A disturbance-based active learning method and system for medical image classification

Abstract

The application discloses a kind of active learning medical image classification method and system based on disturbance;The application evaluates samples according to the influence of feature disturbance on model decision and the degree of disturbance. First, assign a mandatory label to each category to guide the disturbance and obtain the new prediction of the sample. If the model prediction of the sample changes, it is considered uncertain. Then the disturbance distance obtained by the feature disturber is regarded as a diversity vector, which can be used to select valuable samples by clustering. Finally, select the data points closest to each cluster center in the clustering space. The application provides an efficient data utilization scheme for medical image classification research, which helps to accelerate the output of scientific research results, promote technological innovation and clinical application in the field of medical image analysis, and reduce the workload of doctors.

Description

Technical Field

[0001] This invention relates to the field of medical image classification technology, and in particular to a perturbation-based active learning method and system for medical image classification. Background Technology

[0002] Existing medical image classification methods are typically based on supervised learning algorithms, which require a large amount of labeled data to train the model so that it can learn features that distinguish different categories of images.

[0003] However, in the medical field, obtaining high-quality labeled data is a challenging task. The annotation of medical images requires specialized medical knowledge, which is not only time-consuming but also costly.

[0004] To address the aforementioned issues, active learning has gained widespread attention and application in the field of medical image classification in recent years. Its core idea is to strategically select the most "valuable" samples from the unlabeled dataset for manual annotation, and then add these newly annotated data to the training set to iteratively improve model performance.

[0005] Active learning strategies can be mainly divided into uncertainty-based strategies and diversity-based strategies.

[0006] Uncertainty-based active learning methods rely on metrics such as entropy or confidence to assess uncertainty. For example, they evaluate the predictive consistency between unlabeled samples and their data-augmented samples to select samples with lower consistency scores; they measure the difference in model output under different optimization steps to estimate the uncertainty of the samples.

[0007] Diversity-based active learning methods utilize representative information such as feature distribution or gradient direction to select diverse samples. For example, they select samples that are significantly different from labeled samples but are representative among unlabeled samples; they train GCN models to distinguish between labeled and unlabeled samples and select unlabeled samples that are different from labeled samples.

[0008] The above methods mainly focus on measuring uncertainty or representative information, ignoring the distribution characteristics and boundary information of the samples.

[0009] Compared to natural images, medical images exhibit high heterogeneity. Designing more effective sample selection strategies remains to be explored. Therefore, we propose utilizing information from feature perturbations to assess sample uncertainty and diversity. Summary of the Invention

[0010] The purpose of this invention is to overcome the shortcomings and deficiencies of the prior art and provide a perturbation-based active learning medical image classification method and system to further optimize the efficiency and accuracy of medical image classification.

[0011] This invention evaluates samples based on the impact and degree of feature perturbation on model decisions. First, a mandatory label is assigned to each category to guide the perturbation and obtain new predictions for the samples. If the model prediction for a sample changes, it is considered uncertain. Then, the perturbation distance obtained by the feature perturbator is treated as a diversity vector, which can be used to select valuable samples through clustering. Finally, the data points closest to each cluster center are selected in the cluster space, with the same number as the cluster centers.

[0012] This invention is achieved through the following technical solution:

[0013] A perturbation-based active learning method for medical image classification includes the following steps:

[0014] S1, Characteristic perturbation step:

[0015] S1-1, Input the unlabeled samples into the frozen encoder to extract initial features;

[0016] S1-2, Input the initial features into the frozen classifier to obtain the initial predicted values;

[0017] S1-3: Input the initial features into multiple perturbers. Each perturber performs multiple rounds of perturbation to obtain multiple perturbed features. Multiple rounds generally refer to 20 rounds.

[0018] S1-4, input the perturbed features into the frozen classifier to obtain the perturbed predicted value;

[0019] S1-5, all samples whose predicted values ​​after perturbation are not equal to the initial predicted values ​​but are equal to the forced labels will be added to the candidate set;

[0020] The above perturbator is an MLP layer, each perturbator is assigned a forced label, and the loss function is the cross-entropy loss function and the L2 loss function.

[0021] The number of perturbers mentioned above is equal to the number of categories in the classification task, and the forced label is the label for all categories.

[0022] S2, diversity mapping and clustering steps:

[0023] S2-1, For samples in the candidate set, calculate the diversity vector for each sample. The result of each perturber is one element of the vector. If the initial predicted value of each element is the same as the forced label, the element value is set to 0. Otherwise, if the predicted value after perturbation is different from the forced label, the element value is set to -1. If the predicted value after perturbation is the same as the forced label, the element value is set to the Euclidean distance between the initial feature and the perturbation feature.

[0024] S2-2, perform K-means clustering on the diversity vectors, and select the data points closest to each cluster center with the same number of cluster centers.

[0025] A perturbation-based active learning medical image classification system, comprising:

[0026] The uncertainty module for feature perturbation uses a forced label to guide a limited perturbation; samples that are successfully perturbed to the forced label feature space will be selected as candidate samples for further sampling; samples that are not perturbed to the corresponding feature space will be ignored.

[0027] The diversity mapping and clustering module uses the direction and degree of perturbation to construct a diversity vector and cluster it to select diverse samples from candidate samples;

[0028] By repeating the sampling process of the above modules and iteratively using the selected samples for deep learning, a system for medical image classification tasks can be obtained with limited annotation costs.

[0029] Compared with the prior art, the present invention has the following advantages and effects:

[0030] This invention presents a perturbation-based active learning method for medical image classification. The proposed active learning method selects uncertain and diverse samples based on feature perturbation information. It selects the most valuable samples with a limited label budget and obtains a high-performance model.

[0031] This invention can be effectively applied in the field of medical image classification. Compared with traditional active learning methods, this invention, by introducing perturbation analysis, can more accurately identify the key samples that have the greatest influence on model training, thereby significantly improving the efficiency and effectiveness of model training with a limited annotation budget.

[0032] This invention, through an intelligent sample screening strategy, can significantly reduce reliance on professional medical knowledge and manual annotation, effectively reducing the time and economic costs of the data preparation stage.

[0033] This invention provides an efficient data utilization scheme for medical image classification research, which helps to accelerate the output of scientific research results, promote technological innovation and clinical application in the field of medical image analysis, and reduce the workload of doctors. Attached Figure Description

[0034] Figure 1 This is a schematic diagram illustrating the sample selection process of the perturbation-based active learning medical image classification method of the present invention.

[0035] Figure 2 This is a flowchart illustrating the perturbation-based active learning method of the present invention. Detailed Implementation

[0036] The present invention will now be described in further detail with reference to specific embodiments.

[0037] like Figure 1-2 As shown. This invention discloses a perturbation-based active learning method for medical image classification. It selects samples with high uncertainty from unlabeled samples as candidate samples, then filters diverse samples from these candidate samples for labeling and use them for model training, thereby obtaining a high-performance model with limited labeling costs. It mainly consists of two steps:

[0038] (1) Selecting uncertain samples

[0039] In the feature space, samples that are susceptible to small perturbations and cause changes in prediction can be considered uncertain samples. Therefore, an uncertainty module called feature perturbation is proposed to guide finite perturbations using forced labels. Samples successfully perturbed to the forced-label feature space are selected as candidate samples for further sampling, for example... Figure 1 Sample a in (a). Conversely, samples that fail to perturb the corresponding feature space will be ignored, for example, Figure 1 Sample b in (a). Our uncertainty module enhances the model's ability to distinguish between different categories.

[0040] (2) Select a diverse sample

[0041] The perturbation direction and distance of the forced labels reflect the differences between samples. Therefore, a diversity module, called diversity mapping and clustering, is proposed, which uses the direction and degree of perturbation to construct a diversity vector and cluster it to select diverse samples from the candidate samples. Figure 1 In (b), candidate samples c and d with different diversity vectors are considered as diversity samples. The diversity module avoids information redundancy.

[0042] By repeating the above sampling process and iteratively using the selected samples for deep learning, a model with good performance for medical image classification tasks can be obtained with limited annotation costs.

[0043] Figure 2 A flowchart illustrating the workflow of this invention is provided.

[0044] Figure 2 The main process of active learning (A) involves initially selecting a small batch of samples to train the model. Subsequent iterations use a pre-labeled sample pool to fine-tune the model. After fine-tuning, an active learning strategy is used to select another batch of samples, manually label them, and add them to the pre-labeled sample pool for further fine-tuning, until a sufficient number of samples are accumulated. The model mainly consists of an encoder and a classifier.

[0045] Figure 2B demonstrates the perturbation-based active learning method proposed in this invention, which is mainly divided into two modules: feature perturbation, diversity mapping, and clustering.

[0046] Feature perturbation module:

[0047] S1. Input the unlabeled samples into the frozen encoder to extract initial features;

[0048] S2. Input the initial features into the frozen classifier to obtain the initial predicted values;

[0049] S3. Input the initial features into multiple perturbers, each perturber undergoing 20 rounds of perturbation, resulting in multiple perturbed features. Each perturber is an MLP layer, and each perturber is assigned a forced label. The loss functions are cross-entropy loss and L2 loss. The number of perturbers is equal to the number of classes in the classification task, and the forced label is the label for all classes.

[0050] S4. Input the perturbed features into the frozen classifier to obtain the perturbed predicted values;

[0051] S5. Any sample whose predicted value after perturbation is not equal to the initial predicted value but is equal to the forced label will be added to the candidate set;

[0052] Diversity mapping and clustering module:

[0053] S11. For samples in the candidate set, calculate the diversity vector for each sample. The result of each perturber is one element of the vector. For each element: if the initial predicted value is the same as the forced label, the element value is set to 0; otherwise, if the predicted value after perturbation is different from the forced label, the element value is set to -1; if the predicted value after perturbation is the same as the forced label, the element value is set to the Euclidean distance between the initial feature and the perturbated feature.

[0054] S22. Perform K-means clustering on the diversity vectors and select the data points closest to each cluster center with the same number of cluster centers.

[0055] As described above, the present invention can be implemented well.

[0056] The implementation of the present invention is not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A perturbation-based active learning method for medical image classification, characterized in that... Includes the following steps: S1, Characteristic perturbation step: S1-1, Input the unlabeled samples into the frozen encoder to extract initial features; S1-2, Input the initial features into the frozen classifier to obtain the initial predicted values; S1-3: Input the initial features into multiple perturbers, each perturber performs multiple rounds of perturbation, resulting in multiple perturbed features; assign a forced label to each perturber. S1-4, input the perturbed features into the frozen classifier to obtain the perturbed predicted value; S1-5, all samples whose predicted values ​​after perturbation are not equal to the initial predicted values ​​but are equal to the forced labels will be added to the candidate set; S2, diversity mapping and clustering steps: S2-1, For samples in the candidate set, calculate the diversity vector for each sample. The result of each perturber is one element of the vector. If the initial predicted value of each element is the same as the forced label, the element value is set to 0. Otherwise, if the predicted value after perturbation is different from the forced label, the element value is set to -1. If the predicted value after perturbation is the same as the forced label, the element value is set to the Euclidean distance between the initial feature and the perturbation feature. S2-2, perform K-means clustering on the diversity vectors, and select the data points closest to each cluster center with the same number of cluster centers.

2. The perturbation-based active learning medical image classification method according to claim 1, characterized in that: The perturbator described in S1-3 is an MLP layer, and the loss functions are the cross-entropy loss function and the L2 loss function.

3. The perturbation-based active learning medical image classification method according to claim 2, characterized in that: The number of perturbers is equal to the number of categories in the classification task, and the forced label is the label for all categories.

4. A system applied to the perturbation-based active learning medical image classification method according to any one of claims 1-3, characterized in that... include: The uncertainty module of the feature perturbation is used to guide the limited perturbation using forced labels; Samples that are successfully perturbed to the forced label feature space will be selected as candidate samples for further sampling; samples that are not perturbed to the corresponding feature space will be ignored. The diversity mapping and clustering module uses the direction and degree of perturbation to construct a diversity vector and cluster it to select diverse samples from candidate samples; By repeating the sampling process of the above modules and iteratively using the selected samples for deep learning, a system for medical image classification tasks can be obtained with limited annotation costs.

Images