A multi-category lesion segmentation method and system for diabetic retinopathy fundus image data

Abstract

The present invention provides a multi-category lesion segmentation method and system for diabetic retinopathy fundus image data, relating to the field of artificial intelligence. The method is applied to a multi-category lesion segmentation system, and the method includes: downloading an image pre-segmentation model; collecting historical images of diabetic retinopathy to generate a historical image set; performing mobile terminal network integration to generate a refined image segmentation model; using the image pre-segmentation model to determine generalization parameters to obtain a generalized image segmentation model; inputting a test sample set into the generalized image segmentation model to obtain lesion semantic segmentation data; and using a lesion recognition and correction module to obtain multi-category lesion segmentation types. The method solves the technical problem of difficulty in performing fine segmentation of diabetic retinopathy images, which leads to difficulty in performing multi-category accurate identification of lesions, and achieves the technical effect of intelligently optimizing the category lesion segmentation scheme of diabetic retinopathy fundus image data, quickly locating lesion category information, and improving the accuracy of lesion image segmentation and multi-category lesion identification.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a multi-category lesion segmentation method and system for diabetic retinopathy fundus image data. Background Art

[0002] "Diabetic retinopathy" stands for diabetic retinopathy, which refers to retinal microvascular damage caused by diabetes and is one of the most common microvascular complications of diabetes. The fundus image-related data information obtained from diabetic patients during fundus screening is the diabetic retinopathy fundus image data. Early standardized treatment can significantly improve the condition. There are many types of diabetic retinopathy fundus image lesions, which may be pathological changes in the microvessels, or leakage or obstruction lesions in the microvascular system. Directly dividing the diabetic retinopathy fundus image lesions by observation cannot guarantee the reliability of the division results. There is an urgent need for a reasonable lesion category segmentation scheme to accurately segment the lesion type information.

[0003] There is a technical problem in the existing technology that it is difficult to perform fine segmentation of diabetic retinopathy images, which makes it difficult to accurately identify lesions in multiple categories. Summary of the Invention

[0004] This application provides a multi-category lesion segmentation method and system for diabetic retinopathy fundus image data, which solves the technical problem that it is difficult to perform fine segmentation of diabetic retinopathy images, resulting in difficulty in accurately identifying multi-category lesions. It achieves the technical effect of intelligently optimizing the category lesion segmentation scheme of diabetic retinopathy fundus image data, quickly locating lesion category information, and improving the accuracy of lesion image segmentation and its multi-category lesion identification.

[0005] In view of the above problems, the present application provides a multi-category lesion segmentation method and system for diabetic retinopathy fundus image data.

[0006] In the first aspect, the present application provides a multi-category lesion segmentation method for diabetic retinopathy fundus image data, wherein the method is applied to a multi-category lesion segmentation system, the system includes a network integration module and a lesion identification and correction module, and the method includes: downloading an image pre-segmentation model from the multi-category lesion segmentation system, wherein the image pre-segmentation model is embedded with a deep learning network structure; by collecting historical images of diabetic retinopathy, a historical image set can be generated, wherein the historical image set can be custom-divided into a training sample set and a test sample set; using the network integration module, the image pre-segmentation model and the training sample set are subjected to mobile terminal network integration, Used to generate a refined image segmentation model, and the refined image segmentation model has a lightweight structure; using the image pre-segmentation model, loss analysis is performed on the trained refined image segmentation model, and the generalization parameters are determined through the analysis results, so as to generalize the refined image segmentation model and obtain a generalized image segmentation model; the test sample set is input into the generalized image segmentation model for training to obtain lesion semantic segmentation data of the test sample set; using the lesion recognition and correction module, lesion recognition and correction are performed on the lesion semantic segmentation data to obtain multi-category lesion segmentation types of the test sample set.

[0007] In the second aspect, the present application provides a multi-category lesion segmentation system for diabetic retinopathy fundus image data, wherein the system comprises: a data acquisition unit, the data acquisition unit is used to download an image pre-segmentation model from the multi-category lesion segmentation system, wherein the image pre-segmentation model is embedded with a deep learning network structure; a data acquisition unit, the data acquisition unit is used to generate a historical image set by collecting historical images of diabetic retinopathy, wherein the historical image set can be customized to be divided into a training sample set and a test sample set; a data analysis unit, the data analysis unit is used to use a network integration module to perform mobile terminal network integration on the image pre-segmentation model and the training sample set to generate a refined image segmentation model, and the refined image segmentation model is used to generate a refined image segmentation model. The image segmentation model has a lightweight structure; an analysis and processing unit, which is used to use the image pre-segmentation model to perform loss analysis on the trained refined image segmentation model, and determine the generalization parameter through the analysis result, so as to generalize the refined image segmentation model to obtain a generalized image segmentation model; a training integration unit, which is used to input the test sample set into the generalized image segmentation model for training, so as to obtain the lesion semantic segmentation data of the test sample set; an identification and correction unit, which is used to use the lesion identification and correction module to perform lesion identification correction on the lesion semantic segmentation data, so as to obtain multi-category lesion segmentation types of the test sample set.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0009] Due to the adoption of downloading the image pre-segmentation model from the multi-category lesion segmentation system; by collecting historical images of diabetic retinopathy, a historical image set can be generated; using the network integration module, the image pre-segmentation model and the training sample set are integrated on the mobile end to generate a refined image segmentation model; using the image pre-segmentation model, the trained refined image segmentation model is subjected to loss analysis, the generalization parameter is determined by the analysis result, the refined image segmentation model is generalized, and a generalized image segmentation model is obtained; the test sample set is input into the generalized image segmentation model for training to obtain the lesion semantic segmentation data of the test sample set; using the lesion recognition and correction module, the lesion semantic segmentation data is subjected to lesion recognition correction to obtain the multi-category lesion segmentation type of the test sample set. The embodiment of the present application achieves the technical effect of intelligently optimizing the category lesion segmentation scheme of diabetic retinopathy fundus image data, quickly locating lesion category information, and improving the accuracy of lesion image segmentation and its multi-category lesion recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 This is a flowchart of a multi-category lesion segmentation method for diabetic retinopathy fundus image data in this application;

[0011] Figure 2 This is a flow chart of custom segmentation of sample data for a multi-category lesion segmentation method for diabetic retinopathy fundus image data in this application;

[0012] Figure 3 This is a flow chart of determining multi-category lesion segmentation types in a multi-category lesion segmentation method for diabetic retinopathy fundus image data according to the present application;

[0013] Figure 4 This is a structural diagram of a multi-category lesion segmentation system for diabetic retinopathy fundus image data in this application.

[0014] Description of reference numerals: data acquisition unit 11 , data collection unit 12 , data analysis unit 13 , analysis and processing unit 14 , training integration unit 15 , recognition and correction unit 16 . DETAILED DESCRIPTION

[0015] This application provides a multi-category lesion segmentation method and system for diabetic retinopathy fundus image data, which solves the technical problem that it is difficult to perform fine segmentation of diabetic retinopathy images, resulting in difficulty in accurately identifying multi-category lesions. It achieves the technical effect of intelligently optimizing the category lesion segmentation scheme of diabetic retinopathy fundus image data, quickly locating lesion category information, and improving the accuracy of lesion image segmentation and its multi-category lesion identification.

[0016] Example 1

[0017] like Figure 1 As shown, the present application provides a multi-category lesion segmentation method for diabetic retinopathy fundus image data, wherein the method is applied to a multi-category lesion segmentation system, the system including a network integration module and a lesion identification and correction module, and the method includes:

[0018] S100: Downloading an image pre-segmentation model from the multi-category lesion segmentation system, wherein the image pre-segmentation model is embedded with a deep learning network structure;

[0019] Specifically, the image pre-segmentation model is downloaded through the data storage unit of the multi-category lesion segmentation system. The image pre-segmentation model is embedded with a deep learning network structure. After the multi-category lesion segmentation system performs data preprocessing, the preprocessed data information is input into the image pre-segmentation model for data analysis and processing. The deep learning network structure can be combined with the BP back-propagation algorithm for structural construction. The embedding does not represent a specific technical operation, but represents the embedding of algorithm logic. Acquiring data information through the multi-category lesion segmentation system can effectively ensure the reliability of the data information.

[0020] S200: By collecting historical images of diabetic retinopathy, a historical image set can be generated, wherein the historical image set can be customized to be divided into a training sample set and a test sample set;

[0021] Furthermore, if Figure 2 As shown, step S200 in this embodiment of the application further includes:

[0022] S210: Using the historical image set as sample data, and performing a multi-feature traversal search on the sample data to determine coverage features, repetition features, and availability features of the sample data;

[0023] Specifically, the sample data is the underlying logical data information involved in data processing. The historical image set is used as sample data, and a multi-feature traversal search is performed on the historical image set to obtain the coverage feature, repetition feature and availability feature of the sample data. The coverage feature determines whether the data type of the historical image set is comprehensive and has sufficient coverage; the repetition feature determines the repetition between the historical image sets. Data repetition should be avoided to ensure the integrity of the historical image set data. Generally, the higher the clarity of the picture, the larger the amount of corresponding availability feature data; the lower the clarity of the picture, the smaller the amount of corresponding availability feature data. The availability feature is judged by the clarity of the picture, and the coverage feature and repetition feature are combined to express the features of the historical image set as sample data. Combined with multiple data feature parameters, the integrity of the historical image set data is further guaranteed, providing technical support for ensuring the stability of training data and improving the data processing efficiency of the training process.

[0024] S220: Using a weight allocation channel, weights are allocated to the coverage feature, the repetition feature, and the availability feature to determine a coverage weight ratio corresponding to the coverage feature, a repetition weight ratio corresponding to the repetition feature, and an availability weight ratio corresponding to the availability feature;

[0025] Specifically, the weight allocation channel is a functional channel. Specifically, the sum of the weight values allocated by the weight allocation channel is 1. The weight allocation channel has a unified weight allocation condition during the weight allocation process. Weights are allocated to the coverage feature, repetition feature, and availability feature. Generally, the coverage weight allocation condition corresponding to the coverage feature requires: the greater the coverage, the greater the weight ratio; the smaller the coverage, the smaller the weight ratio; the repetition weight allocation condition corresponding to the repetition feature requires: the smaller the repetition, the greater the weight ratio; the higher the repetition, the smaller the weight ratio; the availability feature corresponds to the availability weight allocation condition: the higher the availability, the greater the weight ratio; the lower the availability, the smaller the weight ratio. Generally, in the weight allocation process, the training image clarity requirements are high, the coverage requirements are wide, and the repetition requirements are low. Correspondingly, the availability weight ratio is large, the coverage weight ratio is second only to the availability weight ratio, and the repetition weight ratio is small. The weight allocation process needs to be combined with relevant indicator requirements for comparative analysis, and specifically determined in combination with actual data information, which refines the logic of feature weight allocation and provides technical support for the accurate identification of diabetic retinopathy fundus image data.

[0026] S230: Customize the division of the sample data based on the coverage weight ratio, the duplication weight ratio, and the availability weight ratio.

[0027] Furthermore, step S230 in the embodiment of the present application further includes:

[0028] S231: performing a weighted sum operation on the sample data using the coverage weight ratio, the duplication weight ratio, and the availability weight ratio to obtain a weighted sum result;

[0029] S232: By using the customized partitioning logic of the sample data, the weighted sum result is matched accordingly to determine the target sample data partitioning logic corresponding to the weighted sum result.

[0030] Specifically, the sample data is weighted and summed using the coverage weight ratio, the duplication weight ratio, and the availability weight ratio. The weighted sum can be specifically set in combination with the data feature information to obtain a weighted sum result. Through the custom partitioning logic of the sample data, the custom partitioning logic can be set to: when the weighted sum result meets (50-70), the training sample data volume: the test sample data volume is 70%:30%; when the weighted sum result meets (70-80), the training sample data volume: the test sample data volume is 65%:35%; when the weighted sum result meets (80-100), the training sample data volume: the test sample data volume is 60%:40%. The weighted sum results are matched accordingly to determine the target sample data partitioning logic corresponding to the weighted sum result. The partitioning method of the target sample data partitioning logic is not unique. It is actually refined in combination with the actual data information characteristics to determine the target sample data partitioning logic, providing technical support for ensuring the stability of the data partitioning logic and providing a basis for in-depth mining of data feature information.

[0031] S300: Using the network integration module, performing mobile-end network integration on the image pre-segmentation model and the training sample set to generate a refined image segmentation model, wherein the refined image segmentation model has a lightweight structure;

[0032] Specifically, the network integration module is a data logic function processing module, which performs mobile network integration on the image pre-segmentation model and the training sample set. In simple terms, the network integration is to perform data integration operations through relevant algorithms. The relevant algorithms include but are not limited to k-means algorithm, K-medoids algorithm and other data integration algorithms to generate a refined image segmentation model, and the refined image segmentation model has a lightweight structure. In simple terms, the lightweight structure optimizes the construction process of the refined image segmentation model from the perspective of maintaining accuracy and reducing parameters. For example, NAS is used to search the global network structure by optimizing each network block of the model, and the NetAdapt algorithm is used to search the number of filters in each layer. The construction scheme of the algorithm model with the lightweight structure is not unique, and the example does not impose any restrictions on actual data operation and analysis. The refined image segmentation model based on the lightweight structure can reduce the computational amount of the data analysis and processing process and improve the data processing efficiency of the refined image segmentation model.

[0033] S400: using the image pre-segmentation model, performing loss analysis on the trained refined image segmentation model, and determining a generalization parameter based on the analysis result, so as to generalize the refined image segmentation model and obtain a generalized image segmentation model;

[0034] Furthermore, step S400 in this embodiment of the application further includes:

[0035] S410: Using the image pre-segmentation model as a teacher end and the refined image segmentation model as a student end;

[0036] S420: Perform loss function analysis on the Teacher end to determine the corresponding Teacher end loss parameter L t ;

[0037] S430: Perform loss function analysis on the Student end to determine the corresponding Student end loss parameter L s ;

[0038] S440: Calculate the Teacher-side loss parameter L t and the Student-side loss parameter L s A weighted calculation is performed to determine the loss function L of the refined image segmentation model.

[0039] Specifically, by adopting the Teacher-Student mode, the image pre-segmentation model is used as the Teacher side, and the refined image segmentation model is used as the Student side. The Teacher image pre-segmentation model assists the training of the Student refined image segmentation model and enhances the generalization ability of the refined image segmentation model. The Teacher side loss parameter L t and the Student-side loss parameter L s Perform weighted calculation, and the loss function L of the refined image segmentation model is calculated by the Teacher end loss parameter L t and the Student-side loss parameter L s The weighted operation determines that the information of the image pre-segmentation model is complex and the amount of data is large. Based on the image pre-segmentation model, the training of the refined image segmentation model is assisted to enhance the generalization ability of the refined image segmentation model.

[0040] Furthermore, step S400 in this embodiment of the application further includes:

[0041] S441: Based on the formula: L = αL t +βL s , determine the loss function L of the refined image segmentation model, where α and β are loss weighting coefficients, α = 0.9, β = 0.1;

[0042] S442: Based on the formula: The loss function analysis is performed on the Teacher side, where L t Teacher loss generated by the image pre-segmentation model file, p i ′ is the probability of the i-th category generated after the softmax of the image pre-segmentation model file, v i is the unnormalized probability of the i-th class in the image pre-segmentation model, q i ′ is the probability of the i-th class generated by the softmax of the refined image segmentation model file after learning by the teacher network, z i is the unnormalized probability of the i-th class in the refined image segmentation model;

[0043] S443: Based on the formula: Perform loss function analysis on the Student side, where L s is the student loss generated by the refined image segmentation model file, N is the total number of categories, c i is the true value of the i-th category of the refined image segmentation model, the positive sample is 1, the negative sample is 0, q iis the probability of the i-th category generated by the softmax of the refined image segmentation model file after true value learning,

[0044] Specifically, through the Teacher side loss parameter L t and the Student-side loss parameter L s Weighted operation to determine the loss function L of the refined image segmentation model, the loss function L of the refined image segmentation model = αL t +βL s , where α and β are loss weighting coefficients, α = 0.9, β = 0.1, α and β are determined by multiple experimental calculations, and the numerical values of α and β are not unique.

[0045] Further specific description, the loss function analysis is performed on the Teacher side, and the loss parameter of the Teacher side is The functional relationship is not detailed in specific values. The actual use process may be deformed in the identity. It is determined in combination with the actual data analysis process. Among them, L t is the teacherloss generated by the image pre-segmentation model file, N is the total number of categories (N is a positive integer), p i ′ is the probability of the i-th category generated after the softmax of the image pre-segmentation model file, v i is the unnormalized probability of the i-th class in the image pre-segmentation model, q i ′ is the probability of the i-th class generated by the softmax of the refined image segmentation model file after learning by the teacher network, z i In order to refine the unnormalized probability of the i-th category of the image segmentation model, it is necessary to understand that the data processing scheme is obtained through optimization. The actual data processing process needs to optimize the data processing scheme in combination with data characteristics, which will not be elaborated here.

[0046] Further specific description, the loss function analysis is performed on the Student side, and the loss parameter of the Student side is The functional relationship is not detailed in specific values, and may be deformed in actual use. s is the student loss generated by the refined image segmentation model file, N is the total number of categories (N is a positive integer), c i is the true value of the i-th category of the refined image segmentation model, the positive sample is 1, the negative sample is 0, q i is the probability of the i-th category generated by the softmax of the refined image segmentation model file after true value learning, z iTo refine the probability of unnormalization for the i-th category of the image segmentation model, it is important to understand that the data processing scheme described above is the result of a prioritization process. The actual data processing process requires optimization based on data characteristics. The training process simply involves saving the trained and fitted model file based on the output semantic segmentation error. Training and fitting the mathematical logic of the data based on a functional relationship model can accelerate model training efficiency and provide technical support for the rapid analysis and processing of diabetic retinopathy image lesion information.

[0047] S500: Inputting the test sample set into the generalized image segmentation model for training to obtain lesion semantic segmentation data of the test sample set;

[0048] S600: Utilizing the lesion recognition and correction module, performing lesion recognition and correction on the lesion semantic segmentation data to obtain multi-category lesion segmentation types of the test sample set.

[0049] Specifically, the historical image collection can be customized to be divided into a training sample set and a test sample set. The training sample set is used to construct a training model, and the test sample set is used to test the stability of the constructed model. The test sample set is input into the generalized image segmentation model for model stability testing. The test requires grouping the test sample data first, and the data type of each group of data is consistent. The data is the feature data of the diabetic retinopathy fundus image, which can specifically be color feature information, color block distribution feature information and other related feature data information. Multiple groups of test data are input in sequence for feature processing operations to obtain the lesion semantic segmentation data of the test sample set; the lesion recognition and correction module is used to perform lesion recognition and correction on the lesion semantic segmentation data to obtain multi-category lesion segmentation types of the test sample set, which further ensures the reliability of the category lesion segmentation scheme from a technical perspective and provides technical support for improving the accuracy of lesion segmentation results.

[0050] Furthermore, if Figure 3 As shown, step S600 in this embodiment of the application further includes:

[0051] S610: The lesion recognition and correction module is embedded with a Densenet classification image network;

[0052] S620: Inputting the lesion semantic segmentation data into the Densenet classification image network to perform classification correction to determine the multi-category lesion segmentation type.

[0053] Specifically, the basic structure of the Densenet classification image network is: i =H i ([X0, X1, ..., Xi-1 ]), where X i The output of the i-th layer, H i Indicates the nonlinear transformation of the i-th layer. The symbol "[]" indicates concatenation. In simple terms, it is from X0 to X i-1 All output feature vectors of the layer are combined by channel. In simple terms, the training process is to save the trained fitting model file based on the output classification error, and input the lesion semantic segmentation data into the Densenet classification image network. Specifically, the target diabetic retinopathy fundus image sample is input into the Densenet classification image network and subjected to classification correction. The lesion image is subjected to diabetic retinopathy semantic segmentation, and lesion identification correction is performed to obtain more accurate lesion results and determine multi-category lesion segmentation types. The multi-category lesion segmentation type identification scheme is obtained by performing scheme optimization, which provides technical support for accurately and quickly positioning lesion type information and improving the efficiency of lesion segmentation and identification.

[0054] To further specify, the lesion image is subjected to diabetic retinopathy semantic segmentation, and lesion identification correction is performed to obtain more accurate lesion results. An example is given that since the appearance of small areas of hemorrhage on diabetic retinopathy fundus images is similar to that of microaneurysms, it is difficult to quickly distinguish them. To ensure accuracy, a Densenet classification image network is cascaded after the image pre-segmentation model to distinguish the lesion type of the segmentation network results. After classification correction, the fundus image is subjected to diabetic retinopathy semantic segmentation to obtain fine lesion segmentation results, thereby improving the efficiency of diabetic retinopathy fundus image data analysis and improving the accuracy of multi-category lesion segmentation results.

[0055] Furthermore, the embodiment of the present application also includes:

[0056] S630: Using the image processing monitoring device, monitor the time consumption of any image being processed in the test sample set to obtain a distribution of the time consumption of each image processing;

[0057] S640: traversing the time consumption distribution of each image processing to obtain abnormal processing images, marking them, and performing secondary processing.

[0058] Specifically, the image processing monitoring device is an operation monitoring device of the multi-category lesion segmentation system. The multi-category lesion segmentation system is cascaded with the image processing monitoring device to monitor the time consumption of any image in the test sample set during processing, and obtain the time consumption distribution of each image processing. The time consumption unit needs to be unified. Generally, the time consumption unit can be determined as seconds. The time consumption distribution of each image processing corresponds to the lesion category distribution. The abnormal time consumption of each image processing time consumption distribution is traversed, and the abnormal processing image corresponding to the lesion position with abnormal time consumption is determined. The abnormal processing image is marked, and the lesion with abnormal time consumption is processed secondary to improve the timeliness of the diabetic retinopathy fundus image data analysis.

[0059] In summary, the multi-category lesion segmentation method and system for diabetic retinopathy fundus image data provided by this application has the following technical effects:

[0060] 1. Due to the adoption of downloading the image pre-segmentation model from the multi-category lesion segmentation system; by collecting historical images of diabetic retinopathy, a historical image set can be generated; using the network integration module, the image pre-segmentation model and the training sample set are integrated on the mobile terminal network to generate a refined image segmentation model; using the image pre-segmentation model, the trained refined image segmentation model is subjected to loss analysis, the generalization parameter is determined by the analysis result, the refined image segmentation model is generalized, and a generalized image segmentation model is obtained; the test sample set is input into the generalized image segmentation model for training to obtain the lesion semantic segmentation data of the test sample set; using the lesion recognition and correction module, the lesion semantic segmentation data is subjected to lesion recognition correction to obtain the multi-category lesion segmentation type of the test sample set. This application provides a multi-category lesion segmentation method and system for diabetic retinopathy fundus image data, thereby achieving the technical effect of intelligently optimizing the category lesion segmentation scheme of diabetic retinopathy fundus image data, quickly locating lesion category information, and improving the accuracy of lesion image segmentation and its multi-category lesion recognition.

[0061] 2. Since the image pre-segmentation model is used as the Teacher side and the refined image segmentation model is used as the Student side, the loss function of the Teacher side is analyzed to determine the Teacher side loss parameter L t ; Analyze the loss function on the Student side and determine the loss parameter L on the Student side s ; Calculate the Teacher side loss parameter L t and the Student-side loss parameter L s Perform weighted calculations to determine the loss function L of the refined image segmentation model. The image pre-segmentation model has complex information and large amounts of data. Based on the image pre-segmentation model, the training of the refined image segmentation model is assisted to enhance the generalization ability of the refined image segmentation model.

[0062] 3. Due to the adoption of the formula: L = αL t +βL s , determine the loss function L of the refined image segmentation model; based on the formula: Perform loss function analysis on the Teacher side; based on the formula: Perform loss function analysis on the Student side. Training and fitting the mathematical logic of the data based on the functional relationship model can accelerate model training efficiency and provide technical support for the rapid analysis and processing of fundus diabetic retinopathy image lesions.

[0063] Example 2

[0064] Based on the same inventive concept as the multi-category lesion segmentation method of diabetic retinopathy fundus image data in the aforementioned embodiment, Figure 4 As shown, the present application provides a multi-category lesion segmentation system for diabetic retinopathy fundus image data, wherein the system includes:

[0065] A data acquisition unit 11, wherein the data acquisition unit 11 is used to download an image pre-segmentation model from a multi-category lesion segmentation system, wherein the image pre-segmentation model is embedded with a deep learning network structure;

[0066] A data acquisition unit 12 is configured to collect historical images of diabetic retinopathy to generate a historical image set, wherein the historical image set can be customized to be divided into a training sample set and a test sample set;

[0067] A data analysis unit 13, configured to utilize a network integration module to perform mobile-end network integration on the image pre-segmentation model and the training sample set to generate a refined image segmentation model having a lightweight structure;

[0068] An analysis and processing unit 14 is configured to perform a loss analysis on the trained refined image segmentation model using the image pre-segmentation model, and determine a generalization parameter based on the analysis result, so as to generalize the refined image segmentation model to obtain a generalized image segmentation model.

[0069] A training integration unit 15, wherein the training integration unit 15 is used to input the test sample set into the generalized image segmentation model for training to obtain lesion semantic segmentation data of the test sample set;

[0070] The recognition and correction unit 16 is configured to use a lesion recognition and correction module to perform lesion recognition and correction on the lesion semantic segmentation data to obtain multi-category lesion segmentation types of the test sample set.

[0071] Furthermore, the system includes:

[0072] a feature retrieval unit, configured to use the historical image set as sample data and perform a multi-feature traversal retrieval on the sample data to determine coverage features, repetition features, and availability features of the sample data;

[0073] A weight allocation unit, configured to allocate weights to the coverage feature, the repetition feature, and the availability feature using a weight allocation channel, and to determine a coverage weight ratio corresponding to the coverage feature, a repetition weight ratio corresponding to the repetition feature, and an availability weight ratio corresponding to the availability feature;

[0074] A definition division unit is used to perform custom division on the sample data according to the coverage weight ratio, the repetition weight ratio and the availability weight ratio.

[0075] Furthermore, the system includes:

[0076] a data processing unit, configured to perform a weighted sum operation on the sample data using the coverage weight ratio, the duplication weight ratio, and the availability weight ratio to obtain a weighted sum result;

[0077] A matching determination unit is used to perform a corresponding match on the weighted sum result through a customized partitioning logic of the sample data, and can determine the target sample data partitioning logic corresponding to the weighted sum result.

[0078] Furthermore, the system includes:

[0079] A data analysis unit, configured to use the image pre-segmentation model as a teacher end and the refined image segmentation model as a student end;

[0080] A data analysis unit is used to perform a loss function analysis on the Teacher end to determine the corresponding Teacher end loss parameter L t ;

[0081] The analysis and determination unit is used to perform loss function analysis on the Student end to determine the corresponding Student end loss parameter Ls ;

[0082] Parameter calculation unit, the parameter calculation unit is used to calculate the teacher end loss parameter L t and the Student-side loss parameter L s A weighted calculation is performed to determine the loss function L of the refined image segmentation model.

[0083] Furthermore, the system includes:

[0084] An analysis and determination unit is used to determine the L value based on the formula: t +βL s , determine the loss function L of the refined image segmentation model, where α and β are loss weighting coefficients, α = 0.9, β = 0.1;

[0085] An analysis operation unit is used to analyze the data based on the formula: The loss function analysis is performed on the Teacher side, where L t Teacher loss generated by the image pre-segmentation model file, p i ′ is the probability of the i-th category generated after the softmax of the image pre-segmentation model file, v i is the unnormalized probability of the i-th class in the image pre-segmentation model, q i ′ is the probability of the i-th class generated by the softmax of the refined image segmentation model file after learning by the teacher network, z i is the unnormalized probability of the i-th class in the refined image segmentation model;

[0086] An operation processing unit, wherein the operation processing unit is configured to: Perform loss function analysis on the Student side, where L s is the student loss generated by the refined image segmentation model file, N is the total number of categories, c i is the true value of the i-th category of the refined image segmentation model, the positive sample is 1, the negative sample is 0, q i is the probability of the i-th category generated by the softmax of the refined image segmentation model file after true value learning,

[0087] Furthermore, the system includes:

[0088] An identification and correction unit, wherein the identification and correction unit is used for embedding a Densenet classification image network in the lesion identification and correction module;

[0089] A classification correction unit is used to perform classification correction by inputting the lesion semantic segmentation data into the Densenet classification image network to determine the multi-category lesion segmentation type.

[0090] This specification and drawings are merely illustrative of the present application and may be modified and combined in various ways without departing from the spirit and scope of the present application. To the extent such modifications and variations fall within the scope of the present claims and their equivalents, the present application is intended to include such modifications and variations.

Claims

1. A multi-category lesion segmentation method for diabetic retinopathy fundus image data, characterized in that: The method is applied to a multi-category lesion segmentation system, the system including a network integration module and a lesion identification and correction module, and the method includes: Downloading an image pre-segmentation model from the multi-category lesion segmentation system, wherein the image pre-segmentation model is embedded with a deep learning network structure; By collecting historical images of diabetic retinopathy, a historical image set can be generated, wherein the historical image set can be customized to be divided into a training sample set and a test sample set; Using the network integration module, performing mobile-end network integration on the image pre-segmentation model and the training sample set to generate a refined image segmentation model, wherein the refined image segmentation model has a lightweight structure; Using the image pre-segmentation model, a loss analysis is performed on the trained refined image segmentation model, and a generalization parameter is determined based on the analysis result, so as to generalize the refined image segmentation model and obtain a generalized image segmentation model; The loss analysis of the trained refined image segmentation model includes: using the image pre-segmentation model as the Teacher end and the refined image segmentation model as the Student end; performing loss function analysis on the Teacher end to determine the corresponding Teacher end loss parameter L t ; Perform loss function analysis on the Student side to determine the corresponding Student side loss parameter L s ; Calculate the Teacher end loss parameter L t and the Student-side loss parameter L s Performing weighted calculation to determine the loss function L of the refined image segmentation model; Among them, based on the formula: L=αL t +βL s , determine the loss function L of the refined image segmentation model, where α and β are loss weighting coefficients, α = 0.9, β = 0.1; Based on the formula: The loss function analysis is performed on the Teacher side, where L t Teacher loss, p′, generated by the image pre-segmentation model file i is the probability of the i-th category generated after the softmax of the image pre-segmentation model file, v i is the unnormalized probability of the i-th class in the image pre-segmentation model, q′ i is the probability of the i-th class generated by the softmax of the refined image segmentation model file after learning by the teacher network, z i is the unnormalized probability of the i-th class in the refined image segmentation model; Based on the formula: Perform loss function analysis on the Student side, where L s is the student loss generated by the refined image segmentation model file, N is the total number of categories, c i is the true value of the i-th category of the refined image segmentation model, the positive sample is 1, the negative sample is 0, q i is the probability of the i-th category generated by the softmax of the refined image segmentation model file after true value learning, Inputting the test sample set into the generalized image segmentation model for training to obtain lesion semantic segmentation data of the test sample set; The lesion recognition and correction module is used to perform lesion recognition and correction on the lesion semantic segmentation data to obtain multi-category lesion segmentation types of the test sample set.

2. The method according to claim 1, wherein The method comprises: Taking the historical image set as sample data, and performing multi-feature traversal retrieval on the sample data, the coverage feature, repetition feature, and availability feature of the sample data can be determined; The weight distribution channel is used to distribute weights to the coverage feature, the repetition feature, and the availability feature, so as to determine the coverage weight ratio corresponding to the coverage feature, the repetition weight ratio corresponding to the repetition feature, and the availability weight ratio corresponding to the availability feature; The sample data is customized to be divided according to the coverage weight ratio, the repetition weight ratio and the availability weight ratio.

3. The method according to claim 2, wherein The method comprises: Using the coverage weight ratio, the duplication weight ratio, and the availability weight ratio, a weighted sum operation is performed on the sample data to obtain a weighted sum result; By using the customized partitioning logic of the sample data and performing corresponding matching on the weighted sum results, the target sample data partitioning logic corresponding to the weighted sum results can be determined.

4. The method according to claim 1, wherein The performing lesion identification and correction on the lesion semantic segmentation data includes: The lesion recognition and correction module is embedded with a Densenet classification image network; The lesion semantic segmentation data is input into the Densenet classification image network to perform classification correction to determine the multi-category lesion segmentation type.

5. The method according to claim 4, wherein The multi-category lesion segmentation system is further cascaded with an image processing monitoring device, and the method includes: By means of the image processing monitoring device, time consumption of any image in the test sample set is monitored to obtain a distribution of time consumption of each image processing; By performing abnormal time consumption traversal on the time consumption distribution of each image processing, abnormal processing images are obtained, marked and processed again.

6. A multi-category lesion segmentation system for diabetic retinopathy fundus image data, characterized in that: The system comprises: A data acquisition unit, configured to download an image pre-segmentation model from a multi-category lesion segmentation system, wherein the image pre-segmentation model is embedded with a deep learning network structure; A data acquisition unit, wherein the data acquisition unit is used to collect historical images of diabetic retinopathy to generate a historical image set, wherein the historical image set can be customized to be divided into a training sample set and a test sample set; a data analysis unit configured to perform mobile-end network integration on the image pre-segmentation model and the training sample set using a network integration module to generate a refined image segmentation model having a lightweight structure; an analysis and processing unit, configured to perform a loss analysis on the trained refined image segmentation model using the image pre-segmentation model, and determine a generalization parameter based on the analysis result, so as to generalize the refined image segmentation model to obtain a generalized image segmentation model; The loss analysis of the trained refined image segmentation model includes: using the image pre-segmentation model as the Teacher end and the refined image segmentation model as the Student end; performing loss function analysis on the Teacher end to determine the corresponding Teacher end loss parameter L t ; Perform loss function analysis on the Student side to determine the corresponding Student side loss parameter L s ; Calculate the Teacher end loss parameter L t and the Student-side loss parameter L s Performing weighted calculation to determine the loss function L of the refined image segmentation model; Among them, based on the formula: L=αL t +βL s , determine the loss function L of the refined image segmentation model, where α and β are loss weighting coefficients, α = 0.9, β = 0.1; Based on the formula: The loss function analysis is performed on the Teacher side, where L t Teacher loss, p′, generated by the image pre-segmentation model file i is the probability of the i-th category generated after the softmax of the image pre-segmentation model file, v i is the unnormalized probability of the i-th class in the image pre-segmentation model, q′ i is the probability of the i-th class generated by the softmax of the refined image segmentation model file after learning by the teacher network, z i is the unnormalized probability of the i-th class in the refined image segmentation model; Based on the formula: Perform loss function analysis on the Student side, where L s is the student loss generated by the refined image segmentation model file, N is the total number of categories, c i is the true value of the i-th category of the refined image segmentation model, the positive sample is 1, the negative sample is 0, q i is the probability of the i-th category generated by the softmax of the refined image segmentation model file after true value learning, A training integration unit, configured to input the test sample set into the generalized image segmentation model for training, so as to obtain semantic segmentation data of lesions of the test sample set; An identification and correction unit is used to use a lesion identification and correction module to perform lesion identification and correction on the lesion semantic segmentation data to obtain multi-category lesion segmentation types of the test sample set.

Images