Dual-stage active learning method for fine-grained bronchial airway segmentation

Abstract

The invention relates to a two-stage active learning method for fine-grained bronchial airway segmentation. The method comprises the following steps: S1, generating an embedded vector of an unlabeled medical image data set based on a pre-trained self-adaptive bronchial tube segmentation model; s2, selecting a candidate sample with the maximum diversity and adding the candidate sample into a final sample set; s3, judging whether the sample number of the final sample set reaches a threshold value or not, if yes, executing S4, and if not, continuing to select the candidate sample with the maximum diversity and adding the candidate sample into the final sample set; and S4, labeling the final sample set, taking the labeled sample set as a training set of a pre-trained adaptive bronchial tube segmentation model for training to obtain an optimized adaptive bronchial tube segmentation model, and obtaining a complex bronchial network segmentation result based on the optimized adaptive bronchial tube segmentation model. Compared with the prior art, the method has the advantages of improving the value density of the selected sample set, improving the recognition precision of the fine bronchus and the like.

Description

Technical Field

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

[0002] The current mainstream trend of bronchial airway segmentation research focuses on deep learning technology, among which the 3D-UNet architecture is widely adopted as the core model. Through the carefully designed encoding-decoding layer stack, the precise distinction of bronchial structures is achieved. For example, the CFDA method uniquely introduces a collaborative feature disentanglement and enhancement framework, cleverly uses clean and noisy CT scan data, and encodes the topological information of the clean scan and enhances the feature representation in the noisy data, which greatly improves the generalization performance and segmentation accuracy of the model in the bronchial segmentation task. Faced with the significant sample imbalance problem between high-level and low-level branches in the bronchial structure, the NaviAirway project adopted an innovative strategy, using a customized loss function that is particularly sensitive to fine bronchi, and combined with the model design of cross-level fine learning, through an iterative training strategy, effectively overcame the challenge of sample imbalance and significantly enhanced the segmentation quality. In addition, the contribution of Yang Nan and other scholars lies in proposing an advanced framework that combines fuzzy attention neural networks with composite loss functions. This framework uses a channel-specific fuzzy attention mechanism to cleverly solve the heterogeneity problem between different feature channels. At the same time, by designing a carefully designed comprehensive loss function, it enhances the spatial coherence of the segmentation output, thereby promoting an overall leap in the performance of airway segmentation.

[0003] However, fully supervised deep learning requires the use of a large number of annotated data sets as support. On the one hand, it is difficult to collect medical images because there are few public medical image data sets on the Internet. On the other hand, medical images are relatively complex and difficult to annotate. In many data sets, only some images are annotated by professional doctors. Due to these two problems with medical images, the training and deployment of deep learning models, especially supervised models, face the dual challenges of data acquisition and annotation. Active learning methods can determine the most valuable unlabeled samples in the data set by evaluating and quantifying the informativeness of unlabeled images. By introducing active learning methods in medical imaging models, high-performance medical image analysis can be achieved with less annotation.

[0004] Active learning has been widely used in many medical tasks, but the training strategies and methods have not been fully explored in the bronchial airway segmentation task. Researching active learning methods can further reduce the annotation cost of pulmonary CT images and unleash the performance of related supervision models.

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

[0006] The core of uncertainty-based active learning is that it identifies the "blind spots" of learning by quantifying the impact of the model on individual samples or its impact on the confidence of the model's predictions. This strategy can be further divided into instance uncertainty and model uncertainty. Instance uncertainty focuses on the predicted probability distribution of a single sample, such as using entropy or minimum probability as a metric to identify instances that are difficult for the model to accurately classify. Model uncertainty goes a step further and evaluates the uncertainty of model parameters through methods such as Bayesian active learning, reflecting the model's ability to adapt to new data. This type of method helps to directly strengthen the weak links in model cognition.

[0007] Representative sampling strategies focus on maintaining or increasing the diversity and coverage of data sets. Representative methods ensure that samples are collected from all corners of the data distribution, especially those that can effectively fill the blank areas of the feature space, by measuring the distance between samples, cluster distribution, or core sample selection. This not only promotes the understanding of complex data structures, but also improves the ability of the model to generalize to unseen data. Representative sampling is sometimes combined with cluster analysis to ensure that samples are selected for each cluster to achieve balanced coverage of the data space.

[0008] Currently, there are two imbalance problems in the bronchial dataset: intra-class imbalance and inter-class imbalance. The intra-class imbalance refers to the fact that in lung CT images, the proportion of airway pixels is much smaller than that of background pixels. The inter-class imbalance refers to the difference between thick bronchi and thin bronchi. The bronchi near the end of the alveoli may only occupy a few pixels. Compared with thick bronchi, the sample ratio of such thin bronchi is smaller.

[0009] For the current active learning technology, since only the representation characteristics of a single sample are considered, there are two problems when facing the diverse bronchial appearance structures: 1) It is difficult to construct a discriminative representation of fine bronchi samples, resulting in fine bronchi samples being often confused with negative samples; 2) The data selection method usually prefers to select thick bronchial samples with obvious visual features, resulting in a significant sample size gap between the distribution of thick and thin bronchial samples after sampling. The above defects significantly affect the performance of the segmentation model trained by active learning in the complex bronchial network segmentation task, often resulting in the omission of fine bronchi, reducing the overall segmentation accuracy and completeness. Summary of the invention

[0010] The purpose of the present 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 improve the accuracy of fine bronchi identification.

[0011] The purpose of the present invention can be achieved by the following technical solutions:

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

[0013] S1, Generate embedding vectors of unlabeled medical image datasets based on pre-trained adaptive bronchial tube segmentation model;

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

[0015] S3, determine whether the number of samples in the final sample set reaches the threshold, if so, execute S4, otherwise continue to select the largest candidate sample to add to the final sample set;

[0016] S4. Label the final sample set, and use the labeled sample set as a training set for a pre-trained adaptive bronchial tubule segmentation model to obtain an optimized adaptive bronchial tubule segmentation model, and obtain a complex bronchial network segmentation result based on the optimized adaptive bronchial tubule segmentation model.

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

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

[0019] Furthermore, the pre-trained adaptive bronchial tubule segmentation model includes n network layers connected to each other, each network layer includes a first ordinary convolution layer, a linear deformable convolution, a channel attention module, and a second ordinary convolution layer connected in sequence, and the first ordinary convolution layer and the second ordinary convolution layer are jump-connected.

[0020] Furthermore, the n network layers deploy convolution kernels of multiple different sizes, and the nth network layer outputs an embedding vector.

[0021] Furthermore, among the n network layers, the i-th and ni-th network layers are jump-connected and have the same deployed convolution kernel size, while the convolution kernel sizes deployed in the network layers without jump connection are different.

[0022] Furthermore, the uncertainty is:

[0023]

[0024] Among them, uncertainty represents uncertainty, di and dj represent the distances between the sample and two different cluster centers respectively, x is embedded in the vector, ci represents the cluster center, and uncertainty is the minimum value of the difference between any two distances, that is, traverse all possible distances, for di and dj, calculate the absolute value of the difference, and take the minimum value as uncertainty.

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

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

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

[0028]

[0029] Among them, S represents the current sample set, x i represents the samples in the current sample set, x represents the candidate samples, n represents the total number of candidate samples, and sim(x, S) represents 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:

[0034] A medical image is acquired, and image enhancement transformation is performed on the medical image to obtain an unlabeled medical image dataset, and 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] The present invention realizes the sample screening of multiple bronchioles and thick bronchi by comprehensively evaluating the sample uncertainty, representativeness and diversity at the same time and applies them to the active learning training of the segmentation model. The method breaks the bottleneck of low information gain that may be faced by the traditional active learning strategy based on intuitive image features. By calculating uncertainty and diversity, a two-stage screening process with fine division is realized, which not only enhances the ability to identify the importance of samples, but also significantly reduces the inclusion of redundant samples, effectively improves the value density of the selected sample set, and provides strong support for in-depth mining of key information in the bronchial data set.

[0037] In addition, the adaptive bronchial tubular structure segmentation network of the present invention adopts a comprehensive multi-scale feature extraction and fusion strategy to fuse the features obtained at different scales. The network effectively integrates the advantages of various convolution operations and integrates the cross-scale feature information into one place to form a hierarchical and information-rich feature representation system, which enhances the grasp of the complexity of the bronchial structure, promotes the model's simultaneous sensitivity to subtle and macro features, and improves the segmentation accuracy and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

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

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

[0040] Figure 3 Screening flow chart for Phase 1 active learning;

[0041] Figure 4 Screening flowchart for Phase II active learning;

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

[0043] Figure 6 This is the system architecture diagram;

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

[0045] Figure 8 It is the bronchial airway segmentation module;

[0046] Fig. 9 The inference flow chart of the multi-level airway and learning module;

[0047] Fig.10 Flowchart for system segmentation model training. DETAILED DESCRIPTION

[0048] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0049] The present invention proposes a two-stage active learning method for fine-grained bronchial airway segmentation. By comprehensively evaluating the uncertainty, representativeness, and diversity of samples at the same time, the sample screening of multiple bronchioles and thick bronchi is achieved and applied to the active learning training of the segmentation model. The method breaks the bottleneck of low information gain that may be faced by traditional active learning strategies based on intuitive image features. Through the finely divided two-stage screening process, not only the ability to identify the importance of samples is enhanced, but also the inclusion of redundant samples is significantly reduced, the value density of the selected sample set is effectively improved, and strong support is provided for in-depth mining of key information in the bronchial data set. The steps of this method include:

[0050] S1, Generate embedding vectors of unlabeled medical image datasets based on pre-trained adaptive bronchial tube segmentation model;

[0051] S2, cluster the embedding vectors, calculate the uncertainty of the samples in the clustered data set, sort them based on the uncertainty, select candidate samples, calculate the diversity of the candidate samples, and select the candidate samples with the largest diversity to add to the final sample set;

[0052] S3, determine whether the number of samples in the final sample set reaches the threshold, if so, execute S2, otherwise continue to select the largest candidate sample to add to the final sample set;

[0053] S4. Label the final sample set, and use the labeled sample set as a training set for a pre-trained adaptive bronchial tubule segmentation model to obtain an optimized adaptive bronchial tubule segmentation model, and obtain a complex bronchial network segmentation result based on the optimized adaptive bronchial tubule segmentation model.

[0054] Then, the segmentation system based on active learning was proposed by combining the interactive system, front-end and back-end communication, and the adaptive bronchial tubular structure segmentation network. The system evaluates the uncertainty, representativeness, and diversity of the collected samples, selects key data for annotation and segmentation model training, and optimizes the segmentation effect through multiple rounds of iterations. The system reduces the data size required for bronchial segmentation model training, reduces the annotation cost, and realizes the real-time discovery and reuse of key data by the segmentation system during diagnosis and application.

[0055] The uncertainty is:

[0056]

[0057] Among them, uncertainty represents uncertainty, di and dj represent the distance between the sample and two different cluster centers. By using the clustering algorithm on the embedded vector set, multiple cluster centers c1, c2, ..., cn of the set can be obtained. Further, the distance di between each embedded vector x and the cluster center ci can be obtained by the distance measurement algorithm. Uncertainty is the minimum value of the difference between any two distances, that is, traverse all possible distance pairs di and dj, calculate the absolute value of their difference, and take the minimum value as uncertainty.

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

[0059] The two-stage active learning sample screening method consists of an embedding vector generation module, a first-stage active learning module, a second-stage active learning module and other technical modules. The data is annotated by experts as users to achieve a closed-loop training of active learning of the segmentation network model. The two-stage active learning flow chart is as follows: Figure 1 As shown in the figure, 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 will first process the input CT three-dimensional image data and perform a variety of image enhancement transformations to improve the robustness of the segmentation model; the vector generation network base is a pre-trained network model interface that allows researchers to combine it with the most suitable detachable segmentation network model according to the needs of specific tasks. The working principle of the network base revolves around the deep feature extraction of the detachable segmentation network model. Through multiple layers of nonlinear transformation, the deep feature vectors of the input image are gradually extracted as the embedding vector of the sample. This process starts with the capture of underlying visual features (such as edges and basic textures) and gradually transitions to the recognition of more complex image structures and object contours. The embedding vector generated by the intermediate layer, as the crystallization of this series of abstract processes, condenses the key information of the image and presents it in a high-dimensional and compact form, providing feature representation for the subsequent active learning screening module. The vector generation network base is a component interface compatible with a variety of detachable segmentation network models, which can adapt the input of different segmentation networks and generate a unified embedding vector for the subsequent active learning sample screening stage. In the present invention, an adaptive bronchial tubular structure segmentation network is designed according to the task characteristics as the detachable segmentation network model selected by this method.

[0060] The adaptive bronchial tubular structure segmentation network is aimed at two imbalance problems in the bronchial data set, and the performance of the model is enhanced by introducing domain knowledge. The network introduces a self-attention module and a linear deformable convolution in the basic 3D U-Net architecture to enable it to identify the unique features of the tubular structure, thereby improving the segmentation accuracy of the local area. At the same time, the continuity of the global morphology is maintained by expanding the receptive field of the convolution kernel. The self-attention module enables the model to focus on important areas, thereby extracting richer airway features. The linear deformable convolution is a convolution with learnable position parameters, which can adjust the position of the convolution kernel according to the sample shape. In order to better adapt to the narrow and long shape of the bronchial airway, the movement of the convolution kernel is restricted to linear, so that the adjustment of the convolution kernel is more in line with the structural characteristics of the bronchus. The flexibility of the linear deformable convolution enables it to adaptively capture the subtle and curved local features of the tubular structure, thereby more effectively extracting the key features of the bronchial airway and further improving the accuracy of segmentation. Figure 1 The detachable segmentation network model in , that is, the adaptive bronchial tube segmentation network as Figure 2 The corresponding bronchial airway segmentation module is shown in Figure 8 shown.

[0061] The adaptive bronchial tubular structure segmentation network adopts a comprehensive multi-scale feature extraction and fusion strategy, aiming to enhance the focus and understanding of key anatomical features from multiple perspectives. The network is built on the previously described linear deformable convolution technique. Convolution kernels of various sizes (e.g., 3x3, 5x5, 7x7) are deployed at each network level to achieve multi-dimensional analysis of the input information. Smaller convolution kernels are good at capturing the subtle local structural information of the bronchial airways, while larger linear deformable convolution kernels are better at depicting and adapting to the overall morphology and topology of the bronchial airways. By fusing the features obtained at these different scales, the network effectively integrates the advantages of various convolution operations, integrates the cross-scale feature information into one, and forms a hierarchical and information-rich feature representation system. This strategy not only enhances the grasp of the complexity of bronchial structure, but also promotes the model's sensitivity to both subtle and macro features, improving segmentation accuracy and robustness.

[0062] Figure 1 The first stage screening part in the process is the first stage active learning screening flow chart as shown in Figure 3 shown.

[0063] The core goal of the first stage of active learning screening is to accurately evaluate the stability of the embedding vector of the data sample, focusing on identifying those highly unstable samples located on the decision boundary of the segmentation model. These samples are difficult to clearly classify, so they have become rare resources in the data set, which can significantly enhance the model's ability to classify and segment complex images. To this end, the module groups these feature vectors through a clustering algorithm to form several clusters that reflect different data distributions. Then, the distance between each sample vector and the center of each cluster is further measured. For each sample, the minimum difference between the two distances is used as the "boundary difference" of the sample. The smaller the boundary difference, the closer the sample is to the decision boundary, which brings higher model uncertainty. Based on this, the system calculates and sorts the minimum boundary difference of all samples as its uncertainty indicator. In this stage, a certain number of samples are screened according to the preset threshold to form a preliminary set of candidate samples. In this stage, the stability of the embedding vector of the data sample is measured and sampled, which effectively enriches samples with large information and high uncertainty, but it does not fully consider the diversity requirements of the sample set, resulting in possible redundancy between the selected samples.

[0064] Figure 1 The second stage screening part in the figure is the second stage active learning screening flow chart as shown in Figure 4 shown.

[0065] The second stage aims to optimize the data screening results of the first stage, improve the annotation efficiency and enhance the diversity of the sample set. In the second stage, the candidate samples obtained in the previous stage will be screened according to the diversity quantification mechanism to select the most representative samples as the final selected sample set. In specific implementation, the diversity contribution of each candidate sample is evaluated by measuring the similarity between each candidate sample and all samples in the selected sample set. Each time, the sample with the largest difference from the existing sample set is selected and added to the selected set until the predetermined diversity threshold is reached, so as to construct a highly diverse and information-rich sample set.

[0066] In the two-stage active learning sample screening process, the process starts with the generation of embedding vectors for unlabeled datasets. Through the collaboration of medical image preprocessing and embedding vector generation modules, the original CT images are converted into high-dimensional feature representations. This stage not only uses image enhancement methods to improve the robustness of the model, but also flexibly adopts a variety of different efficient network architectures to meet the needs of specific segmentation tasks. The generated embedding vectors are then used as input for the two-stage active learning screening. First, unstable samples of the decision boundary are identified based on uncertainty metrics to ensure maximum capture of information and uncertainty; then, the comprehensiveness of the sample set is supplemented through a diversity quantification strategy, redundancy is eliminated, and a learning sample set that is both informative and diverse is constructed.

[0067] The two-stage active learning sample screening flow chart is as follows Figure 5 The corresponding two-stage screening module is shown as Figure 7 shown.

[0068] The architecture of the overall segmentation and learning system for controlling the bronchial airway segmentation module and the dual-stage screening module of the present invention is as follows: Figure 6 shown.

[0069] The multi-level airway segmentation and learning system based on the two-stage active learning method integrates a professional 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 fine screening; 3) real-time airway annotation acquisition and adaptive bronchial tubular structure segmentation network learning.

[0070] In each module of the present invention,

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

[0072] The detachable vector generation network base supports the use of a variety of different bronchial airway segmentation network models, and generates high-dimensional feature representations of samples as embedded feature vectors after pre-training;

[0073] The dual-stage active learning sample screening module considers both the amount of information and uncertainty of the data samples and 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 according to the size of the boundary difference;

[0075] The second-stage active learning method calculates the similarity between the sample and the selected sample set based on the candidate sample set through the diversity evaluation module, screens out the sample with the greatest diversity, and forms the final sample set.

[0076] The adaptive bronchial tubular structure segmentation network model introduces a self-attention module and a linear deformable convolution to enable it to recognize the unique features of the tubular structure, thereby improving the segmentation accuracy of the local area.

[0077] The active learning two-stage system uses the two-stage active learning sample screening module to screen valuable samples

[0078] The data pre-processing module pre-processes the received image data, including adjustment of CT window width and window position, feature enhancement and positioning, etc.;

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

[0080] The training process of the active learning segmentation system generates the final sample set and constructs the labeled data set for further model training;

[0081] The inference process of the active learning segmentation system takes as input a bronchial airway image and returns the predicted airway annotations.

[0082] The human-computer interaction interface and the front-end and back-end communication method of the active learning segmentation system are aimed at ensuring the convenience of user interaction and the efficiency of data processing.

[0083] The professional user-adaptive human-computer interaction interface is designed for medical professionals, providing intuitive image visualization tools and smooth interactive interfaces to ensure that doctors can efficiently participate in the annotation process. At the same time, the user interaction interface is designed to be friendly and supports drag-and-drop upload or browse file selection, ensuring that medical professionals can operate easily. The data pre-processing module will perform data pre-processing on the input image, and significantly improve the robustness of subsequent analysis through refined image calibration and enhancement. This process involves: 1) adjusting the 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 transformation, including rotation, translation and scaling, while adding simulation processing such as elastic deformation and artifact injection to cope with complex variables in actual scanning.

[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 samples that have the greatest impact on model training and provide these samples to doctors for annotation. Doctors then obtain annotated data sets and build high-quality training sets, which can not only be used directly for iterative optimization of platform models, but also support export to other model training.

[0085] The instant airway labeling function uses the built-in adaptive bronchial segmentation network to process the unlabeled CT images that users can upload, and then undergoes careful optimization of the post-processing module, including spatial continuity repair, edge smoothing, and restoration of the connection of broken structures, to ultimately output an accurate airway segmentation labeling 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 contents: 1) the reasoning process of the multi-level airway and learning module; 2) the training process of the main multi-level airway and learning module.

[0087] The reasoning flow chart of multi-level airway and learning module is as follows Fig. 9 As shown, it mainly includes 1) the front-end and back-end communication process; 2) the back-end data processing process; 3) the real-time update process of the processed data feedback and the user interaction interface.

[0088] Front-end and back-end communication process: Once the file is uploaded, the front-end transmits the data to the back-end through the MQTT port. The front-end sends the file data to the MQTT topic in the form of a message, including the 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 through MQTT or HTTP. The front-end displays the feedback information on the user interface. The whole process includes error handling and retransmission mechanisms to ensure transmission reliability, and uses SSL / TLS encrypted communication to ensure data security.

[0089] Back-end data processing process: After receiving the request, the processing flow starts immediately, and the front-end interface may display a loading animation or progress bar to give users clear operational feedback and maintain the smoothness of the interaction. The data pre-processing module performs standardized adjustments on the received CT images, automatically sets the window width and window position according to the best observation practices for lung imaging, and the adaptive bronchial tubular structure segmentation network performs bronchial airway segmentation to accurately identify and segment the airway structure. The data post-processing and 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 returning processed data and updating the user interface in real time: After the segmented image is processed by the backend, the data is returned to the frontend in an easy-to-interpret format. The user interface is updated in real time to display the clearly labeled bronchial airway segmentation results to assist doctors in further diagnostic analysis. In addition, the system may also provide interactive tools such as zooming, rotating images, and switching different view modes so that users can carefully review the segmentation results from multiple angles.

[0091] The segmentation model training flow chart is as follows Fig.10 As shown. The active learning segmentation system mainly includes the following processes: First, when the unlabeled data set is passed into the system, the system's two-stage active learning system will generate a network base based on the preset screening accuracy parameters and detachable vectors, screen the data, and filter out a highly valuable subset of the data set, that is, the final sample set. The final sample set will be annotated by doctors or professional annotators to form annotated data sets. After pre-processing, these annotated data sets will be used to train the system's segmentation network model.

[0092] The innovations of the present invention are mainly reflected in two aspects.

[0093] The two-stage active learning sample screening method uses embedding vector generation and uncertainty evaluation in the first stage to screen out samples that contribute most to the model segmentation results, ensuring that samples with larger amounts of information in the data set are labeled first; the second stage further evaluates the diversity of samples, reduces redundant samples, and improves labeling efficiency and model training effects.

[0094] In the adaptive bronchial tubular structure segmentation network, the self-attention module enables the model to focus on important areas and improve the ability to extract airway features. The linear deformable convolution adaptively adjusts the position of the convolution kernel according to the shape of the bronchus to enhance the recognition of bronchioles and curved structures. At the same time, the multi-scale feature extraction and fusion method is used. By using convolution kernels of different sizes for feature extraction and combining feature information of different scales, the global and local feature capture capabilities of the segmentation model are improved, and the accuracy of the model in bronchial airway segmentation is enhanced.

[0095] The multi-level airway segmentation and learning system based on the dual-stage active learning method is designed for medical professionals to support functions related to human-computer collaboration and ensure that doctors can efficiently participate in the labeling process. After rigorous testing and verification, this system has successfully significantly improved the efficiency of labeling work.

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

[0097] The adaptive bronchial tubular structure segmentation network combines the self-attention module and the linear deformable convolution, so that the model can more accurately capture the subtle and complex features of the bronchi, and significantly improves the accuracy of fine-grained bronchial airway segmentation.

[0098] To address the intra-class and inter-class imbalance problems in the bronchial dataset, by introducing domain knowledge and enhancing feature extraction, the model can better handle the imbalance of bronchioles and background pixels, further improving the segmentation effect.

[0099] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A two-stage active learning method for fine-grained bronchial airway segmentation, characterized in that The method comprises the following steps: S1. Generate embedding vectors of unlabeled medical image datasets based on pre-trained adaptive bronchial tube segmentation model; S2, cluster the embedded vectors, calculate the uncertainty of the samples in the clustered data set, sort them based on the uncertainty, select candidate samples, calculate the diversity coefficient for the candidate samples, select the candidate sample with the largest diversity coefficient and add it to the current sample set to obtain the final sample set; S3, determine whether the number of samples in the final sample set reaches the threshold, if so, execute S4, otherwise continue to select the largest candidate sample to add to the final sample set; S4. Label the final sample set, and use the labeled sample set as a training set for a pre-trained adaptive bronchial tubule segmentation model to obtain an optimized adaptive bronchial tubule segmentation model, and obtain a complex bronchial network segmentation result based on the optimized adaptive bronchial tubule segmentation model.

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

3. A two-stage active learning method for fine-grained bronchial airway segmentation according to claim 2, characterized in that: The pre-trained adaptive bronchial tubule segmentation model includes n network layers connected to each other, each network layer includes a first ordinary convolution layer, a linear deformable convolution, a channel attention module and a second ordinary convolution layer connected in sequence, and the first ordinary convolution layer and the second ordinary convolution layer are jump-connected.

4. A two-stage active learning method for fine-grained bronchial airway segmentation according to claim 3, characterized in that: The n network layers deploy convolution kernels of various sizes, and the nth network layer outputs an embedding vector.

5. A two-stage active learning method for fine-grained bronchial airway segmentation according to claim 4, characterized in that: Among the n network layers, the i-th and ni-th network layers are jump-connected and have the same deployed convolution kernel size, while the convolution kernel sizes deployed in the network layers without jump connections are different.

6. A two-stage active learning method for fine-grained bronchial airway segmentation according to claim 1, characterized in that: The uncertainty stated is: Among them, uncertainty represents uncertainty, di and dj represent the distances between the sample and two different cluster centers respectively, x is embedded in the vector, ci represents the cluster center, and uncertainty is the minimum value of the difference between any two distances, that is, traverse all possible distances, for di and dj, calculate the absolute value of the difference, and take the minimum value as uncertainty.

7. A two-stage active learning method for fine-grained bronchial airway segmentation according to claim 6, characterized in that: The specific steps for calculating the diversity of candidate samples and selecting the candidate sample with the largest diversity to add to the current sample set are: The similarity between the candidate sample and all the samples in the current sample set is calculated, and the diversity coefficient ratio is calculated based on the similarity. The size of the diversity coefficient ratio is positively correlated with the difference of the sample, that is, the larger the ratio, the greater the difference between the sample and the current sample set. Subsequently, the sample with the largest diversity is selected from the candidate samples and added to the current sample set. After that, the above steps are repeated to continuously compare, select and add the diversity coefficients of the samples to the current sample set until the number of samples in the current sample set reaches the preset number threshold. Finally, a final sample set that meets the requirements is obtained.

8. A two-stage active learning method for fine-grained bronchial airway segmentation according to claim 7, characterized in that: The similarity between a candidate sample and all samples in the candidate sample set is: Among them, S represents the current sample set, x i represents the samples in the current sample set, x represents the candidate samples, n represents the total number of candidate samples, and sim(x, S) represents the similarity.

9. A two-stage active learning method for fine-grained bronchial airway segmentation according to claim 8, characterized in that: The diversity coefficient is: Where n represents the total number of candidate samples.

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

Images