing processing.
5. The ovarian cancer database construction and intelligent analysis system according to claim 2, characterized in that: The neural network model trained by the model training module includes: a radiological image processing layer and a pathological image processing layer; The imaging image processing layer is used to extract imaging features and identify tumor areas; The pathological image processing layer is used to extract the pathological features of the tumor area and predict the degree of deterioration.
6. The ovarian cancer database construction and intelligent analysis system according to claim 5, characterized in that: The total loss function used by the model training module is: L total =α1L cls +α2L reg +α3L score Among them, α1, α2, and α3 are the weight factors of the corresponding losses; L total is the total loss, L cls is the tumor region classification loss, L reg is the tumor bounding box regression loss, L score Predict the loss for the degree of deterioration; the specific loss functions are as follows: L cls =-a t (1-p t ) γ log(p t ) In the above formula, p t represents the predicted probability, α t is the balance factor of positive and negative samples, and γ is the hyperparameter that controls the balance of difficult and easy samples; In the above formula, x represents the deviation between the predicted boundary and the true bounding box; i represents the index of the predicted bounding box; In the above formula, L score Predict the loss for the deterioration degree, y score is the true value of the deterioration degree, is the predicted value of the degree of deterioration.
7. The ovarian cancer database construction and intelligent analysis system according to claim 5, characterized in that: The neural network model also includes a multimodal fusion layer, which fuses imaging features and pathological features through a fusion network, and uses the labeled tumor location and deterioration degree information to adjust the parameters of the entire network through back propagation.
8. The ovarian cancer database construction and intelligent analysis system according to claim 1, characterized in that: The deployment and application module includes a model storage submodule, a data interface submodule, a visual configuration submodule and a visual output submodule; The storage submodule is used to store the trained model; The data interface submodule is used to configure the data source and input the data received by the data source into the trained model for intelligent analysis; The visual submodule is used to configure corresponding visual expression effects for different deterioration levels; The visualization output submodule is used to output a heat map according to the intelligent analysis results and the configuration parameters of the vision submodule.
9. A model training method, characterized in that: The method is used for the ovarian cancer database construction and intelligent analysis system according to any one of claims 1 to 8, comprising the following steps: S1: inputting the imaging images in the database into the pre-built first model, extracting imaging features through the first model, identifying the tumor area, calculating the loss according to the tumor area label and optimizing the first model, and fixing the model parameters of the first model when the first model meets the preset conditions; S2: Output the tumor region through the trained first model, and take out the corresponding pathological image from the database according to the tumor region and output it to the pre-built second model; S3: extracting pathological features through the second model and calculating the degree of tumor deterioration; calculating the loss and optimizing the second model according to the annotated deterioration degree label, and fixing the model parameters of the second model when the second model meets the preset conditions.
10. A model training method, characterized in that: The method is used for the ovarian cancer database construction and intelligent analysis system according to any one of claims 1 to 8, comprising the following steps: S1: inputting the imaging images in the database into the pre-built first model, and extracting imaging features through the first model to distinguish the tumor area from the background; S2: extracting corresponding pathological images from the database according to the tumor area and outputting them to the pre-built second model; S3: extracting pathological features through the second model and calculating the degree of tumor progression; S4: Calculate the classification loss and bounding box regression loss of tumor region recognition based on the true label of the tumor region in the imaging image; calculate the deterioration degree prediction loss based on the tumor deterioration degree label; S5: According to the total loss, adjust the parameters of the first network model and the second model through back propagation.