A lesion recognition system and method based on image information

Abstract

The application discloses a lesion recognition system and method based on image information, relates to the technical field of image processing, can allow a user to customize a lesion recognition auxiliary recognition task, has a large applicable range, and adopts an adaptive multi-order information fusion search algorithm to learn data relations between sample image information and label information corresponding to the sample image information in the lesion recognition auxiliary recognition task, obtains a lesion auxiliary recognition model, can effectively improve training effect, guarantees recognition accuracy of the lesion auxiliary recognition model, can effectively assist the user in processing image information, avoids misrecognition and missed recognition caused by manual processing, and assists the user in improving work efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to a lesion identification system and method based on image information. BACKGROUND

[0002] With the rapid development of medical imaging technology (such as CT (computed tomography), MRI (magnetic resonance imaging), X-ray, etc.), image data has become an important basis for clinical diagnosis. However, the traditional manual reading method highly depends on the professional knowledge and experience of radiologists, and has problems such as heavy workload, strong subjectivity, and easy fatigue and missed diagnosis. SUMMARY

[0003] The present application provides a lesion identification system and method based on image information, aiming to solve the problems of the current manual reading method which highly depends on the professional knowledge and experience of radiologists, has heavy workload, strong subjectivity, and is easy to fatigue and miss diagnosis.

[0004] The first aspect of the present application provides a lesion identification system based on image information, comprising:

[0005] A first data acquisition module is configured to acquire a user-input lesion identification auxiliary identification task; wherein the lesion identification auxiliary identification task at least includes a data pre-learning instruction, sample image information, and label information corresponding to the sample image information;

[0006] A data response module is configured to respond to the data pre-learning instruction, and learn the data relationship between the sample image information and the label information corresponding to the sample image information by using an adaptive multi-order information fusion search algorithm, to obtain a lesion auxiliary identification model;

[0007] A second data acquisition module is configured to acquire a user-input to-be-identified image information, and dispatch the lesion auxiliary identification model to process the to-be-identified image information, to obtain a lesion auxiliary identification result corresponding to the to-be-identified image information;

[0008] A result feedback module is configured to associate the lesion auxiliary identification result corresponding to the to-be-identified image information with the to-be-identified image information, to form associated result data, and display the associated result data to the user, to complete the lesion auxiliary identification based on image information.

[0009] In a possible implementation, the lesion identification auxiliary identification task further includes a deep learning model specification instruction;

[0010] Further comprising:

[0011] The deep learning model library is configured to determine a target deep learning model corresponding to a deep learning model designation instruction from a plurality of different stored deep learning models in response to the deep learning model designation instruction, and transmit the target deep learning model to a data response module, so that the data response module trains the target deep learning model according to sample image information and label information corresponding to the sample image information, and obtains a lesion auxiliary identification model by using an adaptive multi-order information fusion search algorithm.

[0012] In a possible implementation, the data response module is further configured to:

[0013] determine whether a deep learning model designation instruction exists in a lesion identification auxiliary identification task, and if so, perform data relationship learning by using a target deep learning model corresponding to the deep learning model designation instruction, or if not, perform data relationship learning by using a default deep learning model.

[0014] In a possible implementation, the data response module includes:

[0015] a training initialization sub-module configured to initialize hyperparameters of the target deep learning model or the default deep learning model to form a plurality of different particles, wherein each particle is obtained by encoding the hyperparameters;

[0016] a fitness acquisition sub-module configured to acquire a fitness corresponding to each particle based on the sample image information and the label information corresponding to the sample image information;

[0017] a multi-order optimization sub-module configured to train the particles by using an adaptive multi-order information fusion search algorithm according to the fitness corresponding to each particle, and obtain a lesion auxiliary identification model.

[0018] In a possible implementation, the training of the particles by using the adaptive multi-order information fusion search algorithm according to the fitness corresponding to each particle to obtain the lesion auxiliary identification model includes:

[0019] determining an optimal particle and a worst particle in a current training process according to the fitness corresponding to each particle;

[0020] performing self-information search on each particle by using a self-information gradient decision strategy to obtain particles after self-information search;

[0021] performing multi-directional information fusion search on the particles after self-information search by using an adaptive group information fusion strategy based on the optimal particle and the worst particle to obtain particles after multi-directional information fusion search;

[0022] Based on the optimal particle and the worst particle, a global jump search strategy with adaptive jump range control is used to perform global search on the particles after the multi-direction information fusion search, to obtain particles after global search;

[0023] A training end condition is judged to determine a training phase; wherein the training phase includes training in progress and training completed;

[0024] In a case where the training phase is training in progress, the particles after global search are taken as input corresponding to a next training process, and the step of determining the optimal particle and the worst particle in the current training process is returned;

[0025] In a case where the training phase is training completed, the optimal particle is re-determined according to the particles after global search, to obtain a target optimal particle;

[0026] According to the target optimal particle, a lesion auxiliary recognition model is obtained.

[0027] In a possible implementation, the self-information search on each particle to obtain particles after self-information search includes:

[0028] According to the current training number, the self-information search step length is:

[0029] ;

[0030] wherein, denotes the self-information search step length corresponding to the i th particle, z denotes self-information search, and c denotes the basis value of the search step length, denotes the self-information search range, denotes a sine function, denotes a circular constant, denotes the current total training number, denotes a preset maximum training number, denotes a logarithmic function; denotes the fitness ranking corresponding to the i th particle based on the order from small to large of the fitness; i = 1, 2, …, NP, denotes the total number of particles;

[0031] For any one particle, the gradient information corresponding to the i th particle is obtained as:

[0032] ;

[0033] wherein, denotes the gradient information corresponding to the i th particle, denotes a first random number that is a or-a, represents the i-th particle in the iter-th training process, represents a random particle with each element randomly being a or -a, represents a fitness function, represents a corresponding fitness;

[0034] According to the self-information search step and the gradient information, each particle is subjected to self-information search, and a particle after self-information search is obtained as follows:

[0035] ;

[0036] wherein, represents the i-th particle after self-information search, represents a unit particle with all elements being 1, represents a sign function.

[0037] In a possible implementation, the self-information search after the particle is subjected to multi-direction information fusion search based on the optimal particle and the worst particle by using an adaptive population information fusion strategy, and a particle after multi-direction information fusion search is obtained, including:

[0038] For any one particle after self-information search, a corresponding neighborhood radius of the particle is determined as:

[0039] ;

[0040] wherein, represents the neighborhood radius of the m-th particle after self-information search, m = 1, 2, …, NP, represents the m-th particle after self-information search in the iter-th training process, represents the n-th particle after self-information search in the iter-th training process, represents a Euclidean distance between ;

[0041] For any one particle after self-information search, other particles having a Euclidean distance less than the neighborhood radius with the particle after self-information search are determined, and a population information fusion particle corresponding to each particle after self-information search is obtained;

[0042] According to the population information fusion particle, population information corresponding to the particle after self-information search is obtained as:

[0043] ;

[0044] ;

[0045] ;

[0046] ;

[0047] ;

[0048] ;

[0049] ;

[0050] ;

[0051] ;

[0052] in, This represents the population information corresponding to the m-th self-information search particle during the iter-th training process. This represents the population information corresponding to the m-th self-information search in the (iter+1)-th training process. Indicates the search step size for group information. Indicates the search direction for group information. Indicates inertia weight, This indicates the search direction for particle contributions in group information fusion. This indicates the search direction for the optimal particle contribution. Jm represents the search information contributed by the j-th swarm information fusion particle corresponding to the m-th self-information search particle during the iter-th training process. The total number of particles in the neighborhood of the group information fusion. Let represent the influence of the j-th ensemble information fusion particle corresponding to the m-th particle after the m-th self-information search during the iter-th training process. This represents the fitness of the j-th ensemble particle. This represents the fitness of the worst-performing particle. This represents the fitness of the optimal particle. Let represent the fitness of the particle after the m-th self-information search in the iter-th training process. Let j represent the ensemble information fusion particle corresponding to the m-th particle after the m-th self-information search during the iter-th training process. This represents the disturbance factor. This represents the second random number between (0,1). This represents the influence of the optimal individual. This represents the search information contributed by the optimal individual. Represents the optimal individual;

[0053] According to the group information, the particle after the self-information search is subjected to multi-direction information fusion search, and a particle after the multi-direction information fusion search is obtained as follows:

[0054] ;

[0055] wherein, represents the mth particle after the multi-direction information fusion search.

[0056] In a possible implementation, the particle after the multi-direction information fusion search is subjected to global search by using a global jump search strategy with adaptive jump range control based on the optimal particle and the worst particle, and a particle after the global search is obtained, including:

[0057] The adaptive jump range control parameter is obtained based on the optimal particle and the worst particle as follows:

[0058] ;

[0059] wherein, represents the adaptive jump range control parameter, represents a constant term between 0 and 1, represents the fitness of the kth global jump search strategy of adaptive jump range control, represents the adaptive jump range base number;

[0060] The particle after the multi-direction information fusion search is subjected to global search by using a global jump search strategy with adaptive jump range control based on the adaptive jump range control parameter, and a particle after the global search is obtained as follows:

[0061] ;

[0062] wherein, represents the dth hyper-parameter of the kth particle after the multi-direction information fusion search in the iterth training process, k = 1, 2, …, NP, represents the dth hyper-parameter of the kth particle after the global search, d = 1, 2, …, D, D represents the total dimension of the hyper-parameters of the particle, represents a random value between and .

[0063] In a possible implementation, the training end condition is judged to determine the training phase, including:

[0064] determine the training phase as being in training.

[0065] The second aspect of the application provides a lesion identification method based on image information, comprising:

[0066] Collecting a user input lesion identification auxiliary identification task; wherein the lesion identification auxiliary identification task at least includes data pre-learning instructions, sample image information, and sample image information corresponding label information;

[0067] In response to the data pre-learning instructions, an adaptive multi-order information fusion search algorithm is used to learn the data relationship between the sample image information and the sample image information corresponding label information, and a lesion auxiliary identification model is obtained;

[0068] Collecting user input to be identified image information, and scheduling the lesion auxiliary identification model to process the to-be-identified image information, and obtaining the lesion auxiliary identification result corresponding to the to-be-identified image information;

[0069] Correlate the lesion auxiliary identification result corresponding to the to-be-identified image information and the to-be-identified image information to form an association result data, and display the association result data to the user, and complete the lesion auxiliary identification based on the image information.

[0070] Beneficial effects:

[0071] The application provides a lesion identification system and method based on image information, which can allow users to customize lesion identification auxiliary identification tasks, has a wide range of applications, and uses an adaptive multi-order information fusion search algorithm to learn the data relationship between sample image information and sample image information corresponding label information in the lesion identification auxiliary identification task, and obtains a lesion auxiliary identification model, which can effectively improve the training effect and ensure the identification accuracy of the lesion auxiliary identification model, can effectively assist users in processing image information, avoid misidentification and missed identification caused by manual processing, and assist users in improving work efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0072] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed to be used in the description of the embodiments of the application. Obviously, the drawings in the following description are only some embodiments of the application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0073] Figure 1 is a structural schematic diagram of a lesion identification system based on image information according to an embodiment of the application;

[0074] Figure 2 is a flowchart of a lesion recognition method based on image information according to an embodiment of the present application. DETAILED DESCRIPTION

[0075] The technical solutions in the embodiments of the present application will be clearly and completely described with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0076] As shown in Figure 1 The present application provides a lesion recognition system based on image information, comprising:

[0077] The first data acquisition module 101 is configured to acquire a lesion recognition auxiliary identification task input by a user, wherein the lesion recognition auxiliary identification task at least includes a data pre-learning instruction, sample image information, and label information corresponding to the sample image information.

[0078] A user (for example, a radiologist) uploads a data set containing 1000 labeled lung CT images through the front-end interface of the system. Each CT image is attached with label information labeled by an expert, indicating whether there is a lung nodule and the location of the nodule. The user clicks the "start training" button, which issues a data pre-learning instruction. The first data acquisition module 101 integrates these information (data pre-learning instruction, sample image information, label information) into a lesion recognition auxiliary identification task, and submits it to the subsequent module for processing.

[0079] The lesion recognition system provided by the present application is mainly used to realize image segmentation, so as to assist the user to quickly segment the lesion area in the image information, thereby assisting the user to improve the work efficiency and reduce the probability of misidentification and missed identification. For image segmentation, the label information corresponding to the sample image information is generally the marking box of the lesion area, so that the loss function value between the target marking box obtained by segmentation and the label information corresponding to the sample image information can be obtained in the training process, so as to evaluate the training effect according to the loss function value.

[0080] The data response module 102 is configured to learn the data relationship between the sample image information and the label information corresponding to the sample image information by using an adaptive multi-order information fusion search algorithm in response to the data pre-learning instruction, and obtain a lesion auxiliary identification model.

[0081] A default deep learning model can be pre-set, which can be a U-Net (U-shaped neural network), so as to assist the user to realize the segmentation of the lesion area. In response to the data pre-learning instruction, the default deep learning model can be trained by the sample image information and the label information corresponding to the sample image information, so as to obtain the lesion auxiliary recognition model.

[0082] In the prior art, intelligent optimization algorithms such as genetic algorithms or particle swarm algorithms are usually used to optimize the hyperparameters of U-Net, so as to obtain the lesion auxiliary recognition model, so that the lesion auxiliary recognition model has the ability of image segmentation. However, using intelligent optimization algorithms such as genetic algorithms or particle swarm algorithms to optimize the hyperparameters of U-Net usually leads to poor training results (such as the hyperparameter optimization falling into local optimum, or the hyperparameter optimization precision being poor), which causes the hyperparameters to be unable to be optimized to the optimal state, so that the lesion auxiliary recognition model cannot accurately realize image segmentation. Therefore, the adaptive multi-order information fusion search algorithm is used to learn the data relationship between the sample image information and the label information corresponding to the sample image information in the embodiments of the present application, so as to solve the technical problems existing in the prior art and ensure the segmentation accuracy of the lesion auxiliary recognition model on the image information.

[0083] The second data acquisition module 103 is configured to acquire the user-inputted to-be-recognized image information, and dispatch the lesion auxiliary recognition model to process the to-be-recognized image information, so as to obtain the lesion auxiliary recognition result corresponding to the to-be-recognized image information.

[0084] After the lesion auxiliary recognition model is trained, the image information-based lesion recognition system enters a standby state. When the user uploads a new, undiagnosed lung CT image (to-be-recognized image information), the second data acquisition module 103 dispatches the lesion auxiliary recognition model to process the to-be-recognized image information.

[0085] The result feedback module 104 is configured to associate the lesion auxiliary recognition result corresponding to the to-be-recognized image information with the to-be-recognized image information, form associated result data, and display the associated result data to the user, so as to complete the image information-based lesion auxiliary recognition.

[0086] The application provides a lesion recognition system based on image information, which can allow a user to customize a lesion recognition auxiliary identification task, has a large applicable range, and learns a data relationship between sample image information and label information corresponding to the sample image information in the lesion recognition auxiliary identification task by using an adaptive multi-order information fusion search algorithm, to obtain a lesion auxiliary identification model, which can effectively improve training effect, ensure identification accuracy of the lesion auxiliary identification model, effectively assist the user in processing image information, avoid misidentification and missed identification caused by manual processing, and assist the user in improving work efficiency.

[0087] In a possible implementation, the lesion recognition auxiliary identification task further includes a deep learning model specifying instruction.

[0088] Further include:

[0089] The deep learning model library 105 is configured to determine a target deep learning model corresponding to the deep learning model specifying instruction from a plurality of different deep learning models stored in the deep learning model library 105 in response to the deep learning model specifying instruction, and transmit the target deep learning model to the data response module 102, so that the data response module 102 trains the target deep learning model according to the sample image information and the label information corresponding to the sample image information by using the adaptive multi-order information fusion search algorithm, to obtain the lesion auxiliary identification model.

[0090] Optionally, the plurality of different deep learning models stored in the deep learning model library 105 can include one or more of a FCN (Fully Convolutional Network), a U-Net, a SegNet (Semantic segmentation network) and / or an ENet (Efficient Neural Network).

[0091] In a possible implementation, the data response module 102 is further configured to:

[0092] determine whether the lesion recognition auxiliary identification task includes a deep learning model specifying instruction, and if yes, learn the data relationship by using a target deep learning model corresponding to the deep learning model specifying instruction, or if not, learn the data relationship by using a default deep learning model.

[0093] The deep learning model library 105 of the application embodiment can allow a user to select a deep learning model by himself, and has higher flexibility.

[0094] In a possible implementation, the data response module 102 comprises:

[0095] a training initialization submodule, configured to initialize hyperparameters of a target deep learning model or a default deep learning model to form a plurality of different particles; each particle is obtained by encoding the hyperparameters;

[0096] The encoding manner of the particles can comprise:

[0097] For the hyperparameters (such as connection weights between network layers) of the target deep learning model or the default deep learning model, the hyperparameters have corresponding upper and lower limits, and thus the hyperparameters can be randomly initialized in the upper and lower limits, and the randomly initialized hyperparameters are encoded into a vector to obtain a particle. After repeating a plurality of times, a plurality of different particles can be obtained.

[0098] a fitness acquisition submodule, configured to acquire a fitness corresponding to each particle based on the sample image information and label information corresponding to the sample image information;

[0099] The method for acquiring the fitness of the particles can comprise:

[0100] The hyperparameters in the particle are applied to the target deep learning model or the default deep learning model, and then the loss function value is acquired by taking the sample image information as input and the label information corresponding to the sample image information as expected output. After the loss function value is added to a minimum constant (such as 0.001) and then taken as an inverse, the fitness of the particle is obtained. The greater the fitness is, the better the position of the particle in the solution space is.

[0101] a multi-order optimization submodule, configured to train the particles according to the fitness corresponding to each particle and by using an adaptive multi-order information fusion search algorithm to acquire a lesion auxiliary identification model.

[0102] After the particles are trained by using the adaptive multi-order information fusion search algorithm, a target optimal particle is obtained. The target optimal particle is the optimal hyperparameter combination obtained by training, and can also be regarded as a particle with the best data relationship learning effect between the sample image information and the label information corresponding to the sample image information. The hyperparameters in the target optimal particle can be used as the final hyperparameters of the target deep learning model or the default deep learning model to obtain the lesion auxiliary identification model.

[0103] In a possible implementation, the training of the particles according to the fitness corresponding to each particle and by using the adaptive multi-order information fusion search algorithm to acquire the lesion auxiliary identification model comprises:

[0104] According to the fitness corresponding to each particle, the optimal particle and the worst particle in the current training process are determined.

[0105] adopting a self-information gradient decision strategy to perform self-information search on each particle to obtain a particle after self-information search;

[0106] adopting an adaptive group information fusion strategy to perform multi-direction information fusion search on the particle after self-information search based on the optimal particle and the worst particle to obtain a particle after multi-direction information fusion search;

[0107] adopting a global leap search strategy with adaptive leap range control to perform global search on the particle after multi-direction information fusion search based on the optimal particle and the worst particle to obtain a particle after global search;

[0108] judging a training end condition to determine a training phase; wherein the training phase includes training in progress and training completed;

[0109] in a case where the training phase is training in progress, the particle after global search is taken as an input corresponding to a next training process, and the step of determining the optimal particle and the worst particle in the current training process is returned to;

[0110] in a case where the training phase is training completed, the optimal particle is re-determined according to the particle after global search to obtain a target optimal particle;

[0111] obtaining a lesion auxiliary recognition model according to the target optimal particle.

[0112] At present, mainstream image information processing systems usually adopt deep learning models, especially U-Net, to automatically analyze medical images to realize lesion detection and segmentation. These systems have achieved remarkable results on specific tasks, but their performance is highly dependent on the design and training process of the deep learning model. Deep learning models contain a large number of hyperparameters, and the setting of these hyperparameters is crucial to the final performance of the model. The existing technology currently adopts intelligent optimization algorithms such as genetic algorithms or particle swarm algorithms to optimize the hyperparameters of U-Net, which usually leads to poor training results (such as the hyperparameter optimization falling into a local optimum, or the hyperparameter optimization precision being poor), resulting in the hyperparameters being unable to be optimized to the optimal state, so that the lesion auxiliary recognition model cannot accurately realize image segmentation. Therefore, the adaptive multi-order information fusion search algorithm provided in the embodiments of the present application solves the technical problems existing in the prior art.

[0113] In a possible implementation, the adopting a self-information gradient decision strategy to perform self-information search on each particle to obtain a particle after self-information search includes:

[0114] According to the current training number, the self-information search step length is:

[0115] ;

[0116] wherein, denotes the self-information search step length corresponding to the i th particle, z denotes self-information search, and c denotes the base value of the search step length (such as a random number between (0.01, 0.2), for example, c can be set to 0.1, the smaller the value, the higher the search accuracy, which can be set according to actual needs), denotes the self-information search range (which can be set to a random number between (0.2, 0.4)), denotes a sine function, denotes a circular constant, denotes the current total number of training times, denotes the preset maximum number of training times, denotes a logarithmic function; denotes the fitness ranking corresponding to the i th particle based on the order from small to large fitness; i = 1, 2, …, NP, denotes the total number of particles;

[0117] For any one particle, the gradient information corresponding to the i th particle is obtained as: ;

[0118] wherein, denotes the gradient information corresponding to the i th particle, denotes a first random number randomly a or -a, denotes the i th particle in the iter th training process, denotes a random particle with each element randomly a or -a, denotes a fitness function, denotes corresponding fitness;

[0119] The gradient information is obtained by this pseudo-gradient calculation method, which can determine the search direction and select a better direction for search, thereby improving the training efficiency and training effect of the algorithm.

[0120] According to the self-information search step length and the gradient information, each particle is subjected to self-information search, and the particle after self-information search is obtained as:

[0121] ;

[0122] wherein, denotes the i th particle after self-information search, denotes a unit particle with all elements being 1, denotes a sign function.

[0123] The self-information gradient decision strategy is adopted in the embodiments of the present application to perform self-information search on each particle, so that each particle can independently explore based on the local gradient information of the particle, and the step length and direction of each particle are adaptive, the step length is dynamically adjusted according to the training process and the ranking of the particle, and the direction is guided by simulating gradient descent to move the particle to the direction of fitness improvement, thereby enhancing the local development capability.

[0124] In a possible implementation, the adaptive group information fusion strategy is adopted to perform multi-direction information fusion search on the particles after the self-information search based on the optimal particle and the worst particle, to obtain particles after the multi-direction information fusion search, including:

[0125] For any one particle after the self-information search, the neighborhood radius corresponding to the particle is determined as:

[0126] ;

[0127] wherein, r m represents the neighborhood radius of the mth particle after the self-information search, m = 1, 2, …, NP, represents the mth particle after the self-information search in the iter training process, represents the n th particle after the self-information search in the iter training process, represents the n th particle after the self-information search in the iter training process, represents the Euclidean distance between and

[0128] For any one particle after the self-information search, other particles having a Euclidean distance less than the neighborhood radius with the particle after the self-information search are determined, to obtain the group information fusion particles corresponding to each particle after the self-information search;

[0129] According to the group information fusion particles, the group information corresponding to the particle after the self-information search is obtained as:

[0130] ;

[0131] ;

[0132] ;

[0133] ;

[0134] ;

[0135] ;

[0136] ; ​​

[0137] ;

[0138] ;

[0139] wherein, denotes the group information corresponding to the particle after the mth self-information search in the iter training process, denotes the group information corresponding to the particle after the mth self-information search in the (iter+1)th training process, denotes the search step of the group information (which can be set to 0.45 or 0.55), denotes the search direction of the group information, denotes the inertia weight (which can be set to a fixed constant such as 0.1; or can be set to a number that linearly decreases from 0.3 to 0 as the training number increases), denotes the search direction of the group information fusion particle contribution, denotes the search direction of the optimal particle contribution, denotes the search information contributed by the jth group information fusion particle corresponding to the particle after the mth self-information search in the iter training process, Jm denotes the total number of group information fusion particles in the neighborhood range of denotes the influence corresponding to the jth group information fusion particle corresponding to the particle after the mth self-information search in the iter training process, denotes the fitness corresponding to the jth group information fusion particle, denotes the fitness corresponding to the worst particle, denotes the fitness corresponding to the optimal particle, denotes the fitness of the particle after the mth self-information search in the iter training process, denotes the jth group information fusion particle corresponding to the particle after the mth self-information search in the iter training process, denotes the disturbance factor, denotes a second random number between (0, 1), denotes the influence of the optimal individual, denotes the search information contributed by the optimal individual, denotes the optimal individual;

[0140] According to the group information, the particle after the self-information search is subjected to multi-directional information fusion search to obtain a particle after multi-directional information fusion search:

[0141] ;

[0142] wherein, represents the particle after the mth multi-directional information fusion search.

[0143] The embodiment of the application adopts an adaptive group information fusion strategy to perform multi-directional information fusion search on the particles after the self-information search, simulates social learning behavior, and each particle adjusts its search direction according to the information of other particles in its neighborhood and the information of the global optimal particle. Adaptive definition of the neighborhood radius makes information exchange more efficient, and influence calculation based on fitness difference ensures that the weight of excellent information is greater, thereby guiding the whole group to converge to a more optimal region, increasing the search of unknown regions, and accelerating convergence.

[0144] In a possible implementation, the global search strategy with adaptive jump range control is adopted to perform global search on the particles after the multi-directional information fusion search based on the optimal particle and the worst particle, to obtain particles after global search, including:

[0145] The adaptive jump range control parameter is obtained based on the optimal particle and the worst particle, and is:

[0146] ;

[0147] wherein, represents the adaptive jump range control parameter, represents a constant term between 0 and 1 (which can be set to 0.96), represents the fitness of the global jump search strategy of the kth adaptive jump range control, represents the adaptive jump range base (which can be set to 0.15 or 0.25);

[0148] The global search strategy with adaptive jump range control is adopted to perform global search on the particles after the multi-directional information fusion search according to the adaptive jump range control parameter, to obtain particles after global search:

[0149] ;

[0150] wherein, represents the dth hyperparameter of the particle after the kth multi-directional information fusion search in the iterth training process, k = 1, 2, …, NP, represents the dth hyperparameter of the particle after the kth global search, d = 1, 2, …, D, D represents the total dimension of the hyperparameters of the particle, represents a random value between and .

[0151] The global leap search strategy of adaptive leap range control is adopted to perform global search on the particles after the multi-direction information fusion search, which can effectively prevent the algorithm from falling into local optimum, and the optimal and worst particles are taken as the benchmark, so that each particle performs a leap, and the leap range is adaptive, which balances development and exploration and enhances the global search capability.

[0152] Optionally, the global leap search strategy can be controlled by using a greedy strategy or a simulated annealing algorithm, and only the search with increased fitness is accepted, so as to ensure the training speed of the algorithm. The particles can be processed for boundary crossing after each search strategy is executed, for example, the hyperparameters that cross the boundary are randomly generated within the upper and lower limit range, or the hyperparameters that exceed the upper limit are set to the corresponding upper limit value, and the hyperparameters that exceed the lower limit are set to the corresponding lower limit value.

[0153] The training of the deep learning model is effectively realized by the cooperation of the several strategies, and the segmentation accuracy of the image information is ensured.

[0154] In a possible implementation, the judgment on the training end condition to determine the training phase includes:

[0155] It is judged whether the total training number is greater than or equal to the preset maximum training number, if yes, it is determined that the training phase is complete training, otherwise it is determined that the training phase is training.

[0156] As shown in Figure 2 The application provides a lesion identification method based on image information, which includes:

[0157] S201, a user input lesion identification auxiliary identification task is collected; wherein the lesion identification auxiliary identification task at least includes a data pre-learning instruction, sample image information and label information corresponding to the sample image information;

[0158] S202, in response to the data pre-learning instruction, an adaptive multi-order information fusion search algorithm is used to learn the data relationship between the sample image information and the label information corresponding to the sample image information, and a lesion auxiliary identification model is obtained;

[0159] S203, a user input to-be-identified image information is collected, and the lesion auxiliary identification model is dispatched to process the to-be-identified image information, and a lesion auxiliary identification result corresponding to the to-be-identified image information is obtained;

[0160] S204, the lesion auxiliary identification result corresponding to the to-be-identified image information and the to-be-identified image information are associated to form associated result data, and the associated result data is displayed to the user, and the lesion auxiliary identification based on the image information is completed.

[0161] The lesion recognition method based on image information provided by the embodiment of the application can be executed by the system, and the principle and advantages are similar, which will not be repeated here.

[0162] Each embodiment in the specification is described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same and similar parts between each embodiment can be referred to each other.

[0163] The embodiments of the application are described with reference to flowcharts and / or block diagrams according to the methods, devices, electronic equipment and computer program products of the embodiments of the application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing terminal equipment to produce a machine, so that the instructions executed by the computer or other programmable data processing terminal equipment produce a device for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 The device for implementing the functions specified in one flow or multiple flows and / or blocks. Figure 1 The device for implementing the functions specified in one block or multiple blocks.

[0164] These computer program instructions can also be stored in a computer readable memory that can guide the computer or other programmable data processing terminal equipment to work in a specific way, so that the instructions stored in the computer readable memory produce a product including instruction devices, which implement the functions specified in the flowcharts and / or block diagrams. Figure 1 The device for implementing the functions specified in one flow or multiple flows and / or blocks. Figure 1 The device for implementing the functions specified in one block or multiple blocks.

[0165] These computer program instructions can also be loaded into a computer or other programmable data processing terminal equipment, so that a series of operation steps are performed on the computer or other programmable terminal equipment to produce a computer-implemented process, so that the instructions executed on the computer or other programmable terminal equipment provide a process for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 The device for implementing the functions specified in one flow or multiple flows and / or blocks. Figure 1 The device for implementing the functions specified in one block or multiple blocks.

[0166] Although the preferred embodiments of the application have been described, those skilled in the art can make additional changes and modifications to the embodiments once they know the basic creative concept. Therefore, the appended claims are intended to include the preferred embodiments and all changes and modifications falling within the scope of the embodiments of the application.

[0167] Finally, it needs to be pointed out that in this document, relational terms such as first and second and the like can only be intended to distinguish one entity or operation from another entity or operation without necessarily requiring or implying any actual such relationship or order between such entities or operations. Moreover, the terms "comprising", "comprising" or any other variations thereof are intended to cover non-exclusive inclusions, so that a process, method, article or terminal device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such a process, method, article or terminal device. Without more limitations, the element defined by the statement "comprising a" does not exclude the presence of additional identical elements in the process, method, article or terminal device comprising the element.

[0168] The principles and implementation manners of the present application are described by applying specific examples in this document, and the above example descriptions are only used to help understand the method and its core idea of the present application; meanwhile, for the general technical personnel in the art, the specific implementation manners and application ranges will be changed according to the idea of the present application, and the above description should not be understood as the limitation of the present application.

Claims

1. A lesion identification system based on image information, characterized in that, The method comprises the following steps: A first data acquisition module is configured to acquire a user inputted lesion identification auxiliary identification task; wherein the lesion identification auxiliary identification task at least comprises data pre-learning instructions, sample image information, and label information corresponding to the sample image information; A data response module is configured to learn a data relationship between the sample image information and the label information corresponding to the sample image information by using a self-adaptive multi-order information fusion search algorithm in response to the data pre-learning instructions, and obtain a lesion auxiliary identification model; A second data acquisition module is configured to acquire user inputted to-be-identified image information, and dispatch the lesion auxiliary identification model to process the to-be-identified image information, and obtain a lesion auxiliary identification result corresponding to the to-be-identified image information; A result feedback module is configured to associate the lesion auxiliary identification result corresponding to the to-be-identified image information with the to-be-identified image information, form associated result data, and display the associated result data to a user, thereby completing lesion auxiliary identification based on image information. The data response module comprises: A training initialization submodule is configured to initialize hyperparameters of a target deep learning model or a default deep learning model to form a plurality of different particles; wherein each particle is obtained by encoding the hyperparameters. The method of initializing the hyperparameters of the target deep learning model or the default deep learning model to form the plurality of different particles comprises: randomly initializing the hyperparameters in an upper limit and a lower limit, encoding the randomly initialized hyperparameters into a vector to obtain a particle, and repeating the above steps to obtain a plurality of different particles. An adaptability obtaining submodule is configured to obtain an adaptability corresponding to each particle based on the sample image information and the label information corresponding to the sample image information. The method of obtaining the adaptability of the particle comprises: The method of obtaining the adaptability of the particle comprises: A multi-order optimization submodule is configured to train the particles according to the adaptability corresponding to each particle and by using a self-adaptive multi-order information fusion search algorithm, and obtain a lesion auxiliary identification model.

2. The lesion identification system based on image information according to claim 1, characterized in that, The lesion identification auxiliary identification task further comprises a deep learning model designation instruction. Further comprising: A deep learning model library is configured to determine a target deep learning model corresponding to the deep learning model designation instruction from a plurality of different deep learning models stored in the deep learning model library in response to the deep learning model designation instruction, and transmit the target deep learning model to the data response module, so that the data response module trains the target deep learning model according to the sample image information and the label information corresponding to the sample image information by using a self-adaptive multi-order information fusion search algorithm, and obtains a lesion auxiliary identification model.

3. The lesion identification system based on image information according to claim 2, characterized in that, The data response module is further configured to: The judgment is whether there is a deep learning model specified instruction in the lesion recognition auxiliary recognition task, if yes, the target deep learning model corresponding to the deep learning model specified instruction is used to learn the data relationship, otherwise the default deep learning model is used to learn the data relationship.

4. The lesion identification system based on video information according to claim 3, characterized in that, According to the fitness corresponding to each particle, the particles are trained by using an adaptive multi-order information fusion search algorithm to obtain a lesion auxiliary recognition model, including: According to the fitness corresponding to each particle, the optimal particle and the worst particle in the current training process are determined; Each particle is searched by using a self-information gradient decision strategy to obtain a particle after self-information search; Based on the optimal particle and the worst particle, a particle after multi-direction information fusion search is obtained by using an adaptive group information fusion strategy to perform multi-direction information fusion search on the particle after self-information search; Based on the optimal particle and the worst particle, a particle after global search is obtained by using a global jump search strategy with adaptive jump range control to perform global search on the particle after multi-direction information fusion search; The training end condition is judged to determine a training phase; wherein the training phase includes training and completing training; If the training phase is training, the particle after global search is used as the input corresponding to the next training process, and the step of determining the optimal particle and the worst particle in the current training process is returned; If the training phase is completing training, a target optimal particle is obtained by re-determining the optimal particle according to the particle after global search; According to the target optimal particle, a lesion auxiliary recognition model is obtained.

5. The lesion identification system based on video information according to claim 4, characterized in that, The self-information search step is obtained according to the current training number: For any one particle, the gradient information corresponding to the i-th particle is obtained: ； wherein, represents the self-information search step length corresponding to the i-th particle, z represents the self-information search, and c represents the base value of the search step length, represents the self-information search range, represents a sine function, represents a circular constant, represents the current total number of training, represents the preset maximum number of training, represents a logarithmic function; represents the fitness ranking corresponding to the i-th particle on the basis of the order from small to large fitness; i = 1, 2, …, NP, represents the total number of particles; According to the self-information search step and the gradient information, each particle is searched by using a self-information gradient decision strategy to obtain a particle after self-information search. ； wherein, represents gradient information corresponding to the i-th particle, represents a first random number that is randomly a or -a, represents the i-th particle in the iter-th training process, represents a random particle that is randomly a or -a for each element, represents a fitness function, represents corresponding fitness; The adaptive group information fusion strategy is used to perform multi-direction information fusion search on the particle after self-information search based on the optimal particle and the worst particle to obtain a particle after multi-direction information fusion search. ； wherein, represents the i-th particle after self-information search, represents a unit particle in which all elements are 1, represents a sign function.

6. The lesion identification system based on video information according to claim 5, characterized in that, For any one particle after self-information search, the neighborhood radius corresponding to the particle is determined: For any one particle after self-information search, other particles with a Euclidean distance less than the neighborhood radius from the particle after self-information search are determined to obtain group information fusion particles corresponding to each particle after self-information search; ； wherein, represents the neighborhood radius of the particle after the mth self-information search, m = 1, 2, …, NP, represents the particle after the mth self-information search in the iterth training process, represents the particle after the nth self-information search in the iterth training process, represents the Euclidean distance between and According to the group information fusion particles, the group information corresponding to the particle after self-information search is obtained: According to the group information, the particle after self-information search is searched by using a multi-direction information fusion search strategy to obtain a particle after multi-direction information fusion search. ； ； ； ； ； ； ； ； ； wherein, denotes the group information corresponding to the particle after the mth self-information search in the iter training process, denotes the group information corresponding to the particle after the mth self-information search in the iter+1 training process, denotes the search step of the group information, denotes the search direction of the group information, denotes the inertia weight, denotes the search direction of the group information fusion particle contribution, denotes the search direction of the optimal particle contribution, denotes the search information contributed by the jth group information fusion particle corresponding to the particle after the mth self-information search in the iter training process, Jm denotes the total number of group information fusion particles in the neighborhood range of Jm, denotes the influence corresponding to the jth group information fusion particle corresponding to the particle after the mth self-information search in the iter training process, denotes the fitness corresponding to the jth group information fusion particle, denotes the fitness corresponding to the worst particle, denotes the fitness corresponding to the optimal particle, denotes the fitness of the particle after the mth self-information search in the iter training process, denotes the jth group information fusion particle corresponding to the particle after the mth self-information search in the iter training process, denotes the disturbance factor, denotes the second random number between (0, 1), denotes the influence of the optimal individual, denotes the search information contributed by the optimal individual, denotes the optimal individual; ​ ； wherein, represents the mth multi-directional information fusion search after the particle.

7. The lesion identification system based on video information according to claim 6, characterized in that, The global jump search strategy with adaptive jump range control is used to perform global search on the particles after the multi-direction information fusion search based on the optimal particle and the worst particle, to obtain particles after global search, and the method comprises the steps of: The adaptive jump range control parameters are obtained based on the optimal particle and the worst particle, and the adaptive jump range control parameters are as follows: ； wherein, denotes an adaptive jump range control parameter, denotes a constant term between (0, 1), denotes the fitness of the global jump search strategy of the kth adaptive jump range control, denotes an adaptive jump range base; The global jump search strategy with adaptive jump range control is used to perform global search on the particles after the multi-direction information fusion search based on the adaptive jump range control parameters, to obtain particles after global search, and the adaptive jump range control parameters are as follows: ； wherein, represents the dth dimension hyper-parameter of the particle after the kth multi-directional information fusion search in the iterth training process, k = 1, 2, …, NP, represents the dth dimension hyper-parameter of the particle after the kth global search, d = 1, 2, …, D, D represents the total dimension of the hyper-parameters of the particle, represents a random value between and 8. The lesion identification system based on video information according to claim 7, characterized in that, The training end condition is judged to determine the training phase, and the method comprises the steps of: It is judged whether the total training number is greater than or equal to the preset maximum training number, if yes, the training phase is determined to be complete training, otherwise the training phase is determined to be training.

9. An image information-based lesion identification method capable of being applied to the image information-based lesion identification system according to any one of claims 1 to 8, characterized by, It comprises the steps of: Collecting a user input lesion identification auxiliary identification task; wherein the lesion identification auxiliary identification task at least includes data pre-learning instructions, sample image information and sample image information corresponding label information; In response to the data pre-learning instructions, the adaptive multi-order information fusion search algorithm is used to learn the data relationship between the sample image information and the sample image information corresponding label information, and a lesion auxiliary identification model is obtained; Collecting user input to-be-identified image information, and scheduling the lesion auxiliary identification model to process the to-be-identified image information, and obtaining the lesion auxiliary identification result corresponding to the to-be-identified image information; The lesion auxiliary identification result corresponding to the to-be-identified image information and the to-be-identified image information are associated to form an association result data, and the association result data is displayed to the user, and the lesion auxiliary identification based on the image information is completed.

Images