Image classification-oriented unbalanced perception training-free neural network architecture search method

Abstract

The invention relates to an image classification-oriented unbalanced perception training-free neural network architecture search method, which designs a first proxy index for a long-tail medical data set classification task, introduces three selection and variation mechanisms in a neural architecture search process, and effectively solves the problems of high time overhead, data set imbalance and easiness in falling into local optimum. The method comprises the following steps: performing fast forward statistics on a candidate architecture on a small number of samples through an unbalanced perception data sampling and category weighting method to obtain a training-free proxy score capable of describing tail recognition capability and overall separability, forming uniform architecture fitness through a fusion module, and realizing fast sorting and screening of the candidate structures. Through probability selection and three selections, a variation mechanism is balanced, optimal and explored, and architecture output oriented to clinical deployment can be realized in combination with model complexity constraints. The time overhead can be obviously reduced, and the architecture selection efficiency and reliability in unbalanced medical image classification can be improved.

Description

Technical Field

[0001] This invention belongs to the field of automated machine learning technology, and in particular relates to a method for searching imbalanced perceptual training-free neural network architectures for image classification. Background Technology

[0002] Medical image classification plays a crucial role in early disease screening, lesion localization, and clinical auxiliary diagnosis, significantly improving diagnostic efficiency. Clinical datasets commonly suffer from class imbalance, characterized by a significantly smaller number of lesion-class samples compared to common or negative categories. In image classification tasks, convolutional neural networks and neural architecture search methods demonstrate immense potential due to their powerful inductive bias and automated architecture optimization, respectively.

[0003] Despite the success of these methods in imbalanced medical image classification tasks and the potential of neural architecture search (NAS) methods, the following technical problems remain.

[0004] First, the network architecture of deep and traditional machine learning methods still relies heavily on expert experience for manual design, which requires balancing factors such as pathological characterization, imaging modality, and data imbalance, resulting in a time-consuming and complex process that limits the universality and scalability of the methods.

[0005] Secondly, existing surrogate metrics for training-free NAS are mostly validated on natural images, and their applicability and robustness on medical datasets with imbalanced class distributions are still insufficient. Direct transfer will lead to a low correlation between network ranking and actual performance.

[0006] Furthermore, during the neural architecture search phase, it is susceptible to premature convergence due to exploration exploitation imbalance and candidate diversity decay, thus falling into local optima.

[0007] Therefore, this invention aims to provide a training-free neural network architecture search method for imbalanced perception in image classification, thereby achieving efficient medical image classification. First, a training-free NAS method is proposed, avoiding the manual design process required for CNNs, thus reducing time and effort costs. Then, a zero-cost surrogate index for imbalanced medical image classification is designed to alleviate the problem of insufficient relevance of NAS surrogate in imbalanced scenarios. Furthermore, three selection mutation mechanisms are introduced based on evolutionary algorithms to enhance population diversity during the NAS search process and improve the discovery of high-quality architectures, thereby avoiding the problem of getting trapped in local optima. Summary of the Invention

[0008] The purpose of this invention is to provide an imbalanced perceptual training-free neural network architecture search method for image classification, in order to solve the technical problems existing in the prior art and achieve efficient medical image classification.

[0009] The technical solution adopted in this invention is as follows: A method for searching imbalance-aware, training-free neural network architectures for image classification includes the following steps: S1: Configure the initial candidate architecture population to be randomly generated within the search space, and set the initial values ​​or ranges of the initial individual mutation rate (IMR) and global mutation rate (GMR) for each candidate architecture. S2: Collect samples of a specified batch size from the imbalanced medical image dataset, perform training-free evaluation on each candidate architecture, obtain feature gradient information related to the candidate architecture, and after standardizing and robustly transforming the feature gradient information, perform imbalance-aware weighting to generate surrogate score values ​​for the candidate architectures. ; S3: Agent score based on each candidate architecture Set up a Softmax probability sampling selection mechanism to select the agent score value. The temperature-scaled exponential mapping is converted into selection probabilities, and candidate architectures are randomly sampled according to the selection probabilities to determine the parent population; S4: Set up the mutation control module, based on the parent population and its corresponding agent score value. The individual mutation rate (IMR) of the parent generation and the global mutation rate (GMR) of the current generation are updated, and structural mutation operations are performed on the parent generation candidate architecture under the constraints of the IMR and the GMR to form the next generation population. S5: Repeat steps S2-S4 to perform a cyclical evolutionary search for evaluating, selecting, and mutating candidate architectures until the final generation evaluation is completed, and output the surrogate score. The optimal target architecture.

[0010] Preferably, the specific process of step S1 is as follows: S11: Configure the candidate architecture search space Ω, set the structural encoding method, optional operator set and connection rules of the candidate architecture, so that the candidate architecture can be combined by Normal cell and Reduction cell and expanded according to the preset Stage structure; S12: Set the initial population size N and the random sampling strategy, and randomly sample candidate architectures within the search space Ω to obtain the candidate architecture set. ; S13: For the candidate architecture set Perform structural validity checks and constraint elimination, removing candidate architectures that do not meet the preset structural constraints to form an initial candidate architecture population. ; S14: The initial candidate architecture population Each candidate architecture individual Set individual variation rates separately Initial value or range of values, and set the global mutation rate for the current generation. The initial value or range of values ​​is used for mutation control in subsequent evolutionary searches.

[0011] Preferably, the specific process of step S2 is as follows: S21: From an imbalanced medical image dataset Collect small batches of samples ,in For batch size, For the input image, Category labels; S22: For the aforementioned small batch of sample sets For each candidate architecture Perform a forward propagation to obtain multi-layer feature activations And perform a backpropagation once to obtain the loss gradient corresponding to the feature activation. ,in Representation layer index, L The number of layers participating in the evaluation; S23: For each layer Feature activation Calculate the mean within the layer with intralayer standard deviation : ; ; in, For the first The dimension of the activation channel after the layer space is expanded; S24: Based on step S23 and For each layer Feature activation is normalized using Z-score to obtain normalized activation. : ; in, To prevent division by zero of constants; S25: Based on standardized activation With gradient Calculate the Fisher inner product for each sample in each layer. : ; S26: For the Fisher inner product Perform logarithmic smoothing and absolute value processing to obtain the robust amount required for category-hierarchy aggregation: ; Among them, for categories , This indicates that the label in the current mini-batch is a category. The sample set, ; S27: Set class weights based on the number of samples in each class of the dataset. : ; in, For category The number of samples, The maximum number of samples in the largest category. For hyperparameters; S28: Based on the category weights With the category-hierarchy aggregation amount Agent scores for generating candidate architectures : ; And the agent score The fitness of the candidate architecture is used in subsequent steps S3–S5.

[0012] Preferably, the specific process of step S3 is as follows: S31: Agent scores for each candidate architecture obtained in step S2 Set the temperature coefficient And score the agent. Perform a temperature-scaled exponential mapping to generate selection probabilities. ; S32: Based on the stated selection probability The candidate architectures are randomly sampled using the Softmax probability sampling method to determine the parent population. The selection probability satisfy: ; in, This represents the current generation of candidate architecture population size; S33: Transfer the parent population Output to subsequent mutation stages, while retaining the surrogate score of the parent individual. As input for adaptive variation control.

[0013] Preferably, step S4 specifically includes: S41: Based on the parent population determined in step S3 and its corresponding agent rating The agent score is normalized to obtain a normalized score. : ; in , It is a very small positive constant used to avoid the denominator being zero; S42: Based on the normalized score Update the individual variation rate for each parent individual The individual variation rate is then truncated to a preset range. : ; Where represents the truncation function, and Used as parent individuals The structural variation intensity parameter during offspring generation; S43: Update the global mutation rate based on the current search status. The global mutation rate is truncated to a preset range. : ; in, The global baseline mutation rate varies with algebra and is used to characterize the overall exploration intensity in the current search phase; S44: In the individual variability rate With the global mutation rate Under the common constraints, the structural mutation operator is executed on the parent candidate architecture. Generate a set of child architectures : ; Wherein, the structural mutation operator It includes at least one of the following: operator replacement, connection reconnection, channel number adjustment, or network depth adjustment; S45: Set the sub-architecture collection The population is input into the next round of evaluation and participates in environmental selection together with the parent population to update and obtain the next generation of candidate architecture populations. : ; in, This represents an environment selection operator based on agent rating, used to retain high-rated individuals and maintain population diversity under a given capacity constraint.

[0014] Preferably, the specific process of step S5 is as follows: S51: Set a search termination condition, which includes at least one of the following: the number of generations reaches a preset upper limit. The cumulative number of assessments has reached the preset budget. Or the best agent score in continuous The intra-generation improvement is less than the preset threshold. The optimal agent rating is defined as: ; S52: If the termination condition is not met, iteratively execute the cross-generational update process: for the current generation of candidate architecture population Chinese architecture Calculate imbalance-aware agent scores without weight training. Based on agent scores, the system completes parent selection, offspring generation, and environment selection to obtain the next generation of candidate architecture population. ; S53: After each generation update, maintain a set of cross-generational optimal solution archives. And the best individuals of the present generation will be recorded in the archives: ; And save the above The structure encoding, operator configuration, and connection topology information are processed in a loop until the termination condition is met, and the optimal architecture of the population evolution is output.

[0015] The beneficial effects of this invention include: The imbalance-aware training-free neural network architecture search method for image classification provided by this invention differs from the traditional NAS approach, which requires long-term training on a large number of candidate architectures. This invention does not rely on complete training during the search phase, but instead uses imbalance-aware training-free agent scoring to quickly and comparablely distinguish the merits of candidate architectures.

[0016] Furthermore, in the neural architecture search phase, this invention introduces three selection and mutation mechanisms based on evolutionary algorithms. These three mechanisms mutually constrain each other during the search process, thereby guiding the search for the optimal architecture. This training-free NAS technology theoretically guarantees that this invention can search for the optimal architecture suitable for medical clinical data on imbalanced datasets. It effectively avoids the time overhead of manually designing architectures, stabilizes the correlation between surrogate metric scores and actual architecture performance, and effectively avoids getting trapped in local optima. It is particularly suitable for the technical needs of assisting doctors in diagnosis, such as medical image classification.

[0017] Through the complete technical implementation path described above, this invention successfully achieves a neural network architecture that can find a suitable medical dataset in a simple candidate search space without training, providing a technical solution with strict theoretical guarantees and practical engineering feasibility for medical image classification tasks.

[0018] Other advantages, objectives, and features of the invention will be set forth in part in the detailed description that follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0019] Figure 1 This is a flowchart provided in an embodiment of the present invention.

[0020] Figure 2 This is a graph showing the experimental results of the correlation coefficient on the MedMNIST dataset provided in this embodiment of the invention.

[0021] Figure 3 This is a radar chart showing the multi-index experimental results on the MedMNIST, ISIC, and BUSI datasets provided in this embodiment of the invention. Figure 3 In this context, 'a' represents the experimental results of the dermoscopic skin lesion dataset. Figure 3 In this context, 'b' represents the experimental results of the breast ultrasound dataset. Figure 3 In the figure, 'c' represents the experimental results of a small medical image dataset. Detailed Implementation

[0022] The following is in conjunction with the appendix Figures 1-3 The present invention will be described in further detail. It should be understood that the preferred embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0023] Example 1 See appendix Figure 1 As shown, the imbalanced perceptual no-training neural network architecture search method for image classification includes the following steps: S1: Configure the initial candidate architecture population to be randomly generated within the search space, and set the initial values ​​or ranges of the initial individual mutation rate (IMR) and global mutation rate (GMR) for each candidate architecture.

[0024] S2: Collect samples of a specified batch size from the imbalanced medical image dataset. Perform training-free evaluation processing on each candidate architecture generated in step S1 to obtain feature gradient information related to the candidate architecture. After standardizing and robustly transforming the feature gradient information, apply imbalance-aware weighting to generate surrogate score values ​​for the candidate architecture. .

[0025] S3: Agent score values ​​for each candidate architecture obtained in step S2 Set up a Softmax probability sampling selection mechanism to select the agent score value. The temperature-scaled exponential mapping is converted into selection probabilities, and candidate architectures are randomly sampled according to these selection probabilities to determine the parent population.

[0026] S4: Set up the mutation control module, based on the parent population and its corresponding agent score value. The individual mutation rate (IMR) of the parent generation and the global mutation rate (GMR) of the current generation are updated, and structural mutation operations are performed on the parent generation candidate architecture under the constraints of the IMR and the GMR to form the next generation population.

[0027] S5: Repeat steps S2-S4 to perform a cyclical evolutionary search for evaluating, selecting, and mutating candidate architectures until the final generation evaluation is completed, and output the surrogate score. The optimal target architecture.

[0028] In this embodiment, the specific process of step S1 is as follows: S11: Configure the candidate architecture search space Ω, set the structural encoding method, optional operator set and connection rules of the candidate architecture, so that the candidate architecture can be combined by Normal cell and Reduction cell and expanded according to the preset Stage structure; S12: Set the initial population size N and the random sampling strategy, and randomly sample candidate architectures within the search space Ω to obtain the candidate architecture set. ; S13: For the candidate architecture set Perform structural validity checks and constraint elimination, removing candidate architectures that do not meet the preset structural constraints to form an initial candidate architecture population. ; S14: The initial candidate architecture population Each candidate architecture individual Set individual variation rates separately The initial value or range of values ​​is set, and the global mutation rate of the current generation is set. The initial value or range of values ​​is used for mutation control in subsequent evolutionary searches.

[0029] Example 2 Based on Example 1, the specific process of step S2 is as follows: S21: From an imbalanced medical image dataset Collect small batches of samples ,in For batch size, For the input image, Category labels; S22: For the aforementioned small batch of sample sets For each candidate architecture Perform a forward propagation to obtain multi-layer feature activations And perform a backpropagation once to obtain the loss gradient corresponding to the feature activation. ,in Representation layer index, L The number of layers participating in the evaluation; S23: For each layer Feature activation Calculate the mean within the layer with intralayer standard deviation : ; ; in, For the first The dimension of the activation channel after the layer space is expanded; S24: Based on step S23 and For each layer Feature activation is normalized using Z-score to obtain normalized activation. : ; in, To prevent division by zero of constants; S25: Based on standardized activation With gradient Calculate the Fisher inner product for each sample in each layer. : ; S26: For the Fisher inner product Perform logarithmic smoothing and absolute value processing to obtain the robust amount required for category-hierarchy aggregation: ; Among them, for categories , This indicates that the label in the current mini-batch is a category. The sample set, ; S27: Set class weights based on the number of samples in each class of the dataset. : ; in, For category The number of samples, The maximum number of samples in the largest category. For hyperparameters; S28: Based on the category weights With the category-hierarchy aggregation amount Agent scores for generating candidate architectures : ; And the agent score The fitness of the candidate architecture is used in subsequent steps S3–S5.

[0030] In this embodiment, the specific process of step S3 is as follows: S31: Agent scores for each candidate architecture obtained in step S2 Set the temperature coefficient And score the agent. Perform a temperature-scaled exponential mapping to generate selection probabilities. ; S32: Based on the stated selection probability The candidate architectures are randomly sampled using the Softmax probability sampling method to determine the parent population. The selection probability satisfy: ; in, This represents the current generation of candidate architecture population size; S33: Transfer the parent population Output to subsequent mutation stages, while retaining the surrogate score of the parent individual. As input for adaptive variation control.

[0031] Example 3 Based on Example 1 or Example 2, step S4 specifically includes: S41: Based on the parent population determined in step S3 and its corresponding agent rating The agent score is normalized to obtain a normalized score. : ; in , It is a very small positive constant used to avoid the denominator being zero; S42: Based on the normalized score Update the individual variation rate for each parent individual The individual variation rate is then truncated to a preset range. : ; Where represents the truncation function, and Used as parent individuals The structural variation intensity parameter during offspring generation; S43: Update the global mutation rate based on the current search status. The global mutation rate is truncated to a preset range. : ; in, The global baseline mutation rate varies with algebra and is used to characterize the overall exploration intensity in the current search phase; S44: In the individual variability rate With the global mutation rate Under the common constraints, the structural mutation operator is executed on the parent candidate architecture. Generate a set of child architectures : ; Wherein, the structural mutation operator It includes at least one of the following: operator replacement, connection reconnection, channel number adjustment, or network depth adjustment; S45: Set the sub-architecture collection The population is input into the next round of evaluation and participates in environmental selection together with the parent population to update and obtain the next generation of candidate architecture populations. : ; in, This represents an environment selection operator based on agent rating, used to retain high-rated individuals and maintain population diversity under a given capacity constraint.

[0032] In this embodiment, the specific process of step S5 is as follows: S51: Set a search termination condition, which includes at least one of the following: the number of generations reaches a preset upper limit. The cumulative number of assessments has reached the preset budget. Or the best agent score in continuous The intra-generation improvement is less than the preset threshold. The optimal agent rating is defined as: ; S52: If the termination condition is not met, iteratively execute the cross-generational update process: for the current generation of candidate architecture population Chinese architecture Calculate imbalance-aware agent scores without weight training. Based on agent scores, the system completes parent selection, offspring generation, and environment selection to obtain the next generation of candidate architecture population. ; S53: After each generation update, maintain a set of cross-generational optimal solution archives. And the best individuals of the present generation will be recorded in the archives: ; And save the above The structure encoding, operator configuration, and connection topology information are processed in a loop until the termination condition is met, and the optimal architecture of the population evolution is output.

[0033] This invention organically embeds an imbalance-aware, training-free agent scoring system, designed for long-tail medical data distribution, into the core optimization objective of Neural Architecture Search (NAS) for medical clinical image classification. This achieves a complete technical path from "abstract agent evaluation criteria" to "executable evolutionary search algorithm." Through imbalance-aware data sampling and class weighting, this invention enables rapid forward / backward statistical analysis of candidate architectures using only a small number of samples, without training candidate network weights. This yields a training-free agent score that characterizes tail-end recognition ability and overall separability. Furthermore, probabilistic selection and adaptive mutation mechanisms are introduced to balance "preserving excellence" and "exploration," thereby efficiently screening high-performance network structures suitable for clinical deployment constraints.

[0034] Example 4 Based on Example 1, Example 2, or Example 3, this example provides... A dataset consisting of medical image training samples ; in, For the first Zhang's medical images, For category labels, Let be the number of categories. To adapt to the long-tailed imbalanced distribution commonly found in medical scenarios, this invention first performs statistical analysis on the category distribution, assuming the th category is... The number of samples in each class is Then the prior proportion of the categories is: ; And construct the category weights accordingly. This is used to increase the contribution of tail categories in the agent score, for example, by using: ; in, To prevent division by zero and the resulting stable term. Then, from the dataset... A small batch of samples was drawn from the middle. This batch is used for training-free evaluation of candidate architectures. An imbalanced sensing sampling strategy can be employed to ensure that tail-class samples are included. It is covered.

[0035] To reduce the interference of differences in the numerical scale of features across different layers on the proxy score, this invention performs intra-layer normalization on the intermediate features of the network before calculating the proxy score. Specifically, let the candidate network be at the th layer... Layer to the first The feature activation of each sample is... The standardization result is: ; in, and They represent the first Layer activation in batch Mean and standard deviation within, To prevent small constants from being divided by zero, and to map "training-free evaluation" from abstract criteria to an executable computational process, this invention provides the following definition.

[0036] Definition 1: Tail-perception learnability statistics Let the candidate architecture be Randomly initialized parameters are The network is recorded as The weighted gradient expectation of tail-aware sensing is defined in this invention as: ; in, For sample categories The corresponding weights This indicates the network output regarding the parameters. The gradient (which can be obtained through a single backpropagation in engineering). Furthermore, the energy form of the above statistics is defined as the tail-perceived learnability component: ; This component, without training the weights, can directly reflect the response strength of the candidate architecture to the "supervised signal with class weights", especially emphasizing the learnability of the tail classes.

[0037] Definition 2: Overall expressiveness and separability statistics To characterize the feature representation ability and inter-class separability of candidate architectures under random initialization, this invention extracts standardized features in several key layers. And define the feature covariance of this layer. Inter-class divergence Let the first... Class in The mean characteristics of the layer are The overall mean is Then the inter-class divergence is defined as: ; Based on this, the present invention constructs expressive / separable components: ; in, Represents the largest eigenvalue operator. For balance coefficient, This represents the number of layers involved in the statistics. This component relies solely on forward feature statistics and can be computed without any weight training.

[0038] Definition 3: Imbalance-aware training-free agent scoring (architecture fitness) This invention uses a fusion function to combine the two components obtained from Definition 1 and Definition 2. Synthesize into a unified agent rating: ; in It is a coordinate monotonic fusion function, that is, when Increase or When it increases, This also increases accordingly to ensure that the agent score remains semantically consistent and interpretable.

[0039] The core technological contribution of this invention lies in establishing a systematic mapping mechanism from abstract evaluation indicators such as "tail learnability, overall separability, and expressiveness" to "a proxy score that can be directly obtained through a single forward / backward pass of a small number of samples." Through a carefully designed technical solution, a strict one-to-one correspondence is established between the abstract objects and the data and computational objects in the actual training / search process: wherein the real data distribution consists of the original medical dataset... Bearing; training-free statistical sampling is based on mini-batch sets Bearing; the abstract "tail importance" is determined by class weights. Concrete implementation; the abstract "learnability" is determined by weighted gradient statistics. Concrete implementation; the abstract "expressiveness and separability" are... and The combined statistics are specifically implemented; ultimately, through a fusion function. Forming a single fitness , used for evolutionary search.

[0040] Based on the above-mentioned mapping relationship from theory to implementation, this invention fully embeds the training-free agent scoring into the optimization objective of NAS, transforming the abstract problem of "selecting the best architecture" into the following executable search objective: ; When clinical deployment constraints need to be met simultaneously, complexity constraints are introduced. (For example, the number of parameters, FLOPs, or inference latency), resulting in a constrained optimization form: ; Alternatively, an executable single-objective score can be constructed using a penalty-based approach: ; In order to balance "high score preservation" and "diversity exploration" in the subsequent search phase, and to avoid premature convergence and local optima trapping on imbalanced medical data, this invention further adopts an evolutionary search strategy and introduces three adaptive selection and mutation mechanisms.

[0041] First, there is the selection of Softmax probability sampling, for the th Individuals by temperature Calculate the probability of selection: ; Then, individual adaptive variation intensity control is performed, first by normalizing the scores: ; Remapping yields the offspring mutation rate: ; Secondly, there is the adaptive adjustment of the global baseline variability rate, which adjusts based on the degree of contemporary convergence and structural diversity. Dynamic updates suppress search stagnation and improve global exploration efficiency from a mechanism perspective.

[0042] Unlike the common NAS technique approach in existing technologies that evaluates candidate architectures through full training or long training curves, this invention obtains reliable surrogate scores for imbalanced long-tail medical data through rapid statistical analysis of a small number of samples without weight training in the candidate network. Furthermore, this invention explicitly introduces tail class weights and three adaptive selection / mutation mechanisms during the search process, enabling the search to both stably retain high-scoring architectures and continuously explore potentially superior structures. This effectively reduces time overhead and improves the efficiency and reliability of architecture selection in imbalanced medical image classification tasks.

[0043] Through the complete technical implementation path described above, this invention successfully transforms the training-free agent scoring from a mathematical statistical definition to a practically deployable NAS search algorithm, providing an engineering-feasible and interpretable technical solution for rapidly obtaining high-performance network structures that meet resource constraints in clinical scenarios.

[0044] Figure 2 The correlation coefficients of the present invention on 11 subsets of the MedMNIST dataset are shown, which demonstrate that the agent score and architecture performance of the method of the present invention are strongly correlated on this dataset. Figure 3 The radar charts showing the performance of the method of the present invention compared with other methods on three datasets demonstrate that the method of the present invention achieves effective performance while having the minimum number of parameters.

[0045] In summary, the imbalanced perception training-free neural network architecture search method for image classification provided by this invention rapidly selects high-performance network structures suitable for long-tailed medical data distributions from a predefined search space without training candidate network weights. This invention designs the first surrogate index for long-tailed medical dataset classification tasks and introduces three selection and mutation mechanisms during the neural architecture search process, effectively solving the problems of high time overhead, dataset imbalance, and susceptibility to local optima. Through imbalanced perception data sampling and class weighting, this invention performs rapid forward statistics on candidate architectures on a small number of samples, obtaining a training-free surrogate score that characterizes tail recognition ability and overall separability. A unified architecture fitness is formed through a fusion module, thereby achieving rapid ranking and selection of candidate structures. Based on this, this invention introduces an evolutionary search strategy oriented towards stability and diversity, balancing "preserving excellence" and "exploration" through probabilistic selection and the three selection and mutation mechanisms, and can combine model complexity constraints to achieve architecture outputs suitable for clinical deployment. This invention can significantly reduce time overhead, improve the efficiency and reliability of architecture selection in imbalanced medical image classification, and provide a feasible technical path for rapidly obtaining high-performance model structures in clinical scenarios.

Claims

1. An unbalanced perception-free training-free neural network architecture search method for image classification, characterized in that, The method comprises the following steps: S1: configuring random generation of initial candidate architecture population in search space, setting initial individual mutation rate and initial value or range of global mutation rate for each candidate architecture; S2: collecting a specified size batch of samples in the imbalanced medical image dataset, performing training-free evaluation processing on each candidate architecture, obtaining feature gradient information related to the candidate architecture, and performing standardization and robust transformation on the feature gradient information, and then implementing imbalance perception weighting to generate proxy score values of the candidate architecture; S3: based on the proxy score values of each candidate architecture, setting a Softmax probability sampling selection mechanism, converting the proxy score values into selection probabilities through temperature scaling exponential mapping, and randomly sampling the candidate architectures according to the selection probabilities to determine the parent population; S4: setting a mutation control module, updating the individual mutation rate of the parent individual and the global mutation rate of the current generation according to the parent population and the corresponding proxy score values, and performing structure mutation operation on the parent candidate architecture under the constraint of the individual mutation rate and the global mutation rate of the current generation to form the next generation population; S5: repeating steps S2-S4 to perform cyclic evolution search of candidate architecture evaluation, selection and mutation until the last generation evaluation is completed, and outputting the target architecture with the optimal proxy score value.

2. The image classification oriented unbalanced-aware training-free neural network architecture search method according to claim 1, wherein, The specific process of step S1 is as follows: S11: configuring candidate architecture search space, setting structure encoding mode, optional operator set and connection rule of candidate architecture, so that the candidate architecture can be combined by Normal cell and Reduction cell and expanded according to the preset Stage structure; S12: setting initial population size and random sampling strategy, randomly sampling candidate architectures in the search space to obtain a candidate architecture set; S13: performing structure legality verification and constraint screening on the candidate architecture set, eliminating candidate architectures that do not meet the preset structure constraint to form an initial candidate architecture population; S14: setting the initial value or value range of the individual mutation rate for each candidate architecture individual in the initial candidate architecture population, and setting the initial value or value range of the global mutation rate of the current generation for subsequent evolution search mutation control.

3. The image classification oriented unbalanced-aware training-free neural network architecture search method according to claim 1, wherein, The specific process of step S2 is as follows: S21: collecting a small batch of sample set from the imbalanced medical image dataset; S22: for the small batch of sample set, performing forward propagation on each candidate architecture to obtain multi-layer feature activation, and performing backward propagation to obtain loss gradient corresponding to the feature activation; S23: calculating the intra-layer mean and intra-layer standard deviation of each layer of feature activation; S24: performing Z-score standardization on the feature activation of each layer according to the intra-layer mean and intra-layer standard deviation of each layer obtained in step S23 to obtain standardized activation; S25: calculating the Fisher inner product of each layer and each sample based on the standardized activation and gradient; S26: performing logarithmic smoothing and absolute value processing on the Fisher inner product to obtain the robust quantity required for class-level aggregation; S27: setting class weight according to the number of class samples in the dataset; S28: generate a proxy score of the candidate architecture based on the category weight and the category-hierarchy aggregated quantity, and use the proxy score as the fitness of the candidate architecture for subsequent steps S3-S5.

4. The image classification oriented unbalanced-aware training-free neural network architecture search method according to claim 3, characterized in that, The specific process of step S3 is as follows: S31: based on the proxy score of each candidate architecture obtained in step S2, set a temperature coefficient, and perform temperature scaling exponential mapping on the proxy score to generate a selection probability; S32: randomly sample the candidate architectures according to the selection probability in a Softmax probability sampling manner to determine a parent population; S33: output the parent population to a subsequent mutation stage, and keep the proxy score of the parent individual as an input for adaptive mutation control.

5. The image classification oriented unbalanced-aware training-free neural network architecture search method according to claim 4, characterized in that, Step S4 specifically includes: S41: based on the parent population and the corresponding proxy score determined in step S3, perform normalization processing on the proxy score to obtain a normalized score; S42: update the individual mutation rate for each parent individual according to the normalized score, and truncate the individual mutation rate to a preset interval; S43: update the global mutation rate according to the current generation search state, and truncate the global mutation rate to a preset interval; S44: under the joint constraint of the individual mutation rate and the global mutation rate, perform a structure mutation operator on the parent candidate architecture to generate a child architecture set; S45: input the child architecture set to the next round of evaluation stage, and participate in environment selection together with the parent population to update the next generation candidate architecture population.

6. The image classification oriented unbalanced-aware training-free neural network architecture search method according to claim 5, characterized in that, The specific process of step S5 is as follows: S51: setting a search termination condition, the termination condition including at least one of the following: the number of evolution generations reaches a preset upper limit, the cumulative number of evaluations reaches a preset budget, or the optimal agent score increases by less than a preset threshold within a preset number of consecutive generations. K the optimal agent score increases by less than a preset threshold within a preset number of consecutive generations. S52: when the termination condition is not met, iteratively perform a cross-generation update process: calculate an imbalance-aware proxy score for each architecture in the current generation candidate architecture population without weight training, and complete parent selection, child generation and environment selection based on the proxy score to obtain the next generation candidate architecture population; S53: after each generation update, maintain a cross-generation optimal solution archive set, and write the current optimal individual into the archive, and save the structure encoding, operator configuration and connection topology information of the archive, and repeat the process until the termination condition is met, and output the optimal architecture of the population evolution.

Images