Medical image lesion recognition method and system based on neural network

Abstract

The present application provides a method and system for medical image lesion recognition based on a neural network. First, the target medical image acquisition data is accurately positioned in the lesion area to generate candidate lesion areas, and then a medical image feature relationship network is used, which reveals the medical feature connection between multiple reference lesion areas. Then, a subnetwork corresponding to the candidate lesion area is extracted from the medical image feature relationship network to ensure that the members in the subnetwork are highly correlated with the reference lesion area. Further, through lesion knowledge vector analysis, these subnetworks are converted into lesion knowledge vectors, and finally blended into a comprehensive lesion knowledge vector. Finally, based on this blended lesion knowledge vector, the target medical image acquisition data is accurately identified and classified, which not only improves the accuracy of lesion identification, but also strengthens the understanding of the complex relationship between lesions.

Description

Technical Field

[0001] The present application relates to the field of deep learning technology, and more specifically, to a method and system for identifying medical image lesions based on a neural network. Background Art

[0002] In the field of medical imaging diagnosis, accurate identification and analysis of lesion areas are crucial for early detection and treatment of diseases. With the continuous development of medical imaging technology, a large amount of medical imaging data is generated, which provides a rich source of information for lesion identification and analysis. However, traditional medical imaging lesion identification methods mainly rely on the experience and knowledge of doctors. This method is not only time-consuming and labor-intensive, but also has certain subjectivity and risks of misjudgment.

[0003] In recent years, with the rapid development of artificial intelligence and machine learning technologies, medical image lesion recognition methods based on neural networks have gradually become a research hotspot. Neural networks have powerful feature extraction and classification capabilities, and can automatically learn features in medical images and accurately identify lesions. However, existing neural network-based lesion recognition methods still have some challenges.

[0004] First, accurate positioning of the lesion area is a key issue. Due to the complexity and diversity of medical images, accurate positioning of the lesion area is crucial for subsequent identification and analysis. Secondly, how to effectively use the feature information in medical images is also a difficult problem. Medical images contain a large amount of feature information, and how to extract and use this feature information is of great significance to improving the accuracy of lesion identification.

[0005] In addition, existing lesion recognition methods often only focus on single lesion features, while ignoring the medical feature-related connections between lesions. In fact, there may be complex connections and influences between different lesions, which are of great significance for accurately identifying lesions and judging the progression of the disease. Summary of the invention

[0006] In view of the above-mentioned problems, in combination with the first aspect of the present application, an embodiment of the present application provides a medical image lesion recognition method based on a neural network, the method comprising:

[0007] Positioning the lesion region on the target medical image acquisition data to generate at least one candidate lesion region;

[0008] Obtaining a medical image feature relationship network that characterizes the medical feature correlation between a plurality of reference lesion regions, and extracting medical image feature relationship sub-networks corresponding to each of the candidate lesion regions from the medical image feature relationship network, wherein a network member in the medical image feature relationship network has a unique corresponding relationship with the reference lesion region, the medical image feature relationship sub-network includes a target network member, and the reference lesion region corresponding to the target network member has a feature correlation degree with the corresponding candidate lesion region that is greater than a preset feature correlation degree;

[0009] Performing lesion knowledge vector analysis on each of the medical image feature relationship sub-networks to generate lesion knowledge vectors of each of the medical image feature relationship sub-networks, and blending each of the lesion knowledge vectors to generate a blended lesion knowledge vector of the target medical image acquisition data;

[0010] According to the fused lesion knowledge vector, lesion identification and classification are performed on the target medical image acquisition data to generate a lesion identification and classification result corresponding to the target medical image acquisition data.

[0011] In a possible implementation of the first aspect, the performing lesion region positioning on the target medical image acquisition data to generate at least one candidate lesion region includes:

[0012] Acquire medical image data to be detected, and perform medical feature recognition on the medical image data to be detected to generate a medical feature set of the medical image data to be detected;

[0013] When the medical feature set indicates that the medical image data to be detected meets a preset lesion feature condition, outputting the medical image data to be detected as the target medical image acquisition data;

[0014] Performing image segmentation processing on the target medical image acquisition data to generate a plurality of candidate image regions in the target medical image acquisition data;

[0015] For each of the candidate image regions, when the candidate image region meets a preset lesion feature condition, the candidate image region is output as the candidate lesion region.

[0016] In a possible implementation of the first aspect, after performing medical feature recognition on the to-be-detected medical image data to generate a medical feature set of the to-be-detected medical image data, the method further includes:

[0017] When the medical feature set indicates that the medical image data to be detected does not meet the preset lesion feature condition, for each candidate image region in the medical image data to be detected, respectively extract the candidate medical image feature relationship network corresponding to each candidate image region from the medical image feature relationship network, perform lesion knowledge vector analysis on each candidate medical image feature relationship network, generate candidate lesion knowledge vectors of each candidate medical image feature relationship network, and blend each candidate lesion knowledge vector to generate a candidate blended lesion knowledge vector of the medical image data to be detected;

[0018] According to the candidate fusion lesion knowledge vector, the medical image data to be detected is subjected to lesion recognition and classification, and a target lesion recognition and classification result corresponding to the medical image data to be detected is generated;

[0019] Among them, the target lesion identification and classification result meets the preset lesion characteristic conditions, and the target lesion identification and classification result includes the medical image data to be detected.

[0020] In a possible implementation of the first aspect, extracting the medical image feature relationship subnetworks corresponding to the candidate lesion regions from the medical image feature relationship network respectively includes:

[0021] For each of the candidate lesion regions, the candidate lesion region is respectively compared with each of the reference lesion regions in the medical image feature relationship network to generate comparison details, wherein the comparison details indicate whether there is a reference lesion region identical to the candidate lesion region in the medical image feature relationship network;

[0022] Determining, according to the comparison details, a target reference lesion region corresponding to the candidate lesion region from each of the reference lesion regions in the medical image feature relationship network;

[0023] Outputting the network member corresponding to the target reference lesion region in the medical image feature relationship network as the target network member corresponding to the candidate lesion region, and outputting the network member having a network dependency relationship with the target network member in the medical image feature relationship network as the candidate network member corresponding to the candidate lesion region;

[0024] Outputting the sub-network structure composed of the target network members and the candidate network members in the medical image feature relationship network as a medical image feature relationship sub-network corresponding to the candidate lesion area;

[0025] Wherein, determining, based on the comparison details, from the reference lesion regions in the medical image feature relationship network, a target reference lesion region corresponding to the candidate lesion region comprises:

[0026] When the comparison details represent that there is a reference lesion region identical to the candidate lesion region in the medical image feature relationship network, outputting the reference lesion region identical to the candidate lesion region as a target reference lesion region corresponding to the candidate lesion region;

[0027] When the comparison details indicate that there is no reference lesion region identical to the candidate lesion region in the medical image feature relationship network, obtaining the feature correlation degree between the candidate lesion region and each of the reference lesion regions;

[0028] At least one reference lesion region whose feature correlation degree is greater than the preset feature correlation degree is output as a target reference lesion region corresponding to the candidate lesion region.

[0029] In a possible implementation of the first aspect, the lesion recognition and classification is implemented by a lesion feature attention model and a lesion recognition and classification model, and the target medical image acquisition data is subjected to lesion recognition and classification based on the fused lesion knowledge vector to generate a lesion recognition and classification result corresponding to the target medical image acquisition data, including:

[0030] By using the lesion feature attention model, the blended lesion knowledge vector is subjected to self-attention processing to generate a target lesion knowledge vector of the target medical image acquisition data;

[0031] Through the lesion recognition and classification model, according to the target lesion knowledge vector, the target medical image acquisition data is subjected to lesion recognition and classification, and a lesion recognition and classification result corresponding to the target medical image acquisition data is generated.

[0032] In a possible implementation of the first aspect, the lesion feature attention model includes a plurality of self-attention processing units, and the self-attention processing is performed on the fused lesion knowledge vector through the lesion feature attention model to generate a target lesion knowledge vector of the target medical image acquisition data, including:

[0033] Performing a first round of self-attention processing on the blended lesion knowledge vector through a first self-attention processing unit to generate a first target lesion knowledge vector;

[0034] Traverse x, and perform the xth self-attention processing on the blended lesion knowledge vector according to the x-1th target lesion knowledge vector through the xth self-attention processing unit to generate the xth target lesion knowledge vector; wherein 1<x<X, X represents the number of the self-attention processing units in the lesion feature attention model;

[0035] The Xth target lesion knowledge vector is output as the target lesion knowledge vector of the target medical image acquisition data.

[0036] In a possible implementation of the first aspect, before performing self-attention processing on the fusion lesion knowledge vector through the lesion feature attention model to generate a target lesion knowledge vector of the target medical image acquisition data, the method further includes:

[0037] Obtaining a sample lesion feature attention model including a plurality of sample self-attention processing units, sample medical image acquisition data, and a lesion label knowledge vector of lesion label data;

[0038] Positioning the lesion region on the sample medical image acquisition data to generate at least one example candidate lesion region, and extracting example medical image feature relationship networks corresponding to each of the example candidate lesion regions from the medical image feature relationship network;

[0039] Performing lesion knowledge vector analysis on each of the example medical image feature relationship networks to generate example lesion knowledge vectors for each of the example medical image feature relationship networks, and blending each of the example lesion knowledge vectors to generate an example blended lesion knowledge vector for the sample medical image acquisition data;

[0040] Sequentially performing self-attention processing on the example blended lesion knowledge vector through each of the example self-attention processing units of the example lesion feature attention model, and generating example target lesion knowledge vectors corresponding to each of the example self-attention processing units;

[0041] Based on each of the example target lesion knowledge vectors and the lesion label knowledge vector, a training error parameter of the sample lesion feature attention model is determined, and based on the training error parameter, the sample lesion feature attention model is optimized to generate the lesion feature attention model.

[0042] In a possible implementation of the first aspect, the sample lesion feature attention model includes X sample self-attention processing units, the lesion label knowledge vector and the vector dimension of each of the example target lesion knowledge vectors are consistent, and determining the training error parameter of the sample lesion feature attention model based on each of the example target lesion knowledge vectors and the lesion label knowledge vector includes:

[0043] Traversing y, determining the difference knowledge vector between the example target lesion knowledge vector corresponding to the y+1th sample self-attention processing unit and the example target lesion knowledge vector corresponding to the yth sample self-attention processing unit, generating the difference knowledge vector corresponding to the yth sample self-attention processing unit, and outputting the mapping value of the feature matrix size of the difference knowledge vector as the unit training error parameter corresponding to the yth sample self-attention processing unit, where 1≤y≤X;

[0044] Adding the unit training error parameters to generate a first training error parameter;

[0045] Determine the difference knowledge vector of the example target lesion knowledge vector corresponding to the X-th sample self-attention processing unit and the lesion label knowledge vector to generate a difference knowledge vector, and output the mapping value of the feature matrix size of the difference knowledge vector as the second training error parameter of the sample lesion feature attention model;

[0046] The first training error parameter and the second training error parameter are weightedly fused to generate the training error parameter.

[0047] In a possible implementation of the first aspect, performing lesion identification and classification on the target medical image acquisition data according to the fused lesion knowledge vector to generate a lesion identification and classification result corresponding to the target medical image acquisition data includes:

[0048] When the lesion recognition category of the target medical image acquisition data is an recognition category for determining the existence of a lesion in the target medical image acquisition data, performing lesion recognition on the target medical image acquisition data according to the fused lesion knowledge vector, and generating a conclusion on whether a lesion exists in the target medical image acquisition data;

[0049] When the lesion recognition category of the target medical image acquisition data is an recognition category used to determine the specific type of lesions in the target medical image acquisition data, the target medical image acquisition data is lesion classified according to the fused lesion knowledge vector to obtain the specific type of lesions in the target medical image acquisition data.

[0050] On the other hand, an embodiment of the present application also provides a medical image lesion recognition system based on a neural network, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0051] Based on the above aspects, by locating the lesion area of ​​the target medical image acquisition data, at least one candidate lesion area can be generated, and the medical image feature relationship network can be used to characterize the medical feature correlation between multiple reference lesion areas. By extracting the medical image feature relationship subnetwork corresponding to each candidate lesion area from the medical image feature relationship network, the medical features of the lesion can be fully captured, and the lesion knowledge vector of each medical image feature relationship subnetwork is parsed to generate the lesion knowledge vector of each subnetwork, and the lesion knowledge vector is further blended to form the blended lesion knowledge vector of the target medical image acquisition data, which not only realizes the deep extraction of lesion features, but also strengthens the global expression of lesion features through vector fusion. Based on the blended lesion knowledge vector, the target medical image acquisition data can be accurately identified and classified for lesions, and the corresponding lesion identification and classification results can be generated. Since the blended lesion knowledge vector is rich in comprehensive and in-depth lesion feature information, it can support high-precision lesion identification and classification, providing an important basis for medical diagnosis and treatment. As a result, the accuracy and efficiency of medical image lesion recognition have been significantly improved through precise lesion localization, comprehensive feature capture, effective lesion knowledge vector analysis and fusion, and precise lesion identification and classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a schematic diagram of the execution flow of the neural network-based medical image lesion recognition method provided in an embodiment of the present application.

[0053] Figure 2 It is a schematic diagram of the hardware architecture of a medical image lesion recognition system based on a neural network provided in an embodiment of the present application. DETAILED DESCRIPTION

[0054] The present application will be described in detail below with reference to the accompanying drawings. Figure 1 This is a flow chart of a medical image lesion recognition method based on a neural network provided by an embodiment of the present application. The medical image lesion recognition method based on a neural network is introduced in detail below.

[0055] Step S110 , locating the lesion region on the target medical image acquisition data, and generating at least one candidate lesion region.

[0056] The target medical image acquisition data may refer to medical image data that needs to be used for lesion area location, such as medical images such as CT, MRI or X-ray. For example, a patient undergoes a lung CT scan, and the resulting CT image is the target medical image acquisition data.

[0057] Lesion area localization is to find the possible lesion location in medical images. For example, in lung CT images, image processing techniques such as edge detection and threshold segmentation are used to mark the location of possible lung nodules or tumors.

[0058] The candidate lesion region may refer to a region that is initially identified as a possible lesion location in the target medical image acquisition data. For example, in a CT image, a marked suspected pulmonary nodule region is a candidate lesion region.

[0059] That is, in this embodiment, the server receives a batch of medical image acquisition data, which comes from a CT scanning device in a hospital. The server first analyzes these medical image acquisitions using a pre-trained lesion area localization model, and identifies the areas in the medical image acquisition that may contain lesions through a neural network algorithm.

[0060] For example, in a lung CT image, an abnormal nodule region is identified, which has significant differences in density and texture compared with the surrounding lung tissue. The server marks this abnormal nodule region as a candidate lesion region and saves its coordinate position in memory for subsequent processing.

[0061] Step S120, obtaining a medical image feature relationship network that characterizes the medical feature correlation between multiple reference lesion areas, and extracting medical image feature relationship sub-networks corresponding to each of the candidate lesion areas from the medical image feature relationship network, wherein network members in the medical image feature relationship network have a unique corresponding relationship with the reference lesion area, and the medical image feature relationship sub-network includes a target network member, and the reference lesion area corresponding to the target network member has a feature correlation degree with the corresponding candidate lesion area that is greater than a preset feature correlation degree.

[0062] In detail, the medical image feature relationship network characterizes the medical feature related connections between multiple reference lesion regions. Each member (or node) in the medical image feature relationship network represents a reference lesion region, and the connection between members represents the feature relationship between these reference lesion regions. It can be imagined as a graph consisting of multiple points (representing lesion regions) and lines (representing the relationship between these regions). For example, if two lesion regions have similar shape, size or texture features in a medical image, then they will be connected in the medical image feature relationship network.

[0063] The medical image feature relationship sub-network is a partial network related to a specific candidate lesion region extracted from the entire medical image feature relationship network, and the medical image feature relationship sub-network includes reference lesion regions that have feature associations with the candidate lesion region. For example, assuming there is a candidate lesion region A, in the medical image feature relationship network, reference lesion regions B, C, and D that have similar features to A will be extracted and together with A form a sub-network.

[0064] Then, the network member represents a point or node of a reference lesion region in the medical image feature relationship network. That is, in the medical image feature relationship network, each reference lesion region is represented as a network member.

[0065] The feature correlation is an indicator to measure the similarity between two lesion regions. If the two lesion regions are similar in medical features such as shape, size, texture, etc., then their feature correlation is high. For example, suppose there are two images of lung nodules. If they are very similar in shape, density and texture, then the feature correlation between the two nodules is very high.

[0066] That is, in this embodiment, the server then accesses a medical image feature relationship network database, which stores the medical image features of many known lesion areas and the relationship network between these medical image features. Each network member represents a lesion feature of a specific reference lesion area, such as shape, size, edge sharpness, etc.

[0067] The server uses the features of the candidate lesion region located in step S110 to compare with the reference lesion regions in the medical image feature relationship network database. By calculating the feature similarity, the server finds several reference lesion regions that are most similar to the candidate lesion region features and determines the feature correlation between them.

[0068] Then, the server extracts a medical image feature relationship sub-network related to these reference lesion regions from the entire medical image feature relationship network, and the medical image feature relationship sub-network contains medical features that are highly correlated with the candidate lesion regions and their mutual relationships.

[0069] Step S130, performing lesion knowledge vector analysis on each of the medical image feature relationship sub-networks to generate lesion knowledge vectors for each of the medical image feature relationship sub-networks, and blending each of the lesion knowledge vectors to generate a blended lesion knowledge vector for the target medical image acquisition data.

[0070] In detail, the lesion knowledge vector is a numerical and vectorized representation of the characteristics of the lesion area, which contains various characteristic information about the lesion, such as shape, size, texture, location, etc., and encodes this information into a multi-dimensional vector. For example, for a lung nodule, its lesion knowledge vector may include information of multiple dimensions such as the sphericity, burr degree, and calcification degree of the nodule, and each dimension is a component of the vector.

[0071] Fusion here refers to merging or integrating multiple lesion knowledge vectors to form a more comprehensive and integrated lesion feature representation, which involves weighted averaging, feature splicing or other forms of fusion of different lesion knowledge vectors. For example, assuming there are two lesion knowledge vectors A and B of candidate lesion areas, fusion can be to weighted average the corresponding features in A and B to form a new lesion knowledge vector C that integrates the features of A and B.

[0072] Then, the fused lesion knowledge vector is a comprehensive feature vector obtained by blending multiple lesion knowledge vectors, integrating the feature information of multiple candidate lesion regions, and is used to more accurately describe the lesion features in the target medical image. For example, in a lung CT image, if there are multiple areas suspected of being lung nodules, a lesion knowledge vector can be generated for each area, and then these vectors are blended to obtain a fused lesion knowledge vector for subsequent lesion identification and classification.

[0073] Thus, in this embodiment, the server continues to process each extracted medical image feature relationship sub-network. In detail, the lesion knowledge contained in the medical image feature relationship sub-network can be converted into a vector form, and the conversion process takes into account the relationship and weight between the various features in the sub-network to ensure that important features are properly reflected in the vector. The server blends these lesion knowledge vectors, that is, combines the information of multiple lesion knowledge vectors to form a more comprehensive and richer blended lesion knowledge vector, which not only contains the direct features of the candidate lesion area, but also reflects the position of these features in the medical knowledge network and the relationship with other features.

[0074] Step S140 , performing lesion recognition and classification on the target medical image acquisition data according to the fused lesion knowledge vector, and generating a lesion recognition and classification result corresponding to the target medical image acquisition data.

[0075] In this embodiment, lesion recognition and classification refers to the process of automatically classifying and identifying lesions in the target medical image according to the fused lesion knowledge vector, and the type of lesion (such as benign, malignant, etc.) can be determined according to the fused lesion knowledge vector.

[0076] The lesion recognition and classification result refers to the output result obtained after the lesion recognition and classification process, indicating the type or nature of the lesion in the target medical image. The result can be one or more of multiple possibilities such as benign, malignant, inflammatory, tuberculosis, etc. For example, in the lesion recognition and classification of lung CT images, the output result may be a label (such as "malignant"), or a probability distribution (such as the probability of malignancy is 0.7, the probability of benign is 0.3, etc.), which is used to instruct doctors to make further diagnosis and treatment decisions.

[0077] That is, in the end, the server uses the fusion lesion knowledge vector to identify and classify the original medical image acquisition data. For example, the server passes the fusion lesion knowledge vector as input to the corresponding neural network model. The neural network model determines the type and severity of the lesions in the target medical image acquisition data by analyzing the information in the fusion lesion knowledge vector. For example, in the context of a lung CT image, it can be identified whether the nodule is benign or malignant, or the type of nodule and the corresponding probability. The server saves the output results and sends them back to the hospital's information system for the doctor's reference and decision-making. In this way, through automated lesion identification and classification, the efficiency and accuracy of diagnosis can be greatly improved.

[0078] Based on the above steps, by locating the lesion area of ​​the target medical image acquisition data, at least one candidate lesion area can be generated, and the medical image feature relationship network can be used to characterize the medical feature correlation between multiple reference lesion areas. By extracting the medical image feature relationship subnetwork corresponding to each candidate lesion area from the medical image feature relationship network, the medical features of the lesion can be fully captured, and the lesion knowledge vector of each medical image feature relationship subnetwork is parsed to generate the lesion knowledge vector of each subnetwork, and the lesion knowledge vector is further blended to form the blended lesion knowledge vector of the target medical image acquisition data, which not only realizes the deep extraction of lesion features, but also strengthens the global expression of lesion features through vector fusion. According to the blended lesion knowledge vector, the target medical image acquisition data can be accurately identified and classified, and the corresponding lesion identification and classification results can be generated. Since the blended lesion knowledge vector is rich in comprehensive and in-depth lesion feature information, it can support high-precision lesion identification and classification, providing an important basis for medical diagnosis and treatment. As a result, the accuracy and efficiency of medical image lesion recognition have been significantly improved through precise lesion localization, comprehensive feature capture, effective lesion knowledge vector analysis and fusion, and precise lesion identification and classification.

[0079] In a possible implementation, step S110 may include:

[0080] Step S111 : acquiring medical image data to be detected, and performing medical feature recognition on the medical image data to be detected to generate a medical feature set of the medical image data to be detected.

[0081] In this embodiment, the server automatically receives a set of new CT scan data from the hospital's picture archiving and communication system (PACS). These CT scan data are high-resolution images of the patient's lungs and are used for preliminary screening of lung diseases.

[0082] The server uses deep learning algorithms, such as convolutional neural networks (CNN), to extract features from the lung CT scan data to be tested. During this process, it automatically identifies and extracts key medical features in the image, such as the structure, texture, and density changes of the lungs, and generates a set containing these features.

[0083] Step S112: when the medical feature set indicates that the medical image data to be detected meets a preset lesion feature condition, the medical image data to be detected is output as the target medical image acquisition data.

[0084] In this embodiment, the server compares the extracted medical features with pre-set lesion feature conditions, which may be established based on a large number of historical cases and medical expert knowledge bases to identify potential lesion areas. If the extracted feature set highly matches the pre-set conditions, such as abnormal nodular structures, density changes, etc., it is determined that the medical image data to be detected may contain lesions.

[0085] After comparison, the server determines that the group of CT scan data meets the preset lesion characteristic conditions, so it marks it as target medical imaging acquisition data and prepares for the next detailed analysis.

[0086] Step S113: performing image segmentation processing on the target medical image acquisition data to generate a plurality of candidate image regions in the target medical image acquisition data.

[0087] The server uses image segmentation algorithms, such as region-based segmentation or threshold segmentation, to process the target medical image acquisition data. This process aims to separate different tissues, organs and potential lesion areas in the image to form multiple candidate image regions.

[0088] Step S114: for each of the candidate image regions, when the candidate image region meets a preset lesion feature condition, output the candidate image region as the candidate lesion region.

[0089] For each candidate image region, the server again uses the preset lesion feature conditions for screening, which may include the size, shape, density, edge features, etc. Only when the candidate region meets these conditions will it be considered as a potential lesion region.

[0090] After careful screening and comparison, the server finally identified several candidate areas that met the preset lesion characteristics and marked these areas as candidate lesion areas. These candidate lesion areas may be potential lung nodules, tumors or other abnormal structures, which require further diagnosis and treatment planning by medical experts.

[0091] In a possible implementation manner, after step S111, the method further includes:

[0092] Step A110, when the medical feature set represents that the medical image data to be detected does not meet the preset lesion feature conditions, for each candidate image area in the medical image data to be detected, the candidate medical image feature relationship network corresponding to each candidate image area is extracted from the medical image feature relationship network, and the lesion knowledge vector of each candidate medical image feature relationship network is parsed to generate a candidate lesion knowledge vector of each candidate medical image feature relationship network, and the candidate lesion knowledge vectors are blended to generate a candidate blended lesion knowledge vector of the medical image data to be detected.

[0093] Step A120: performing lesion recognition and classification on the medical image data to be detected based on the candidate fusion lesion knowledge vector, and generating a target lesion recognition and classification result corresponding to the medical image data to be detected.

[0094] Among them, the target lesion identification and classification result meets the preset lesion characteristic conditions, and the target lesion identification and classification result includes the medical image data to be detected.

[0095] In this embodiment, the server has performed medical feature recognition on the medical image data to be detected through an algorithm and obtained a medical feature set. Now, the server compares this medical feature set with the preset lesion feature conditions. The preset lesion feature conditions may include specific features such as shape, size, density or texture, which are set by medical experts based on historical cases and clinical experience to identify potential lesion areas.

[0096] After comparison, the server finds that the medical feature set does not fully meet the preset lesion feature conditions, which means that based on the preliminary feature recognition, the medical image data to be tested does not directly show an obvious lesion area. However, in order not to give up any possible lesions, the server can proceed to the next step of in-depth analysis.

[0097] For each candidate image region in the medical image data to be detected, the server extracts the corresponding candidate medical image feature relationship network from the previously constructed medical image feature relationship network. This process is similar to finding subgraphs associated with specific regions in a large network graph. These subgraphs (i.e., candidate medical image feature relationship networks) contain feature relationships of regions that are similar to each candidate image region and have been historically confirmed to be lesions.

[0098] For each candidate medical image feature relationship network, the server can perform lesion knowledge vector analysis, which is a process of converting the feature relationship in the network into a numerical vector representation to facilitate subsequent computer processing and analysis. Then, the server can blend all candidate lesion knowledge vectors, that is, combine the information of all vectors to generate a comprehensive candidate blended lesion knowledge vector, which integrates the feature information of multiple candidate image regions and provides a more comprehensive lesion description.

[0099] Based on the candidate fusion lesion knowledge vectors, the server can use machine learning or deep learning models to perform lesion recognition and classification on the medical image data to be detected. This process is similar to the classification task in image recognition, but is more focused on the lesion features in medical images. The classification result will be marked as the target lesion recognition classification result. This result meets the preset lesion feature conditions and explicitly includes the medical image data to be detected. In other words, even if the preliminary set of medical features does not directly show the lesion, through in-depth network analysis and vector fusion, the server can still identify potential lesion areas and classify them into specific lesion types (such as tumors, nodules, etc.). This process greatly improves the sensitivity and accuracy of lesion detection.

[0100] In a possible implementation, step S120 may include:

[0101] Step S121, for each of the candidate lesion regions, the candidate lesion region is compared with each of the reference lesion regions in the medical image feature relationship network to generate comparison details, wherein the comparison details indicate whether there is a reference lesion region identical to the candidate lesion region in the medical image feature relationship network.

[0102] In this embodiment, the server first obtains multiple candidate lesion regions, which may be abnormal regions detected in medical images. Then, the server can obtain a series of reference lesion regions from the previously constructed medical image feature relationship network. Then, for each candidate lesion region, the server can compare it one by one with the reference lesion region in the medical image feature relationship network. The comparison process will generate a comparison detail, which records which reference lesion regions the candidate lesion region is similar to, as well as the specific value of the similarity and other information.

[0103] Step S122 , determining a target reference lesion region corresponding to the candidate lesion region from each of the reference lesion regions in the medical image feature relationship network according to the comparison details.

[0104] In this embodiment, based on the comparison details generated in the previous step, the server can analyze and determine the reference lesion region that is most similar to the candidate lesion region, and the selected reference lesion region is called the target reference lesion region. The process of determining the target reference lesion region may be based on a comprehensive consideration of multiple factors such as similarity score, shape matching, and texture features.

[0105] Step S123, outputting the network member corresponding to the target reference lesion area in the medical image feature relationship network as the target network member corresponding to the candidate lesion area, and outputting the network member having a network dependency relationship with the target network member in the medical image feature relationship network as the candidate network member corresponding to the candidate lesion area.

[0106] Once the target reference lesion area is determined, the server can further extract network members directly associated with the target reference lesion area from the medical image feature relationship network. These network members are called target network members, which may include tissues, blood vessels, lymph nodes, etc. directly adjacent to the lesion area. In addition, the server will also identify those network members that have direct or indirect network dependencies with the target network members, which are called candidate network members. These candidate network members may be functionally or structurally associated with the target reference lesion area.

[0107] Step S124, outputting the sub-network structure composed of the target network members and the candidate network members in the medical image feature relationship network as a medical image feature relationship sub-network corresponding to the candidate lesion area.

[0108] Finally, the server can combine the target network members and the candidate network members to form a sub-network structure, which retains the feature relationship most directly related to the target reference lesion area, and is therefore called a medical imaging feature relationship sub-network. This sub-network can provide important reference information for subsequent lesion analysis and diagnosis, helping doctors to more accurately understand the nature, location and possible impact of the lesions.

[0109] Wherein, step S122 may include:

[0110] Step S1221, when the comparison details represent that there is a reference lesion region identical to the candidate lesion region in the medical image feature relationship network, the reference lesion region identical to the candidate lesion region is output as a target reference lesion region corresponding to the candidate lesion region.

[0111] Step S1222, when the comparison details indicate that there is no reference lesion region identical to the candidate lesion region in the medical image feature relationship network, obtain the feature correlation between the candidate lesion region and each of the reference lesion regions.

[0112] Step S1223: output at least one reference lesion region whose feature correlation degree is greater than the preset feature correlation degree as a target reference lesion region corresponding to the candidate lesion region.

[0113] In this embodiment, after receiving the candidate lesion area, the server will first use the previously generated comparison details to determine whether there is a reference lesion area that is exactly the same in the medical imaging feature relationship network. This judgment process may be completed by comparing the medical features such as shape, size, position, texture, etc. of the two.

[0114] Scenario A:

[0115] If the server finds that there is indeed a reference lesion region that is exactly the same as the candidate lesion region in the medical image feature relationship network, then the server directly outputs the same reference lesion region as the target reference lesion region corresponding to the candidate lesion region.

[0116] Case B:

[0117] If the server cannot find a reference lesion region that is exactly the same as the candidate lesion region in the medical image feature relationship network, then the next step of the processing flow is entered.

[0118] When no reference lesion region that is exactly the same as the candidate lesion region is found in the medical image feature relationship network, the server can start calculating the feature correlation between the candidate lesion region and each reference lesion region in the network. The feature correlation is calculated by comparing the medical features of the two, and may include shape similarity, texture similarity, positional relationship and other aspects.

[0119] After calculating the feature correlation between the candidate lesion area and each reference lesion area, the server will compare these feature correlations with a preset feature correlation threshold. Only when the feature correlation between a reference lesion area and the candidate lesion area is greater than the preset threshold, the reference lesion area will be output by the server as the target reference lesion area corresponding to the candidate lesion area. This process is to find reference lesion areas that are highly similar to the candidate lesion area in medical features, although not completely the same, so as to provide more reference information for subsequent analysis and diagnosis.

[0120] In a possible implementation, the lesion recognition and classification is implemented by a lesion feature attention model and a lesion recognition and classification model, and step S140 may include:

[0121] Step S141 , performing self-attention processing on the blended lesion knowledge vector through the lesion feature attention model to generate a target lesion knowledge vector of the target medical image acquisition data.

[0122] In this embodiment, the server first calls the lesion feature attention model, which is specially trained to identify and process lesion features. When the server inputs the fusion lesion knowledge vectors into the lesion feature attention model, the lesion feature attention model processes these vectors using the self-attention mechanism.

[0123] The process of self-attention can be imagined as the lesion feature attention model examining these intertwined lesion knowledge vectors and focusing on the features that are most critical for lesion identification. In this way, the model can extract the most important information in the intertwined lesion knowledge vector, that is, the information that best represents the lesion characteristics in the target medical image acquisition data.

[0124] After being processed by the lesion feature attention model, the server can obtain a new target lesion knowledge vector, which is more refined and focused and mainly contains key information for lesion identification.

[0125] Step S142, using the lesion recognition and classification model, based on the target lesion knowledge vector, the target medical image acquisition data is subjected to lesion recognition and classification, and a lesion recognition and classification result corresponding to the target medical image acquisition data is generated.

[0126] Next, the server may input the target lesion knowledge vector into a lesion recognition and classification model, which is specifically used for classification based on lesion features.

[0127] The lesion recognition and classification model analyzes the target lesion knowledge vector and classifies the target medical imaging data according to the lesion feature information contained therein. The classification process may be based on deep learning algorithms, such as convolutional neural networks (CNNs) or other machine learning classifiers.

[0128] Finally, the lesion recognition and classification model will output a classification result, which clearly indicates the type of lesions present in the target medical image acquisition data. For example, it may determine whether the lesions in the image data are benign or malignant, or what specific disease type they belong to.

[0129] The server can save the classification results and possibly send them to the terminal devices of medical personnel for further analysis and diagnosis. In this way, medical personnel can formulate corresponding treatment plans based on the classification results to help patients recover their health.

[0130] In a possible implementation, the lesion feature attention model includes a plurality of self-attention processing units, and step S141 may include:

[0131] Step S1411, performing a first round of self-attention processing on the blended lesion knowledge vector through a first self-attention processing unit to generate a first target lesion knowledge vector.

[0132] In this embodiment, the server first starts the first self-attention processing unit in the lesion feature attention model, and the self-attention processing unit receives the fusion lesion knowledge vectors as input, and these fusion lesion knowledge vectors contain the original lesion feature information extracted from the medical image.

[0133] The first self-attention processing unit will perform the first round of self-attention processing on these fused lesion knowledge vectors. This process is similar to a screening or weighted operation. The first self-attention processing unit will focus on those features that are crucial for lesion identification, while reducing the attention to unimportant features.

[0134] After the first round of self-attention processing, the server can obtain the first target lesion knowledge vector, which has optimized and refined the fusion lesion knowledge vector and further highlights the key features.

[0135] Step S1412, traverse x, and perform the xth self-attention processing on the blended lesion knowledge vector according to the x-1th target lesion knowledge vector through the xth self-attention processing unit to generate the xth target lesion knowledge vector. Wherein, 1<x<X, X represents the number of the self-attention processing units in the lesion feature attention model.

[0136] Step S1413: output the Xth target lesion knowledge vector as the target lesion knowledge vector of the target medical image acquisition data.

[0137] Next, the server can enter a loop process, traversing each subsequent self-attention processing unit in the lesion feature attention model. For the xth (1<x<X) self-attention processing unit, the x-1th target lesion knowledge vector is received as input. Each self-attention processing unit will perform further self-attention processing on the input lesion knowledge vector. During this process, the features will continue to be optimized and refined, gradually focusing on the information that is most critical for lesion identification. After each self-attention processing unit, the server will obtain an updated target lesion knowledge vector, which is getting closer and closer to the core information that best represents the lesion characteristics during the gradual refinement process.

[0138] After the server traverses all the self-attention processing units, it obtains the Xth target lesion knowledge vector, which is the final result obtained after multiple rounds of self-attention processing and contains highly refined and optimized lesion feature information.

[0139] The server can output this Xth target lesion knowledge vector as the target lesion knowledge vector of the target medical image acquisition data. The target lesion knowledge vector will serve as an important input of the subsequent lesion recognition and classification model to further improve the accuracy and efficiency of lesion recognition.

[0140] In a possible implementation manner, before step S141, the method further includes:

[0141] Step S1401, obtaining a sample lesion feature attention model including multiple sample self-attention processing units, sample medical image acquisition data, and a lesion label knowledge vector of lesion label data.

[0142] In this embodiment, the server first obtains the sample lesion feature attention model, sample medical image acquisition data, and the lesion label knowledge vector of the corresponding lesion label data from the database. These data are important resources for training and optimizing the lesion feature attention model.

[0143] The sample lesion feature attention model contains multiple sample self-attention processing units, which will be used in the subsequent training and optimization process. The sample medical image acquisition data is a series of annotated medical image data, which will be input into the sample lesion feature attention model as training samples. The lesion label knowledge vector is the real lesion label corresponding to these medical image data, which is used to calculate the training error of the sample lesion feature attention model.

[0144] Step S1402, locating the lesion region on the sample medical image acquisition data, generating at least one example candidate lesion region, and extracting example medical image feature relationship networks corresponding to each of the example candidate lesion regions from the medical image feature relationship network.

[0145] The server will then locate the lesion area on the sample medical image acquisition data, which is usually achieved through image processing and analysis techniques, such as edge detection, threshold segmentation, etc. After the positioning is completed, the server can obtain at least one example candidate lesion area.

[0146] Subsequently, the server may extract example medical image feature relationship networks corresponding to these example candidate lesion regions from the previously constructed medical image feature relationship networks, respectively. These example medical image feature relationship networks include various medical features related to the candidate lesion regions and the relationships between them.

[0147] Step S1403, performing lesion knowledge vector analysis on each of the example medical image feature relationship networks to generate example lesion knowledge vectors for each of the example medical image feature relationship networks, and blending each of the example lesion knowledge vectors to generate an example blended lesion knowledge vector for the sample medical image acquisition data.

[0148] For each example medical image feature relationship network, the server can perform lesion knowledge vector parsing, which is a process of converting the feature information in the network example medical image feature relationship network into the form of a numerical vector for subsequent mathematical calculations and model processing. After the parsing is completed, the server can obtain the example lesion knowledge vectors of each example medical image feature relationship network. Then, the server can blend these vectors, that is, merge them into a more comprehensive example blended lesion knowledge vector, which will be used as the input for subsequent self-attention processing.

[0149] Step S1404, performing self-attention processing on the example blended lesion knowledge vector respectively through each of the example self-attention processing units of the example lesion feature attention model in turn, and generating example target lesion knowledge vectors corresponding to each of the example self-attention processing units respectively.

[0150] Next, the server can perform self-attention processing on the example fusion lesion knowledge vector in turn through each sample self-attention processing unit in the sample lesion feature attention model. During this process, each sample self-attention processing unit will pay attention to and emphasize different feature parts in the vector to generate the corresponding example target lesion knowledge vector.

[0151] After completing the processing of all sample self-attention processing units, the server can calculate the training error parameters of the model based on these example target lesion knowledge vectors and lesion label knowledge vectors. This is usually achieved through a loss function, such as mean square error, cross entropy, etc.

[0152] Step S1405, determining the training error parameters of the sample lesion feature attention model based on each of the example target lesion knowledge vectors and the lesion label knowledge vector, and optimizing the sample lesion feature attention model based on the training error parameters to generate the lesion feature attention model.

[0153] Finally, the server can optimize the sample lesion feature attention model based on the calculated training error parameters. The optimization process may include adjusting the model parameters, changing the network structure, updating the learning rate, etc. The goal of the optimization is to reduce the training error and improve the prediction accuracy of the model. After multiple iterations and optimizations, the server can obtain a lesion feature attention model with better performance. The model will be able to process the fusion lesion knowledge vector more accurately and generate a more accurate target lesion knowledge vector, providing a more reliable basis for subsequent lesion identification and classification.

[0154] In a possible implementation, the sample lesion feature attention model includes X sample self-attention processing units, the lesion label knowledge vector and the vector dimension of each of the example target lesion knowledge vectors are consistent, and step S1405 includes:

[0155] Step S1405-1, traverse y, determine the difference knowledge vector between the example target lesion knowledge vector corresponding to the y+1th sample self-attention processing unit and the example target lesion knowledge vector corresponding to the yth sample self-attention processing unit, generate the difference knowledge vector corresponding to the yth sample self-attention processing unit, and output the mapping value of the feature matrix size of the difference knowledge vector as the unit training error parameter corresponding to the yth sample self-attention processing unit, where 1≤y≤X.

[0156] In this embodiment, the server first starts to traverse the X sample self-attention processing units in the sample lesion feature attention model. For each sample self-attention processing unit, the server determines the difference knowledge vector.

[0157] Specifically, when traversing to the y-th sample self-attention processing unit (1≤y≤X), the server can obtain the example target lesion knowledge vector corresponding to the y+1-th sample self-attention processing unit and the example target lesion knowledge vector corresponding to the y-th sample self-attention processing unit. These two vectors represent the lesion feature information after processing by the y+1-th and y-th self-attention processing units, respectively.

[0158] The server then calculates the difference between the two example target lesion knowledge vectors and generates a difference knowledge vector corresponding to the y-th sample self-attention processing unit, which reflects the error or information loss that may be introduced by the y-th unit during the processing process.

[0159] The server can then calculate the size of the feature matrix of this difference knowledge vector and map it to a specific value, which is the unit training error parameter corresponding to the y-th sample self-attention processing unit, which measures the accuracy of the y-th unit in processing lesion features.

[0160] Step S1405-2: Add the training error parameters of each unit to generate a first training error parameter.

[0161] After the server traverses all sample self-attention processing units and calculates the unit training error parameters of each sample self-attention processing unit, these unit training error parameters can be added to generate a total first training error parameter, which reflects the overall error of the entire sample lesion feature attention model when processing lesion features.

[0162] Step S1405-3, determine the difference knowledge vector of the example target lesion knowledge vector corresponding to the Xth sample self-attention processing unit and the lesion label knowledge vector, generate a difference knowledge vector, and output the mapping value of the feature matrix size of the difference knowledge vector as the second training error parameter of the sample lesion feature attention model.

[0163] Next, the server can pay special attention to the output of the last (i.e., the Xth) sample self-attention processing unit. For example, the example target lesion knowledge vector corresponding to this sample self-attention processing unit can be compared with the true lesion label knowledge vector, and the difference knowledge vector between them can also be calculated.

[0164] The server can then calculate the size of the feature matrix of the difference knowledge vector and map it to a specific value, which is the second training error parameter of the sample lesion feature attention model. The second training error parameter measures the gap between the final output and the true label.

[0165] Step S1405-4: weightedly fuse the first training error parameter and the second training error parameter to generate the training error parameter.

[0166] Finally, the server can perform weighted fusion of the first training error parameter and the second training error parameter to generate a comprehensive training error parameter. This training error parameter takes into account both the error accumulation of the model during the processing (reflected by the first training error parameter) and the degree of consistency between the final output of the model and the true label (reflected by the second training error parameter). This comprehensive training error parameter will provide an important reference for subsequent model optimization.

[0167] In a possible implementation, step S140 may further include:

[0168] Step S143, when the lesion recognition category of the target medical image acquisition data is an recognition category used to determine the existence of lesions in the target medical image acquisition data, lesion recognition is performed on the target medical image acquisition data based on the fused lesion knowledge vector to generate a conclusion on whether a lesion exists in the target medical image acquisition data.

[0169] In this embodiment, when the server receives target medical image acquisition data, and the lesion recognition category of the target medical image acquisition data is used to determine the existence of a lesion, the server may perform a lesion recognition operation.

[0170] Specifically, the server can first load the fusion lesion knowledge vectors, which have been highly refined and optimized through the previous steps and contain rich lesion feature information. Then, the server can compare and analyze these fusion lesion knowledge vectors with the target medical image acquisition data.

[0171] In this process, the server can use image processing technology and machine learning algorithms to find and match the feature patterns in the fusion lesion knowledge vector to determine whether there are similar lesion features in the target medical image acquisition data. If a feature pattern with a high degree of matching is found, the server will conclude that there is a lesion in the target medical image acquisition data.

[0172] For example, if the fusion lesion knowledge vector contains characteristic information of a certain tumor, the server will search for these characteristics in the target medical imaging data. If a similar characteristic pattern is found, it will be determined that the tumor exists in the data.

[0173] Step S144, when the lesion recognition category of the target medical image acquisition data is an recognition category used to determine the specific type of lesions in the target medical image acquisition data, the target medical image acquisition data is classified into lesions according to the fused lesion knowledge vector to obtain the specific type of lesions in the target medical image acquisition data.

[0174] If the lesion recognition category of the target medical image acquisition data is used to determine the specific type of the lesion, the server will perform a lesion classification operation.

[0175] In this process, the server will also load the fusion lesion knowledge vector and compare and analyze it with the target medical image acquisition data. However, the difference is that this time the server needs to more finely match the characteristic patterns in the fusion lesion knowledge vector to determine the specific type of lesion.

[0176] To accomplish this task, the server can use more complex machine learning algorithms, such as deep learning models, to extract and analyze subtle features in the target medical image acquisition data. By accurately matching these features with the features in the fusion lesion knowledge vector, the server can accurately determine the specific type of lesion.

[0177] For example, if the fusion lesion knowledge vector contains multiple different types of tumor feature information, the server will determine the type of tumor in the target medical imaging data through precise feature matching and analysis, such as lung cancer, liver cancer, etc. In this way, doctors can provide patients with more accurate diagnosis and treatment plans based on the server's classification results.

[0178] Figure 2 The hardware structure of the neural network-based medical image lesion recognition system 100 for implementing the above-mentioned neural network-based medical image lesion recognition method provided in the embodiment of the present application is shown as follows: Figure 2 As shown, the neural network-based medical image lesion recognition system 100 may include a processor 110 , a machine-readable storage medium 120 , a bus 130 , and a communication unit 140 .

[0179] In a possible design, the medical image lesion recognition system 100 based on a neural network can be a single server or a server group. The server group can be centralized or distributed (for example, the medical image lesion recognition system 100 based on a neural network can be a distributed system). In some embodiments, the medical image lesion recognition system 100 based on a neural network can be local or remote. For example, the medical image lesion recognition system 100 based on a neural network can access information and / or data stored in a machine-readable storage medium 120 via a network. For another example, the medical image lesion recognition system 100 based on a neural network can be directly connected to a machine-readable storage medium 120 to access stored information and / or data. In some embodiments, the medical image lesion recognition system 100 based on a neural network can be implemented on a medical image lesion recognition system based on a neural network. As an example only, the medical image lesion recognition system based on a neural network can include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, etc. or any combination thereof.

[0180] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data acquired from an external terminal. In some embodiments, the machine-readable storage medium 120 may store data and / or instructions that the neural network-based medical image lesion recognition system 100 uses to execute or use to complete the exemplary method described in this application.

[0181] During the specific implementation process, at least one processor 110 executes computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the medical image lesion recognition method based on neural network in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0182] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned neural network-based medical image lesion recognition system 100. The implementation principles and technical effects are similar and will not be repeated in this embodiment.

[0183] In addition, an embodiment of the present application also provides a readable storage medium, in which computer executable instructions are preset. When a processor executes the computer executable instructions, the above-mentioned medical image lesion recognition method based on neural network is implemented.

[0184] It should be noted that in order to simplify the description disclosed in this application and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, multiple features are sometimes combined into one embodiment, drawings, or descriptions thereof. Similarly, it should be noted that in order to simplify the description disclosed in this application and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, multiple features are sometimes combined into one embodiment, drawings, or descriptions thereof.

Claims

1. A medical image lesion recognition method based on neural network, characterized in that: The method comprises: Positioning the lesion region on the target medical image acquisition data to generate at least one candidate lesion region; Obtaining a medical image feature relationship network that characterizes the medical feature correlation between a plurality of reference lesion regions, and extracting medical image feature relationship sub-networks corresponding to each of the candidate lesion regions from the medical image feature relationship network, wherein a network member in the medical image feature relationship network has a unique corresponding relationship with the reference lesion region, the medical image feature relationship sub-network includes a target network member, and the reference lesion region corresponding to the target network member has a feature correlation degree with the corresponding candidate lesion region that is greater than a preset feature correlation degree; Performing lesion knowledge vector analysis on each of the medical image feature relationship sub-networks to generate lesion knowledge vectors of each of the medical image feature relationship sub-networks, and blending each of the lesion knowledge vectors to generate a blended lesion knowledge vector of the target medical image acquisition data; According to the fused lesion knowledge vector, the target medical image acquisition data is subjected to lesion recognition and classification, and a lesion recognition and classification result corresponding to the target medical image acquisition data is generated; The lesion recognition and classification is implemented by a lesion feature attention model and a lesion recognition and classification model. The lesion recognition and classification is performed on the target medical image acquisition data based on the fused lesion knowledge vector to generate a lesion recognition and classification result corresponding to the target medical image acquisition data, including: By using the lesion feature attention model, the blended lesion knowledge vector is subjected to self-attention processing to generate a target lesion knowledge vector of the target medical image acquisition data; Using the lesion recognition and classification model, the target medical image acquisition data is subjected to lesion recognition and classification according to the target lesion knowledge vector, and a lesion recognition and classification result corresponding to the target medical image acquisition data is generated; The lesion feature attention model includes a plurality of self-attention processing units, and the lesion feature attention model is used to perform self-attention processing on the fused lesion knowledge vector to generate a target lesion knowledge vector of the target medical image acquisition data, including: Performing a first round of self-attention processing on the blended lesion knowledge vector through a first self-attention processing unit to generate a first target lesion knowledge vector; Traverse x, and perform the xth self-attention processing on the blended lesion knowledge vector according to the x-1th target lesion knowledge vector through the xth self-attention processing unit to generate the xth target lesion knowledge vector; wherein 1<x<X, X represents the number of the self-attention processing units in the lesion feature attention model; Outputting the Xth target lesion knowledge vector as the target lesion knowledge vector of the target medical image acquisition data; Before performing self-attention processing on the blended lesion knowledge vector through the lesion feature attention model to generate the target lesion knowledge vector of the target medical image acquisition data, the method further includes: Obtaining a sample lesion feature attention model including a plurality of sample self-attention processing units, sample medical image acquisition data, and a lesion label knowledge vector of lesion label data; Positioning the lesion region on the sample medical image acquisition data to generate at least one example candidate lesion region, and extracting example medical image feature relationship networks corresponding to each of the example candidate lesion regions from the medical image feature relationship network; Performing lesion knowledge vector analysis on each of the example medical image feature relationship networks to generate example lesion knowledge vectors for each of the example medical image feature relationship networks, and blending each of the example lesion knowledge vectors to generate an example blended lesion knowledge vector for the sample medical image acquisition data; Sequentially performing self-attention processing on the example blended lesion knowledge vector through each of the example self-attention processing units of the example lesion feature attention model, and generating example target lesion knowledge vectors corresponding to each of the example self-attention processing units; Determining a training error parameter of the sample lesion feature attention model according to each of the example target lesion knowledge vectors and the lesion label knowledge vector, and optimizing the sample lesion feature attention model according to the training error parameter to generate the lesion feature attention model; The sample lesion feature attention model includes X sample self-attention processing units, the lesion label knowledge vector and the vector dimension of each of the example target lesion knowledge vectors are consistent, and the training error parameter of the sample lesion feature attention model is determined based on each of the example target lesion knowledge vectors and the lesion label knowledge vector, including: Traversing y, determining the difference knowledge vector between the example target lesion knowledge vector corresponding to the y+1th sample self-attention processing unit and the example target lesion knowledge vector corresponding to the yth sample self-attention processing unit, generating the difference knowledge vector corresponding to the yth sample self-attention processing unit, and outputting the mapping value of the feature matrix size of the difference knowledge vector as the unit training error parameter corresponding to the yth sample self-attention processing unit, where 1≤y≤X; Adding the unit training error parameters to generate a first training error parameter; Determine the difference knowledge vector of the example target lesion knowledge vector corresponding to the X-th sample self-attention processing unit and the lesion label knowledge vector to generate a difference knowledge vector, and output the mapping value of the feature matrix size of the difference knowledge vector as the second training error parameter of the sample lesion feature attention model; The first training error parameter and the second training error parameter are weightedly fused to generate the training error parameter.

2. The method for identifying medical image lesions based on a neural network according to claim 1, characterized in that: The step of locating the lesion region on the target medical image acquisition data to generate at least one candidate lesion region includes: Acquire medical image data to be detected, and perform medical feature recognition on the medical image data to be detected to generate a medical feature set of the medical image data to be detected; When the medical feature set indicates that the medical image data to be detected meets a preset lesion feature condition, outputting the medical image data to be detected as the target medical image acquisition data; Performing image segmentation processing on the target medical image acquisition data to generate a plurality of candidate image regions in the target medical image acquisition data; For each of the candidate image regions, when the candidate image region meets a preset lesion feature condition, the candidate image region is output as the candidate lesion region.

3. The method for identifying medical image lesions based on a neural network according to claim 2, characterized in that: After performing medical feature recognition on the medical image data to be detected and generating a medical feature set of the medical image data to be detected, the method further includes: When the medical feature set indicates that the medical image data to be detected does not meet the preset lesion feature condition, for each candidate image region in the medical image data to be detected, respectively extract the candidate medical image feature relationship network corresponding to each candidate image region from the medical image feature relationship network, perform lesion knowledge vector analysis on each candidate medical image feature relationship network, generate candidate lesion knowledge vectors of each candidate medical image feature relationship network, and blend each candidate lesion knowledge vector to generate a candidate blended lesion knowledge vector of the medical image data to be detected; According to the candidate fusion lesion knowledge vector, the medical image data to be detected is subjected to lesion recognition and classification, and a target lesion recognition and classification result corresponding to the medical image data to be detected is generated; Among them, the target lesion identification and classification result meets the preset lesion characteristic conditions, and the target lesion identification and classification result includes the medical image data to be detected.

4. The method for identifying medical image lesions based on a neural network according to claim 1, characterized in that: The extracting, from the medical image feature relationship network, the medical image feature relationship sub-networks corresponding to the candidate lesion regions respectively includes: For each of the candidate lesion regions, the candidate lesion region is respectively compared with each of the reference lesion regions in the medical image feature relationship network to generate comparison details, wherein the comparison details indicate whether there is a reference lesion region identical to the candidate lesion region in the medical image feature relationship network; Determining, according to the comparison details, a target reference lesion region corresponding to the candidate lesion region from each of the reference lesion regions in the medical image feature relationship network; Outputting the network member corresponding to the target reference lesion region in the medical image feature relationship network as the target network member corresponding to the candidate lesion region, and outputting the network member having a network dependency relationship with the target network member in the medical image feature relationship network as the candidate network member corresponding to the candidate lesion region; Outputting the sub-network structure composed of the target network members and the candidate network members in the medical image feature relationship network as a medical image feature relationship sub-network corresponding to the candidate lesion area; Wherein, determining, based on the comparison details, from the reference lesion regions in the medical image feature relationship network, a target reference lesion region corresponding to the candidate lesion region comprises: When the comparison details represent that there is a reference lesion region identical to the candidate lesion region in the medical image feature relationship network, outputting the reference lesion region identical to the candidate lesion region as a target reference lesion region corresponding to the candidate lesion region; When the comparison details indicate that there is no reference lesion region identical to the candidate lesion region in the medical image feature relationship network, obtaining the feature correlation degree between the candidate lesion region and each of the reference lesion regions; At least one reference lesion region whose feature correlation degree is greater than the preset feature correlation degree is output as a target reference lesion region corresponding to the candidate lesion region.

5. The method for identifying medical image lesions based on a neural network according to claim 1, characterized in that: The step of performing lesion identification and classification on the target medical image acquisition data based on the fused lesion knowledge vector to generate a lesion identification and classification result corresponding to the target medical image acquisition data includes: When the lesion recognition category of the target medical image acquisition data is an recognition category for determining the existence of a lesion in the target medical image acquisition data, performing lesion recognition on the target medical image acquisition data according to the fused lesion knowledge vector, and generating a conclusion on whether a lesion exists in the target medical image acquisition data; When the lesion recognition category of the target medical image acquisition data is an recognition category used to determine the specific type of lesions in the target medical image acquisition data, the target medical image acquisition data is lesion classified according to the fused lesion knowledge vector to obtain the specific type of lesions in the target medical image acquisition data.

6. A medical image lesion recognition system based on neural network, characterized in that: The neural network-based medical image lesion recognition system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the neural network-based medical image lesion recognition method described in any one of claims 1 to 5.

Images