A method and device for training an algorithm model applied to medical image analysis, an electronic device and a medium

Abstract

This invention relates to an algorithm model training method, apparatus, electronic device, and medium. The algorithm model training method includes the following steps: acquiring an algorithm model to be trained; determining the hyperparameters of the distribution function of the algorithm model to be trained, the hyperparameters including at least the standard deviation coefficient and kurtosis coefficient of the distribution function; obtaining an intermediate algorithm model based on a preset standard deviation coefficient value; performing a discretized iterative search on the kurtosis coefficient of the intermediate algorithm model to determine a target kurtosis coefficient value; performing a discretized iterative search on the standard deviation coefficient of the intermediate algorithm model based on the determined target kurtosis coefficient value to determine a target standard deviation coefficient value; and training the algorithm model to be trained using training set data based on the target standard deviation coefficient value and the target kurtosis coefficient value to obtain a target algorithm model. This method can obtain a better model initialization distribution function, achieving higher accuracy of the trained algorithm model without changing the algorithm model.

Description

Technical Field

[0001] This invention relates to the field of algorithm model training technology, and in particular to an algorithm model training method, device, electronic device and medium for medical image analysis. Background Technology

[0002] Medical image analysis is a major application area for deep learning algorithms in the medical field. In tasks such as lesion detection, segmentation, and recognition, deep learning algorithms far surpass traditional machine learning algorithms in accuracy, reaching or even exceeding the level of human experts and physicians. However, when training algorithm models, the training process is often only transferred from the training process for natural image vision tasks, without targeted optimization for the specific data distribution characteristics of medical images. Medical image analysis tasks often target specific data modalities and human body parts. The data modality is determined by the physical imaging method, so for a specific analysis task, the input data has characteristics such as fixed distribution and high similarity. However, for different tasks, due to differences in data modalities and human body parts, the data distribution varies greatly between tasks, which determines that the training and optimization process of medical image algorithms has a strong task relevance. Deep learning parameter optimization is non-convex optimization; the initialization of parameters determines the starting position of the algorithm optimization, directly affecting the local optimum space and thus having a crucial impact on the final accuracy of the algorithm model.

[0003] In general algorithm model training, the Kaiming initialization method is typically used to derive the variance of the algorithm model, thereby determining the initialization parameters. However, due to the inclusion of batch normalization (BN) layers and residual structures in the network architecture of modern deep learning algorithms, the variance of feature vectors during forward propagation and the variance of gradients during backpropagation are constrained by the BN layers and residual structures, making them less prone to divergence or vanishing. Furthermore, when using the Kaiming initialization method, differences in the kurtosis of the distribution function can also affect the model's accuracy. Therefore, parameter initialization methods based on variance constraints are no longer universally applicable, making it difficult to determine the optimal initialization parameters for the algorithm model. This results in insufficient accuracy of the trained algorithm model, thus impacting the accuracy of medical image analysis. Summary of the Invention

[0004] To overcome the problems existing in related technologies, embodiments of the present invention provide an algorithm model training method, device, electronic device and storage medium for medical image analysis. The algorithm model training method can search for hyperparameters of the algorithm model to obtain a better model initialization distribution function, thereby achieving higher accuracy of the trained algorithm model without changing the algorithm model.

[0005] According to a first aspect of the present invention, a method for training an algorithm model for medical image analysis is disclosed, comprising the following steps:

[0006] Obtain the algorithm model to be trained, and determine the hyperparameters of the distribution function of the algorithm model to be trained. The hyperparameters include at least the standard deviation coefficient and kurtosis coefficient of the distribution function. The algorithm model to be trained is an algorithm model applied to medical image analysis. Its medical image analysis task is for a specific data modality and human body part. The data modality is determined by the physical imaging method.

[0007] An intermediate algorithm model is obtained based on a preset standard deviation coefficient value. The kurtosis coefficient of the intermediate algorithm model is discretized and iteratively searched to determine the target kurtosis coefficient value.

[0008] Based on the determined target kurtosis coefficient value, the standard deviation coefficient of the intermediate algorithm model is discretized and iteratively searched to determine the target standard deviation coefficient value.

[0009] Based on the target standard deviation coefficient and the target kurtosis coefficient, the target algorithm model is obtained by training the algorithm model to be trained using medical image training set data.

[0010] According to a second aspect of the present invention, an algorithm model training device for medical image analysis is disclosed, comprising:

[0011] The hyperparameter determination module is used to obtain the algorithm model to be trained and determine the hyperparameters of the distribution function of the algorithm model to be trained. The hyperparameters include at least the standard deviation coefficient and kurtosis coefficient of the distribution function. The algorithm model to be trained is an algorithm model applied to medical image analysis. Its medical image analysis task is for a specific data modality and human body part. The data modality is determined by the physical imaging method.

[0012] The kurtosis coefficient determination module is used to obtain an intermediate algorithm model based on a preset standard deviation coefficient value, and to perform discretized iterative search on the kurtosis coefficient of the intermediate algorithm model to determine the target kurtosis coefficient value.

[0013] The standard deviation coefficient determination module is used to perform discretized iterative search on the standard deviation coefficient of the intermediate algorithm model based on the determined target kurtosis coefficient value to determine the target standard deviation coefficient value.

[0014] The training module is used to train the algorithm model to be trained using medical image training data based on the target standard deviation coefficient value and the target kurtosis coefficient value to obtain the target algorithm model.

[0015] According to a third aspect of the present invention, an electronic device is disclosed, comprising: a processor and a memory; the memory is electrically connected to the processor via a communication bus;

[0016] The memory stores a computer program adapted to be loaded by the processor and executed as described in any of the above embodiments for training the algorithm model.

[0017] According to a fourth aspect of the present invention, a computer-readable storage medium is disclosed having a computer program stored thereon, which, when executed by a processor, implements the algorithm model training method as described in any of the above embodiments.

[0018] By applying the above-described technical solution of this invention, the optimal values ​​of the standard deviation coefficient and kurtosis coefficient of the algorithm model to be trained are determined through discretized iterative search within a preset search space. Then, based on the target standard deviation coefficient and target kurtosis coefficient values, the algorithm model to be trained is trained using training set data to obtain the target algorithm model. This results in a better distribution function and thus a more accurate algorithm model. Compared to existing technologies, the algorithm model training method of this invention can search for the hyperparameters of the algorithm model to obtain a better model initialization distribution function. This allows for higher accuracy of the trained algorithm model without changing the algorithm model itself, thus contributing to improved accuracy in medical image analysis.

[0019] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention.

[0020] To better understand and implement this invention, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating an algorithm model training method according to an embodiment of the present invention;

[0022] Figure 2 This is a schematic diagram illustrating the hyperparameter search process of an algorithm model according to an embodiment of the present invention;

[0023] Figure 3 This is a schematic diagram of the sampling curve of the syn distribution shown in one embodiment of the present invention;

[0024] Figure 4 This is a flowchart illustrating step S2 of the algorithm model training method according to an embodiment of the present invention;

[0025] Figure 5 This is a flowchart illustrating step S23 of the algorithm model training method according to an embodiment of the present invention;

[0026] Figure 6 This is a flowchart illustrating a method for obtaining the latest value sequence number according to an embodiment of the present invention;

[0027] Figure 7 This is a schematic diagram illustrating the process of fine-grained search of hyperparameters of an algorithm model according to an embodiment of the present invention;

[0028] Figure 8 This is a flowchart illustrating an algorithm model training method according to another embodiment of the present invention;

[0029] Figure 9 This is a flowchart illustrating a search algorithm according to another embodiment of the present invention;

[0030] Figure 10 This is a schematic diagram of the structure of the algorithm model training device shown in an embodiment of the present invention;

[0031] Figure 11 This is a schematic diagram of the structure of an electronic device according to an optional embodiment of the present invention. Detailed Implementation

[0032] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.

[0033] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The singular forms “a,” “the,” and “the” used in this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0034] It should be understood that although the terms first, second, third, etc., may be used in this invention to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first information may also be referred to as second information without departing from the scope of this invention, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0035] According to a first aspect of the present invention, an algorithm model training method is provided. (See also...) Figure 1 and Figure 2 , Figure 1 This is a flowchart illustrating an algorithm model training method according to an embodiment of the present invention; please refer to [link / reference]. Figure 2 , Figure 2 This is a schematic diagram illustrating the hyperparameter search process of an algorithm model according to an embodiment of the present invention.

[0036] like Figure 1 As shown, the algorithm model training method includes the following steps:

[0037] S1: Obtain the algorithm model to be trained and determine the hyperparameters of the algorithm model to be trained. The hyperparameters include at least the standard deviation coefficient and the kurtosis coefficient.

[0038] In this embodiment, the algorithm model to be trained can be the original algorithm model or an intermediate algorithm model after partial hyperparameter optimization. The algorithm model to be trained must contain at least one hyperparameter to be determined, which may include the standard deviation coefficient and the kurtosis coefficient. The algorithm model to be trained can be an algorithm model applied to medical image analysis, which is a type of algorithm model that is highly sensitive to the distribution of parameter initialization; that is, using different parameter initialization strategies for the same algorithm model will result in significant differences in accuracy.

[0039] In machine learning, hyperparameters are parameters whose values ​​are set before the learning process begins, rather than parameters obtained through training. Typically, hyperparameters need to be optimized to select an optimal set for the learning machine, thereby improving learning performance and effectiveness.

[0040] In this context, standard deviation is the arithmetic square root of variance, which, in probability theory and statistics, measures the dispersion of a random variable or a set of data. In probability theory, variance measures the degree of deviation between a random variable and its expected value (mean). In statistics, variance (sample variance) is the average of the squared differences between each sample value and the mean of all sample values.

[0041] Based on a specific variance, such as when the variance is 1, there will be a... By controlling The value of can yield distribution functions with different kurtosis, then this This is the kurtosis coefficient.

[0042] In an optional embodiment, a random variable Y is defined as shown in Formula 1:

[0043] ;

[0044] Where N, U, and L are random variables, following Gaussian, uniform, and Laplace distributions respectively, and are mutually independent, with an expected value of 0 and a variance of . Assuming the random variable Y follows a combination distribution syn, then using Formula 1, we can deduce that the expected value of Y is 0 and the variance is... Kurtosis coefficient is .

[0045] If take =1, then you can get The sampling curves of the random variable Y ~ syn(0, 1) for different values ​​(the random variable Y satisfies the syn distribution, that is, syn is the distribution function).

[0046] like Figure 3 As shown, When syn = -1, syn is uniformly distributed. When syn = 0, the distribution is Gaussian. When syn = 1, it follows a Laplace distribution. Therefore, by controlling... By taking different values ​​of , distribution functions with different kurtosis can be obtained. Discretizing the search space yields A={ , ...}, during the search .

[0047] In the embodiments of this application, a kurtosis coefficient with better model evaluation coefficients is obtained through a search. This allows us to obtain a distribution function with controllable kurtosis.

[0048] In an optional embodiment, the algorithm model to be trained may include several algorithm models to be trained.

[0049] Obtain the algorithm model to be trained, including:

[0050] Obtain the original algorithm model, discretize the original algorithm model to obtain the overall model and several algorithm modules, and use the algorithm modules as the algorithm models to be trained.

[0051] Discretization of the model means dividing it into multiple modules. The control coefficients of the parameter distribution functions of different modules can be searched independently, meaning the hyperparameters of different modules can be searched vertically. The discretized model can be represented as M={M0, M1, M2, ..., Mn}, where M0 represents the overall model, and M1-Mn represent the n modules corresponding to the discretized model. If discretization is not required, then M={M0}, meaning the model is initialized using uniform distribution control coefficients.

[0052] S2: Obtain an intermediate algorithm model based on the preset standard deviation coefficient value, and perform discretization iterative search on the kurtosis coefficient of the intermediate algorithm model to determine the target kurtosis coefficient value.

[0053] In this embodiment, the target kurtosis coefficient is the value that the algorithm model can obtain with the highest model evaluation coefficient based on the preset standard deviation coefficient value.

[0054] Please see Figure 4 , Figure 4 This is a flowchart illustrating step S2 of the algorithm model training method according to an embodiment of the present invention. Optionally, step S2 includes:

[0055] S21: Substitute the preset standard deviation coefficient values ​​into the algorithm model to be trained to obtain the intermediate algorithm model.

[0056] The preset standard deviation coefficient value can be a value set by the inventor of this application based on business experience. After the preset standard deviation coefficient value is set, the variance can be determined. Then, based on the determined variance, other hyperparameters can be searched, such as the kurtosis coefficient.

[0057] In an optional embodiment, the method for presetting the standard deviation coefficient value is as follows: a unique standard deviation coefficient value is determined based on the Kaiming initialization method, and this unique standard deviation coefficient value is used as the preset standard deviation coefficient value.

[0058] Following the derivation of the Kaiming He method, the variance of the model parameters is initialized. Satisfying Formula 1:

[0059] ;

[0060] in, This indicates the number of input / output neurons; here, we will use the number of input neurons, i.e. =ck 2 c represents the number of channels in the input feature map, and k represents the size of the convolution kernel. For two-dimensional convolution, =ck 2 For 3D convolution, =ck 3 ; The first part of the model in Formula 1 represents the... layer, For the model in Formula 1 Layer parameters.

[0061] Based on the above variance formula one, design the variance Formula 2,

[0062] ;

[0063] in, This indicates the number of input / output neurons; here, we will use the number of input neurons for reference. This indicates scaling the standard deviation coefficient *s* by a factor of *s*. When *s*=1, this represents the variance used in the Kaiming He method derivation. Let *S*={s1, s2,...}, and discretize the search space of *s*, i.e., during the search... .

[0064] In this embodiment, the preset standard deviation coefficient is set to 1, that is... =1, s=1.

[0065] In other embodiments, the preset standard deviation coefficient value can be determined by other methods, which are not limited in this application.

[0066] S22: Construct a kurtosis coefficient search space, and discretize the kurtosis coefficient search space to obtain the set of kurtosis coefficient values.

[0067] In this embodiment, the kurtosis coefficient search space is A, and the set of kurtosis coefficient values ​​obtained after discretization is A={ , ,...},right After discretizing the search space, i.e. during the search... Set A contains a finite number of kurtosis coefficient values.

[0068] S23: Substitute each kurtosis coefficient value in the set of kurtosis coefficient values ​​into the intermediate algorithm model for iterative search to determine the intermediate algorithm model with the highest model evaluation coefficient.

[0069] Please see Figure 5 , Figure 5 This is a flowchart illustrating step S23 of the algorithm model training method according to an embodiment of the present invention. Optionally, step S23 includes:

[0070] S231: Number each kurtosis coefficient value in the set of kurtosis coefficient values ​​in order to obtain the value sequence number corresponding to the kurtosis coefficient value.

[0071] S232: Select the first kurtosis coefficient value from the set of kurtosis coefficient values ​​and record its corresponding value number. Substitute the first kurtosis coefficient value into the intermediate algorithm model to obtain the algorithm model to be evaluated.

[0072] In an optional embodiment, the value sequence number can be set to b, and several value sequences are b1, b2, b3, b4, b5, ... During the determination of the kurtosis coefficient value, the kurtosis coefficient value with value sequence number b1 can be used as the first kurtosis coefficient value. The kurtosis coefficient value corresponding to b1 can be substituted into the intermediate algorithm model to obtain the algorithm model to be evaluated; then, the kurtosis coefficient values ​​corresponding to b2 or other value sequences can be substituted into the intermediate algorithm model to obtain the algorithm model to be evaluated.

[0073] S233: The model evaluation coefficient is obtained by evaluating the algorithm model to be evaluated. The latest value number is determined based on the model evaluation coefficient, the value number b corresponding to the first kurtosis coefficient, and the first preset search step size d.

[0074] The first preset search step size determines the pace and efficiency of the search. In an optional embodiment, the first preset search step size can be greater than 2. During the coarse search phase, the first preset search step size is a larger value.

[0075] Optionally, the model evaluation coefficients can be obtained by evaluating the algorithm model using AUC (area under the curve). AUC is a model evaluation metric; AUC is the area under the ROC curve. The model evaluation coefficients can then be obtained after the evaluation.

[0076] In an optional embodiment, please refer to Figure 6 , Figure 6 This is a flowchart illustrating a method for obtaining the latest value sequence number according to an embodiment of the present invention.

[0077] Step S233 also includes the following steps:

[0078] S2331: Evaluate the algorithm model to be evaluated to obtain the current model evaluation coefficient.

[0079] S2332: If the current model evaluation coefficient is greater than or equal to the previous model evaluation coefficient, then the direction of the first preset search step is determined to be positive, that is, move forward and continue searching; if the current model evaluation coefficient is less than the previous model evaluation coefficient, then the direction of the first preset search step is determined to be negative, that is, step back and continue searching.

[0080] In this embodiment, determining the search direction can improve search efficiency by continuously searching in directions with higher model accuracy, thereby reducing search steps and time.

[0081] S2333: Determine the latest value sequence number based on the value sequence number b corresponding to the first kurtosis coefficient, the first preset search step size d, and the direction (positive or negative) of the first preset search step size.

[0082] S234: Substitute the second kurtosis coefficient value corresponding to the latest value into the intermediate algorithm model for re-evaluation to obtain the model evaluation coefficient. Stop the search when the latest value exceeds the maximum value in the set of kurtosis coefficient values.

[0083] This allows for a finite number of searches within the set of kurtosis coefficients. After each search, the corresponding kurtosis coefficient value and the training set data are substituted into the algorithm model for evaluation, resulting in several model evaluation coefficients.

[0084] In an optional embodiment, S234: Substitute the second kurtosis coefficient value corresponding to the latest value index into the intermediate algorithm model for re-evaluation to obtain the model evaluation coefficient, and stop the search when the latest value index exceeds the maximum value index in the set of kurtosis coefficient values. The process further includes the following steps:

[0085] S235: Obtain the set of fine search kurtosis values, reduce the first preset search step size, search again in the set of fine search kurtosis values ​​and evaluate, and obtain several model evaluation coefficients.

[0086] Please see Figure 7 , Figure 7 This is a schematic diagram illustrating the process of fine-grained search of hyperparameters of an algorithm model according to an embodiment of the present invention.

[0087] S235: Obtain the set of fine search kurtosis values, reduce the first preset search step size, search again in the set of fine search kurtosis values ​​and evaluate, obtain several model evaluation coefficients, thereby improving the accuracy of hyperparameter search and improving the precision of the algorithm model.

[0088] Optionally, step S235 includes:

[0089] S2351: When the search stops because the latest value index exceeds the maximum value index in the set of kurtosis coefficient values, obtain the first value index of the kurtosis coefficient corresponding to the algorithm model with the highest current model evaluation coefficient. This is to help determine the upper boundary of the set of kurtosis values ​​for fine-grained search.

[0090] S2352: Determine the second and third value sequences adjacent to the first value sequence number based on the first value sequence number and the first preset search step size.

[0091] S2353: Obtain the model evaluation coefficients of the algorithm model corresponding to the second and third value indices, and determine the value indices with larger model evaluation coefficients as the lower boundary of the fine search kurtosis value set.

[0092] S2354: Use the first value index as the upper boundary of the fine search kurtosis value set, and determine the range of the fine search kurtosis value set based on the upper and lower boundaries.

[0093] S2355: Reduce the first preset search step size to the second preset search step size. Based on the second preset search step size, search and evaluate the kurtosis coefficient values ​​in the fine search kurtosis value set again to obtain all model evaluation coefficients. The setting of the second preset search step size determines the depth and accuracy of the hyperparameter search. The second preset search step size is 1 or other values.

[0094] S236: Compare all the model evaluation coefficients obtained from the evaluation to determine the intermediate algorithm model with the highest model evaluation coefficient.

[0095] S24: Determine the target kurtosis coefficient value based on the intermediate algorithm model with the highest model evaluation coefficient.

[0096] In this embodiment, the target kurtosis coefficient value can be determined by the intermediate algorithm model with the highest model evaluation coefficient.

[0097] S3: Based on the determined target kurtosis coefficient value, the standard deviation coefficient of the intermediate algorithm model is discretized and iteratively searched to determine the target standard deviation coefficient value.

[0098] In this embodiment, the target kurtosis coefficient value obtained in the above embodiment is substituted into the algorithm model to be trained to obtain an intermediate algorithm model. Then, the standard deviation coefficient of the intermediate algorithm model is discretized and iteratively searched to determine the target standard deviation coefficient value.

[0099] In an optional embodiment, the order of steps S2 and S3 can be reversed. That is, first, an intermediate algorithm model is obtained based on a preset kurtosis coefficient value, and the standard deviation coefficient of the intermediate algorithm model is discretized and iteratively searched to determine the target standard deviation coefficient value; then, based on the determined target standard deviation coefficient value, the kurtosis coefficient of the intermediate algorithm model is discretized and iteratively searched to determine the target kurtosis coefficient value.

[0100] S4: Based on the target standard deviation coefficient and the target kurtosis coefficient, the target algorithm model is obtained by training the algorithm model to be trained using the training set data. This results in an algorithm model with controllable kurtosis and high accuracy.

[0101] In an optional embodiment, the algorithm model to be trained includes several models to be trained; when optimizing each algorithm model to be trained, preset hyperparameter values ​​are substituted into one of the algorithm models to be trained, and hyperparameter search is performed on the other algorithm models to be trained until all the algorithm models to be trained are optimized, thereby determining the optimal values ​​of each hyperparameter.

[0102] In an optional embodiment, the preset hyperparameter values ​​include preset standard deviation coefficient values ​​and preset kurtosis coefficient values.

[0103] Please see Figure 8 and Figure 9 , Figure 8 This is a flowchart illustrating an algorithm model training method according to another embodiment of the present invention; Figure 9 This is a flowchart illustrating a search algorithm according to another embodiment of the present invention.

[0104] The preset hyperparameters are substituted into one of the algorithm models to be trained, and hyperparameter search is performed on the other algorithm models to be trained until all the algorithm models to be trained are optimized. The steps are as follows:

[0105] S801: Determine the current algorithm model to be trained, substitute the preset standard deviation coefficient value into the current algorithm model to be trained, and search for the kurtosis coefficient value to determine the first target kurtosis coefficient value.

[0106] In other embodiments, a preset kurtosis coefficient value can be substituted into the current algorithm model to be trained, and the standard deviation coefficient value can be searched to determine the first standard deviation coefficient value.

[0107] S802: Based on the value of the first target kurtosis coefficient, search for the standard deviation coefficient value of the current algorithm model to be trained to determine the value of the first target standard deviation coefficient.

[0108] S803: The first algorithm sub-model is obtained by training the current model to be trained based on the first target standard deviation coefficient value, the first target kurtosis coefficient value, and the training set data.

[0109] S804: Based on the first algorithm sub-model, optimize the other algorithm models to be trained until all algorithm models to be trained are optimized.

[0110] In an optional embodiment, when optimizing each algorithm model to be trained, the Kaiming initialization method is used to determine the hyperparameters of the overall model. Based on the overall model with determined hyperparameters, the hyperparameters of other algorithm modules are searched and determined to optimize other algorithm modules.

[0111] By applying the above-mentioned technical solution of this invention, the optimal values ​​of the standard deviation coefficient and kurtosis coefficient of the algorithm model to be trained are determined through discretized iterative search within a preset search space. Then, based on the target standard deviation coefficient and target kurtosis coefficient values, the algorithm model to be trained is trained using training set data to obtain the target algorithm model. This results in a distribution function with controllable kurtosis, leading to a more accurate algorithm model. Compared to existing technologies, the algorithm model training method of this invention can search for the hyperparameters of the algorithm model to obtain a better model distribution function. Without changing the algorithm model, the trained algorithm model achieves higher accuracy, which helps improve the accuracy of medical image analysis.

[0112] According to a second aspect of the present invention, an algorithm model training apparatus is disclosed. This apparatus can be used to execute the algorithm model training method described in Embodiment 1 of this application, and has corresponding functions and beneficial effects. For details not disclosed in the embodiments of the algorithm model training apparatus of this application, please refer to the content of the algorithm model training method of this application.

[0113] Please see Figure 10 , Figure 10 This is a schematic diagram of the structure of the algorithm model training device shown in an embodiment of the present invention.

[0114] Algorithm model training device 100, comprising:

[0115] The hyperparameter determination module 101 is used to obtain the algorithm model to be trained and determine the hyperparameters of the distribution function of the algorithm model to be trained. The hyperparameters include at least the standard deviation coefficient and the kurtosis coefficient of the distribution function.

[0116] The kurtosis coefficient determination module 102 is used to obtain an intermediate algorithm model based on a preset standard deviation coefficient value, and to perform discretization iterative search on the kurtosis coefficient of the intermediate algorithm model to determine the target kurtosis coefficient value.

[0117] The standard deviation coefficient determination module 103 is used to perform discretized iterative search on the standard deviation coefficient of the intermediate algorithm model based on the determined target kurtosis coefficient value to determine the target standard deviation coefficient value.

[0118] The training module 104 is used to train the algorithm model to be trained using training set data based on the target standard deviation coefficient value and the target kurtosis coefficient value to obtain the target algorithm model.

[0119] In this embodiment, each module performs a discretized iterative search within a preset search space to determine the optimal values ​​of the standard deviation coefficient and kurtosis coefficient of the algorithm model to be trained. Then, based on the target standard deviation coefficient and target kurtosis coefficient values, the algorithm model to be trained is trained using training set data to obtain the target algorithm model. This results in a better distribution function and a more accurate algorithm model. Compared to existing technologies, the algorithm model training device of this invention can search for the hyperparameters of the algorithm model to obtain a model distribution function with controllable kurtosis. This allows for higher accuracy of the trained algorithm model without changing the algorithm model itself, thus contributing to improved accuracy in medical image analysis.

[0120] It is worth noting that the algorithm model training device provided in the above embodiments is only illustrated by the division of the above functional modules when executing the algorithm model training method. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the algorithm model training device and the algorithm model training method embodiments provided in the above embodiments belong to the same concept, and the implementation process is detailed in the embodiments, which will not be repeated here.

[0121] According to a third aspect of the present invention, an electronic device is disclosed, see below. Figure 11 , Figure 11 This is a schematic diagram of the structure of an electronic device according to an optional embodiment of the present invention.

[0122] Electronic device 900 includes: at least one processor 901 and at least one memory 902;

[0123] The memory 902 is used to store one or more computer programs, which are adapted to be loaded by the processor and executed as in any of the above embodiments of the algorithm model training method.

[0124] The electronic device 900 also includes at least one network interface, a user interface, a memory, and at least one communication bus. The communication bus is used to enable communication between these components.

[0125] The user interface may include an interface for connecting to a display screen and an interface for connecting to a camera. Optional user interfaces may also include standard wired interfaces and wireless interfaces.

[0126] The network interface can be selected from standard wired interfaces and wireless interfaces (such as WIFI interfaces).

[0127] The processor 901 may include one or more processing cores. The processor 901 connects to various parts within the electronic device 900 using various interfaces and lines. It executes various functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored within the processor 901, and by calling data stored in the memory 902. Optionally, the processor 901 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 901 may integrate one or more of the following: a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), and a modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content to be displayed on the screen; and the modem handles wireless communication. It is understood that the modem may also be implemented as a separate chip without being integrated into the processor 901.

[0128] The memory 902 may include random access memory (RAM) or read-only memory. Optionally, the memory 902 may include a non-transitory computer-readable storage medium. The memory 902 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 902 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 902 may also be at least one storage device located remotely from the aforementioned processor 901. Figure 11 As shown, the memory 902, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and operating applications for smart devices.

[0129] exist Figure 11In the electronic device 900 shown, the user interface is mainly used to provide an input interface for the user, obtain the user input data, and provide a video input interface for the camera to obtain image signals; while the processor 901 can be used to call the operating application of the smart device stored in the memory 902 and execute the relevant operations in the image quality adjustment method in the above embodiment.

[0130] The aforementioned intelligent device can be used to execute the algorithm model training method of the corresponding embodiment of this application, and has corresponding functions and beneficial effects.

[0131] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored. When executed by a processor, the computer program implements the relevant operations in the algorithm model training method described in any of the foregoing embodiments, and has corresponding functions and beneficial effects. The computer-readable medium includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium, which can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient media, such as modulated data signals and carrier waves.

[0132] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0133] Note that the above are merely preferred embodiments and the technical principles employed in this application. Those skilled in the art will understand that this application is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of this application. Therefore, although this application has been described in detail through the above embodiments, this application is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of this application, the scope of which is determined by the scope of the appended claims.

Claims

1. A method for training an algorithm model applied to medical image analysis, characterized in that, Includes the following steps: Obtain the algorithm model to be trained, determine the hyperparameters of the distribution function of the algorithm model to be trained, the hyperparameters include at least the standard deviation coefficient and the kurtosis coefficient of the distribution function, and the distribution function with different kurtosis can be obtained by adjusting the kurtosis coefficient based on the standard deviation coefficient. The algorithm model to be trained is an algorithm model applied to medical image analysis, and its medical image analysis task is for a specific data modality and human body part, wherein the data modality is determined by the physical imaging method. An intermediate algorithm model is obtained based on a preset standard deviation coefficient value. The kurtosis coefficient of the intermediate algorithm model is then discretized and iteratively searched to determine the target kurtosis coefficient value. This includes: substituting the preset standard deviation coefficient value into the algorithm model to be trained to obtain an intermediate algorithm model; constructing a kurtosis coefficient search space and discretizing the search space to obtain a set of kurtosis coefficient values; substituting each kurtosis coefficient value in the set into the intermediate algorithm model for iterative search to determine the intermediate algorithm model with the highest evaluation coefficient; and determining the target kurtosis coefficient value based on the intermediate algorithm model with the highest evaluation coefficient. The method for determining the preset standard deviation coefficient value is as follows: a unique standard deviation coefficient value is determined based on the Kaiming initialization method, and this unique standard deviation coefficient value is used as the preset standard deviation coefficient value. According to the derivation of the Kaiming method, the variance of the initialized model parameters satisfies the following formula: This represents the number of input / output neurons; here, we'll use the number of input neurons for reference. For two-dimensional convolution, =ck 2 For 3D convolution, =ck 3 c is the number of channels in the input feature map, and k is the size of the convolution kernel; The model represents the first layer, For the model's first Layer parameters, This indicates that the standard deviation coefficient s is scaled by a factor of s, and S = {s1, s2, ...} is used to discretize the search space of s, i.e., during the search... ; Based on the determined target kurtosis coefficient value, the standard deviation coefficient of the intermediate algorithm model is discretized and iteratively searched to determine the target standard deviation coefficient value. Based on the target standard deviation coefficient and the target kurtosis coefficient, the target algorithm model is obtained by training the algorithm model to be trained using medical image training set data.

2. The algorithm model training method for medical image analysis according to claim 1, characterized in that, The step of substituting each kurtosis coefficient value from the set of kurtosis coefficient values ​​into the intermediate algorithm model for iterative search to determine the intermediate algorithm model with the highest evaluation coefficient includes: Each kurtosis coefficient value in the set of kurtosis coefficient values ​​is numbered sequentially to obtain the value sequence number corresponding to the kurtosis coefficient value; Select a first kurtosis coefficient value from the set of kurtosis coefficient values ​​and record its corresponding value number. Substitute the first kurtosis coefficient value into the intermediate algorithm model to obtain the algorithm model to be evaluated. The model to be evaluated is evaluated to obtain the model evaluation coefficient. The latest value sequence number is determined based on the model evaluation coefficient, the value sequence number corresponding to the first kurtosis coefficient, and the first preset search step size. Substitute the second kurtosis coefficient value corresponding to the latest value into the intermediate algorithm model for re-evaluation to obtain the model evaluation coefficient. The search stops when the latest value exceeds the maximum value in the set of kurtosis coefficient values. The evaluation coefficients of all models obtained from the evaluation are compared to determine the intermediate algorithm model with the highest evaluation coefficient.

3. The algorithm model training method for medical image analysis according to claim 2, characterized in that, After the step of substituting the second kurtosis coefficient value corresponding to the latest value index into the intermediate algorithm model for re-evaluation to obtain model evaluation coefficients, and stopping the search when the latest value index exceeds the maximum value index in the set of kurtosis coefficient values, the method further includes: Obtain a set of fine-grained search kurtosis values, reduce the first preset search step size, and search and evaluate again within the set of fine-grained search kurtosis values ​​to obtain several model evaluation coefficients, including: When the search stops when the latest value index exceeds the maximum value index in the set of kurtosis coefficient values, the first value index of the kurtosis coefficient corresponding to the algorithm model with the highest current model evaluation coefficient is obtained. The second and third value sequences adjacent to the first value sequence are determined based on the first value sequence number and the first preset search step size; Obtain the model evaluation coefficients of the algorithm models corresponding to the second and third value indices, and determine the value indices with larger model evaluation coefficients as the lower boundary of the fine search kurtosis value set; The first value sequence number is used as the upper boundary of the fine search kurtosis value set, and the range of the fine search kurtosis value set is determined based on the upper boundary and the lower boundary. The first preset search step size is reduced to the second preset search step size. Based on the second preset search step size, the kurtosis coefficient values ​​in the fine search kurtosis value set are searched and evaluated again to obtain all model evaluation coefficients.

4. The algorithm model training method for medical image analysis according to claim 2, characterized in that, The step of evaluating the algorithm model to be evaluated to obtain model evaluation coefficients, and determining the latest value sequence number based on the model evaluation coefficients, the value sequence number corresponding to the first kurtosis coefficient, and the first preset search step size, includes: The evaluation coefficients of the current model are obtained by evaluating the algorithm model to be evaluated; If the current model evaluation coefficient is greater than or equal to the previous model evaluation coefficient, the direction of the first preset search step is determined to be positive; if the current model evaluation coefficient is less than the previous model evaluation coefficient, the direction of the first preset search step is determined to be negative. The latest value number is determined based on the value sequence number corresponding to the first kurtosis coefficient value, the first preset search step size, and the direction of the first preset search step size.

5. The algorithm model training method for medical image analysis according to claim 1, characterized in that, The algorithm model to be trained includes several models to be trained; when optimizing each of the algorithm models to be trained, the preset hyperparameter values ​​are substituted into one of the algorithm models to be trained, and hyperparameter search is performed on the other algorithm models to be trained until all the algorithm models to be trained are optimized.

6. The algorithm model training method for medical image analysis according to claim 5, characterized in that, The preset hyperparameter values ​​include preset standard deviation coefficient values ​​and preset kurtosis coefficient values; The process involves substituting preset hyperparameters into one of the algorithm models to be trained, and then performing hyperparameter search on the other algorithm models to be trained until all algorithm models to be trained are optimized. Determine the current algorithm model to be trained, substitute the preset standard deviation coefficient value into the current algorithm model to be trained, and search for the kurtosis coefficient value to determine the first target kurtosis coefficient value; Based on the first target kurtosis coefficient value, the standard deviation coefficient value is searched for the current algorithm model to be trained to determine the first target standard deviation coefficient value; The first algorithm sub-model is obtained by training the current algorithm model based on the first target standard deviation coefficient value, the first target kurtosis coefficient value, and the training set data. Based on the first algorithm sub-model, other algorithm models to be trained are optimized until all algorithm models to be trained are optimized.

7. The algorithm model training method for medical image analysis according to claim 1, characterized in that, The algorithm model to be trained includes several models to be trained; obtaining the algorithm model to be trained includes: Obtain the original algorithm model, discretize the original algorithm model to obtain the overall model and several algorithm modules, and use the algorithm modules as the algorithm model to be trained.

8. The algorithm model training method for medical image analysis according to claim 7, characterized in that, When optimizing each of the algorithm models to be trained, the Kaiming initialization method is used to determine the hyperparameters of the overall model. Based on the overall model determined by the hyperparameters, the hyperparameters of the other algorithm modules are searched and determined to optimize the other algorithm modules.

9. An algorithm model training device for medical image analysis, characterized in that, include: The hyperparameter determination module is used to obtain the algorithm model to be trained and determine the distribution function hyperparameters of the algorithm model to be trained. The hyperparameters include at least the standard deviation coefficient and the kurtosis coefficient of the distribution function. Based on the standard deviation coefficient, the distribution function with different kurtosis can be obtained by adjusting the kurtosis coefficient. The algorithm model to be trained is an algorithm model applied to medical image analysis. Its medical image analysis task is for a specific data modality and human body part. The data modality is determined by the physical imaging method. The kurtosis coefficient determination module is used to obtain an intermediate algorithm model based on a preset standard deviation coefficient value, and to perform a discretized iterative search on the kurtosis coefficient of the intermediate algorithm model to determine a target kurtosis coefficient value. This includes: substituting the preset standard deviation coefficient value into the algorithm model to be trained to obtain an intermediate algorithm model; constructing a kurtosis coefficient search space, and discretizing the kurtosis coefficient search space to obtain a set of kurtosis coefficient values; substituting each kurtosis coefficient value in the set of kurtosis coefficient values ​​into the intermediate algorithm model for iterative search to determine the intermediate algorithm model with the highest model evaluation coefficient; and determining the target kurtosis coefficient value based on the intermediate algorithm model with the highest model evaluation coefficient. The method for determining the preset standard deviation coefficient value is as follows: a unique standard deviation coefficient value is determined based on the Kaiming initialization method, and this unique standard deviation coefficient value is used as the preset standard deviation coefficient value. According to the derivation of the Kaiming method, the variance of the initialized model parameters satisfies the following formula: This represents the number of input / output neurons; here, we'll use the number of input neurons for reference. For two-dimensional convolution, =ck 2 For 3D convolution, =ck 3 c is the number of channels in the input feature map, and k is the size of the convolution kernel; The model represents the first layer, For the model's first Layer parameters, This indicates that the standard deviation coefficient s is scaled by a factor of s, and S = {s1, s2, ...} is used to discretize the search space of s, i.e., during the search... ; The standard deviation coefficient determination module is used to perform discretized iterative search on the standard deviation coefficient of the intermediate algorithm model based on the determined target kurtosis coefficient value to determine the target standard deviation coefficient value. The training module is used to train the algorithm model to be trained using medical image training data based on the target standard deviation coefficient value and the target kurtosis coefficient value to obtain the target algorithm model.

10. An electronic device, characterized in that, include: The processor and memory; the memory is electrically connected to the processor via a communication bus; The memory stores a computer program adapted to be loaded by the processor and executed as described in any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the algorithm model training method as described in any one of claims 1 to 8.

Images