Target object recognition model, model parameter determination method, device and equipment

Abstract

The invention relates to the technical field of medical image processing, and discloses a target object recognition model, a model parameter determination method, device and equipment, and the model comprises a target object recognition network which is used for recognizing a region of interest from a medical image of a target medical structure, and recognizing a target object from the region of interest; the meta learning network is used for determining parameter values of the target object recognition network based on the support sample set and the query sample set; wherein each support sample in the support sample set comprises training data of a second object recognition network obtained by training the first object recognition network; the second object identification network and the target object identification network have the same structure and partial or all parameters are different; in different support samples, the identification tasks of the second object identification network are different. According to the method, the accuracy of the target object recognition model in a cross-domain small sample learning scene or even a zero sample learning scene is greatly improved, and target object recognition with high sensitivity and extremely low false positive is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a target object recognition model, a model parameter determination method, a device and equipment. Background Art

[0002] When identifying target objects in human structures based on medical images, since human structures and target objects have strong individual heterogeneity and complex morphological and image representations, the signals within the target objects also vary significantly due to differences in blood flow velocity, size, etc., and the voxel intensity may even be much lower than the vascular lumen or other related tissue structures, making it difficult to identify them as regions of interest. Therefore, relying solely on supervised learning of limited data sets, it is still inevitable that the detection model will encounter local small-sample learning or even zero-sample learning scenarios when faced with vascular anatomical structures, regions of interest, lumen density distribution, target object morphology, etc. that are far from the distribution of training data. In this scenario, the detection accuracy of the detection model is low. Summary of the invention

[0003] In view of this, the present invention provides a target object recognition model, a model parameter determination method, an apparatus and a device to solve the problem of low accuracy in target object recognition based on medical images.

[0004] In a first aspect, the present invention provides a target object recognition model, the model comprising:

[0005] A target object recognition network, used to identify a region of interest from a medical image of a target medical structure, and to identify a target object from the region of interest;

[0006] A meta-learning network, used to determine parameter values ​​of the target object recognition network based on a support sample set and a query sample set;

[0007] Among them, the supporting sample set includes multiple supporting samples, each of the supporting samples includes training data of the second object recognition network obtained by training the first object recognition network; the structures of the first object recognition network, the second object recognition network and the target object recognition network are the same, and some or all of the parameters are different; in different supporting samples, the recognition tasks of the second object recognition network are different.

[0008] In an optional embodiment, the meta-learning network is a domain adversarial network, including a domain classifier, a feature extractor, and a label predictor;

[0009] Wherein, the domain classifier includes a dual-channel context-aware meta-sample enhancement module.

[0010] In an optional implementation, the medical image is a cerebral artery image;

[0011] The target object is a cerebral aneurysm;

[0012] In the supporting sample, the object identified by the second object recognition network is a normal blood vessel or a blood vessel tortuosity or a bifurcation point or a stenotic blood vessel, or a cerebral aneurysm in a different region of interest from the cerebral aneurysm identified by the target object recognition network.

[0013] In an optional implementation, the target object recognition network includes an infrastructure network and a parallel graph convolutional network, the infrastructure network adopts a multi-supervisory network of an encoding-decoding framework, and the parallel graph convolutional network is used to extract key semantic knowledge in multiple layers;

[0014] The input of the graph convolutional network is a fusion graph of target medical structure data;

[0015] The input of the basic structure network is the medical image and the first multi-level target medical structure-knowledge fusion data, the target medical structure has multiple levels, and the target object is a part or all of the structure on one level in the target medical structure.

[0016] In a second aspect, the present invention provides a model parameter determination method, which is applied to any target object recognition model described in the first aspect, and the method comprises:

[0017] Using a meta-learning network, similarity judgment is performed on the recognition process of the meta-training task corresponding to each of the supporting samples in the supporting sample set and the recognition process of the target object recognition task;

[0018] Determining a first parameter value of each parameter in the target object recognition network based on a judgment result of the similarity judgment by using the meta-learning network;

[0019] The first parameter value is fine-tuned based on the query sample set using the meta-learning network to obtain final values ​​of various parameters of the target object recognition network.

[0020] In an optional implementation, before using the meta-learning network to perform similarity judgment on the recognition process of the meta-training task corresponding to each of the supporting samples in the supporting sample set and the recognition process of the target object recognition task, the method further includes:

[0021] For each of the plurality of meta-training tasks, meta-training the first object recognition network is performed to obtain a second object recognition network;

[0022] The training data of each of the meta-training tasks is obtained respectively to form the support samples.

[0023] In an optional implementation, the recognition objects of the meta-training task corresponding to the supporting samples in the supporting sample set include: structural objects in different regions, different positions and / or different types on the medical image.

[0024] In an optional implementation, the medical image is a cerebral artery image;

[0025] The recognition objects of the meta-training task corresponding to the supporting samples in the supporting sample set include: cerebral artery structure objects of different segments, different partitions and / or different types of cerebral arteries.

[0026] In an optional implementation, the target object recognition network includes: an infrastructure network and a parallel graph convolutional network;

[0027] The infrastructure network adopts a multi-supervisory network of an encoding-decoding framework, and the parallel graph convolutional network is used to extract key semantic knowledge in multiple layers;

[0028] The input of the graph convolutional network is a fusion graph of target medical structure data;

[0029] The input of the basic structure network is the medical image and the first multi-level target medical structure-knowledge fusion data;

[0030] The training samples of the meta-training task include pseudo-noise data, where the pseudo-noise data is formed based on target medical structure segmentation data and an optimal knowledge representation set, where the optimal knowledge representation set is a knowledge quantization representation data set guided by target medical structure knowledge.

[0031] In an optional implementation, the optimal knowledge representation set is generated by the following method:

[0032] Acquire a medical image including a target medical structure, annotated data of a target object, and knowledge of the target medical structure, wherein the knowledge of the target medical structure includes structural hierarchy information, distribution information, and an identification standard of the target object;

[0033] Performing knowledge quantification on the medical image of the target medical structure, the annotation data of the target object, and the knowledge of the target medical structure to obtain a variety of knowledge representation data;

[0034] From the multiple knowledge representation data, some knowledge representation data are selected to form an optimal knowledge representation set.

[0035] In an optional implementation, selecting part of the knowledge representation data from the multiple knowledge representation data to form an optimal knowledge representation set includes:

[0036] Using similarity measurement to cluster and group the knowledge representation data;

[0037] According to the knowledge representation data grouping, representative knowledge representation data are extracted respectively;

[0038] evaluating the bootstrapping performance of the representative knowledge representation data;

[0039] Select N knowledge representation methods with the best guiding performance to form the optimal knowledge representation set.

[0040] In an optional implementation, the first multi-level target medical structure-knowledge fusion data is obtained by the following method:

[0041] stratifying the target medical structure knowledge according to the hierarchy of the target medical structure;

[0042] stratifying the medical image data of the target medical structure according to a correspondence between a knowledge source of the target medical structure knowledge and the target medical structure;

[0043] The target medical structure knowledge of each layer and the medical image data of the target medical structure are fused accordingly to form the first multi-level target medical structure-knowledge fusion data.

[0044] In an optional implementation, the hierarchical fusion strategy of the first multi-level target medical structure-knowledge fusion data is determined according to the following method:

[0045] Adopting a plurality of different hierarchical fusion methods, the target medical structure knowledge of each layer and other medical image data of the target medical structure are correspondingly fused to form a plurality of second multi-level target medical structure-knowledge fusion data;

[0046] The plurality of the second multi-level target medical structure-knowledge fusion data are subjected to at least one layer of convolution to form an output, and a three-dimensional information entropy is output based on mutual information calculation, and the three-dimensional information entropy is used to perform hierarchical clustering of the second multi-level target medical structure-knowledge fusion data to form a plurality of different second multi-level target medical structure-knowledge fusion data groups corresponding to different hierarchical fusion methods;

[0047] Using multiple baseline networks, using different second multi-level target medical structure-knowledge fusion data groups, model training is performed, and statistical performance of the trained networks is performed;

[0048] For different second multi-level target medical structure-knowledge fusion data groups, based on the performance of the multiple networks obtained through training, calculating a comprehensive guide performance score;

[0049] According to the ranking of the guidance performance comprehensive score, selecting a fused data in the second multi-level target medical structure-knowledge fusion data group as the optimal fused data;

[0050] The hierarchical fusion method used by the optimal fusion data is determined as the hierarchical fusion strategy of the first multi-level target medical structure-knowledge fusion data.

[0051] In a third aspect, the present invention provides a model parameter determination device for training and obtaining any one of the target object recognition models described in the first aspect, the device comprising:

[0052] A similarity judgment module, used to use the meta-learning network to perform similarity judgment on the recognition process of the meta-training task and the recognition process of the target object recognition task corresponding to each of the supporting samples in the supporting sample set;

[0053] A network parameter determination module, used to determine a first parameter value of each parameter in the target object recognition network based on a judgment result of the similarity judgment by the meta-learning network;

[0054] A network parameter adjustment module is used to use the meta-learning network to fine-tune the first parameter value based on the query sample set to obtain the final value of each parameter of the target object recognition network.

[0055] In a fourth aspect, the present invention provides a computer device, comprising: a memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, and the processor executing the model parameter determination method of the above-mentioned second aspect or any corresponding embodiment thereof by executing the computer instructions.

[0056] In a fifth aspect, the present invention provides a computer-readable storage medium having computer instructions stored thereon, the computer instructions being used to enable a computer to execute the model parameter determination method of the above-mentioned second aspect or any corresponding embodiment thereof.

[0057] In a sixth aspect, the present invention provides a computer program product, comprising computer instructions for causing a computer to execute the model parameter determination method of the first aspect or any corresponding embodiment thereof.

[0058] The target object recognition model in the embodiment of the present invention includes a meta-learning network, so that the model can not only learn the sample of the target object itself, but also learn the entire process of discovering the target object, and transform the original process of classifying the positive / negative of the area of ​​interest into an autonomous exploration of a larger area. The exploration experience of different areas is integrated and utilized in the exploration of new areas, which greatly improves the accuracy of the model in cross-domain small sample learning or even zero-sample learning scenarios, and realizes high-sensitivity and extremely low false positive target object recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the related technologies, the drawings required for use in the specific embodiments or the related technical descriptions will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0060] Figure 1 is a structural block diagram of a target object recognition model according to an embodiment of the present invention;

[0061] Figure 2 is a schematic diagram of a model parameter determination process and a domain adversarial network according to an embodiment of the present invention;

[0062] Figure 3 is a schematic diagram of the structure and data processing process of a target object recognition model according to an embodiment of the present invention;

[0063] Figure 4 is a schematic diagram of a parameter adjustment process of a target object recognition model according to an embodiment of the present invention;

[0064] Figure 5 is a flow chart of a method for determining model parameters according to an embodiment of the present invention;

[0065] Figure 6 is a schematic diagram of a process of forming an optimal knowledge representation set and a process of acquiring optimal fusion data according to an embodiment of the present invention;

[0066] Figure 7 is a structural block diagram of a device for determining model parameters according to an embodiment of the present invention;

[0067] Figure 8 It is a schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0068] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are 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 those skilled in the art without creative work are within the scope of protection of the present invention.

[0069] The embodiment of the present invention provides a target object recognition model, such as Figure 1 As shown, the target object recognition model includes:

[0070] The target object recognition network 101 is used to identify a region of interest from a medical image of a target medical structure, and to identify a target object from the region of interest; the target object is a partial structure in the target medical structure. The target object recognition network 101 may also be referred to as a target object recognition model or a target object recognition module, etc. The target object recognition network 101 is the basic network of the target object recognition model. For example, it may be a CEDNet network (Channel Enhanced DenseConnection Network, i.e., a lightweight feature extraction network based on dense connections).

[0071] The medical image in the embodiment of the present invention may be a cerebral artery image, for example, a Time-of-Flight Magnetic Resonance Angiography (TOF-MRA) image of the brain. Accordingly, the target object may be a cerebral aneurysm, etc., and the region of interest is the arterial blood vessel partition or segment where the cerebral aneurysm is located. Of course, the medical image may also be an image of other medical structures, such as a cardiovascular image, a pulmonary blood vessel image, a pulmonary trachea (including a bronchus), etc., and the target object is a partial structure in the corresponding medical structure. The following mainly uses a cerebral aneurysm as an example for the target object. The recognition process of other target objects is similar and will not be described in detail.

[0072] A meta-learning network 102, used to determine parameter values ​​of the target object recognition network based on a support sample set and a query sample set;

[0073] The supporting sample set includes multiple supporting samples, each of which includes training data of a second object recognition network obtained by training a first object recognition network; the first object recognition network, the second object recognition network and the target object recognition network have the same structure, and some or all parameters are different; in different supporting samples, the recognition tasks of the second object recognition network are different. The first object recognition network can be a pre-trained recognition network.

[0074] When the medical image is a cerebral artery image and the target object is a cerebral aneurysm, the object identified by the second object recognition network is a normal blood vessel, a vascular tortuosity, a bifurcation, or a stenotic blood vessel, or a cerebral aneurysm in a different region of interest from the cerebral aneurysm identified by the target object recognition network.

[0075] In some specific implementations, such as Figure 2 As shown, the meta-learning network is a domain adversarial network, including a domain classifier, a feature extractor, and a label predictor;

[0076] The domain classifier includes a dual-channel context-aware meta-sample enhancement module (ie, a dual-channel CaMSE (Contextualized Meta-Sample Enhancement) module).

[0077] In the embodiment of the present invention, the target object recognition network is used as the basic model to construct a meta-learning framework, and a domain adversarial neural network is used to perform similarity discrimination on the tasks of training and testing. Specifically, during the meta-testing process, the type of query samples for the meta-test (such as aneurysms, vascular tortuosity, various regions of cerebral blood vessels, etc.) is generated according to the similarity of the similarity discrimination. The judgment of similarity is aimed at the similarity between the process of the query sample data passing through the test model and the process of the sample set constructed during training passing through the test model. The tasks of meta-training and meta-testing are to first segment the region of interest (ROI) and then identify the lesions in the ROI. This task is the same, except that the specific ROI types and lesion types are different, and the type range of meta-training is generally considered to include the types that will be encountered in the meta-test, so the type of query samples for the meta-test can be generated through the above-mentioned similarity judgment. The linear output of the dual-channel CaMSE module is used to fine-tune the linear head parameters of the target object recognition network to achieve the optimal effect.

[0078] Domain Adversarial Networks (DANs) are used to solve the problem of domain adaptation. Domain adaptation refers to how to train a model to perform well in the target domain (here, the network model after meta-testing, i.e., the target object recognition network) when there is a large amount of labeled data in the source domain (Source Domain) (here, the meta-trained network, i.e., the second object recognition network) and only a small amount or no labeled data in the target domain (Target Domain). Meta-learning can be used to train models to quickly adapt to new tasks, while domain adversarial networks can help models reduce the differences between the source domain and the target domain. Meta-learning combined with domain adversarial networks can further improve the generalization ability of the model in different tasks and domains.

[0079] Specifically, the domain classifier is used to distinguish the data in the source domain and the target domain. The goal of the domain classifier is to minimize the difference between the source domain and the target domain, so that the feature extractor can learn domain-invariant feature representations. The feature extractor is responsible for extracting useful features from the input data. This module is usually shared and used to process data from different domains. Adversarial training is performed between the feature extractor and the domain classifier. The feature extractor attempts to generate domain-invariant features, while the domain classifier attempts to accurately distinguish the data in the source domain and the target domain. Through this adversarial training, the model can perform better on the target domain. The label predictor is used to output the classification result of the final attribution of the test data (i.e., the sample data in the query sample set).

[0080] The contextualized sample enhancement (CaSE) module is used to enhance sample representation by introducing contextual information to improve the feature representation of samples, thereby helping the model to learn and generalize better.

[0081] The embodiment of the present invention refers to the CaSE module to construct a dual-channel CaMSE module embedded in the domain adversarial network to perform quantitative evaluation of the support sample set (also referred to as the support set) and the query sample set (also referred to as the query set) and fine-tune the network parameters to generate a high-precision target object recognition model.

[0082] In some other optional specific implementations, the target object recognition network includes: an infrastructure network and a parallel graph convolutional network, the infrastructure network adopts a multi-supervisory network of an encoding-decoding framework, and the parallel graph convolutional network is used to extract key semantic knowledge in multiple layers;

[0083] The input of the graph convolutional network is a fusion graph of target medical structure data;

[0084] The input of the infrastructure network is the medical image and the first multi-level target medical structure-knowledge fusion data. The target medical structure has multiple levels, the medical image includes the target medical structure, and the target object is a part or all of the structure at one level in the target medical structure. The medical image can be a preprocessed medical image that only includes the target medical structure. The first multi-level target medical structure-knowledge fusion data can be fusion data of the target medical structure in the medical image and the medical knowledge of the target medical structure.

[0085] As described above, the medical image may be an image of a cerebral artery, and the target medical structure is the cerebral artery. The medical image may also be an image of other medical structures.

[0086] In the embodiment of the present invention, the target object recognition network may be a target object recognition network guided by graph convolution knowledge.

[0087] For example, if Figure 3 As shown, the basic structure of the target object recognition network adopts a multi-supervised network of the encoding-decoding framework, and uses a parallel graph convolutional network to extract key semantic knowledge in multiple dimensions. The input of the graph convolutional network is the cerebral artery data fusion graph, and the aneurysm detection is taken as the goal, and the semantic cross entropy is used as the loss function. Among them, the cerebral artery data fusion graph is a graph data, which consists of nodes and edges, and the nodes are connected by edges. Specifically, the structure of the cerebral artery can be used as a way to decompose the nodes, and the knowledge representation (such as the optimal knowledge representation obtained below), image information, etc. can be assigned as data to different nodes or edges to obtain the cerebral artery data fusion graph. The input of the backbone network, that is, the infrastructure network, is the optimal vascular preprocessing result (for example, an image that only includes cerebral arteries obtained by processing TOF-MRA images) and the first multi-level (arterial) vascular-knowledge fusion data. In the encoding stage, the network receptive field adopts a multi-scale receptive field in the vascular space with the vascular centerline as the axis, which is established based on the cerebral artery segmentation data and the vascular structure stratification. The receptive field scale includes local blood vessels, vascular segments, vascular partitions, overall blood vessels, and the entire image. In the encoding process, multi-branch long connections are adopted, and continuous residual blocks are used to suppress the influence of another branch. The effective transmission of shallow features is enhanced through the spatial attention mechanism to realize the long connection mechanism of hierarchical feature decoupling.

[0088] The recognition process of the target object recognition network is as follows: after the new cerebral artery data has been segmented, the data fusion graph has been generated, and the first multi-level vascular-knowledge fusion data has been generated, the data is input into Figure 3 The graph convolutional network and convolutional network in the model will output the aneurysm location annotation information after model detection. The annotation is usually binary three-dimensional data, with the position where the voxel value is 1 representing the aneurysm and the other parts being 0 as the result of aneurysm recognition.

[0089] The process of obtaining the first multi-level target medical structure-knowledge fusion data is described in the following embodiment. Regardless of the meta-training stage, the meta-testing stage or the subsequent practical application stage, the same hierarchical fusion strategy is used to obtain the multi-level target medical structure-knowledge fusion data.

[0090] In some optional specific implementations, the medical image is a cerebral artery image;

[0091] The target object is a cerebral aneurysm;

[0092] In the supporting sample, the object identified by the second object recognition network is a normal blood vessel or a blood vessel tortuosity or a bifurcation point or a stenotic blood vessel, or a cerebral aneurysm in a different region of interest from the cerebral aneurysm identified by the first object recognition network.

[0093] Specifically, Figure 4 As shown in the figure, based on the labels of normal blood vessels, vascular tortuosity, bifurcation points, vascular stenosis, etc. encountered in aneurysm detection, as well as the partitioning and segmentation standards of cerebral arteries, the meta-training task is designed. First, the task is divided into two cascaded stages: segmentation of cerebral arteries (specifically, segmenting cerebral artery images from original brain images) and detection of abnormal points. Then, the second-stage task is divided into aneurysm detection in different regions and different vascular segments, as well as detection of different types of abnormal points such as aneurysms and vascular tortuosity, and detection of healthy blood vessels is added. The input of the second-stage task can be the overall cerebral artery map, or the partitioned and segmented cerebral artery map. For example, if the second-stage task is aneurysm detection in a certain partition, then the input at this time can be the cerebral artery map of the partition.

[0094] The construction of specific meta-training tasks can be carried out according to actual conditions, such as the number of training samples that can actually be obtained.

[0095] The target object recognition model in the embodiment of the present invention includes a meta-learning network, so that the model can not only learn the sample of the target object itself, but also learn the entire process of discovering the target object, and transform the original process of classifying the positive / negative of the area of ​​interest into an autonomous exploration of a larger area. The exploration experience of different areas is integrated and utilized in the exploration of new areas, which greatly improves the accuracy of the model in cross-domain small sample learning or even zero-sample learning scenarios, and realizes high-sensitivity and extremely low false positive target object recognition.

[0096] In addition, the embodiment of the present invention not only takes target object detection as a detection task, but also adds region of interest segmentation as a preceding cascade task, so that the task of the meta-learning framework is implemented as region of interest recognition and target object detection in a cascaded sequence, thereby realizing a high-precision target object detection model with autonomous perception capability.

[0097] Meta-learning, also known as "learning to learn", is a machine learning method that allows models to quickly adapt to new tasks or environments. In the field of image recognition, meta-learning can be used to train a model to quickly learn new categories of images with a small number of samples.

[0098] The process of meta-learning includes:

[0099] 1. Data preparation: prepare a dataset containing multiple tasks, each of which usually corresponds to a specific classification problem. For example, in image recognition, each task may involve identifying new categories from a set of categories.

[0100] 2. Inner Loop: For each task, the data is divided into a support set and a query set. The support set is used to simulate a small amount of labeled data, while the query set is used to evaluate the performance of the model on this task. The support set is used to perform one or more rapid adaptation training on the model. The goal of this stage is to enable the model to learn the key features of the new task based on a small number of examples.

[0101] 3. Outer Loop: Update the initial parameters of the model based on the performance of the model on the query sample set. This is done by optimizing the performance over the entire task distribution, with the goal of enabling the model to quickly adapt to new tasks and achieve good performance.

[0102] 4. Iterative optimization: The inner and outer loops are repeated until the model can learn quickly and effectively on unseen tasks.

[0103] 5. Testing: Test the model on completely new tasks that have never appeared during the training process to verify the model's generalization ability and rapid learning ability.

[0104] According to an embodiment of the present invention, an embodiment of a method for determining model parameters is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of executable computer instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0105] In this embodiment, a model parameter determination method is provided, which can be used in various computer devices. Figure 5 is a flow chart of a method for determining model parameters according to an embodiment of the present invention. The method for determining model parameters can be applied to the target object recognition model described in the above embodiment. Figure 5 As shown, the process includes the following steps:

[0106] Step S501, using a meta-learning network to perform similarity judgment on the recognition process of the meta-training task corresponding to each of the supporting samples in the supporting sample set and the recognition process of the target object recognition task;

[0107] Step S502, determining a first parameter value of each parameter in the target object recognition network based on a judgment result of the similarity judgment by using the meta-learning network;

[0108] Step S503: Using the meta-learning network to fine-tune the first parameter value based on the query sample set to obtain final values ​​of various parameters of the target object recognition network.

[0109] In some optional specific implementations, before using the meta-learning network to perform similarity judgment on the recognition process of the meta-training task corresponding to each of the supporting samples in the supporting sample set and the recognition process of the target object recognition task, the method further includes:

[0110] For each of the plurality of meta-training tasks, meta-training the first object recognition network is performed to obtain a second object recognition network;

[0111] The training data of each meta-training task is obtained respectively to form the supporting samples; the training data includes training process data (including recognition process data) and related data involved in the training process.

[0112] The meta-learning network is used to perform similarity judgment on the recognition process of the meta-training task corresponding to each of the supporting samples in the supporting sample set and the recognition process of the target object recognition task, including:

[0113] The image sample data in the support sample and the query sample corresponding to the target object recognition task are judged for similarity.

[0114] In some optional specific implementations, the recognition objects of the meta-training task corresponding to the supporting samples in the supporting sample set include: structural objects in different regions, different positions and / or different types on the medical image.

[0115] In one optional specific implementation manner, the medical image is a cerebral artery image;

[0116] The recognition objects of the meta-training task corresponding to the supporting samples in the supporting sample set include: cerebral artery structure objects of different segments, different partitions and / or different types of cerebral arteries.

[0117] In the embodiment of the present invention, different segments of cerebral arteries and different types of structures on arteries are used as task decomposition basis to design cascaded diversified meta-training tasks, form a distribution of related tasks for cerebral aneurysm detection, and normalize and weight the correlation of different tasks. Thus, the detection process of cerebral aneurysms is transformed from the original process of classifying the positive / negative of the region of interest to an autonomous exploration of a larger region, and the exploration experience of different regions is integrated and utilized in the exploration of new regions.

[0118] In some optional specific implementations, the target object recognition network includes: a basic structure network and a parallel graph convolution network;

[0119] The infrastructure network adopts a multi-supervisory network of an encoding-decoding framework, and the parallel graph convolutional network is used to extract key semantic knowledge in multiple layers;

[0120] The input of the graph convolutional network is a fusion graph of target medical structure data;

[0121] The input of the basic structure network is the medical image and the first multi-level medical structure-knowledge fusion data;

[0122] The training samples of the meta-training task include pseudo-noise data, where the pseudo-noise data is formed based on target medical structure segmentation data and an optimal knowledge representation set, where the optimal knowledge representation set is a knowledge quantization representation data set used for target medical structure knowledge guidance.

[0123] Specifically, the pseudo noise data can be formed by using a point spread function processing method. The point spread function is a Gaussian distribution function, which is used to traverse the image using a function core (such as 3*3) or design a specific walking trajectory to add noise to the entire image.

[0124] Taking cerebral aneurysm recognition as an example, we can use cerebral artery segmentation data and the optimal knowledge representation set, through the knowledge representation on the cerebral artery, combined with the point spread function processing method, to form a pseudo-noise adding meta-training process.

[0125] In the embodiment of the present invention, when constructing training samples for meta-training, pseudo noise can be formed and added to the training samples, thereby alleviating the overfitting problem of meta-learning. After constructing the training samples for meta-training, meta-training is performed based on the constructed training samples, and the meta-training process and the related data involved in the meta-training process can be formed as support samples. In addition, in the process of constructing the query sample set, pseudo noise data can also be added according to the above method.

[0126] In summary, during the meta-training process, different meta-training tasks correspond to different training samples (different training samples have corresponding labels), and the parameters of the trained model (i.e., the second object recognition network) are also different. The test process of these meta-training tasks and the related data will constitute a support sample set, which is used to compare with the newly input data (i.e., the query sample in the query sample set) in the meta-test.

[0127] In some optional specific implementations, such as Figure 6 As shown, the optimal knowledge representation set is generated by the following method:

[0128] Step 1: Acquire a medical image including a target medical structure, annotation data of a target object, and knowledge of the target medical structure, wherein the knowledge of the target medical structure includes structural hierarchy information, distribution information, and an identification standard of the target object.

[0129] For example, a medical image may be a brain TOF-MRA image, and the cerebral artery in the brain TOF-MRA image is the target medical structure. The target object may be a cerebral aneurysm. The annotation data may be manually annotated data of a cerebral aneurysm. The knowledge of the cerebral artery may be the imaging and morphological description of the cerebral artery course, division, segmentation, and diagnostic criteria for a cerebral aneurysm. The knowledge of the target medical structure and the target object may be knowledge in text form.

[0130] Step 2: Perform knowledge quantification on the medical image of the target medical structure, the annotation data of the target object, and the knowledge of the target medical structure to obtain a variety of knowledge representation data.

[0131] For the knowledge quantification method, the image preprocessing algorithm set is used to form multiple related algorithm combinations corresponding to each knowledge according to the knowledge text description and the algorithm processing type. For TOF-MRA images and manually annotated data, three processing methods are used: image processing alone, annotation processing alone, and joint processing and fusion. Grid-based parameter search is used to form a combination of algorithm combinations and processing methods, generate multiple knowledge representation data, and construct a knowledge quantification representation dataset for cerebral aneurysm knowledge guidance.

[0132] For example, if the knowledge is that the target has a clear outline, then the processing algorithm is the processing result of the Sobel operator + the processing result of the Canny operator. Both operators are edge feature extraction operators. Then, another piece of knowledge describes that the brightness of the same target is high, so the median and maximum value of the grayscale value of the target area are directly extracted as features. The processing result of the Sobel operator, the processing result of the Canny operator, and the median and maximum value of the grayscale value of the target area, these four data constitute a set of knowledge representation data.

[0133] Step three: Select some knowledge representation data from the multiple knowledge representation data to form an optimal knowledge representation set.

[0134] In some optional specific implementations, step three, i.e., selecting part of the knowledge representation data from the multiple knowledge representation data to form an optimal knowledge representation set, includes:

[0135] Step three (1), using similarity measurement to cluster and group the knowledge representation data. Among them, for the knowledge representation data in image form, the similarity of image texture can be used to cluster and group them. For the knowledge representation data in image form, for example, the median of a 3*3 area of ​​the original image is extracted and traversed, then the knowledge representation data in image form can be generated. The similarity measure of image texture can be, for example, a gray-level co-occurrence matrix (Gray-Level Co-occurrence Matrix, GLCM), a local binary pattern (Local Binary Pattern, LBP), a structural similarity index (Structural Similarity Index, SSIM), etc.

[0136] Step three (2), for each knowledge representation data group, representative knowledge representation data are extracted. Representative knowledge representation data may be, for example, knowledge representation data that is relatively close to the cluster center.

[0137] In other implementations, other methods besides clustering grouping may also be used to extract representative knowledge representation data.

[0138] Step three (3), evaluating the guidance performance of the representative knowledge representation data.

[0139] For example, a dual-channel full convolution target object detection network can be used to evaluate the guiding performance of representative knowledge representation data. Although the dual-channel full convolution target object detection network has a different structure from the above target object detection network, the target objects to be detected can be the same, for example, both are cerebral aneurysms, and the inputs are both cerebral artery vascular images. The baseline network of the dual-channel full convolution target object detection network can adopt a dual-channel SE-3D Unet (SE-3D U-Net is a deep learning architecture that combines the characteristics of U-Net with the advantages of the Squeeze-and-Excitation (SE) module), and the baseline guiding data can be the Sobel contour of the target medical structure (such as cerebral artery). Specifically, the extracted representative knowledge representation data can be directly used as the second channel input of the dual-channel full convolution target object detection network, and then the guiding performance of the representative knowledge representation data can be determined based on the detection performance of the dual-channel full convolution target object detection network. By comparing the guiding performance with the baseline guiding data, it can be determined whether the guiding performance of the representative knowledge representation data is better. The detection performance of the dual-channel full convolution target object detection network can be determined, for example, by indicators such as detection sensitivity and specificity.

[0140] Step three (4), select N knowledge representation methods with the best guiding performance to form an optimal knowledge representation set.

[0141] Furthermore, different similarity measures can be used to cluster and group various knowledge representation data. For each grouping, representative knowledge representation data is extracted according to the above method, and the guiding performance of the representative knowledge representation data is evaluated. Then, from the representative knowledge representation data extracted from multiple groupings, N knowledge representation methods with the top guiding performance are selected to form an optimal knowledge representation set. That is, the guiding performance of the representative knowledge representation data extracted from each grouping is horizontally quantitatively compared, and finally the top N knowledge representation methods are selected according to the ranking to form an optimal knowledge representation set. For example, the top 10-20 knowledge representation methods are selected to form an optimal knowledge representation set.

[0142] In an embodiment of the present invention, in an aneurysm detection scenario, based on a preprocessing algorithm set of TOF-MRA and manually labeled data, a mapping quantitative relationship between supervised knowledge such as cerebral artery description and aneurysm diagnostic criteria and the preprocessing algorithm is established, and a self-supervised clustering network is designed to optimize the knowledge quantitative representation in the cerebral aneurysm detection task to obtain a knowledge quantitative representation data set with a certain redundancy.

[0143] In some optional specific implementations, the first multi-level target medical structure-knowledge fusion data is obtained by the following method:

[0144] Step 1: Layer the target medical structure knowledge according to the hierarchy of the target medical structure. For example, when the target medical structure is the cerebral artery, the cerebral artery knowledge is layered according to the vascular structure, and is divided into the whole brain, vascular tree, arterial region, arterial segmentation, and lesion local from global to local. The arterial region is divided into the anterior cerebral artery ACA, the middle cerebral artery MCA, the posterior cerebral artery PCA, the internal carotid artery ICA, the vertebral artery VA, the basilar artery BA, the anterior cerebral circulation ACOM, the posterior cerebral circulation PCOM, the cavernous sinus AchA and other regions, and the arterial segmentation is a refined segmentation performed in the above regions, such as the anterior cerebral artery ACA can be further divided into vascular segments such as A1, A2, A3, and A4.

[0145] The target medical structure knowledge may be a knowledge representation in the optimal knowledge representation set described in the above embodiment.

[0146] Step 2: According to the correspondence between the knowledge source of the target medical structure knowledge and the target medical structure, the medical image data of the target medical structure is layered. That is to say, the layering method of the target medical structure knowledge is the same as the layering method of the medical image data of the target medical structure. If the target medical structure knowledge is a knowledge representation in the optimal knowledge representation set described in the above embodiment, then according to the correspondence between the knowledge source of the knowledge representation and the target medical structure (e.g., arterial vascular structure), the medical image data of the target medical structure is layered in the same layering method as the target medical structure knowledge.

[0147] The medical image of the target medical structure can be obtained by segmenting the original medical image including the target medical structure, that is, segmenting the image of the target medical structure from the original medical image, that is, the image of the target medical structure. The knowledge source of the target medical structure knowledge can be clinical knowledge, books, experience, etc.

[0148] Step 3: Fusing the target medical structure knowledge of each layer and the medical image data of the target medical structure accordingly to form the first multi-level target medical structure-knowledge fusion data.

[0149] Specifically, data fusion of different layers can be performed in different ways, such as using data point multiplication weighting for global data, using bounding box assignment for regional data, using segmented splicing for segmented data, and using inter-layer interpolation for local lesion data.

[0150] In an embodiment of the present invention, based on the correspondence between the knowledge source of the target medical structure knowledge (such as the knowledge representation in the optimal knowledge representation set described in the above embodiment) and the target medical structure, the target medical structure is automatically hierarchically segmented according to corresponding partitioning, stratification, and segmentation standards, and the target medical structure knowledge can be integrated accordingly according to the different levels.

[0151] In some optional specific implementations, such as Figure 6 As shown, the hierarchical fusion strategy of the first multi-level target medical structure-knowledge fusion data is determined according to the following method:

[0152] Step 3 (1) adopts a plurality of different hierarchical fusion methods to fuse the target medical structure knowledge of each layer and other medical image data of the target medical structure accordingly to form a plurality of second multi-layer target medical structure-knowledge fusion data. The other medical image data may not be the medical image data input when performing target object recognition.

[0153] Step 3 (2), subjecting a plurality of the second multi-level target medical structure-knowledge fusion data to at least one layer of convolution to form an output, and outputting a three-dimensional information entropy based on mutual information calculation, and performing hierarchical clustering of the second multi-level target medical structure-knowledge fusion data using the three-dimensional information entropy to form a plurality of different second multi-level target medical structure-knowledge fusion data groups corresponding to different said hierarchical fusion methods.

[0154] Specifically, the shallow convolution layer of the fully convolutional backbone network can be used to perform 1-2 layers of shallow convolution on the second multi-level target medical structure-knowledge fusion data.

[0155] Step 3 (3), using multiple baseline networks, using different second multi-level target medical structure-knowledge fusion data groups, to perform model training, and statistically analyze the performance of the trained network.

[0156] Among them, network performance includes sensitivity, number of false positives, specificity and other aspects.

[0157] For example, DeepMedic, SE-3D UNet, ViT can be used as baseline networks.

[0158] Step 3 (4): for different groups of the second multi-level target medical structure-knowledge fusion data, based on the performance of the multiple networks obtained through training, calculate the comprehensive score of the guided performance.

[0159] The comprehensive guidance performance score is the performance score of the second multi-level target medical structure-knowledge fusion data. The better the guidance performance of the second multi-level target medical structure-knowledge fusion data, the better the prediction performance of the baseline network guided by the second multi-level target medical structure-knowledge fusion data, and the higher the guidance performance score of the second multi-level target medical structure-knowledge fusion data. In the embodiment of the present invention, multiple baseline networks are used to evaluate the guidance performance of the second multi-level target medical structure-knowledge fusion data, and the performance evaluation of each baseline network is performed from multiple aspects, so the obtained guidance performance score is a comprehensive guidance performance score. The comprehensive guidance performance score can be obtained, for example, by weighted summation.

[0160] Step 3 (5): According to the ranking of the guidance performance comprehensive score, select one of the fused data in the second multi-level target medical structure-knowledge fusion data group as the optimal fused data.

[0161] Step 3 (6): determine the hierarchical fusion method used for the optimal fusion data as the hierarchical fusion strategy for the first multi-level target medical structure-knowledge fusion data.

[0162] In the embodiment of the present invention, during the detection of cerebral aneurysms, the fusion method of both the original data and the labeled data with the knowledge representation is explored based on the partitioning, layering and segmentation of the cerebral artery tree structure. The shallow convolution feature map of the multi-level cerebral artery-knowledge fusion data is output and unsupervised grouping based on cross-correlation is performed, and the diversified layered fusion strategies are cyclically tested. The multi-level fusion method of the knowledge representation and the cerebral artery data is optimized, and finally a multi-level fusion data of optimal cerebral artery knowledge, TOF-MRA images and manual annotations is formed.

[0163] In addition, the hierarchical fusion method used by the optimal fusion data is the optimal hierarchical fusion strategy. During the meta-training and meta-testing processes, the optimal hierarchical fusion strategy is used to fuse the target medical structure knowledge and the medical image data of the target medical structure.

[0164] In this embodiment, a model parameter determination device is also provided, which is used to implement the above-mentioned model parameter determination method embodiment and preferred implementation mode, which have been described and will not be repeated. As used below, the term "module" can implement a combination of software and / or hardware of a predetermined function. Although the device described in the following embodiments is preferably implemented in software, the implementation of hardware, or a combination of software and hardware is also possible and conceivable.

[0165] This embodiment provides a model parameter determination device, which is used to train and obtain any of the above-mentioned target object recognition models, such as Figure 7 As shown, the device comprises:

[0166] A similarity judgment module 701 is used to use the meta-learning network to perform similarity judgment on the recognition process of the meta-training task and the recognition process of the target object recognition task corresponding to each of the supporting samples in the supporting sample set;

[0167] A network parameter determination module 702, configured to determine a first parameter value of each parameter in the target object recognition network by using the meta-learning network based on the judgment result of the similarity judgment;

[0168] The network parameter adjustment module 703 is used to use the meta-learning network to fine-tune the first parameter value based on the query sample set to obtain the final value of each parameter of the target object recognition network.

[0169] In some optional specific implementations, the model parameter determination device further includes:

[0170] A meta-training module, configured to perform meta-training on the first object recognition network for each of the plurality of meta-training tasks to obtain a second object recognition network;

[0171] The supporting sample generation module is used to respectively obtain the training data of each of the meta-training tasks to form the supporting samples.

[0172] In some optional specific implementations, the model parameter determination device further includes an optimal knowledge representation set generation module, and the optimal knowledge representation set generation module includes:

[0173] A knowledge acquisition unit, configured to acquire a medical image including a target medical structure, annotation data of a target object, and knowledge of the target medical structure, wherein the knowledge of the target medical structure includes structural hierarchy information, distribution information, and an identification standard of the target object;

[0174] A knowledge quantification unit, configured to perform knowledge quantification on the medical image of the target medical structure, the annotation data of the target object, and the knowledge of the target medical structure, respectively, to obtain a plurality of knowledge representation data;

[0175] The selection unit is used to select part of the knowledge representation data from the multiple knowledge representation data to form an optimal knowledge representation set.

[0176] In some optional specific implementations, the selection unit is specifically used to cluster and group the knowledge representation data using a similarity measure; extract representative knowledge representation data for each knowledge representation data group; evaluate the guidance performance of the representative knowledge representation data; and select N knowledge representation methods with the highest guidance performance to form an optimal knowledge representation set.

[0177] In some optional specific embodiments, the model parameter determination device also includes: a first multi-level target medical structure-knowledge fusion data acquisition module, which is used to stratify the target medical structure knowledge according to the level of the target medical structure; stratify the medical image data of the target medical structure according to the correspondence between the knowledge source of the target medical structure knowledge and the target medical structure; and respectively fuse the target medical structure knowledge of each layer and the medical image data of the target medical structure to form the first multi-level target medical structure-knowledge fusion data.

[0178] In some optional specific embodiments, the model parameter determination device also includes: a hierarchical fusion strategy determination module, which is used to adopt a plurality of different hierarchical fusion methods to fuse the target medical structure knowledge of each layer and other medical image data of the target medical structure accordingly to form a plurality of second multi-level target medical structure-knowledge fusion data; the plurality of second multi-level target medical structure-knowledge fusion data are subjected to at least one layer of convolution to form an output, and a three-dimensional information entropy is output based on mutual information calculation, and the second multi-level target medical structure-knowledge fusion data is hierarchically clustered with the three-dimensional information entropy to form a plurality of different second multi-level target medical structure-knowledge fusion data. According to the grouping, corresponding to the different hierarchical fusion methods; using multiple baseline networks, using different second multi-level target medical structure-knowledge fusion data groupings, model training is carried out, and the performance of the trained network is statistically analyzed; for different second multi-level target medical structure-knowledge fusion data groupings, based on the performance of multiple networks obtained by training, a guided performance comprehensive score is calculated; according to the ranking of the guided performance comprehensive score, a fusion data within the second multi-level target medical structure-knowledge fusion data group is selected as the optimal fusion data; the hierarchical fusion method used by the optimal fusion data is determined as the hierarchical fusion strategy of the first multi-level target medical structure-knowledge fusion data.

[0179] The further functional description of each of the above modules and units is the same as that of the above corresponding embodiments and will not be repeated here.

[0180] The model parameter determination device in this embodiment is presented in the form of a functional unit, where the unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that executes one or more software or fixed programs, and / or other devices that can provide the above functions.

[0181] The embodiment of the present invention also provides a computer device having the above Figure 7 The model parameter determination device shown.

[0182] See also Figure 8 , Figure 8 is a schematic diagram of the structure of a computer device provided by an optional embodiment of the present invention, such as Figure 8As shown, the computer device includes: one or more processors 10, a memory 20, and interfaces for connecting various components, including high-speed interfaces and low-speed interfaces. Various components are connected to each other using different buses for communication, and can be installed on a common mainboard or installed in other ways as needed. The processor can process the instructions executed in the computer device, including instructions stored in or on the memory to display the graphical information of the GUI on an external input / output device (such as, a display device coupled to the interface). In some optional embodiments, if necessary, multiple processors and / or multiple buses can be used together with multiple memories and multiple memories. Similarly, multiple computer devices can be connected, and each device provides some necessary operations (for example, as a server array, a group of blade servers, or a multi-processor system). Figure 8 A processor 10 is taken as an example.

[0183] The processor 10 may be a central processing unit, a network processor or a combination thereof. The processor 10 may further include a hardware chip. The hardware chip may be a dedicated integrated circuit, a programmable logic device or a combination thereof. The programmable logic device may be a complex programmable logic device, a field programmable gate array, a general purpose array logic or any combination thereof.

[0184] The memory 20 stores instructions executable by at least one processor 10, so that the at least one processor 10 executes the method shown in the above embodiment.

[0185] The memory 20 may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created according to the use of the computer device, etc. In addition, the memory 20 may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some optional embodiments, the memory 20 may optionally include a memory remotely arranged relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0186] The memory 20 may include a volatile memory, such as a random access memory; the memory may also include a non-volatile memory, such as a flash memory, a hard disk or a solid state drive; the memory 20 may also include a combination of the above types of memory.

[0187] The computer device also includes an input device 30 and an output device 40. The processor 10, the memory 20, the input device 30 and the output device 40 may be connected via a bus or other means. Figure 8 The example of connecting through bus is taken in the following.

[0188] The input device 30 can receive input digital or character information, and generate key signal input related to the user settings and function control of the computer device, such as a touch screen, a keypad, a mouse, a track pad, a touch pad, an indicator bar, one or more mouse buttons, a trackball, a joystick, etc. The output device 40 may include a display device, an auxiliary lighting device (e.g., an LED) and a tactile feedback device (e.g., a vibration motor), etc. The above-mentioned display device includes but is not limited to a liquid crystal display, a light emitting diode, a display and a plasma display. In some optional embodiments, the display device can be a touch screen.

[0189] The computer device also includes a communication interface, which is used for the computer device to communicate with other devices or a communication network.

[0190] The embodiment of the present invention also provides a computer-readable storage medium. The method according to the embodiment of the present invention can be implemented in hardware, firmware, or can be implemented as a computer code that can be recorded in a storage medium, or can be implemented as a computer code that is originally stored in a remote storage medium or a non-temporary machine-readable storage medium and will be stored in a local storage medium through a network download, so that the method described herein can be stored in such software processing on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only storage memory, a random access memory, a flash memory, a hard disk or a solid-state hard disk, etc.; further, the storage medium can also include a combination of the above types of memories. It can be understood that a computer, a processor, a microprocessor controller, or programmable hardware includes a storage component that can store or receive software or computer code. When the software or computer code is accessed and executed by a computer, a processor, or hardware, the method shown in the above embodiment is implemented.

[0191] A part of the present invention may be applied as a computer program product, such as a computer program instruction, which, when executed by a computer, can call or provide the method and / or technical solution according to the present invention through the operation of the computer. Those skilled in the art should understand that the existence of the computer program instruction in a computer-readable medium includes, but is not limited to, a source file, an executable file, an installation package file, etc., and accordingly, the way in which the computer program instruction is executed by the computer includes, but is not limited to: the computer directly executes the instruction, or the computer compiles the instruction and then executes the corresponding compiled program, or the computer reads and executes the instruction, or the computer reads and installs the instruction and then executes the corresponding installed program. Here, the computer-readable medium may be any available computer-readable storage medium or communication medium accessible to the computer.

[0192] Although the embodiments of the present invention have been described in conjunction with the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention, and such modifications and variations are all within the scope defined by the appended claims.

Claims

1. A target object recognition model, characterized in that: The target object recognition model includes: A target object recognition network, used to identify a region of interest from a medical image of a target medical structure, and to identify a target object from the region of interest; A meta-learning network, used to determine parameter values ​​of the target object recognition network based on a support sample set and a query sample set; Among them, the supporting sample set includes multiple supporting samples, each of the supporting samples includes training data of the second object recognition network obtained by training the first object recognition network; the structures of the first object recognition network, the second object recognition network and the target object recognition network are the same, and some or all of the parameters are different; in different supporting samples, the recognition tasks of the second object recognition network are different.

2. The model according to claim 1, characterized in that The meta-learning network is a domain adversarial network, including a domain classifier, a feature extractor, and a label predictor; Wherein, the domain classifier includes a dual-channel context-aware meta-sample enhancement module.

3. The model according to claim 1, characterized in that The medical image is a cerebral artery image; The target object is a cerebral aneurysm; In the supporting sample, the object identified by the second object recognition network is a normal blood vessel or a blood vessel tortuosity or a bifurcation point or a stenotic blood vessel, or a cerebral aneurysm in a different region of interest from the cerebral aneurysm identified by the target object recognition network.

4. The model according to claim 1, characterized in that The target object recognition network includes an infrastructure network and a parallel graph convolutional network, wherein the infrastructure network adopts a multi-supervisory network of an encoding-decoding framework, and the parallel graph convolutional network is used to extract key semantic knowledge in multiple layers; The input of the graph convolutional network is a fusion graph of target medical structure data; The input of the basic structure network is the medical image and the first multi-level target medical structure-knowledge fusion data, the target medical structure has multiple levels, and the target object is a part or all of the structure on one level in the target medical structure.

5. A method for determining model parameters, characterized in that: The target object recognition model applied to any one of claims 1 to 4, the method comprising: Using a meta-learning network, similarity judgment is performed on the recognition process of the meta-training task corresponding to each of the supporting samples in the supporting sample set and the recognition process of the target object recognition task; Determining a first parameter value of each parameter in the target object recognition network based on a judgment result of the similarity judgment by using the meta-learning network; The first parameter value is fine-tuned based on the query sample set using the meta-learning network to obtain final values ​​of various parameters of the target object recognition network.

6. The method according to claim 5, characterized in that Before using the meta-learning network to perform similarity judgment on the recognition process of the meta-training task corresponding to each of the supporting samples in the supporting sample set and the recognition process of the target object recognition task, the method further includes: For each of the plurality of meta-training tasks, meta-training the first object recognition network is performed to obtain a second object recognition network; The training data of each of the meta-training tasks is obtained respectively to form the support samples.

7. The method according to claim 5 or 6, characterized in that: The recognition objects of the meta-training task corresponding to the supporting samples in the supporting sample set include: structural objects in different regions, different positions and / or different types on the medical image.

8. The method according to claim 7, characterized in that The medical image is a cerebral artery image; The recognition objects of the meta-training task corresponding to the supporting samples in the supporting sample set include: cerebral artery structure objects of different segments, different partitions and / or different types of cerebral arteries.

9. The method according to claim 5, characterized in that The target object recognition network includes: a basic structure network and a parallel graph convolution network; The infrastructure network adopts a multi-supervisory network of an encoding-decoding framework, and the parallel graph convolutional network is used to extract key semantic knowledge in multiple layers; The input of the graph convolutional network is a fusion graph of target medical structure data; The input of the basic structure network is the medical image and the first multi-level target medical structure-knowledge fusion data; The training samples of the meta-training task include pseudo-noise data, where the pseudo-noise data is formed based on target medical structure segmentation data and an optimal knowledge representation set, where the optimal knowledge representation set is a knowledge quantization representation data set used for target medical structure knowledge guidance.

10. The method according to claim 9, characterized in that The optimal knowledge representation set is generated by the following method: Acquire a medical image including a target medical structure, annotated data of a target object, and knowledge of the target medical structure, wherein the knowledge of the target medical structure includes structural hierarchy information, distribution information, and an identification standard of the target object; Performing knowledge quantification on the medical image of the target medical structure, the annotation data of the target object, and the knowledge of the target medical structure to obtain a variety of knowledge representation data; From the multiple knowledge representation data, some knowledge representation data are selected to form an optimal knowledge representation set.

11. The method according to claim 10, characterized in that The selecting part of the knowledge representation data from the multiple knowledge representation data to form an optimal knowledge representation set includes: Using similarity measurement to cluster and group the knowledge representation data; According to the knowledge representation data grouping, representative knowledge representation data are extracted respectively; evaluating the bootstrapping performance of the representative knowledge representation data; Select N knowledge representation methods with the best guiding performance to form the optimal knowledge representation set.

12. The method according to any one of claims 9 to 11, characterized in that: The first multi-level target medical structure-knowledge fusion data is obtained by the following method: stratifying the target medical structure knowledge according to the hierarchy of the target medical structure; stratifying the medical image data of the target medical structure according to a correspondence between a knowledge source of the target medical structure knowledge and the target medical structure; The target medical structure knowledge of each layer and the medical image data of the target medical structure are fused accordingly to form the first multi-level target medical structure-knowledge fusion data.

13. The method according to claim 12, characterized in that The hierarchical fusion strategy of the first multi-level target medical structure-knowledge fusion data is determined according to the following method: Adopting a plurality of different hierarchical fusion methods, the target medical structure knowledge of each layer and other medical image data of the target medical structure are correspondingly fused to form a plurality of second multi-level target medical structure-knowledge fusion data; The plurality of the second multi-level target medical structure-knowledge fusion data are subjected to at least one layer of convolution to form an output, and a three-dimensional information entropy is output based on mutual information calculation, and the three-dimensional information entropy is used to perform hierarchical clustering of the second multi-level target medical structure-knowledge fusion data to form a plurality of different second multi-level target medical structure-knowledge fusion data groups corresponding to different hierarchical fusion methods; Using multiple baseline networks, using different second multi-level target medical structure-knowledge fusion data groups, model training is performed, and statistical performance of the trained networks is performed; For different second multi-level target medical structure-knowledge fusion data groups, based on the performance of the multiple networks obtained through training, calculating a comprehensive guide performance score; According to the ranking of the guidance performance comprehensive score, selecting a fused data in the second multi-level target medical structure-knowledge fusion data group as the optimal fused data; The hierarchical fusion method used for the optimal fusion data is determined as the hierarchical fusion strategy for the first multi-level target medical structure-knowledge fusion data.

14. A model parameter determination device, characterized in that: The device is used for training a target object recognition model according to any one of claims 1 to 4, comprising: A similarity judgment module, used to use the meta-learning network to perform similarity judgment on the recognition process of the meta-training task and the recognition process of the target object recognition task corresponding to each of the supporting samples in the supporting sample set; A network parameter determination module, used to determine a first parameter value of each parameter in the target object recognition network based on a judgment result of the similarity judgment by the meta-learning network; A network parameter adjustment module is used to use the meta-learning network to fine-tune the first parameter value based on the query sample set to obtain the final value of each parameter of the target object recognition network.

15. A computer device, characterized in that: include: A memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the method according to any one of claims 5 to 13 by executing the computer instructions.

16. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a computer to execute the method according to any one of claims 5 to 13.

Images