Clustering-enhanced active learning method to improve the efficiency of brain tumor classification and annotation

Abstract

This invention provides a clustering-enhanced active learning method to improve the efficiency of brain tumor classification and labeling. It involves acquiring multi-class brain tumor MRI image samples to obtain an unlabeled dataset of brain tumor MRI images; obtaining preprocessed MRI images; using a rotation prediction self-supervised learning module to extract features from the unlabeled brain tumor images to obtain feature vectors; constructing a clustering-enhanced multi-class data pool; selecting representative samples using a dynamic sampling strategy; adding the labeled representative samples to the labeled sample set; obtaining a trained brain tumor classification model; using the trained brain tumor classification model to predict the unlabeled sample pool; and obtaining the final brain tumor classification model. This method can achieve high-efficiency, high-precision, and low-cost automatic brain tumor classification, effectively improving the model's discrimination performance and stability in complex brain imaging scenarios.

Description

Technical Field

[0001] This invention relates to a clustering-enhanced active learning method to improve the efficiency of brain tumor classification and labeling, belonging to the field of brain tumor classification technology. Background Technology

[0002] Brain tumors are among the most serious central nervous system diseases threatening human life and health. Clinically, brain tumors present complex pathological types, commonly including astrocytomas, oligodendrogliomas, and glioblastomas multiforme, among others. Their morphology, location, and growth rate vary significantly, posing considerable challenges to diagnosis and treatment. Accurate identification and classification of brain tumors are not only crucial for developing individualized treatment plans but also a key step in assessing patient prognosis.

[0003] With the development of deep learning technology, significant progress has been made in brain tumor classification methods based on convolutional neural networks. Typical models such as ResNet, DenseNet, and EfficientNet have demonstrated high classification accuracy on publicly available MRI datasets and have provided strong support for intelligent clinical image analysis. However, these supervised learning methods still have many limitations in practical applications:

[0004] (1) High annotation cost and insufficient samples: Medical image annotation relies on professional radiologists, which is time-consuming and expensive, resulting in a limited number of training samples for the model, especially for some rare tumor types, which leads to serious class imbalance problems.

[0005] (2) Weak cross-domain generalization ability: Due to different imaging equipment, scanning parameters and individual patient differences, the model’s performance drops significantly across different datasets and lacks adaptability to heterogeneous data.

[0006] (3) Fixed sampling mechanism lacks flexibility: Traditional active learning frameworks usually select a fixed number of samples for labeling in each iteration, ignoring the changes in model learning needs at different stages, resulting in the rich information samples not being fully utilized.

[0007] In summary, existing brain tumor classification methods still have significant shortcomings in areas such as annotation cost control, handling inter-class imbalance, and cross-domain generalization ability. How to achieve full utilization of feature representation and dynamic balance between sample sampling through self-supervised feature learning and active sample selection mechanisms under limited annotation resources has become a key problem to be solved in the field of intelligent brain tumor classification. Summary of the Invention

[0008] The purpose of this invention is to provide a clustering-enhanced active learning method to improve the efficiency of brain tumor classification and annotation, thereby addressing the problems of insufficient labeled data, class imbalance, and low annotation efficiency in existing technologies.

[0009] The technical solution of this invention is:

[0010] A clustering-enhanced active learning method to improve the efficiency of brain tumor classification and annotation includes the following steps:

[0011] S1. Obtain multi-category brain tumor MRI image samples to obtain an unlabeled dataset of brain tumor MRI images;

[0012] S2. Preprocessing of brain tumor MRI image samples includes image size unification, pixel normalization, and data augmentation operations to obtain preprocessed MRI images;

[0013] S3. Extract feature vectors from the preprocessed MRI images;

[0014] S4. Perform K-Means clustering on the feature vectors obtained in step S3, and construct a multi-class unlabeled sample pool with enhanced clustering.

[0015] S5. For the cluster-enhanced multi-class unlabeled sample pool, a dynamic sampling strategy is used to select representative samples;

[0016] S6. Label the representative samples to obtain labeled representative samples, and add the labeled representative samples to the labeled sample set;

[0017] S7. After training the brain tumor classification model using the labeled sample set, the trained brain tumor classification model is obtained.

[0018] S8. Use the trained brain tumor classification model to predict the multi-class unlabeled sample pool;

[0019] S9. Repeat steps S5 to S8 until the active learning round ends or the model performance converges to obtain the final brain tumor classification model.

[0020] S10. Input the MRI image of the brain tumor to be classified into the final brain tumor classification model to obtain the classification result.

[0021] Furthermore, in step S2, the data augmentation operation specifically includes:

[0022] S21. Enhanced operations including random rotation, horizontal flip, vertical flip, brightness adjustment, contrast perturbation, and Gaussian noise injection are employed to simulate different scanning directions, intensity changes, and acquisition noise.

[0023] S22. Use the pseudo-color mapping method to convert grayscale MRI images into three-channel RGB images to adapt to the input format of the brain tumor classification model;

[0024] S23. Adopt an adaptive scaling strategy to maintain the aspect ratio of the image, scale the shortest side to 224 pixels proportionally, and fill in the edges of images that are less than 224×224 pixels, and unify the input size to 224×224 pixels.

[0025] Furthermore, in step S3, the preprocessed MRI image is subjected to a rotation prediction self-supervised learning module to extract features and obtain feature vectors.

[0026] Further, in step S4, K-Means clustering is performed on the feature vectors obtained in step S3, and a cluster-enhanced multi-class unlabeled sample pool is constructed, specifically as follows:

[0027] S41. Use the K-Means clustering algorithm to perform cluster analysis on the feature vectors obtained in step S3, and divide the entire unlabeled dataset of brain tumor MRI images into several category clusters. Each category cluster corresponds to a category center that describes the average semantic features of the brain tumor MRI image samples of that category cluster.

[0028] S42. Based on the distance between each brain tumor MRI image sample and the category center of its respective category cluster, the brain tumor MRI image samples are assigned to the corresponding category clusters to form an initial sample pool;

[0029] S43. To characterize the uncertainty of brain tumor MRI image samples, calculate the distance difference between the brain tumor MRI image sample and the category center of its class cluster and the neighboring category centers. The smaller the distance difference of the brain tumor MRI image sample, the higher the degree of uncertainty of the brain tumor MRI image sample.

[0030] S44. For each category cluster, sort the brain tumor MRI image samples according to the degree of uncertainty, and arrange them from high to low uncertainty to obtain the sorted brain tumor MRI image sample pool.

[0031] S45. The sorted brain tumor MRI image sample pool is divided into I batches according to the rounds of active learning to obtain a cluster-enhanced multi-class unlabeled sample pool.

[0032] Furthermore, in step S45, the samples of the i-th batch out of the I batches... :

[0033]

[0034] in, Let be the number of samples in the d-th cluster. For sorted brain tumor MRI image sample pool The kth sample in the dataset.

[0035] Furthermore, in step S5, a dynamic sampling strategy is used to select representative samples from the cluster-enhanced multi-class unlabeled sample pool. Specifically,

[0036] S51. Set a variable number of representative samples K for each active learning cycle;

[0037] S52. In each round of active learning, the dynamic sampling strategy selects samples from all unlabeled sample pools of the current batch, prioritizing the retention of samples with lower confidence levels as representative samples. Specifically, the sampler... From various cluster-enhanced multi-class unlabeled sample pools Select the one with the lowest confidence level. A sample is used as a representative sample, where D is the total number of unlabeled multi-class samples in the clustering enhancement of step S4.

[0038] Furthermore, in step S51, in the i-th iteration, the number of representative samples, K, is: , in, This indicates the labeling ratio, where N is the number of unlabeled samples in the multi-class pool for cluster enhancement. is the sampling factor.

[0039] Furthermore, in step S7, the brain tumor classification model is trained using a residual neural network ResNet and a cross-entropy loss function with an Adam optimizer.

[0040] The beneficial effects of this invention are:

[0041] I. This clustering-enhanced active learning method improves the efficiency of brain tumor classification and annotation, enabling high-efficiency, high-precision, and low-cost automatic brain tumor classification. Based on a self-supervised feature learning and active learning framework, this invention introduces a clustering-enhanced multi-class unlabeled sample pool construction mechanism and a dynamic sampling strategy. While maintaining sample diversity, it dynamically adjusts the sampling ratio of uncertain samples, effectively improving the model's discrimination performance and stability in complex brain imaging scenarios.

[0042] II. This clustering-enhanced active learning method for improving the efficiency of brain tumor classification and annotation reduces reliance on manual labels through self-supervised feature learning, enhances sample diversity coverage through clustering enhancement mechanisms, and balances uncertainty and representativeness selection through dynamic sampling strategies. This effectively alleviates class imbalance and improves the overall robustness and cross-dataset generalization of the model. This method can achieve efficient and stable brain tumor classification modeling with limited annotation costs, demonstrating strong practicality and scalability. It is suitable for intelligent auxiliary diagnostic systems for brain tumors in clinical scenarios and can also be extended to stroke detection, neurodegenerative disease identification, and other medical image analysis tasks. Attached Figure Description

[0043] Figure 1 This is a flowchart illustrating the clustering-enhanced active learning method for improving the efficiency of brain tumor classification and labeling according to an embodiment of the present invention.

[0044] Figure 2 This is a schematic diagram illustrating the active learning framework in the embodiment;

[0045] Figure 3 This is an illustrative diagram illustrating the construction of a cluster-enhanced multi-class unlabeled sample pool in the embodiment;

[0046] Figure 4 This is a schematic diagram illustrating the dynamic sampling strategy in the embodiment. Detailed Implementation

[0047] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0048] An example of a clustering-enhanced active learning method to improve the efficiency of brain tumor classification and labeling, such as... Figure 1 and Figure 2 This includes the following steps:

[0049] S1. Obtain MRI image samples of multiple categories of brain tumors to obtain an unlabeled dataset of brain tumor MRI images.

[0050] S2. Preprocessing of brain tumor MRI image samples includes image size unification, pixel normalization, and data augmentation operations to obtain preprocessed MRI images.

[0051] In step S2, the data augmentation operation specifically includes:

[0052] S21. Enhanced operations including random rotation, horizontal flip, vertical flip, brightness adjustment, contrast perturbation, and Gaussian noise injection are employed to simulate different scanning directions, intensity changes, and acquisition noise.

[0053] S22. Use the pseudo-color mapping method to convert grayscale MRI images into three-channel RGB images to adapt to the input format of the brain tumor classification model;

[0054] S23. Adopt an adaptive scaling strategy to maintain the aspect ratio of the image, scale the shortest side to 224 pixels proportionally, and fill in the edges of images that are less than 224×224 pixels, and unify the input size to 224×224 pixels.

[0055] In step S2, to improve the model's robustness to brain tumors of different morphologies, a series of data augmentation and standardization preprocessing operations were performed on the original MRI images. Augmentation methods included random rotation, mirror flipping, center cropping, and brightness perturbation. During image preprocessing, an adaptive scaling method was used to proportionally scale the images to a shorter side of 224 pixels, and edge padding was used to uniformly adjust all images to a 224×224 pixel input size. Furthermore, all pixel intensities were normalized to the [0,1] interval. Finally, the three datasets were divided into training and testing sets in an 8:2 ratio for model training, tuning, and final evaluation.

[0056] S3. Extract feature vectors from unlabeled brain tumor images.

[0057] In step S3, a rotation prediction self-supervised learning module is used to extract feature vectors from unlabeled brain tumor images. This module rotates the preprocessed MRI image by 0°, 90°, 180°, and 270° to generate four viewpoint samples, which are then input into a convolutional neural network (CNN) for encoding, extracting spatial and textural features. The module predicts the rotation angle category using a rotation angle classification head and optimizes the prediction using a cross-entropy loss function to minimize the prediction error. As training progresses, the model gradually learns rotation-invariant high-level semantic features, enabling it to capture the spatial structure and boundary features of brain tumor images under unlabeled conditions. The rotation-invariant features learned through self-supervised learning capture the shape and boundary information of the tumor region, providing a stable and reliable feature representation for subsequent clustering and active learning modules.

[0058] S4. Perform K-Means clustering on the feature vectors obtained in step S3, and construct a cluster-enhanced multi-class unlabeled sample pool.

[0059] S41. Use the K-Means clustering algorithm to perform cluster analysis on the feature vectors obtained in step S3, and divide the entire unlabeled dataset of brain tumor MRI images into several category clusters. Each category cluster corresponds to a category center that describes the average semantic features of the brain tumor MRI image samples of that category cluster.

[0060] S42. Based on the distance between each brain tumor MRI image sample and the category center of its respective category cluster, the brain tumor MRI image samples are assigned to the corresponding category clusters to form an initial sample pool;

[0061] S43. To characterize the uncertainty of brain tumor MRI image samples, calculate the distance difference between the brain tumor MRI image sample and the category center of its class cluster and the neighboring category centers. The smaller the distance difference of the brain tumor MRI image sample, the higher the degree of uncertainty of the brain tumor MRI image sample.

[0062] S44. For each category cluster, sort the brain tumor MRI image samples according to the degree of uncertainty, and arrange them from high to low uncertainty to obtain the sorted brain tumor MRI image sample pool.

[0063] S45. The sorted brain tumor MRI image sample pool is divided into I batches according to the active learning rounds, resulting in a cluster-enhanced multi-class unlabeled sample pool, such as... Figure 3 .

[0064] Within each category cluster, the samples are sorted in ascending order of distance difference to obtain the sorted brain tumor MRI image sample pool. Subsequently, each sorted pool of brain tumor MRI images is divided into I batches, corresponding to the rounds of active learning. Taking I=10 as an example, the sorted pool of brain tumor MRI images is uniformly divided into 10 batches, and the samples in the i-th batch of each I batch... :

[0065]

[0066] in, Let be the number of samples in the d-th cluster. For sorted brain tumor MRI image sample pool The kth sample in the dataset.

[0067] The clustering-enhanced multi-class unlabeled sample pool constructed through steps S41-S45 above can simultaneously maintain the distribution diversity between different classes and the uncertain structure within each class, providing a high-quality sample candidate set for subsequent dynamic sampling and labeling. Its effectiveness was verified in ablation experiments, and the experimental results are shown in Table 1.

[0068] Table 1. Comparison of classification accuracy (%) of different methods under different annotation ratios.

[0069]

[0070] S5. For the cluster-enhanced multi-class unlabeled sample pool, a dynamic sampling strategy is used to select representative samples.

[0071] S51. Set a variable number of representative samples K for each active learning cycle;

[0072] In step S51, the number of representative samples K in the i-th iteration is:

[0073] , in, This indicates the labeling ratio, where N is the number of unlabeled samples in the multi-class pool for cluster enhancement. is the sampling factor.

[0074] In step S51, to address the issue of high-value sample omissions caused by fixed sampling size in traditional active learning, this invention proposes a dynamic sampling mechanism that dynamically adjusts the sampling quantity K based on the training phase. The sampling quantity K of representative samples is determined by a sampling factor. The sampling factor is determined by both the sampling rate and the labeling ratio, thus forming a flexible sample selection mechanism. The sampling factor can be adaptively determined by maximizing the performance of the brain tumor classification model. Based on the prediction results of the previous training model, the confidence distribution of unlabeled samples is calculated, and the sampling factor is determined based on the uncertainty of the samples; when the sampling factor... When the sampling factor is positive, high-uncertainty samples are prioritized, thereby enhancing the model's discriminative ability in early iterations; when the sampling factor is positive... When the value is negative, samples with strong representativeness and diverse categories are prioritized to improve the model's generalization ability. This achieves a dynamic balance between the number of samples and the feature distribution during the active learning phase, enabling the model to continuously optimize with limited annotation costs.

[0075] S52. In each round of active learning, the dynamic sampling strategy selects samples from all unlabeled sample pools of the current batch, prioritizing the retention of samples with lower confidence levels as representative samples. Specifically, the sampler... From various cluster-enhanced multi-class unlabeled sample pools Select the one with the lowest confidence level. Each sample forms a new labeled set. : , where D is the total number of unlabeled multi-class sample pools for cluster enhancement in step S4.

[0076] In step S5, this strategy allows the model to automatically adjust the balance of sample selection at different learning stages: in the early stages, it focuses more on representative samples with higher uncertainty to quickly converge the classification boundary; in the later stages, it collects more representative samples with lower uncertainty to improve data distribution coverage. By introducing an adjustable sampling factor and a cross-round sample size adaptive mechanism, this dynamic sampling strategy effectively avoids the information waste problem caused by a fixed sampling scheme, achieving a dynamic balance between uncertainty utilization and data diversity under limited annotation costs. Its effectiveness is verified in ablation experiments, and the experimental results are shown in Table 2.

[0077] Table 2. Comparison of classification accuracy (%) of different methods under different annotation ratios.

[0078]

[0079] S6. Label the representative samples to obtain labeled representative samples, and add the labeled representative samples to the labeled sample set.

[0080] In step S6, the new annotation set from step S52 is... Update the labeled sample set for: ,in, This is the labeled sample set from the previous round.

[0081] S7. After training the brain tumor classification model using the labeled sample set, the trained brain tumor classification model is obtained.

[0082] In step S7, the brain tumor classification model uses a ResNet structure as the feature extraction backbone network, including convolutional layers, batch normalization layers, residual blocks, global average pooling layers, and softmax classification layers. The feature extraction part is used to obtain multi-scale tumor morphological features, and the classification layer outputs tumor category predictions. During the training phase, the cross-entropy loss function and Adam optimizer are used, with the initial learning rate set to 1e-4, and the cosine annealing scheduling function is used to achieve adaptive decay. At the same time, an early stopping mechanism and an L2 regularization term are introduced to prevent overfitting and enhance the model's generalization ability.

[0083] S8. Use the trained brain tumor classification model to predict the multi-class unlabeled sample pool.

[0084] S9. Repeat steps S5 to S8 until the active learning round ends or the model performance converges to obtain the final brain tumor classification model.

[0085] S10. Input the MRI image of the brain tumor to be classified into the final brain tumor classification model to obtain the classification result.

[0086] This clustering-enhanced active learning method improves the efficiency of brain tumor classification and annotation, enabling high-efficiency, high-precision, and low-cost automatic brain tumor classification. Based on a self-supervised feature learning and active learning framework, this invention introduces a clustering-enhanced multi-class unlabeled sample pool construction mechanism and a dynamic sampling strategy. While maintaining sample diversity, it dynamically adjusts the sampling ratio of uncertain samples, effectively improving the model's discriminative performance and stability in complex brain imaging scenarios.

[0087] This clustering-enhanced active learning method, designed to improve the efficiency of brain tumor classification and annotation, reduces reliance on manual labels through self-supervised feature learning, enhances sample diversity coverage through clustering enhancement mechanisms, and balances uncertainty and representativeness selection through dynamic sampling strategies. This effectively alleviates class imbalance and improves the overall robustness and cross-dataset generalization of the model. The method achieves efficient and stable brain tumor classification modeling with limited annotation costs, demonstrating strong practicality and scalability. It is suitable for intelligent auxiliary diagnostic systems for brain tumors in clinical settings and can also be extended to stroke detection, neurodegenerative disease identification, and other medical image analysis tasks.

[0088] This clustering-enhanced active learning method for improving the efficiency of brain tumor classification and annotation firstly introduces a rotation prediction self-supervised feature learning module into the active learning framework to reduce reliance on manually labeled samples and improve the generalization of feature representation. This module performs multi-angle rotation on the input brain tumor MRI image and predicts its rotation angle category, enabling the model to learn the structural semantic features and morphological patterns of the image under unsupervised conditions. The self-supervised features obtained after pre-training are used to construct a high-dimensional semantic space, providing a stable and discriminative representation foundation for subsequent clustering analysis and sample selection. Secondly, to improve the sample utilization efficiency in the active learning stage, this invention designs a mechanism combining a clustering-enhanced multi-class unlabeled sample pool with a dynamic sampling strategy. This mechanism first performs K-Means clustering on the unlabeled samples in the self-supervised feature space, dividing the samples into multiple clusters and assigning a cluster label to each sample. Subsequently, within each cluster, the cluster center distance and model confidence distribution of the samples are calculated, and the active learning candidate set is generated through the joint metric of these two factors to comprehensively reflect the uncertainty and diversity of the samples. In each round of active learning iteration, this invention introduces an adjustable sampling factor, which dynamically adjusts the sampling ratio of samples in different class clusters according to the model confidence level. This allows for the priority selection of high uncertainty samples in the early stage to accelerate convergence, and the addition of representative samples in the later stage to improve model stability and generalization ability.

[0089] This clustering-enhanced active learning method, which improves the efficiency of brain tumor classification and annotation, performs forward prediction on the unlabeled sample pool after each round of active learning, calculates the confidence distribution, and selects the Top-k samples with the lowest confidence to add to the next round of annotation set. This strategy can gradually reduce the uncertainty region of the model's decision boundary and accelerate model convergence. In the early stage when there are few labeled samples, the proportion of Top-k samples is increased to quickly improve the model's classification and discrimination performance, while in the later stage, the proportion of Top-k samples is gradually reduced to stabilize the model's performance.

[0090] This clustering-enhanced active learning method, which improves the efficiency of brain tumor classification and annotation, enhances the diversity of sample selection through a clustering enhancement mechanism and increases the probability of minority class samples appearing in the sampling pool through a dynamic sampling strategy to alleviate class imbalance. At the same time, self-supervised pre-training reduces the dependence on manual annotation, enabling the model to learn discriminative features even in small sample scenarios. By jointly measuring feature space clustering and sample uncertainty, it effectively avoids the cumulative bias of high-confidence misclassified samples and improves the overall data utilization rate.

[0091] This invention can significantly reduce the cost of manual annotation, enhance the robustness of the model to inter-class differences and noise interference, is suitable for intelligent auxiliary diagnostic systems for clinical brain tumors, and can be extended to other medical image classification tasks.

[0092] The clustering-enhanced active learning method for improving the efficiency of brain tumor classification and labeling was experimentally validated as follows:

[0093] Two publicly available brain tumor MRI image datasets were obtained: the DS1 dataset and the DS2 dataset. The DS1 dataset consists of three subsets: Figshare, SARTAJ, and Br35H, containing a total of 7,023 MRI images. The image categories include glioma, meningioma, pituitary adenoma, and no tumor. The training set contains 5,712 images, and the test set contains 1,311 images, showing a relatively balanced class distribution. The DS2 dataset, publicly released by Cheng et al., contains 3,064 MRI images from 233 patients. The images are categorized into glioma, meningioma, and pituitary adenoma, exhibiting class imbalance. The number of glioma samples is approximately twice that of the other two categories. This dataset is split with 80% training on Figshare and 20% testing on Figshare.

[0094] The performance of the trained models was tested on the test set, using labeled data trained at 20%, 30%, 40%, 50%, and 60% of the data, respectively. The performance of the comparison method and the example method is summarized in Table 3.

[0095] Table 3. Accuracy (%) of the comparative method and the example method at different labeling ratios

[0096] In summary, the results in Table 3 demonstrate that the accuracy of all methods continuously improves with increasing labeled data ratios. Compared to recent state-of-the-art models such as ARM-Net, MedViTV2, MedMamba, and MobileNetV4, the proposed method in this embodiment achieves higher accuracy across all label ratios (LRs) on the DS1 and DS2 datasets. On DS1, the method in this embodiment achieves 96.3% accuracy using only 20% labeled data, outperforming ARM-Net (95.2%) and MedViTV2 (94.1%). When the label ratio increases to 60%, the method in this embodiment reaches 99.1%, maintaining its consistent advantage.

[0097] Compared to semi-supervised learning (SSL) methods such as MixMatch and FixMatch, the Implementation Method achieves more substantial improvements on various LRs. For example, on DS1, with an LR of 20%, the Implementation Method outperforms MixMatch and FixMatch by 3.0% and 2.6%, respectively.

[0098] Compared to other active learning (AL) frameworks such as PT4AL, BAL, PAL, and VAAL, the proposed method achieves the highest accuracy across all label ratios, with particularly significant gains on lower logistic regressions (LRs). On DS1 with a 20% LR, the implementation method outperforms PAL and BAL by 2.9% and 1.6%, respectively. On the smaller, more imbalanced DS2, the implementation method maintains good performance. At a 20% LR, the implementation method outperforms BAL by 1.4% and PAL by 2.2%. At a 30% LR, these improvements increase to 1.3% and 4.4%, respectively. These results demonstrate that the implementation method effectively addresses data imbalance and training bias, thereby achieving more robust model learning with a limited annotation budget.

[0099] During the validation phase, to evaluate the model's generalization ability, cross-dataset testing was employed. Cross-validation was performed on brain tumor MRI datasets from different sources and with different magnetic resonance sequences (such as T1, T2, and FLAIR). Classification accuracy, macro-average F1 score, and AUC were statistically analyzed to quantify the model's stability on heterogeneous datasets. The results show that the implementation method maintains stable performance across devices and centers, achieving high robustness and generalizability in clinical settings. Furthermore, the framework can be extended to stroke detection, neurodegenerative disease identification, and multimodal image fusion tasks. By combining transfer learning and active sampling strategies, high-precision medical diagnosis can be achieved with low annotation costs.

[0100] The above are merely embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A clustering-enhanced active learning method to improve the efficiency of brain tumor classification and labeling, characterized in that: Includes the following steps, S1. Obtain multi-category brain tumor MRI image samples to obtain an unlabeled dataset of brain tumor MRI images; S2. Preprocessing of brain tumor MRI image samples includes image size unification, pixel normalization, and data augmentation operations to obtain preprocessed MRI images; S3. Extract feature vectors from the preprocessed MRI images; S4. Perform K-Means clustering on the feature vectors obtained in step S3, and construct a multi-class unlabeled sample pool with enhanced clustering. S5. For the cluster-enhanced multi-class unlabeled sample pool, a dynamic sampling strategy is used to select representative samples; S51. Set a variable number of representative samples K for each active learning cycle; In step S51, the number of representative samples K in the i-th iteration is: in, This indicates the labeling ratio, where N is the number of unlabeled samples in the multi-class pool for cluster enhancement. S52. In each round of active learning, the dynamic sampling strategy selects samples from all unlabeled sample pools of the current batch, prioritizing the retention of samples with lower confidence as representative samples. Specifically, the sampler... From various cluster-enhanced multi-class unlabeled sample pools Select the one with the lowest confidence level. 1 sample is used as a representative sample, where D is the total number of unlabeled multi-class samples in the clustering enhancement of step S4; S6. Label the representative samples to obtain labeled representative samples, and add the labeled representative samples to the labeled sample set; S7. After training the brain tumor classification model using the labeled sample set, the trained brain tumor classification model is obtained. S8. Use the trained brain tumor classification model to predict the multi-class unlabeled sample pool; S9. Repeat steps S5 to S8 until the active learning round ends or the model performance converges to obtain the final brain tumor classification model. S10. Input the MRI image of the brain tumor to be classified into the final brain tumor classification model to obtain the classification result.

2. The clustering-enhanced active learning method for improving the efficiency of brain tumor classification and labeling as described in claim 1, characterized in that: In step S2, the data augmentation operation specifically includes: S21. Enhanced operations including random rotation, horizontal flip, vertical flip, brightness adjustment, contrast perturbation, and Gaussian noise injection are employed to simulate different scanning directions, intensity changes, and acquisition noise. S22. Use the pseudo-color mapping method to convert grayscale MRI images into three-channel RGB images to adapt to the input format of the brain tumor classification model; S23. Adopt an adaptive scaling strategy to maintain the aspect ratio of the image, scale the shortest side to 224 pixels proportionally, and fill in the edges of images that are less than 224×224 pixels, and unify the input size to 224×224 pixels.

3. The clustering-enhanced active learning method for improving the efficiency of brain tumor classification and labeling as described in claim 1, characterized in that: In step S3, the preprocessed MRI image is subjected to a rotation prediction self-supervised learning module to extract features and obtain feature vectors.

4. The clustering-enhanced active learning method for improving the efficiency of brain tumor classification and labeling as described in any one of claims 1-3, characterized in that: In step S4, K-Means clustering is performed on the feature vectors obtained in step S3, and a cluster-enhanced multi-class unlabeled sample pool is constructed. Specifically, S41. Use the K-Means clustering algorithm to perform cluster analysis on the feature vectors obtained in step S3, and divide the entire unlabeled dataset of brain tumor MRI images into several category clusters. Each category cluster corresponds to a category center that describes the average semantic features of the brain tumor MRI image samples of that category cluster. S42. Based on the distance between each brain tumor MRI image sample and the category center of its respective category cluster, the brain tumor MRI image samples are assigned to the corresponding category clusters to form an initial sample pool; S43. To characterize the uncertainty of brain tumor MRI image samples, calculate the distance difference between the brain tumor MRI image sample and the category center of its class cluster and the neighboring category centers. The smaller the distance difference of the brain tumor MRI image sample, the higher the degree of uncertainty of the brain tumor MRI image sample. S44. For each category cluster, sort the brain tumor MRI image samples according to the degree of uncertainty, and arrange them from high to low uncertainty to obtain the sorted brain tumor MRI image sample pool. S45. The sorted brain tumor MRI image sample pool is divided into I batches according to the rounds of active learning to obtain a cluster-enhanced multi-class unlabeled sample pool.

5. The clustering-enhanced active learning method for improving the efficiency of brain tumor classification and labeling as described in claim 4, characterized in that: In step S45, the samples of the i-th batch out of the I batches : in, For sorted brain tumor MRI image sample pool The kth sample in the dataset.

6. The clustering-enhanced active learning method for improving the efficiency of brain tumor classification and labeling as described in any one of claims 1-3, characterized in that: In step S7, the brain tumor classification model uses a residual neural network ResNet and is trained with a cross-entropy loss function and an Adam optimizer.

Images