Self-adaptive drainage basin growth lung segment boundary accurate positioning method

Abstract

The invention relates to the technical field of medical image processing, in particular to an adaptive drainage basin growth lung segment boundary accurate positioning method, which comprises the following steps: acquiring lung CT image data; performing image preprocessing based on the lung CT image data to obtain a preprocessed lung image; according to the preprocessed lung image, determining a main trunk point set of a lung segment artery; based on the trunk point set, adopting a sub-point intelligent arrangement strategy to determine initial seed points; according to the initial seed point, adaptive drainage basin growth based on point competition is executed, and a lung segment boundary preliminary positioning result is obtained; based on the initial positioning result of the lung segment boundary, boundary conflict detection and processing are carried out, and an optimized lung segment boundary is obtained; the optimized lung segment boundary is output as a final accurate positioning result of the lung segment boundary, by dynamically adjusting growth parameters and introducing a point competition mechanism, the method can more accurately capture the complex form of the lung segment boundary, and the accuracy and robustness of segmentation are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and more specifically to a method for accurately locating lung segment boundaries by adaptive watershed growth. Background Art

[0002] The precise localization of lung segment boundaries is of great significance in the diagnosis of lung diseases, surgical planning, and radiotherapy. With the advancement of medical imaging technology, especially the widespread use of computed tomography (CT), the visualization of lung anatomical structures has been greatly improved. However, due to the complexity of lung structure and individual differences, accurately localizing lung segment boundaries remains a challenging task.

[0003] Traditional lung segment boundary localization methods mainly rely on manual segmentation or semi-automatic segmentation techniques. These methods are not only time-consuming and labor-intensive, but also easily affected by the subjective judgment of the operator, resulting in poor consistency and repeatability of the results. In recent years, with the development of computer vision and artificial intelligence technology, some automated lung segmentation methods have been proposed. These methods can be roughly divided into anatomical knowledge-based methods, image feature-based methods, and deep learning-based methods.

[0004] Methods based on anatomical knowledge usually use anatomical landmarks such as lobar fissures and vascular trees to locate lung segment boundaries. The advantage of such methods is good interpretability, but their accuracy is often limited by the recognition accuracy of anatomical landmarks. Especially in some pathological conditions, such as patients with emphysema or pulmonary fibrosis, anatomical landmarks may become unclear or distorted, resulting in the failure of such methods.

[0005] Methods based on image features, such as region growing and level set methods, perform segmentation by analyzing image features such as grayscale and texture. These methods perform well when dealing with normal lung structures, but are not robust enough when dealing with lesion areas or situations with large individual differences. In addition, these methods usually require a large number of parameters to be set manually, and the choice of parameters has a significant impact on the results, limiting their widespread application in clinical practice.

[0006] In recent years, deep learning-based methods, especially convolutional neural networks (CNNs), have made significant progress in the field of medical image segmentation. These methods are able to automatically learn image features and even surpass the performance of human experts in some tasks. However, deep learning methods also face some challenges. First, they usually require a large amount of training data with precise annotations, which is expensive and time-consuming in the medical field. Second, deep learning models are often black-box and lack interpretability, which may raise some ethical and legal issues in medical applications. Finally, the generalization ability of these methods in dealing with edge cases and rare cases still needs to be improved.

[0007] In the existing technologies, both traditional methods and deep learning methods have a common problem, that is, it is difficult to balance accuracy, efficiency and adaptability at the same time. In particular, when dealing with different patients, different scanning devices and various pathological conditions, existing methods often have difficulty maintaining stable performance. In addition, most methods lack effective mechanisms to integrate expert knowledge and automated algorithms, and cannot fully utilize the experience of clinical experts to continuously improve system performance. Summary of the invention

[0008] The present invention aims to solve the above technical problems and provide a method for accurately locating lung segment boundaries with adaptive watershed growth. The method innovatively combines image processing, machine learning and deep learning technologies to achieve high-precision, high-efficiency and high-adaptability positioning of lung segment boundaries.

[0009] In order to solve the above technical problems, the present invention adopts the following technical solutions: The method for accurately locating lung segment boundaries by adaptive watershed growing includes: The acquisition steps include: Acquire lung CT image data; Processing steps include: Based on the lung CT image data, image preprocessing is performed to obtain a preprocessed lung image; Determining a set of main points of the pulmonary segmental artery according to the preprocessed lung image; Based on the trunk point set, an initial seed point is determined by adopting a sub-point intelligent arrangement strategy; According to the initial seed points, performing adaptive watershed growth based on point competition to obtain a preliminary positioning result of the lung segment boundary; Based on the preliminary positioning result of the lung segment boundary, boundary conflict detection and processing are performed to obtain an optimized lung segment boundary; Output steps include: The optimized lung segment boundary is output as the final lung segment boundary precise positioning result.

[0010] Preferably, the image preprocessing specifically includes: Performing bilateral filtering on the lung CT image data to obtain a denoised lung image; Based on the denoised lung image, blood vessel extraction is performed to obtain a lung blood vessel mask.

[0011] Preferably, the sub-point intelligent arrangement strategy specifically includes: Based on the anatomical structure characteristics of the pulmonary artery in the human body, determine the starting point and branching point of the aorta; Determine the anatomical landmarks based on the connection points between the main pulmonary artery and the left and right pulmonary arteries and the left and right hilum; Based on the aorta starting point, branching point and anatomical landmark point, the position of the initial seed point is determined.

[0012] Preferably, the adaptive watershed growth based on point competition specifically includes: Based on the initial seed point, construct a point group; Using Euclidean distance as a similarity parameter, unsupervised clustering is performed on the point group; According to the clustering results, determine the watershed corresponding to each class; Based on the watershed, adaptive seed growing is performed, which includes: Define the point group growth rules based on the distance between seed points, including adaptive step size and distance threshold parameters; According to the average distance between the class points and the seed points, the adaptive convergence criterion is set; By comparing the competitive scores of each point, boundaries are drawn between adjacent point classes.

[0013] Preferably, the boundary conflict detection and processing specifically includes: Calculate the seed competition score between the boundary area point class and the adjacent point class area; identifying boundary conflict regions between a plurality of watersheds based on the seed competition scores; For detected boundary conflicts, the narrowest channel between two seed points is found in the conflict area; According to the narrowest channel, the point cluster in the conflict area is divided into two parts to resolve the boundary conflict.

[0014] As a preference, it also includes: Based on the deep learning model, the optimized lung segment boundaries are further optimized, wherein: A convolutional neural network model is used to compare the model-predicted boundaries with the standard results; The boundary correction model is optimized through the loss function to obtain the final accurate positioning result of the lung segment boundary.

[0015] Preferably, the adaptive seed growth further comprises: Anisotropic tensor fields are used to simulate the directional constraints of local expansion; Based on the anisotropic tensor field, the control point group grows along the blood vessel direction.

[0016] As a preference, it also includes: Based on the precise positioning result of the lung segment boundary, lung tumor analysis is performed; Using the lung segment boundary as a reference boundary for tumor staging; Combined with imaging omics analysis tools, it provides tumor pathological feature analysis information.

[0017] Preferably, the output step further comprises: Visually displaying the precise positioning result of the lung segment boundary; A three-dimensional model including the lung segment boundaries is generated.

[0018] As a preference, it also includes: A user interaction interface is provided to allow a doctor to fine-tune the precise positioning result of the lung segment boundary; Based on the fine-tuning results, the training data set of the deep learning model is updated.

[0019] The method of the present invention has the following significant technical effects: First, the adaptive watershed growing model proposed in this paper can effectively adapt to the anatomical differences and various pathological conditions of different patients. By dynamically adjusting the growth parameters and introducing the point competition mechanism, this method can more accurately capture the complex morphology of the lung segment boundary and significantly improve the accuracy and robustness of segmentation.

[0020] Secondly, the present invention cleverly combines the advantages of traditional image processing technology and deep learning methods. The initial segmentation based on image features provides a good starting point, while the introduction of deep learning models further optimizes the segmentation results. This hybrid approach not only improves the performance of the algorithm, but also enhances its interpretability, making it easier to gain the trust and acceptance of clinical experts.

[0021] Furthermore, the boundary conflict detection and processing mechanism introduced by the present invention effectively solves the boundary fuzziness problem between multiple lung segments. This innovation greatly improves the consistency and reliability of the segmentation results, providing more accurate anatomical information for subsequent clinical applications such as surgical planning and radiotherapy.

[0022] In addition, the user interaction interface and continuous learning mechanism designed by the present invention establish a good collaborative relationship between clinical experts and AI systems. This not only allows experts to fine-tune the results based on clinical experience, but also feeds these corrections back into the model to achieve continuous optimization of system performance. This human-computer collaborative approach greatly improves the practicality and reliability of the system.

[0023] Finally, the method of the present invention has broad application prospects. In addition to the precise positioning of lung segment boundaries, the method can also be extended to multiple fields such as lung tumor analysis, surgical planning and radiotherapy. By generating accurate three-dimensional models and combining imaging genomics analysis, the present invention provides strong technical support for personalized medicine.

[0024] In general, the adaptive watershed growth method for precise positioning of lung segment boundaries provided by the present invention innovatively solves the problem of balancing accuracy, efficiency and adaptability, provides a more reliable and accurate tool for the diagnosis and treatment of lung diseases, and is expected to play an important role in improving medical quality and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 The figure is an overall flow chart of the method of the present invention.

[0026] Figure 2 The flowchart of the adaptive watershed growth based on point competition of the present invention.

[0027] Figure 3 Flowchart of boundary conflict detection and processing of the present invention.

[0028] Figure 4 Flowchart for optimizing the deep learning model of the present invention. DETAILED DESCRIPTION

[0029] like Figure 1-4 As shown, the present invention provides a method for accurately locating lung segment boundaries with adaptive watershed growth. The method achieves high-precision positioning of lung segment boundaries by innovatively combining image processing, machine learning and deep learning technologies. The specific implementation methods of the present invention will be described in detail below.

[0030] The method for accurately locating lung segment boundaries by adaptive watershed growth provided by the present invention comprises the following steps: First, in the acquisition step, the method acquires lung CT image data. These CT image data are usually acquired by a CT scanner in a medical institution and contain detailed anatomical information of the patient's lungs. Preferably, the resolution of the CT image should be no less than 512x512 pixels and the layer thickness should not exceed 1mm to ensure sufficient image quality and details.

[0031] Next, in the processing step, the method of the present invention first performs image preprocessing based on the acquired lung CT image data to obtain a preprocessed lung image. Image preprocessing is a key step, which can effectively improve the accuracy and efficiency of subsequent processing. In one embodiment of the present invention, image preprocessing may include operations such as noise removal, contrast enhancement, and image normalization. For example, a Gaussian filter can be used to remove high-frequency noise in an image, and a histogram equalization method can be used to enhance image contrast.

[0032] After the preprocessing is completed, the method determines the main point set of the pulmonary segmental artery based on the preprocessed lung image. This step is the basis for subsequent precise positioning. Preferably, a threshold-based segmentation method combined with morphological operations can be used to extract the pulmonary vascular structure, and then a skeletonization algorithm is used to obtain the center line of the blood vessel to determine the main point set. For example, the CT value threshold can be set between -600 and 400 HU, and this range can usually effectively segment the pulmonary vascular structure.

[0033] After determining the set of trunk points, the method of the present invention uses a sub-point intelligent placement strategy to determine the initial seed points based on the trunk point set. The purpose of this step is to provide a good starting point for the subsequent watershed growth algorithm. In a specific embodiment, the sub-point intelligent placement strategy can determine the seed points based on the spatial distribution of the trunk points and the local vascular morphological characteristics. For example, the seed points can be placed at the midpoint between the trunk points, and the bifurcation of the blood vessels can be considered, and additional seed points can be added at the bifurcation.

[0034] Next, the method performs adaptive watershed growth based on point competition according to the determined initial seed point to obtain the preliminary positioning result of the lung segment boundary. This is one of the core steps of the present invention, which realizes the adaptive growth of the lung segment boundary by simulating the competition relationship between points. Specifically, the following formula can be used to calculate the competition intensity between points: , in, Indicate point and The competition intensity between them, I(p) represents the image intensity at point p, is an adjustable parameter used to control the decay rate of competition intensity. Preferably, Can be set between 10% and 20% of the image intensity range.

[0035] After obtaining the preliminary positioning result of the lung segment boundary, the method of the present invention performs boundary conflict detection and processing based on the result to obtain the optimized lung segment boundary. This step is intended to solve the problem of boundary overlap or gap between different lung segments. In one embodiment, morphological operations can be used to detect and process boundary conflicts. For example, the preliminary positioning result can be subjected to dilation and erosion operations, and then the results before and after the operations can be compared to identify areas where conflicts may exist.

[0036] Finally, in the output step, the method outputs the optimized lung segment boundary as the final lung segment boundary accurate positioning result. This result can be presented in various forms such as binary image, contour point set or three-dimensional model to meet different application requirements.

[0037] The method of the present invention may also include other optimization steps. For example, a deep learning model may be introduced to further optimize the boundary localization results. In this case, a classic image segmentation network such as U-Net may be used, with the preliminary localization results as input and a more refined boundary output. Training such a network usually requires a large amount of annotated data, and it is recommended to use at least 1,000 lung segment CT images with expert annotations for training.

[0038] In general, the method for accurately locating lung segment boundaries by adaptive watershed growth provided by the present invention achieves high-precision positioning of lung segment boundaries by comprehensively using image processing, machine learning and deep learning technologies. The method has the advantages of strong adaptability, high precision and good robustness, and can provide important technical support for the diagnosis and treatment of lung diseases.

[0039] The method of the present invention specifically comprises the following steps when performing adaptive watershed growth based on point competition: First, a point group is constructed based on the initial seed point. This step is intended to provide basic data for subsequent clustering and watershed growth. Preferably, the point group can be constructed using the k-nearest neighbor (k-NN) algorithm, with each seed point as the center and the k nearest points around it selected. In one embodiment of the present invention, the k value can be set between 50 and 100, which can usually achieve a good balance between computational efficiency and coverage.

[0040] Next, the method uses Euclidean distance as the similarity parameter to perform unsupervised clustering on the constructed point group. The purpose of this step is to classify similar points into the same watershed. Preferably, the K-means clustering algorithm can be used. The advantage of the K-means algorithm is that it is simple, efficient, and easy to implement. In practical applications, the number of clusters K can be determined based on the number of initial seed points, and can usually be set to 1.2 to 1.5 times the number of seed points to allow a certain degree of over-segmentation for subsequent boundary optimization.

[0041] Based on the clustering results, the method determines the watershed corresponding to each class. This step converts the clustering results into preliminary lung segment boundaries. In a preferred embodiment of the present invention, the following formula can be used to define the watershed: , in, represents the i-th watershed, p represents any point in the image, and denote the i-th and j-th cluster centers respectively, represents the Euclidean distance.

[0042] Based on the determined watershed, the method of the present invention performs adaptive seed growth. This is one of the core steps of the present invention, which achieves adaptive segmentation of lung segments of different shapes and sizes by dynamically adjusting growth parameters. Specifically, adaptive seed growth includes the following key points: First, define the point group growth rule based on the distance between seed points, including the adaptive step size and distance threshold parameters. Preferably, the adaptive step size can be calculated using the following formula: , in, represents the step size of the i-th iteration, is the initial step size (can be set from 1 to 3 pixels), is the distance from the current point to the nearest seed point, is the maximum distance from all points to their nearest seed point. This formula can make the area close to the seed point grow slower, while the area far from the seed point grows faster, thus adapting to lung segments of different sizes.

[0043] Secondly, the method sets an adaptive convergence criterion based on the average distance between the class point and the seed point. The purpose of this step is to obtain appropriate segmentation results in lung segments of different sizes. In one embodiment of the present invention, the following formula can be used to determine whether convergence has occurred: , Where N is the number of points in the current class, is the i-th point in the class, c is the center point of the class, is a small positive number representing the convergence threshold. Preferably, Can be set from 1 to 3 pixels.

[0044] Finally, this method divides the boundaries between adjacent point classes by comparing the competition scores of each point. The purpose of this step is to obtain more accurate lung segment boundaries. The competition score can be calculated using the following formula: , Where S(p) is the competition score of point p, N(p) is the neighborhood of p, w(p,q) is the weight between p and q (which can be defined based on distance or image intensity), is the indicator function, when q belongs to the same region as p 1 if the value is 0, otherwise it is 0.

[0045] The method of the present invention specifically includes the following steps when performing boundary conflict detection and processing: First, the seed competition score between the boundary area point class and the adjacent point class area is calculated. The purpose of this step is to quantify the competition relationship between different areas. In a preferred embodiment of the present invention, the competition score can be calculated using the following formula: , in, Indicates area and The competition score between represents the boundary point set of two regions, and Respectively represent the point p to the region and degree of belonging.

[0046] Next, the method identifies the boundary conflict areas between multiple watersheds based on the calculated seed competition scores. Preferably, a threshold T can be set. >T, the area and There is a boundary conflict between them. The choice of threshold T can be determined by experiment and can usually be set to 1.5 to 2 times the average competition score.

[0047] For the detected boundary conflict, the method of the present invention searches for the narrowest channel between the two seed points in the conflict area. The purpose of this step is to find the most reasonable boundary division position. In one embodiment of the present invention, the Dijkstra algorithm can be used to find the narrowest channel, and the image gradient is used as the weight of the edge.

[0048] Finally, this method divides the point clusters in the conflicting area into two parts according to the narrowest channel found, thereby resolving the boundary conflict. This step can effectively eliminate the boundary overlap or gap problem between different lung segments and improve the accuracy of segmentation.

[0049] In a preferred embodiment of the present invention, a deep learning model can be introduced to further optimize the optimized lung segment boundaries. Specifically, a convolutional neural network model can be used to compare the model-predicted boundaries with the standard results. Classic image segmentation network structures such as U-Net or DeepLab can be used here. The input of the network can be a fusion of the original CT image and the preliminary segmentation result, and the output is the optimized segmentation mask.

[0050] In order to train such a deep learning model, a large number of lung segment CT images with expert annotations are required. Preferably, the size of the training set should be no less than 1,000 cases to ensure the generalization ability of the model. During the training process, the Dice coefficient can be used as the loss function, which can effectively balance the contribution of objects of different sizes. The loss function can be expressed as: , Among them, X is the predicted segmentation result, Y is the true label, Indicates the size of the collection.

[0051] By optimizing the boundary correction model with a loss function, this method finally obtains an accurate lung segment boundary localization result. This result not only takes into account the local features of the image, but also integrates the deep learning model's understanding of global semantic information, thus achieving high accuracy.

[0052] In another embodiment of the present invention, adaptive seed growth also includes using anisotropic tensor fields to simulate directional constraints of local expansion. This method can better adapt to the complex structure of pulmonary blood vessels. Specifically, the Hessian matrix can be used to construct anisotropic tensor fields: , in, , , etc. are the second-order partial derivatives of the image in the corresponding direction. By analyzing the eigenvalues ​​and eigenvectors of the Hessian matrix, the main direction of the local structure can be obtained. Based on the constructed anisotropic tensor field, this method controls the point group to grow along the blood vessel direction. This can be achieved by modifying the adaptive step size formula mentioned above: , in, is the unit vector of the current growth direction, is the diffusion tensor derived from the Hessian matrix, is a parameter that controls the degree of anisotropy. Preferably, Can be set between 0.5 and 2.

[0053] By introducing anisotropic tensor fields, the method of the present invention can better adapt to the complex anatomical structure of the lungs and improve the accuracy of segmentation, especially in small blood vessels and bifurcations. This is of great significance for subsequent clinical diagnosis and treatment planning. The method of the present invention can not only accurately locate the boundaries of lung segments, but can also be further applied to lung tumor analysis. In a preferred embodiment of the present invention, lung tumor analysis can be performed based on the precise positioning results of the lung segment boundaries obtained. This analysis method uses the lung segment boundaries as reference boundaries for tumor staging, and combines imaging genomics analysis tools to provide comprehensive tumor pathology feature analysis information.

[0054] Specifically, the method first uses the precisely located lung segment boundaries as anatomical references to determine the exact location and extent of the tumor. Preferably, deep learning-based target detection algorithms, such as FasterR-CNN or YOLO, can be used to automatically detect and locate lung tumors. These algorithms can use lung segment boundary information as additional feature input during training to improve detection accuracy.

[0055] After determining the location of the tumor, the method of the present invention combines imaging omics analysis tools to extract multi-dimensional features of the tumor. These features may include but are not limited to: 1. Morphological characteristics: such as tumor volume, surface area, maximum diameter, sphericity, etc.

[0056] 2. Density characteristics: such as average CT value, CT value distribution histogram, etc.

[0057] 3. Texture features: such as gray-level co-occurrence matrix (GLCM) features, run-length matrix (RLRM) features, etc.

[0058] 4. Wavelet features: multi-scale features extracted through wavelet transform.

[0059] Preferably, feature selection methods such as principal component analysis (PCA) or LASSO can be used to select the most discriminative feature subset from a large number of features. This can effectively reduce feature redundancy and improve the efficiency and accuracy of subsequent analysis.

[0060] Based on the extracted features, the method can construct a machine learning model to predict the malignancy, stage and possible prognosis of the tumor. In one embodiment of the present invention, a prediction model can be constructed using algorithms such as random forest or support vector machine (SVM). The input of these models is the extracted radiomics features, and the output can be the TNM stage, malignancy score, etc. of the tumor.

[0061] In order to improve the reliability of prediction, the method of the present invention can also integrate clinical information, such as the patient's age, gender, smoking history, etc. This information can be input into the prediction model as additional features to obtain more comprehensive and accurate analysis results.

[0062] Through the above steps, the method of the present invention can not only accurately locate the boundaries of lung segments, but also provide important reference information for the diagnosis and treatment decisions of lung tumors. This method combining anatomical positioning and imaging genomics analysis is expected to significantly improve the accuracy of lung cancer diagnosis and the effectiveness of individualized treatment.

[0063] In another embodiment of the present invention, the output step further includes visually displaying the precise positioning result of the lung segment boundary and generating a three-dimensional model including the lung segment boundary. Such visualization and three-dimensional modeling can not only intuitively display the segmentation result, but also provide important reference for doctors' diagnosis and surgical planning.

[0064] Specifically, the visualization display can use multi-planar reconstruction (MPR) technology to simultaneously display the lung segment boundaries in the axial, coronal and sagittal planes. Preferably, different colors can be used to identify different lung segments to enhance the visual effect. For example, the following color scheme can be used: Right upper lung segment: red; Right middle lung segment: green; Right lower lung segment: blue; Left upper lung segment: yellow; Left lower lung segment: purple; On the display interface, an interactive adjustment function can be provided to allow the user to adjust the window width and window position to suit different observation requirements. In addition, a measurement tool can be provided to enable the doctor to accurately measure the size and distance of the area of ​​interest.

[0065] For the generation of the three-dimensional model, the method of the present invention preferably uses a surface reconstruction technique. First, the surface point cloud of each lung segment is extracted based on the segmentation result. Then, an initial triangular mesh model is generated using an algorithm such as Marching Cubes. In order to reduce the complexity of the model and smooth the surface, mesh simplification and smoothing algorithms such as Laplacian smoothing can be further applied.

[0066] The generated 3D model should support rotation, scaling, and translation operations, so that users can observe the morphology of the lung segment from any angle. Preferably, a semi-transparent display function can also be provided to make the internal structure visible. In addition, other relevant information can be superimposed on the 3D model, such as the location and size of the tumor, the direction of the main blood vessels and bronchi, etc.

[0067] Through this visualization and three-dimensional modeling approach, the present invention not only provides accurate lung segment boundary localization results, but also presents these results in an intuitive and interactive manner, greatly improving the usability and clinical value of the results.

[0068] In another embodiment of the present invention, the method further includes setting a user interaction interface to allow doctors to fine-tune the precise positioning results of the lung segment boundaries, and update the training data set of the deep learning model based on the fine-tuning results. This interactive feedback mechanism can continuously improve the performance of the system and make it better adapt to the image characteristics of different hospitals and different equipment.

[0069] The user interface preferably adopts an intuitive graphical design, which mainly includes the following functions: 1. Boundary editing: allows doctors to modify lung segment boundaries directly on the image using a mouse or stylus. Tools such as brushes and erasers can be provided to make the editing process more convenient.

[0070] 2. Marking tools: Provide a series of marking tools, such as arrows, text annotations, etc., so that doctors can mark special anatomical structures or lesion areas.

[0071] 3. Comparative display: The original segmentation results and the manually corrected results are displayed simultaneously to facilitate comparison and evaluation by doctors.

[0072] 4. Undo / Redo: Provides multi-level undo and redo functions to make the editing process more flexible.

[0073] 5. Automatic optimization: Based on the doctor’s local modifications, the system can automatically optimize the segmentation results of adjacent areas and reduce the workload of manual operations.

[0074] When the doctor completes the fine-tuning, the system will record the original segmentation results, the manual correction results, and the corresponding original image data. These data will be used to update the training data set of the deep learning model. Preferably, online learning or incremental learning methods can be used to enable the model to continuously adapt to new data.

[0075] The process of updating the training dataset can be done as follows: 1. Data enhancement: Perform appropriate data enhancement on the results of manual correction, such as rotation, scaling, adding noise, etc., to increase the diversity of the data.

[0076] 2. Sample weight adjustment: Assign higher weights to newly added samples so that the model pays more attention to these new samples that have been corrected by experts.

[0077] 3. Model fine-tuning: Use the updated dataset to fine-tune the existing model. A smaller learning rate can be used to maintain the stability of the model.

[0078] 4. Verification and deployment: Verify the fine-tuned model to ensure its performance is improved before deploying it to actual applications.

[0079] Through this interactive feedback and continuous learning mechanism, the method of the present invention can continuously improve its accuracy and robustness, and better adapt to various situations in clinical practice. This not only improves the practicality of the system, but also establishes a benign cooperative relationship between doctors and AI systems, which is conducive to the promotion and application of AI-assisted diagnosis technology.

[0080] The above description is only a preferred specific implementation manner of the present invention; however, the protection scope of the present invention is not limited thereto; any person familiar with the art who, within the scope disclosed by the present invention, makes equivalent replacements or changes based on the scheme and improved concepts of the present invention shall be covered within the protection scope of the present invention.

Claims

1. A lung segment boundary accurate positioning method based on adaptive watershed growing, characterized in that: include: The acquisition steps include: Acquire lung CT image data; Processing steps include: Based on the lung CT image data, image preprocessing is performed to obtain a preprocessed lung image; Determining a set of main points of the pulmonary segmental artery according to the preprocessed lung image; Based on the trunk point set, an initial seed point is determined by adopting a sub-point intelligent arrangement strategy; According to the initial seed points, performing adaptive watershed growth based on point competition to obtain a preliminary positioning result of the lung segment boundary; Based on the preliminary positioning result of the lung segment boundary, boundary conflict detection and processing are performed to obtain an optimized lung segment boundary; Output steps include: The optimized lung segment boundary is output as the final lung segment boundary precise positioning result.

2. The method according to claim 1, characterized in that The image preprocessing specifically includes: Performing bilateral filtering on the lung CT image data to obtain a denoised lung image; Based on the denoised lung image, blood vessel extraction is performed to obtain a lung blood vessel mask.

3. The method according to claim 1, characterized in that The sub-point intelligent arrangement strategy specifically includes: Based on the anatomical structure characteristics of the pulmonary artery in the human body, determine the starting point and branching point of the aorta; Determine the anatomical landmarks based on the connection points between the main pulmonary artery and the left and right pulmonary arteries and the left and right hilum; Based on the aorta starting point, branching point and anatomical landmark point, the position of the initial seed point is determined.

4. The method according to claim 1, characterized in that: The adaptive watershed growth based on point competition specifically includes: Based on the initial seed point, construct a point group; Using Euclidean distance as a similarity parameter, unsupervised clustering is performed on the point group; According to the clustering results, determine the watershed corresponding to each class; Based on the watershed, adaptive seed growing is performed, which includes: Define the point group growth rules based on the distance between seed points, including adaptive step size and distance threshold parameters; According to the average distance between the class points and the seed points, the adaptive convergence criterion is set; By comparing the competitive scores of each point, boundaries are drawn between adjacent point classes.

5. The method according to claim 1, characterized in that The boundary conflict detection and processing specifically includes: Calculate the seed competition score between the boundary area point class and the adjacent point class area; identifying boundary conflict regions between a plurality of watersheds based on the seed competition scores; For detected boundary conflicts, the narrowest channel between two seed points is found in the conflict area; According to the narrowest channel, the point cluster in the conflict area is divided into two parts to resolve the boundary conflict.

6. The method according to claim 1, characterized in that Also includes: Based on the deep learning model, the optimized lung segment boundaries are further optimized, wherein: A convolutional neural network model is used to compare the model-predicted boundaries with the standard results; The boundary correction model is optimized through the loss function to obtain the final accurate positioning result of the lung segment boundary.

7. The method according to claim 4, characterized in that The adaptive seed growth further comprises: Anisotropic tensor fields are used to simulate the directional constraints of local expansion; Based on the anisotropic tensor field, the control point group grows along the blood vessel direction.

8. The method according to claim 1, characterized in that Also includes: Based on the precise positioning result of the lung segment boundary, lung tumor analysis is performed; Using the lung segment boundary as a reference boundary for tumor staging; Combined with imaging omics analysis tools, it provides tumor pathological feature analysis information.

9. The method according to claim 1, characterized in that: The output step further comprises: Visually displaying the precise positioning result of the lung segment boundary; A three-dimensional model including the lung segment boundaries is generated.

10. The method according to claim 1, characterized in that Also includes: A user interaction interface is provided to allow a doctor to fine-tune the precise positioning result of the lung segment boundary; Based on the fine-tuning results, the training data set of the deep learning model is updated.

Images