A Two-Stage Active Learning Method for Fine-Grained Bronchoalveolar Segmentation

Abstract

This invention relates to a two-stage active learning method for fine-grained bronchial airway segmentation. The method includes the following steps: S1, generating embedding vectors for an unlabeled medical image dataset based on a pre-trained adaptive bronchial tubule segmentation model; S2, selecting the most diverse candidate samples and adding them to the final sample set; S3, determining whether the number of samples in the final sample set has reached a threshold. If so, proceed to S4; otherwise, continue selecting the most diverse candidate samples to add to the final sample set; S4, labeling the final sample set, and using the labeled sample set as the training set for the pre-trained adaptive bronchial tubule segmentation model to obtain an optimized adaptive bronchial tubule segmentation model. Based on the optimized adaptive bronchial tubule segmentation model, a complex bronchial network segmentation result is obtained. Compared with existing technologies, this invention has advantages such as improving the value density of the selected sample set and increasing the accuracy of fine bronchial identification.

Description

Technical Field

[0001] This invention relates to the field of bronchial airway segmentation technology, and in particular to a two-stage active learning method for fine-grained bronchial airway segmentation. Background Technology

[0002] Current mainstream research on bronchial airway segmentation focuses on deep learning technology, with the 3D-UNet architecture widely adopted as a core model. Through carefully designed encoder-decoder layer stacking, it achieves precise differentiation of bronchial structures. For example, the CFDA method uniquely introduces a collaborative feature de-entanglement and enhancement framework, cleverly utilizing clean and noisy CT scan data. By encoding the topological information of clean scans and enhancing the feature representation in noisy data, it significantly improves the model's generalization performance and segmentation accuracy in bronchial segmentation tasks. Addressing the significant sample imbalance problem between high-level and low-level branches in bronchial structures, the NaviAirway project adopted an innovative strategy. It employed a customized loss function particularly sensitive to fine bronchi and combined it with a cross-level fine-tuning learning model design. Through iterative training, it effectively overcame the challenge of sample imbalance and significantly improved segmentation quality. Furthermore, Yang Nan et al. contributed by proposing an advanced framework combining fuzzy attention neural networks and composite loss functions. This framework cleverly solves the heterogeneity problem between different feature channels by utilizing a channel-specific fuzzy attention mechanism. At the same time, by designing a well-thought-out comprehensive loss function, it enhances the spatial coherence of the segmentation output, thereby driving a leap in the overall performance of airway segmentation.

[0003] However, fully supervised deep learning requires a large amount of labeled datasets. On the one hand, publicly available medical image datasets are scarce, making image collection difficult. On the other hand, medical images are complex, making annotation challenging. Many datasets only have a portion of the images annotated by professional physicians. These two issues with medical images present a dual challenge for the training and deployment of deep learning models, especially supervised models, in terms of both data acquisition and annotation. Active learning methods, however, can evaluate and quantify the informativeness of unlabeled images, identifying the most valuable unlabeled samples in the dataset. By introducing active learning methods into medical image models, high-performance medical image analysis can be achieved with less annotation.

[0004] Active learning has been widely applied in many medical tasks, but its training strategies and methods for bronchoalveolar segmentation have not been fully explored. Researching active learning methods can further reduce the annotation costs of intrapulmonary CT images and unlock the performance potential of existing supervised models.

[0005] Existing active learning sampling methods can be mainly divided into three categories: uncertainty-based methods, representativeness-based methods, and a fusion strategy of the two.

[0006] The core of uncertainty-based active learning lies in identifying learning "blind spots" by quantifying the model's impact on individual samples or on the model's prediction confidence. This strategy can be further subdivided into instance uncertainty and model uncertainty. Instance uncertainty focuses on the predicted probability distribution of individual samples, using metrics such as entropy or minimum probability to identify instances that the model struggles to classify precisely. Model uncertainty goes a step further, assessing the uncertainty of model parameters through methods like Bayesian active learning, reflecting the model's adaptability to new data. These methods help to directly address and strengthen weaknesses in the model's understanding.

[0007] Representative sampling strategies focus on maintaining or increasing the diversity and coverage of the dataset. Representative methods ensure that samples are collected from all corners of the data distribution, especially those instances that effectively fill gaps in the feature space, by using techniques such as measuring distance between samples, cluster distribution, or core sample selection. This not only facilitates the understanding of complex data structures but also improves the model's ability to generalize to unseen data. Representative sampling is sometimes combined with cluster analysis to ensure that samples are selected from each cluster, achieving balanced coverage of the data space.

[0008] Currently, the bronchial dataset suffers from two imbalance problems: intra-class imbalance and inter-class imbalance. Intra-class imbalance refers to the situation where, in lung CT images, the proportion of airway pixels is significantly smaller than the proportion of background pixels. Inter-class imbalance refers to the difference between large and small bronchioles; bronchioles near the ends of pulmonary vesicles may occupy only a few pixels, resulting in a smaller sample proportion compared to large bronchioles.

[0009] Current active learning techniques, considering only the representational characteristics of a single sample, suffer from two main problems when dealing with diverse bronchial structures: 1) It is difficult to construct discriminative representations of fine bronchial samples, leading to confusion between fine bronchial samples and negative samples; 2) Data selection methods often favor coarse bronchial samples with obvious visual features, resulting in a significant difference in sample size between coarse and fine bronchial samples after sampling. These shortcomings significantly affect the performance of actively learned segmentation models in complex bronchial network segmentation tasks, frequently leading to the omission of fine bronchials and reducing overall segmentation accuracy and completeness. Summary of the Invention

[0010] The purpose of this invention is to provide a two-stage active learning method for fine-grained bronchial airway segmentation in order to improve the value density of the selected sample set and enhance the accuracy of fine bronchial identification.

[0011] The objective of this invention can be achieved through the following technical solutions:

[0012] A two-stage active learning method for fine-grained bronchial airway segmentation, comprising the following steps:

[0013] S1. An embedded vector of an unlabeled medical image dataset is generated based on a pre-trained adaptive bronchial tubular segmentation model.

[0014] S2. Cluster the embedded vectors, calculate the uncertainty of the samples in the clustered dataset, sort them based on the uncertainty, select candidate samples, calculate the diversity coefficient of the candidate samples, select the candidate sample with the largest diversity coefficient to add to the current sample set, and obtain the final sample set.

[0015] S3. Determine whether the number of samples in the final sample set has reached the threshold. If yes, execute S4; otherwise, continue to select the largest candidate sample to add to the final sample set.

[0016] S4. The final sample set is labeled, and the labeled sample set is used as the training set for the pre-trained adaptive bronchial tubular segmentation model to obtain an optimized adaptive bronchial tubular segmentation model. Based on the optimized adaptive bronchial tubular segmentation model, the segmentation result of the complex bronchial network is obtained.

[0017] Furthermore, the specific steps of S1 are as follows:

[0018] Obtain an unlabeled medical image dataset, input the unlabeled dataset into a pre-trained adaptive bronchial tubular segmentation model for feature extraction, and obtain the embedding vector.

[0019] Furthermore, the pre-trained adaptive bronchial tubular segmentation model includes n interconnected network layers. Each network layer includes a first ordinary convolutional layer, a linear deformable convolutional layer, a channel attention module, and a second ordinary convolutional layer connected in sequence, with skip connections between the first ordinary convolutional layer and the second ordinary convolutional layer.

[0020] Furthermore, the n network layers are deployed with convolutional kernels of various sizes, and the nth network layer outputs an embedding vector.

[0021] Furthermore, in the n network layers, the i-th and ni-th network layers are skip-connected and have the same convolutional kernel size, while the non-skipping network layers have different convolutional kernel sizes.

[0022] Furthermore, the uncertainty is:

[0023]

[0024] Here, uncertainty represents the uncertainty, di and dj represent the distances between the sample and two different cluster centers, x is the embedding vector, ci represents the cluster center, and uncertainty is the minimum difference between any two distances. That is, by iterating through all possible distances, for di and dj, the absolute value of their difference is calculated, and the minimum value is taken as uncertainty.

[0025] Furthermore, the specific steps for calculating the diversity of candidate samples and selecting the candidate sample with the highest diversity to add to the current sample set are as follows:

[0026] The similarity between a candidate sample and all samples in the current sample set is calculated. Based on the similarity, a diversity coefficient ratio is calculated. The size of the diversity coefficient ratio is positively correlated with the difference between the samples. That is, the larger the ratio, the greater the difference between the sample and the current sample set. Then, the sample with the highest diversity is selected from the candidate samples and added to the current sample set. After that, the above steps are repeated to continuously compare the diversity coefficients of the samples, select and add samples to the current sample set until the number of samples in the current sample set reaches a preset threshold. Finally, a final sample set that meets the requirements is obtained.

[0027] Furthermore, the similarity between a candidate sample and all samples within the candidate sample set is:

[0028]

[0029] Where S represents the current sample set, x i Let S represent a sample in the current sample set, x represent a candidate sample, n represent the total number of candidate samples, and sim(x, S) represent the similarity.

[0030] Furthermore, the diversity coefficient is:

[0031]

[0032] Where n represents the total number of candidate samples.

[0033] Furthermore, the specific steps of S1 are as follows:

[0034] Medical images are acquired, and image enhancement transformations are performed on the medical images to obtain an unlabeled medical image dataset. An embedding vector of the unlabeled medical image dataset is generated based on a pre-trained adaptive bronchial tubular segmentation model.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] This invention achieves sample selection for both fine and coarse bronchioles by simultaneously evaluating sample uncertainty, representativeness, and diversity, and applies this selection to the active learning training of a segmentation model. This method overcomes the low information gain bottleneck that traditional active learning strategies based on intuitive image features may face. By calculating uncertainty and diversity, a two-stage screening process for fine segmentation is implemented, which not only enhances the ability to distinguish the importance of samples but also significantly reduces the inclusion of redundant samples, effectively increasing the value density of the selected sample set and providing strong support for in-depth mining of key information in bronchial data sets.

[0037] Furthermore, the adaptive bronchial tubular structure segmentation network of this invention adopts a comprehensive multi-scale feature extraction and fusion strategy, which fuses features obtained at different scales. The network effectively integrates the advantages of various convolution operations, integrates cross-scale feature information into one place, and forms a hierarchical and information-rich feature representation system. This enhances the grasp of the complexity of bronchial structure, promotes the model's simultaneous sensitivity to subtle and macroscopic features, and improves segmentation accuracy and robustness. Attached Figure Description

[0038] Figure 1 This is a flowchart of the two-stage active learning process of the present invention;

[0039] Figure 2 An adaptive bronchial tubular segmentation network;

[0040] Figure 3 This is a flowchart of the first phase of proactive learning and screening.

[0041] Figure 4 The flowchart for the second phase of active learning screening;

[0042] Figure 5 Flowchart for two-stage active learning sample selection;

[0043] Figure 6 System architecture diagram;

[0044] Figure 7 This is a two-stage screening module;

[0045] Figure 8 This is a bronchial airway segmentation module;

[0046] Figure 9 A flowchart illustrating the reasoning process of a multi-level airway and learning module;

[0047] Figure 10 The flowchart shows the training process for the system segmentation model. Detailed Implementation

[0048] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0049] This invention proposes a two-stage active learning method for fine-grained bronchial airway segmentation. By simultaneously evaluating sample uncertainty, representativeness, and diversity, it achieves sample selection for multiple fine and coarse bronchi and applies this selection to the active learning training of the segmentation model. This method overcomes the low information gain bottleneck that traditional active learning strategies based on intuitive image features may face. Through a finely divided two-stage selection process, it not only enhances the ability to discern the importance of samples but also significantly reduces the inclusion of redundant samples, effectively increasing the value density of the selected sample set and providing strong support for in-depth mining of key information in bronchial data sets. The steps of this method include:

[0050] S1. An embedded vector of an unlabeled medical image dataset is generated based on a pre-trained adaptive bronchial tubular segmentation model.

[0051] S2. Cluster the embedded vectors, calculate the uncertainty of the samples in the clustered dataset, sort them based on the uncertainty, select candidate samples, calculate the diversity of the candidate samples, and select the candidate sample with the greatest diversity to add to the final sample set.

[0052] S3. Determine whether the number of samples in the final sample set has reached the threshold. If so, execute S2; otherwise, continue to select the largest candidate sample to add to the final sample set.

[0053] S4. The final sample set is labeled, and the labeled sample set is used as the training set for the pre-trained adaptive bronchial tubular segmentation model to obtain an optimized adaptive bronchial tubular segmentation model. Based on the optimized adaptive bronchial tubular segmentation model, the segmentation result of the complex bronchial network is obtained.

[0054] Subsequently, combining an interactive system, front-end and back-end communication, and an adaptive bronchial tubular structure segmentation network, an active learning-based segmentation system is proposed. This system evaluates the uncertainty, representativeness, and diversity of collected samples, selects key data for annotation and segmentation model training, and iteratively optimizes the segmentation effect through multiple rounds. The system reduces the data scale required for bronchial segmentation model training, lowers annotation costs, and enables real-time discovery and reuse of key data during diagnosis and application.

[0055] The uncertainty is:

[0056]

[0057] Here, uncertainty represents the degree of uncertainty, and di and dj represent the distances between a sample and two different cluster centers, respectively. By applying a clustering algorithm to the set of embedding vectors, multiple cluster centers c1, c2, ..., cn can be obtained. Further, a distance metric algorithm can be used to obtain the distance di between each embedding vector x and a cluster center ci. Uncertainty is the minimum difference between any two distances; that is, by iterating through all possible distance pairs di and dj, calculating the absolute values ​​of their differences, and taking the minimum of these differences as uncertainty.

[0058] This invention mainly includes two key technologies: 1) a two-stage active learning sample screening method; 2) a bronchial segmentation system based on two-stage active learning.

[0059] The two-stage active learning sample selection method consists of embedding vector generation, a one-stage active learning module, and a two-stage active learning module. Combined with expert data annotation by the user, it achieves a closed-loop training mechanism for the active learning of the segmentation network model. The two-stage active learning flowchart is shown below. Figure 1 As shown, the embedding vector generation module consists of a medical image preprocessing method, a vector generation network base, and a detachable segmentation network model. The medical image preprocessing method preprocesses the input CT 3D image, performing various image enhancement transformations to improve the robustness of the segmentation model. The vector generation network base is a pre-trained network model interface, allowing researchers to combine it with the most suitable detachable segmentation network model according to the needs of specific tasks. The working principle of this network base revolves around the deep feature extraction of the detachable segmentation network model. Through multi-layer nonlinear transformations, it progressively extracts deep-level feature vectors from the input image, serving as the embedding vectors for that sample. This process begins with capturing low-level visual features (such as edges and basic textures), gradually transitioning to recognizing more complex image structures and object contours. The embedding vectors generated in the intermediate layers, as the culmination of this series of abstraction processes, condense the key information of the image, presenting it in a high-dimensional and compact form, providing feature representations for the subsequent active learning filtering module. The vector generation network base is a component interface compatible with various detachable segmentation network models. This interface can adapt to the inputs of different segmentation networks and generate a unified embedding vector for the subsequent active learning sample selection stage. In this invention, an adaptive bronchial tubular structure segmentation network was designed as the detachable segmentation network model selected for this method, tailored to the characteristics of the task.

[0060] An adaptive segmentation network for bronchial tubular structures addresses two imbalance problems in bronchial datasets by incorporating domain knowledge to enhance model performance. This network introduces a self-attention module and linearly deformable convolutions into the basic 3D U-Net architecture, enabling it to recognize the unique features of tubular structures and thus improve segmentation accuracy in local regions. Simultaneously, it maintains global morphological continuity by expanding the receptive field of the convolutional kernels. The self-attention module allows the model to focus on important regions, thereby extracting richer airway features. The linearly deformable convolutions are position-parameter-learnable convolutions that adjust the kernel position according to the sample shape. To better adapt to the elongated shape of the bronchial airways, the movement of the convolutional kernels is restricted to linearity, allowing the kernel adjustment to better fit the structural characteristics of the bronchus. The flexibility of the linearly deformable convolutions enables them to adaptively capture subtle and curved local features of the tubular structures, thereby more effectively extracting key features of the bronchial airways and further improving segmentation accuracy. Figure 1 The detachable segmentation network model in the text, namely the adaptive bronchial tubular segmentation network, is as follows: Figure 2 As shown. The corresponding bronchial airway segmentation module is as follows. Figure 8 As shown.

[0061] An adaptive bronchial tubular structure segmentation network employs a comprehensive multi-scale feature extraction and fusion strategy, aiming to enhance the focus and understanding of key anatomical features from multiple perspectives. Built upon the previously described linear deformable convolution technique, each network layer deploys convolutional kernels of various sizes (e.g., 3x3, 5x5, 7x7) to achieve multi-dimensional analysis of the input information. Smaller convolutional kernels excel at capturing subtle local structural information of the bronchial airways, while larger linear deformable convolutional kernels are better at depicting and adapting to the overall morphology and topology of the bronchial airways. By fusing features obtained at these different scales, the network effectively integrates the advantages of various convolutional operations, consolidating cross-scale feature information into a hierarchical and information-rich feature representation system. This strategy not only enhances the grasp of the complexity of bronchial structures but also improves the model's simultaneous sensitivity to subtle and macroscopic features, thereby enhancing segmentation accuracy and robustness.

[0062] Figure 1 The first stage of the screening process, also known as the first stage of active learning screening, is shown in the flowchart below. Figure 3 As shown.

[0063] The core objective of the first stage of active learning screening is to accurately assess the stability of the embedding vectors of data samples, focusing on identifying highly unstable samples located on the decision boundary of the segmentation model. These samples, difficult to classify definitively, become rare resources in the dataset, significantly enhancing the model's ability to classify and segment complex images. To this end, this module uses a clustering algorithm to group these feature vectors, forming several clusters reflecting different data distributions. Then, it further measures the distance between each sample vector and the center of each cluster. For each sample, the minimum difference between two of these distances is taken as the "boundary difference." The smaller the boundary difference, the closer the sample is to the decision boundary, thus leading to higher model uncertainty. Based on this, the system calculates and ranks the minimum boundary difference of all samples as its uncertainty index. This stage filters a certain number of samples according to a preset threshold, forming a preliminary candidate sample set. This stage measures and samples the stability of the embedding vectors of data samples, effectively enriching samples with high information content and high uncertainty. However, it does not fully consider the diversity requirements of the sample set, leading to potential redundancy among the selected samples.

[0064] Figure 1 The second stage of the selection process, namely the second stage of active learning selection, is illustrated in the flowchart below. Figure 4 As shown.

[0065] The second stage aims to optimize the data screening results of the first stage, improve annotation efficiency, and enhance the diversity of the sample set. From the candidate samples obtained in the previous stage, the second stage filters the data according to a diversity quantification mechanism, selecting the most representative samples as the final selected sample set. Specifically, the diversity contribution is evaluated by measuring the similarity between each candidate sample and all samples in the selected sample set. Each time, the sample with the greatest difference from the existing sample set is added to the selected set until a predetermined diversity threshold is reached, thereby constructing a highly diverse and information-rich sample set.

[0066] In the two-stage active learning sample selection process, the process begins with the generation of embedding vectors from the unlabeled dataset. Through the collaboration of medical image preprocessing and embedding vector generation modules, the original CT images are transformed into high-dimensional feature representations. This stage not only utilizes image enhancement methods to improve model robustness but also flexibly adopts various efficient network architectures to adapt to specific segmentation task requirements. The generated embedding vectors then serve as input for the two-stage active learning selection. First, unstable samples at the decision boundary are identified based on uncertainty metrics to ensure maximum capture of both information and uncertainty. Then, a diversity quantification strategy is used to supplement the comprehensiveness of the sample set, eliminate redundancy, and construct a learning sample set that is both informative and diverse.

[0067] The flowchart for the two-stage active learning sample selection process is as follows: Figure 5 As shown. The corresponding two-stage screening module is as follows. Figure 7 As shown.

[0068] The architecture of the overall segmentation and learning system of the bronchial airway segmentation module and the two-stage screening module of this invention is as follows: Figure 6 As shown.

[0069] The multi-level airway segmentation and learning system based on a two-stage active learning method integrates a user-adaptive human-computer interaction interface, an active learning two-stage screening module, and a multi-level airway segmentation and learning module. The system can achieve three functions: 1) real-time airway annotation acquisition; 2) active learning-driven sample screening; and 3) real-time airway annotation acquisition and adaptive segmentation network learning of bronchial tubular structures.

[0070] In the various modules of this invention,

[0071] The embedding vector generation module first preprocesses the medical image and uses the detachable vector generation network base to generate the embedding feature vector of the sample.

[0072] The detachable vector generation network base supports the use of various different bronchial airway segmentation network models, and the high-dimensional feature representation of the samples generated after pre-training is used as the embedded feature vector.

[0073] The dual-stage active learning sample selection module considers both the information content and uncertainty of the data samples, as well as the diversity of the samples.

[0074] The first-stage active learning method clusters the representative feature vectors of the samples, calculates the boundary difference of each sample, and selects the samples with the highest uncertainty as the candidate sample set based on the magnitude of the boundary difference.

[0075] The second-stage active learning method uses the diversity assessment module to calculate the similarity between the sample and the selected sample set based on the candidate sample set, and selects the sample with the greatest diversity to form the final sample set.

[0076] The adaptive segmentation network model for bronchial tubular structures incorporates self-attention modules and linear deformable convolutions, enabling it to identify the unique features of tubular structures and thus improve the segmentation accuracy of local regions.

[0077] The active learning two-stage system uses the two-stage active learning sample selection module to select valuable samples.

[0078] The data preprocessing module preprocesses the received image data, including adjusting the CT window width and window level, feature enhancement and localization, etc.

[0079] The data post-processing module processes the segmentation results, including spatial continuity restoration and structural optimization;

[0080] The training process of the active learning segmentation system involves generating a final sample set and constructing a labeled dataset for further model training.

[0081] The reasoning process of the actively learned segmentation system involves inputting bronchial airway images and returning predicted airway labels.

[0082] The human-computer interaction interface and front-end / back-end communication method of the active learning segmentation system are designed to ensure both user interaction convenience and data processing efficiency.

[0083] The user-adaptive human-computer interface is designed specifically for medical professionals, providing intuitive image visualization tools and a smooth interactive interface to ensure doctors can efficiently participate in the annotation process. The user interface is also user-friendly, supporting drag-and-drop uploads and file selection, ensuring easy operation for medical professionals. The data preprocessing module performs data preprocessing on the input images, significantly improving the robustness of subsequent analysis through refined image calibration and enhancement. This process involves: 1) adjusting CT window settings according to clinical standards to ensure accurate visualization of tissue density; 2) applying Hessian filtering to highlight tubular structure features for easy bronchial identification; 3) precise ROI positioning and flexible spatial transformations, including rotation, translation, and scaling, while incorporating simulation processing such as elastic deformation and artifact injection to handle complex variables in actual scans.

[0084] The two-stage screening module is based on the two-stage active learning sample screening method. By learning from bronchial airway samples, the system can screen out the samples that have the greatest impact on model training and provide these samples to doctors for annotation, thereby obtaining an annotated dataset and constructing a high-quality training set. This set can be used directly for iterative optimization of the platform model and can also be exported to train other models.

[0085] The real-time airway annotation function uses a built-in adaptive bronchial segmentation network to process unannotated CT images uploaded by users. After meticulous optimization by the post-processing module, including spatial continuity restoration, edge smoothing, and connection restoration of broken structures, it finally outputs an accurate airway segmentation and annotation map, providing strong support for clinical decision-making.

[0086] The multi-level airway segmentation and learning module integrates efficient data processing, intelligent image analysis technology, and intuitive user interaction design, forming a seamless workflow from input to output to improve the efficiency and accuracy of bronchial CT image analysis. This module mainly includes two technical components: 1) the inference process of the multi-level airway and learning module; and 2) the training process of the main multi-level airway and learning module.

[0087] The reasoning flowchart of the multi-level airway and learning module is as follows: Figure 9 As shown, it mainly includes 1) the front-end and back-end communication process; 2) the back-end data processing process; and 3) the process of sending back processed data and updating the user interface in real time.

[0088] Front-end and back-end communication process: Once a file is uploaded, the front-end transmits the data to the back-end via the MQTT port. The front-end sends the file data as a message to the MQTT topic, containing file metadata and content. The MQTT server receives the message and forwards it to the back-end service. The back-end obtains the data by subscribing to the topic, decodes and processes the file, and then returns the processing result to the front-end via MQTT or HTTP. The front-end displays the feedback information on the user interface. The entire process includes error handling and retransmission mechanisms to ensure reliable transmission, and uses SSL / TLS encrypted communication to guarantee data security.

[0089] Backend data processing: Upon receiving a request, the processing flow begins immediately. The frontend interface may display a loading animation or progress bar, providing clear feedback to the user and maintaining smooth interaction. The data preprocessing module performs standardization adjustments on the received CT images, automatically setting the window width and window level based on optimal observation practices for lung imaging. The adaptive bronchial tubular structure segmentation network performs bronchial airway segmentation, accurately identifying and segmenting airway structures. The data postprocessing module performs a series of post-processing enhancement steps on the segmentation results to ensure the smoothness and accuracy of the segmentation edges.

[0090] The process of data transmission and real-time updates to the user interface after processing: After the segmented images are processed on the backend, the data is transmitted back to the frontend in an easily interpretable format. The user interface updates instantly, displaying clearly labeled bronchial airway segmentation results to assist doctors in further diagnostic analysis. In addition, the system may provide interactive tools such as zooming and rotating images, and switching between different view modes, allowing users to examine the segmentation results in detail from multiple angles.

[0091] The segmentation model training flowchart is as follows: Figure 10 As shown, the active learning segmentation system mainly includes the following process: First, when an unlabeled dataset is input into the system, the system's two-stage active learning system filters the data based on preset filtering accuracy parameters and detachable vectors to generate a network base, thus obtaining a subset of highly valuable datasets, i.e., the final sample set. This final sample set is then labeled by doctors or domain annotators, forming a labeled dataset. After preprocessing, these labeled datasets are used to train the system's segmentation network model.

[0092] The innovation of this invention is mainly reflected in two aspects.

[0093] The proposed two-stage active learning sample selection method first uses embedding vector generation and uncertainty assessment to select samples that contribute the most to the model segmentation results, ensuring that samples with more information in the dataset are labeled first; the second stage further evaluates the diversity of samples, reduces redundant samples, and improves labeling efficiency and model training effect.

[0094] In the adaptive bronchial tubular structure segmentation network, the self-attention module enables the model to focus on important regions, improving airway feature extraction capabilities. Linear deformable convolution adaptively adjusts the convolution kernel position according to the shape of the bronchus, enhancing the recognition ability of thin bronchioles and curved structures. Simultaneously, a multi-scale feature extraction and fusion method is used, employing convolution kernels of different sizes for feature extraction and combining feature information from different scales. This improves the global and local feature capture capabilities of the segmentation model, enhancing the accuracy of bronchial airway segmentation.

[0095] The multi-level airway segmentation and learning system based on a two-stage active learning method is designed for medical professionals. It supports human-machine collaboration functions, ensuring that doctors can participate in the annotation process efficiently. After rigorous testing and verification, this system has successfully and significantly improved the efficiency of annotation work.

[0096] The two-stage active learning method of the present invention effectively selects the most valuable samples, reduces the number of labeled samples, and lowers the cost and difficulty of medical image labeling.

[0097] The adaptive bronchial tubular structure segmentation network combines self-attention modules and linear deformable convolutions, enabling the model to capture the subtle and complex features of the bronchi more accurately, and significantly improving the segmentation accuracy of fine-grained bronchial airways.

[0098] To address the imbalance between intra- and inter-class segments in the bronchial dataset, the model can better handle the imbalance between fine bronchial and background pixels by introducing domain knowledge and enhancing feature extraction, thus further improving the segmentation performance.

[0099] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A two-stage active learning method for fine-grained bronchial airway segmentation, characterized in that, The method includes the following steps: S1. An embedded vector of an unlabeled medical image dataset is generated based on a pre-trained adaptive bronchial tubular segmentation model. S2. Cluster the embedded vectors, calculate the uncertainty of the samples in the clustered dataset, sort them based on the uncertainty, select candidate samples, calculate the diversity coefficient of the candidate samples, select the candidate sample with the largest diversity coefficient to add to the current sample set, and obtain the final sample set. S3. Determine whether the number of samples in the final sample set has reached the threshold. If yes, execute S4; otherwise, continue to select the largest candidate sample to add to the final sample set. S4. The final sample set is labeled, and the labeled sample set is used as the training set for the pre-trained adaptive bronchial tubular segmentation model to obtain an optimized adaptive bronchial tubular segmentation model. Based on the optimized adaptive bronchial tubular segmentation model, the segmentation result of the complex bronchial network is obtained. The uncertainty is: Here, uncertainty represents uncertainty, di and dj represent the distances between the sample and two different cluster centers, x is the embedding vector, ci represents the cluster center, and uncertainty is the minimum difference between any two distances. That is, by iterating through all possible distances, for di and dj, the absolute value of their difference is calculated, and the minimum value is taken as uncertainty. The specific steps for calculating the diversity of candidate samples and selecting the candidate sample with the highest diversity to add to the current sample set are as follows: The similarity between a candidate sample and all samples in the current sample set is calculated. Based on the similarity, a diversity coefficient ratio is calculated. The size of the diversity coefficient ratio is positively correlated with the difference between the samples. That is, the larger the ratio, the greater the difference between the sample and the current sample set. Then, the sample with the highest diversity is selected from the candidate samples and added to the current sample set. After that, the above steps are repeated to continuously compare the diversity coefficients of the samples, select and add samples to the current sample set until the number of samples in the current sample set reaches the preset number threshold. Finally, the final sample set that meets the requirements is obtained. The similarity between a candidate sample and all samples within the candidate sample set is: Where S represents the current sample set, x i Let sim(x) represent a sample in the current sample set, x represent a candidate sample, n represent the total number of candidate samples, and sim( x , S () indicates similarity; The diversity coefficient is: Where n represents the total number of candidate samples.

2. The two-stage active learning method for fine-grained bronchial airway segmentation according to claim 1, characterized in that, The specific steps of S1 are as follows: Obtain an unlabeled medical image dataset, input the unlabeled dataset into a pre-trained adaptive bronchial tubular segmentation model for feature extraction, and obtain the embedding vector.

3. The two-stage active learning method for fine-grained bronchial airway segmentation according to claim 2, characterized in that, The pre-trained adaptive bronchial tubular segmentation model comprises x interconnected network layers. Each network layer includes a first ordinary convolutional layer, a linear deformable convolutional layer, a channel attention module, and a second ordinary convolutional layer connected in sequence, with skip connections between the first and second ordinary convolutional layers.

4. The two-stage active learning method for fine-grained bronchial airway segmentation according to claim 3, characterized in that, The x network layers are deployed with convolutional kernels of various sizes, and the xth network layer outputs an embedding vector.

5. The two-stage active learning method for fine-grained bronchial airway segmentation according to claim 4, characterized in that, In x network layers, the y-th and xy-th network layers are skipped connections and have the same convolutional kernel size, while the non-skipped network layers have different convolutional kernel sizes.

6. The two-stage active learning method for fine-grained bronchial airway segmentation according to claim 1, characterized in that, The specific steps of S1 are as follows: Medical images are acquired, and image enhancement transformations are performed on the medical images to obtain an unlabeled medical image dataset. An embedding vector of the unlabeled medical image dataset is generated based on a pre-trained adaptive bronchial tubular segmentation model.

Images