A HiPC method based on adaptive loss optimization

Abstract

The application provides a HiPC method based on adaptive loss optimization, and belongs to the technical field of medical image processing.The method is designed with a robust loss function combining positive labels and negative labels, dynamically adjusts sample weights through a sample reweighting strategy, and reduces noise sample interference; meanwhile, the range of negative labels is expanded, additional supervision information is provided through negative learning loss, the learning dimension is expanded, and the robustness of the model to noise labels is jointly enhanced.

Description

Technical Field

[0001] This invention introduces a HiPC method based on adaptive loss optimization, 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 methods based on single training epoch information attempt to alleviate this problem, but due to the instability of prediction results, these methods may cause bias accumulation, further weakening model effectiveness. While the improved historical information-driven partial label loss joint correction method (HiPC) enhances the robustness of noisy label handling, it still suffers from underfitting due to the robust loss function. Summary of the Invention

[0005] The purpose of this invention is to provide a HiPC method based on adaptive loss optimization. On the basis of the original HiPC method, a negative learning loss mechanism is introduced, which splits the label space into positive learning labels and negative learning labels. By using the indirect supervision information provided by the negative labels and combining it with the partially corrected sample reweighted loss function calculated by the positive labels, the training effect of the model is further improved.

[0006] To achieve the above-mentioned objectives, this invention provides a HiPC method based on adaptive loss optimization, the method comprising the following steps:

[0007] S1 proposes a robust loss function that dynamically adjusts sample weights by analyzing historical prediction discrepancies, aiming to mitigate the negative impact of noisy labels on model training.

[0008] For the corrected dataset C x The samples in the sample use the corrected labels. Perform training and set the weights for its loss calculation to 1. For the uncorrected dataset U x The samples in the dataset use historical prediction labels. Replace the original noise label The sample loss weights are dynamically adjusted based on the degree of ambiguity in historical predictions.

[0009] Specifically, this method designs a robust loss function L that adjusts sample weights based on the degree of ambiguity in historical predictions. hpc As in formula (1):

[0010]

[0011] Where, α w This represents the sample loss weight coefficient, which is dynamically adjusted based on the degree of ambiguity in historical predictions, reflecting the reliability of the sample. For sample x i Historical prediction tags, p j (x i ;θ) represents the model for sample x i Predicted probability for category j.

[0012] Although robust loss functions can reduce the impact of noise labels by adjusting the sample loss weights, they may lead to model underfitting in some cases.

[0013] S2. This method introduces a negative label learning (NLL) strategy, which randomly selects any non-given label from the label space as negative labels. These negative labels provide indirect supervision information to the model, thereby optimizing the model parameters. However, in the presence of noisy labels, the true label may appear among the non-given negative labels, causing the selection of negative labels to incorrectly include the true label, thus interfering with the model's learning.

[0014] This method improves upon the original negative learning loss function by proposing complementary negative label learning, with the complementary negative learning loss function L... NL As shown in formula (2):

[0015]

[0016] in, Indicates sample x i Negative labels on category j For the model on sample x i The predicted probability for category j. This loss function indirectly improves the predicted probability of the positive label category by optimizing the probability value associated with the negative label to be close to zero, thus aligning with the objective of negative learning.

[0017] Finally, the complete target loss function of this method is calculated by formula (3):

[0018] L = L hpc +βL NL (3)

[0019] Where β is a hyperparameter used to adjust the negative learning loss L NL Its influence.

[0020] Compared with the prior art, the present invention has the following advantages and effects: Based on the original HiPC, this method dynamically adjusts the sample weights by modifying the sample reweighting strategy, reduces the interference of noisy samples, expands the range of negative labels, provides additional supervision information through negative learning loss, expands the learning dimension, and jointly enhances the robustness of the model to noisy labels. Attached Figure Description

[0021] To more clearly illustrate the specific embodiments of the present invention, the accompanying drawings used in the specific embodiments will be briefly described below.

[0022] Figure 1 This invention presents the HiPC method model based on adaptive loss optimization. Detailed Implementation

[0023] 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.

[0024] 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.

[0025] like Figure 1As shown, the HiPC method based on adaptive loss optimization is presented, where: (a) the original HiPC method, (b) the robust loss function designed in this method, and (c) the complementary negative learning loss function designed in this method. This method includes the following steps:

[0026] S1. In medical image classification tasks, traditional sample selection methods have limitations. These methods typically discard the remaining samples identified as noise after selecting clean samples. However, the scarcity and high cost of acquiring medical images necessitate making full use of every sample. Meanwhile, directly using the original noise labels may introduce erroneous information. Therefore, a robust loss function is proposed, which dynamically adjusts sample weights by analyzing historical prediction discrepancies, aiming to mitigate the negative impact of noise labels on model training.

[0027] Specifically, during training, the model employs different sample processing strategies. For the corrected dataset C... x The samples in the sample use the corrected labels. Perform training and set the weights for its loss calculation to 1. For the uncorrected dataset U x The samples in the dataset use historical prediction labels. Replace the original noise label The sample loss weights are dynamically adjusted based on the degree of ambiguity in historical predictions.

[0028] Specifically, this method designs a robust loss function L that adjusts sample weights based on the degree of ambiguity in historical predictions. hpc :

[0029]

[0030] Where, α w This represents the sample loss weight coefficient, which is dynamically adjusted based on the degree of ambiguity in historical predictions, reflecting the reliability of the sample. For sample x i Historical prediction tags, p j (x i ;θ) represents the model for sample x i Predicted probability for category j.

[0031] Although robust loss functions can reduce the impact of noise labels by adjusting the sample loss weights, they may lead to model underfitting in some cases.

[0032] S2. Although the robust loss function can reduce the impact of noisy labels by adjusting the sample loss weights, it may lead to underfitting in some cases. This is because the gradient direction of the model may become unclear during the dynamic adjustment of the loss weights, thus affecting the optimization effect. To address this, this method introduces a negative label learning (NLL) strategy, which randomly selects any non-given label from the label space as a negative label. These negative labels provide indirect supervision information to the model, thereby optimizing the model parameters. However, in the presence of noisy labels, the true label may appear among the non-given negative labels, causing the selection of negative labels to incorrectly include the true label, thus interfering with the model's learning.

[0033] This method improves upon the original negative learning loss function by proposing complementary negative label learning, with the complementary negative learning loss function L... NL :

[0034]

[0035] in, Indicates sample x i Negative labels on category j For the model on sample x i The predicted probability for category j. This loss function indirectly improves the predicted probability of the positive label category by optimizing the probability value associated with the negative label to be close to zero, thus aligning with the objective of negative learning.

[0036] Complementary negative label learning has two main advantages: First, it employs a more efficient supplementary label strategy. In the original method, the selection range of supplementary labels is limited, and there is a risk of incorrectly selecting potential true labels, leading to low training efficiency. This method expands the selection range of supplementary labels while reducing the risk of misselecting true labels, thus improving training efficiency and accelerating model convergence. Second, it alleviates the underfitting problem of the loss function. This method increases the diversity of model learning by dynamically adjusting the supplementary labels (i.e., selecting different supplementary labels in each training round), effectively mitigating the underfitting phenomenon caused by weight adjustments.

[0037] Finally, the complete objective loss function of this method is:

[0038] L = L hpc +βL NL

[0039] Where β is a hyperparameter used to adjust the negative learning loss L NL Its influence.

[0040] 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 HiPC method based on adaptive loss optimization, characterized in that, Based on the original HiPC method, a negative learning loss mechanism is introduced, which splits the label space into positive learning labels and negative learning labels. By using the indirect supervision information provided by the negative labels and combining it with the partially corrected sample reweighted loss function calculated by the positive labels, the training effect of the model is further improved.

2. As described in claim 1, characterized in that, A robust loss function is proposed, which dynamically adjusts sample weights by analyzing historical prediction discrepancies, aiming to mitigate the negative impact of noisy labels on model training. For the corrected dataset C x The samples in the sample use the corrected labels. Perform training and set the weights for its loss calculation to 1; for the uncorrected dataset U x The samples in the dataset use historical prediction labels. Replace the original noise label And dynamically adjust the sample loss weights based on the degree of ambiguity in historical predictions; Specifically, this method designs a robust loss function L that adjusts sample weights based on the degree of ambiguity in historical predictions. hpc : Where, α w This represents the sample loss weighting coefficient, which is dynamically adjusted based on the degree of ambiguity in historical predictions, reflecting the reliability of the sample. For sample x i Historical prediction tags, p j (x i ;θ) represents the model for sample x i Predicted probability for category j.

3. As described in claim 1, characterized in that, The negative label learning (NLL) strategy is introduced, which randomly selects any non-given label in the label space as a negative label and uses these negative labels to provide indirect supervision information for the model, thereby optimizing the model parameters. An improvement to the original negative learning loss function was proposed, namely complementary negative label learning, with the complementary negative learning loss function L... NL : in, Indicates sample x i Negative labels on category j For the model on sample x i The loss function optimizes the probability value associated with the negative label to be close to zero, thereby indirectly improving the predicted probability of the positive label category, which is consistent with the goal of negative learning. Finally, the complete objective loss function of this method is: L = L hpc +βL NL Where β is a hyperparameter used to adjust the negative learning loss L NL Its influence.

Images