A partial label loss joint correction method based on historical information driving
Patent Information
- Application Number
- CN202510219912.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-26
- Publication Date
- 2026-08-28
AI Technical Summary
[0004]传统深度学习方法在处理噪声标签时往往缺乏有效机制，容易受到噪声干扰，导致模型性能下降
[0039] Compared with the prior art, the present invention has the following advantages and effects: This method utilizes historical information accumulated during multiple rounds of training to dynamically filter and correct some sample labels, and at the same time designs a reweighted loss function for samples with uncorrected labels, effectively alleviating the model's memory effect on noisy labels and improving the model's robustness and classification accuracy under different noise levels.
Smart Images
Figure CN122657532A_ABST
Abstract
Description
Technical Field
[0001] This invention introduces a joint correction method for partial label loss based on historical information, belonging to the field of medical image processing technology. Background Technology
[0002] Medical image classification, as a fundamental task in intelligent medical image analysis, is a crucial prerequisite for achieving advanced analytical tasks such as keypoint localization, lesion segmentation, and lesion detection. Thanks to its efficiency and objectivity, automated medical image classification technology has received widespread attention from academia and industry in recent years. Although the subtle differences in the presentation of diseased tissues in images pose a significant challenge to medical image classification, deep learning-based automated classification methods have made groundbreaking progress in disease grading and judgment, demonstrating outstanding performance.
[0003] However, the performance of deep learning methods is highly dependent on the support of large-scale, high-quality labeled datasets. Constructing medical image datasets faces several unique challenges: patient privacy protection limits dataset size, while professional annotation processes are time-consuming, labor-intensive, and may introduce subjective differences from experts. Current mainstream annotation solutions, including crowdsourced annotation, automatic parsing of clinical reports, and learning with a small number of labels, while each with its advantages, inevitably introduce label noise, i.e., mislabeling or inaccurate labeling. This label noise significantly reduces the model's classification performance, forming the so-called label noise problem. Therefore, how to effectively train models on medical datasets with noisy labels has become a key technical problem driving the clinical application of deep learning models.
[0004] Traditional deep learning methods often lack effective mechanisms for handling noisy labels, making them susceptible to noise interference and leading to performance degradation. Current noise labeling methods based on single training epoch information attempt to alleviate this problem; however, due to the instability of prediction results, these methods may cause bias accumulation, further weakening the model's effectiveness. Therefore, designing more robust noise labeling strategies to reduce the negative impact of label noise on model training has become a crucial problem urgently needing to be solved in the field of medical image classification. Summary of the Invention
[0005] The purpose of this invention is to provide a joint correction method for partial label loss based on historical information. This method uses historical information accumulated during multiple training rounds to dynamically filter and correct partial sample labels. At the same time, it designs a reweighted loss function for samples with uncorrected labels, thereby effectively alleviating the model's memory effect on noisy labels and improving the model's robustness and classification accuracy under different noise levels.
[0006] To achieve the above-mentioned objectives, this invention provides a method for joint correction of partial label loss based on historical information, the method comprising the following steps:
[0007] In S1, the warm-up phase, the model maintains a first-in-first-out (FIFO) queue through the historical information learning module, dynamically recording the predicted probabilities of samples and their corresponding cross-entropy losses in each training round. Simultaneously, a hybrid loss function combining confidence penalty and cross-entropy is used for initial training to suppress model overconfidence and alleviate class imbalance.
[0008] S1-1, for each training sample x i Maintain a dynamically updated historical information H(x) i This contains two key sequences: the historical prediction sequence P. i and historical loss sequence L i ,Right now P i Record the network's performance on sample x over the past T rounds of training. i The predicted label probability, L i Record the corresponding standard cross-entropy loss value. In each round of training, the network performs training on sample x. i Predicted probability and loss value Calculate according to formulas (1) and (2) respectively:
[0009]
[0010] in, This indicates that the network performs a test on sample x in round t. i In all categories with noise labels The predicted probability distribution on. This indicates that the network performs a test on sample x in round t. i The softmax prediction probability value on the j-th class. The original noise labels for the samples are used. The historical information warehouse H(x) adopts a first-in-first-out (FIFO) queue structure and maintains the training information of the most recent T rounds through dynamic updates. The total storage space is 2NT, where the optimal value of the historical length T is determined through experiments.
[0011] S1-2 introduces a confidence penalty term based on information entropy during the warm-up phase to suppress the model's tendency to overconfident during prediction. The model evaluates the sample x... i The information entropy loss is defined by formula (3):
[0012]
[0013] Where, p j (xi ;θ) represents the network for sample x i The softmax prediction probability value for class j. By introducing this penalty term, the model can better balance the learning of different classes, mitigate the impact of class imbalance, and thus learn more effective knowledge during the warm-up phase.
[0014] Combining formula (2), the total cross-entropy loss of the model is defined by formula (4):
[0015]
[0016] Therefore, the total loss during the warm-up phase can be expressed as formula (5):
[0017] L warm =L ce -λL cp (5)
[0018] Where λ is the balance coefficient, used to adjust the weights of the cross-entropy loss and the confidence penalty term.
[0019] In S2, the robust training phase, the dual-screening label correction module first constructs a Gaussian mixture model (GMM) based on the historical loss mean to perform basic screening to distinguish potential clean samples from noisy samples; then, it uses the predicted flip frequency to perform secondary screening to further identify reliable samples; and finally, it corrects the labels of the samples that pass the screening based on the historical predicted dominant category.
[0020] S2-1 proposes a dual-selection label self-correction mechanism based on Gaussian mixture model and predicted flip frequency. It filters and corrects the labels of samples through multiple rounds of historical information, thereby obtaining a more reliable and clean sample set.
[0021] First, we utilize the historical loss mean dataset of all samples during T rounds of training. As a screening criterion, its calculation is as follows: (6)
[0022]
[0023] in, Indicates sample x i The historical mean loss over T rounds of training. Compared to single-round loss values, the historical mean loss dataset... It can more stably reflect the trend of loss changes in the sample, thereby improving the accuracy of GMM in fitting the loss distribution.
[0024] S2-2, using a binary Gaussian mixture model (K=2) to calculate the historical loss mean dataset of potentially clean samples and potentially noisy samples. The distribution of is modeled, and its probability density function is defined by formula (7):
[0025]
[0026] Where, π k It is the mixing coefficient of each Gaussian component. The mean is μ k The variance is Gaussian component distribution.
[0027] After optimizing the GMM parameters using the Expectation-Maximization (EM) algorithm, the latent variable z is calculated. i That is, sample x i Historical average loss The posterior probability of belonging to a certain Gaussian component can be interpreted as... The probability of belonging to the potential clean or noisy category is calculated as shown in formula (8):
[0028]
[0029] S2-3, In a binary GMM, clean samples typically have a lower historical loss mean compared to noisy samples, therefore the mean μ k Smaller Gaussian components (i.e., k = 1) correspond to clean sample classes. Based on this, we set a threshold τ for the latent variable z. i Perform screening: when the z-value of the sample i When the value is greater than τ, it indicates that the sample is a potentially clean sample and should be screened for further processing.
[0030] S2-4, A flipping frequency threshold was added after GMM screening. As a secondary screening condition. Sample x i The flip frequency δ(x) i The definition is as follows:
[0031]
[0032] in, This is an indicator function. When sample x i Historical predictions for round t and round t+1 and When inconsistent, the number of flips fl(x) i Increase by 1. For the potentially clean samples x selected by GMM... i Calculate its flip frequency δ(x) i When δ(x) i Less than the threshold When the prediction for a sample is considered to be stable, label correction can be performed.
[0033] Label correction is based on the category that appears most frequently in historical predictions. Specifically, statistical sample x i The historical label set S(x) is constructed by counting the occurrences of each category in the historical prediction sequence. i ):
[0034]
[0035] in, Indicates sample x i In historical predictions, it was predicted to be a label of type j. j The number of times.
[0036] S2-5, Select the category that appears most frequently in historical predictions as the corrected sample label. When the model's predictions for a sample tend to plateau, its prediction results exhibit significant consistency. Based on this characteristic, a label correction mechanism based on historical prediction consistency is proposed. Specifically, the frequency of each category appearing in the historical prediction sequence is statistically analyzed, and the category with the highest frequency is selected as the corrected label. This method utilizes the predictive consistency demonstrated by the model across multiple rounds of historical predictions to determine the dominant category of the sample, thereby effectively eliminating ambiguity introduced by noisy labels and improving the reliability of label correction. Formally, for sample x... i Its correction label Defined as the dominant category driven by its history, as shown in formula (11):
[0037]
[0038] In summary, the dual-screening mechanism initially filters potentially clean samples using the GMM model and further confirms sample reliability by incorporating the flipping frequency. For samples that pass the dual-screening, their labels are corrected based on their historically predicted dominant class, and the corrected samples are stored in the corrected dataset C. x The remaining samples are retained in the uncorrected dataset U. x This process is dynamically iterative, with the model continuously improving from U in subsequent training rounds. x The system identifies and corrects new, reliable samples, thereby gradually improving the label quality of the dataset.
[0039] Compared with the prior art, the present invention has the following advantages and effects: This method utilizes historical information accumulated during multiple rounds of training to dynamically filter and correct some sample labels, and at the same time designs a reweighted loss function for samples with uncorrected labels, effectively alleviating the model's memory effect on noisy labels and improving the model's robustness and classification accuracy under different noise levels. Attached Figure Description
[0040] To more clearly illustrate the specific embodiments of the present invention, the accompanying drawings used in the specific embodiments will be briefly described below.
[0041] Figure 1 This is the model of the joint correction method for partial label loss based on historical information driven by the present invention.
[0042] Figure 2 This is a dual-selection label self-correction module based on historical loss information. Detailed Implementation
[0043] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0044] To make the description of this disclosure more detailed and complete, illustrative descriptions of embodiments and specific examples of the present invention are provided below; however, these are not the only forms of implementing or utilizing the specific examples of the present invention. The embodiments cover features of multiple specific examples and methods and steps for constructing and operating these specific examples, and their order. However, other specific examples may also be used to achieve the same or equivalent functions and order of steps.
[0045] like Figure 1 and Figure 2 As shown, the joint correction method for partial label loss based on historical information includes the following steps:
[0046] In S1, the warm-up phase, the model maintains a first-in-first-out (FIFO) queue through the historical information learning module, dynamically recording the predicted probabilities of samples and their corresponding cross-entropy losses in each training round. Simultaneously, a hybrid loss function combining confidence penalty and cross-entropy is used for initial training to suppress model overconfidence and alleviate class imbalance.
[0047] S1-1, for each training sample x i Maintain a dynamically updated historical information H(x) i This contains two key sequences: the historical prediction sequence P. i and historical loss sequence L i ,Right now P i Record the network's performance on sample x over the past T rounds of training. i The predicted label probability, L i Record the corresponding standard cross-entropy loss value. In each round of training, the network performs training on sample x. i Predicted probability and loss value Calculate according to the following formulas respectively:
[0048]
[0049]
[0050] in, This indicates that the network performs a test on sample x in round t. i In all categories with noise labels The predicted probability distribution on. This indicates that the network performs a test on sample x in round t. i The softmax prediction probability value on the j-th class. The original noise labels for the samples are used. The historical information warehouse H(x) adopts a first-in-first-out (FIFO) queue structure and maintains the training information of the most recent T rounds through dynamic updates. The total storage space is 2NT, where the optimal value of the historical length T is determined through experiments.
[0051] S1-2 introduces a confidence penalty term based on information entropy during the warm-up phase to suppress the model's tendency to overconfident during prediction. The model evaluates the sample x... i Information entropy loss is defined as:
[0052]
[0053] Where, p j (x i ;θ) represents the network for sample x i The softmax prediction probability value for class j. By introducing this penalty term, the model can better balance the learning of different classes, mitigate the impact of class imbalance, and thus learn more effective knowledge during the warm-up phase.
[0054] According to formula (2), the total cross-entropy loss of the model is defined as:
[0055]
[0056] Therefore, the total loss during the warm-up phase can be expressed as:
[0057] L warm =L ce -λL cp
[0058] Where λ is the balance coefficient, used to adjust the weights of the cross-entropy loss and the confidence penalty term.
[0059] In S2, the robust training phase, the dual-screening label correction module first constructs a Gaussian mixture model (GMM) based on the historical loss mean to perform basic screening to distinguish potential clean samples from noisy samples; then, it uses the predicted flip frequency to perform secondary screening to further identify reliable samples; and finally, it corrects the labels of the samples that pass the screening based on the historical predicted dominant category.
[0060] S2-1 proposes a dual-selection label self-correction mechanism based on Gaussian mixture model and predicted flip frequency. It filters and corrects the labels of samples through multiple rounds of historical information, thereby obtaining a more reliable and clean sample set.
[0061] First, we utilize the historical loss mean dataset of all samples during T rounds of training. As a screening criterion, it is calculated as follows: in, Indicates sample x i The historical mean loss over T rounds of training. Compared to single-round loss values, the historical mean loss dataset... It can more stably reflect the trend of loss changes in the sample, thereby improving the accuracy of GMM in fitting the loss distribution.
[0062] S2-2, using a binary Gaussian mixture model (K=2) to calculate the historical loss mean dataset of potentially clean samples and potentially noisy samples. The distribution is modeled, and its probability density function is defined as:
[0063]
[0064] Where, π k It is the mixing coefficient of each Gaussian component. The mean is μ k The variance is Gaussian component distribution.
[0065] After optimizing the GMM parameters using the Expectation-Maximization (EM) algorithm, the latent variable z is calculated. i That is, sample x i Historical average loss The posterior probability of belonging to a certain Gaussian component can be interpreted as... The probability of belonging to either the potential clean or noisy category is calculated as follows:
[0066]
[0067] S2-3, In a binary GMM, clean samples typically have a lower historical loss mean compared to noisy samples, therefore the mean μ k Smaller Gaussian components (i.e., k = 1) correspond to clean sample classes. Based on this, we set a threshold τ for the latent variable z.i Perform screening: when the z-value of the sample i When the value is greater than τ, it indicates that the sample is a potentially clean sample and should be screened for further processing.
[0068] S2-4, A flipping frequency threshold was added after GMM screening. As a secondary screening condition. Sample x i The flip frequency δ(x) i ) is defined as: in, This is an indicator function. When sample x i Historical predictions for round t and round t+1 and When inconsistent, the number of flips fl(x) i Increase by 1. For the potentially clean samples x selected by GMM... i Calculate its flip frequency δ(x) i When δ(x) i Less than the threshold When the prediction for a sample is considered to be stable, label correction can be performed.
[0069] Label correction is based on the category that appears most frequently in historical predictions. Specifically, statistical sample x i The historical label set is constructed by counting the occurrences of each category in the historical prediction sequence. in, Indicates sample x i In historical predictions, it was predicted to be a label of type j. j The number of times.
[0070] S2-5, Select the category that appears most frequently in historical predictions as the corrected sample label. When the model's predictions for a sample tend to plateau, its prediction results exhibit significant consistency. Based on this characteristic, a label correction mechanism based on historical prediction consistency is proposed. Specifically, the frequency of each category appearing in the historical prediction sequence is statistically analyzed, and the category with the highest frequency is selected as the corrected label. This method utilizes the predictive consistency demonstrated by the model across multiple rounds of historical predictions to determine the dominant category of the sample, thereby effectively eliminating ambiguity introduced by noisy labels and improving the reliability of label correction. Formally, for sample x... i Its correction label Defined as:
[0071]
[0072] In summary, the dual-screening mechanism initially filters potentially clean samples using the GMM model and further confirms sample reliability by incorporating the flipping frequency. For samples that pass the dual-screening, their labels are corrected based on their historically predicted dominant class, and the corrected samples are stored in the corrected dataset C. x The remaining samples are retained in the uncorrected dataset U. x This process is dynamically iterative, with the model continuously improving from U in subsequent training rounds. x The system identifies and corrects new, reliable samples, thereby gradually improving the label quality of the dataset.
[0073] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for joint correction of partial label loss based on historical information, characterized in that, It includes two stages: S1, the warm-up phase, maintains a first-in-first-out (FIFO) queue through the historical information learning module, dynamically recording the predicted probability of samples and their corresponding cross-entropy loss in each round of training; at the same time, a hybrid loss function combining confidence penalty and cross-entropy is used for initial training to suppress model overconfidence and alleviate class imbalance. In S2, the robust training phase, the dual-screening label correction module first constructs a Gaussian mixture model (GMM) based on the historical loss mean to perform basic screening to distinguish potential clean samples from noisy samples; then, it uses the predicted flip frequency to perform secondary screening to further identify reliable samples; and finally, it corrects the labels of the samples that pass the screening based on the historical predicted dominant category.
2. As described in claim 1, characterized in that, The warm-up phase includes the following steps: S1-1, for each training sample x i Maintain a dynamically updated historical information H(x) i ), which contains two key sequences: the historical prediction sequence P i and historical loss sequence L i ,Right now P i Record the network's performance on sample x over the past T rounds of training. i The predicted label probability, L i Record the corresponding standard cross-entropy loss value. In each round of training, the network performs training on sample x. i Predicted probability and loss value They are respectively: in, This indicates that the network performs a test on sample x in round t. i In all categories with noise labels The predicted probability distribution on. This indicates that the network performs a test on sample x in round t. i The softmax prediction probability value on the j-th class. The original noise label of the sample; S1-2 introduces a confidence penalty term based on information entropy during the warm-up phase to suppress the model's overconfidence tendency during prediction. The model evaluates the sample x... i Information entropy loss is defined as: Where, p j (x i ;θ) represents the network for sample x i The softmax predicted probability value for class j; The total cross-entropy loss of the model is defined as: Therefore, the total loss during the warm-up phase can be expressed as: Lwarm=Lce-λLcp Where λ is the balance coefficient, used to adjust the weights of the cross-entropy loss and the confidence penalty term.
3. As described in claim 1, characterized in that, The robust training phase includes the following steps: S2-1, using the historical loss mean dataset of all samples during T rounds of training. As a screening criterion in, Indicates sample x i The historical average loss during T rounds of training; S2-2, using a binary Gaussian mixture model (K=2) to calculate the historical loss mean dataset of potentially clean samples and potentially noisy samples. The distribution is modeled, and its probability density function is defined as follows: Where, π k It is the mixing coefficient of each Gaussian component. The mean is μ k The variance is Gaussian component distribution; After optimizing the GMM parameters using the Expectation-Maximization (EM) algorithm, the latent variable z is calculated. i That is, sample x i Historical average loss The posterior probability of belonging to a certain Gaussian component. S2-3, Set a threshold τ for the latent variable z i Perform screening: when the z-value of the sample i When the value is greater than τ, it indicates that the sample is a potentially clean sample and should be screened for further processing. S2-4, A flipping frequency threshold was added after GMM screening. As a secondary screening condition, sample x i The reversal frequency δ)x i ) is defined as: δ(x) i )=fl(x i ) / T, in, As an indicator function, when sample x i Historical predictions for round t and round t+1 and When inconsistent, the number of flips fl(x) i Increase by 1 for the potential clean samples x selected by GMM. i Calculate its flip frequency When δ(x) i Less than the threshold When the prediction for the sample is considered to be stable, label correction can be performed. Statistical sample x i The historical label set is constructed by counting the occurrences of each category in the historical prediction sequence. Among them, |P i t =y j | represents sample x i In historical predictions, it was predicted to be a label of type j. j The number of times; S2-5, Select the category that appears most frequently in historical predictions as the corrected sample label. For sample x i Its correction label Defined as the dominant category driven by its history, as shown in formula (11):