Method for processing overlapping regions of lung segments in multi-scale watershed analysis

Abstract

The present invention relates to the field of medical image processing technology, and more specifically, to a method for processing overlapping lung segments using multi-scale watershed analysis, comprising: acquiring a three-dimensional lung CT image; preprocessing the three-dimensional lung CT image to obtain an initial morphology of a lung segment model; setting multiple sampling distances based on the initial morphology of the lung segment model; extracting sampling points on the initial morphology of the lung segment model according to the multiple sampling distances; performing region growing based on the sampling points to form sub-watersheds of the lung segment model; performing multi-scale data fusion based on the sub-watersheds obtained at different sampling distances; performing boundary optimization on the results of the multi-scale data fusion; and outputting the optimized processing results of overlapping lung segments. By introducing a multi-scale analysis strategy, this method can simultaneously capture both large-scale structures and small-scale details of the lung. This multi-scale strategy not only improves segmentation accuracy but also enhances the robustness of the algorithm, achieving precise segmentation of overlapping lung segments.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and more particularly to a method for processing lung segment overlapping regions in multi-scale watershed analysis. Background Art

[0002] With the rapid advancement of medical imaging technology, computed tomography (CT) has become an essential tool for diagnosing and planning treatment for lung diseases. However, the complexity of lung structure, particularly the overlapping regions between lung segments, has been a challenge for accurate segmentation and analysis. Traditional lung segmentation methods primarily rely on single-scale image analysis techniques, which often struggle to handle complex lung structures.

[0003] Common lung segmentation methods currently used include threshold-based segmentation, region growing, and level set methods. These methods achieve good results when processing single lung segments or images of simple lung structures. However, their limitations become apparent when faced with complex lung structures, especially those with pathological changes or anatomical variations.

[0004] For example, while threshold-based segmentation methods are simple and fast, they are highly sensitive to image quality and noise, making it difficult to accurately process transition regions between lung segments. Region growing methods, while effective at processing continuous structures, are prone to overgrowth or undergrowth when dealing with overlapping lung segments, leading to inaccurate segmentation results. While level set methods can handle complex topological structures, they are computationally complex and highly sensitive to the choice of initial contours, making it difficult to achieve stable results in practical applications.

[0005] Furthermore, most existing methods only consider image features at a single scale, failing to fully capture the multi-scale characteristics of lung structure. Lung structure exhibits distinct multi-scale characteristics, from large-scale lung lobes to small-scale bronchioles and blood vessels. Single-scale analysis methods struggle to simultaneously maintain the integrity of large-scale structure and the accuracy of small-scale details, especially when dealing with overlapping lung segments.

[0006] Another common problem is that existing methods often lack sufficient utilization of prior knowledge of the lung's anatomy. Anatomical features such as the lung's vascular tree structure and bronchial distribution are crucial for accurately identifying and segmenting lung segments, but existing methods do not adequately consider this aspect.

[0007] Therefore, there is an urgent need for a segmentation method that can comprehensively consider the multi-scale characteristics of the lung, fully utilize anatomical prior knowledge, and effectively handle overlapping lung segments. Such a method should be robust and adaptable while ensuring segmentation accuracy, and be able to cope with various complex clinical scenarios. Summary of the Invention

[0008] This paper addresses these technical issues by proposing a method for processing overlapping lung segments using multi-scale flow analysis. This method innovatively combines multi-scale analysis with an improved region growing algorithm, a probabilistic model, and boundary optimization techniques to effectively address the difficulties faced by existing technologies in processing complex lung structures, particularly overlapping lung segments.

[0009] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0010] The method for processing overlapping areas of lung segments in multi-scale watershed analysis includes:

[0011] The acquisition steps include:

[0012] Acquire three-dimensional lung CT images;

[0013] Preprocessing the three-dimensional lung CT image to obtain an initial shape of a lung segment model;

[0014] Processing steps include:

[0015] Setting a plurality of sampling distances based on the initial shape of the lung segment model;

[0016] extracting sampling points on the initial form of the lung segment model according to the multiple sampling distances;

[0017] Based on the sampling points, region growing is performed to form a sub-watershed of the lung segment model;

[0018] Multi-scale data fusion is performed based on the sub-basins obtained at different sampling distances;

[0019] Output steps include:

[0020] performing boundary optimization on the result of the multi-scale data fusion;

[0021] Output the optimized lung segment overlapping area processing results.

[0022] Preferably, preprocessing the three-dimensional lung CT image specifically includes:

[0023] performing filtering and smoothing processing on the three-dimensional lung CT image;

[0024] Based on the CT value, the three-dimensional lung CT image is binarized using the maximum inter-class variance method to obtain a lung tissue mask;

[0025] Blood vessels are extracted from the three-dimensional lung CT image to obtain a preliminary blood vessel casting model.

[0026] Preferably, the setting of multiple sampling distances specifically includes:

[0027] Set the first sampling ratio and the second sampling ratio;

[0028] The first sampling ratio and the second sampling ratio are adjusted according to the average radius of different bronchi in the initial form of the lung segment model.

[0029] Preferably, performing region growing specifically includes:

[0030] Each sampling point is used as a seed point;

[0031] Based on a preset growth rule, performing region growing on the seed point;

[0032] A sub-flow domain of the lung segment model including the seed point is formed.

[0033] Preferably, the preset growth rules include:

[0034] Expand the sampling points according to different radii to obtain growth points;

[0035] determining whether the growth point is included in the lung segment model;

[0036] If it is not included in the lung segment model, it will continue to be expanded according to different radii;

[0037] If included in the lung segment model, the sub-watershed of the lung segment model is obtained.

[0038] Preferably, the multi-scale data fusion specifically includes:

[0039] Acquire a sub-watershed set based on a first sampling ratio and a sub-watershed set based on a second sampling ratio;

[0040] Performing an intersection operation on the sub-watershed set of the first sampling ratio and the sub-watershed set of the second sampling ratio;

[0041] The intersection of sub-basins under the joint action of the two sampling ratios is obtained.

[0042] Preferably, performing boundary optimization on the result of the multi-scale data fusion specifically includes:

[0043] Construct an energy function that includes parameters of region similarity, fusion degree, and region boundary smoothness;

[0044] Based on the energy function, performing boundary optimization calculation;

[0045] Get the optimal region boundary.

[0046] Preferably, the energy function is defined as:

[0047] ,

[0048] Among them, C is the region boundary, R(C) is the region similarity, F(C) is the degree of regional fusion, S(C) is the boundary smoothness, and α, β, and γ are control factors.

[0049] Preferably, the method further includes the step of constructing a probabilistic vascular tree model:

[0050] Based on probability functions and Boolean algebra set theory, the pulmonary vascular tree is transformed into a probability space;

[0051] Conduct probability distribution analysis on the diameter, shape and distribution of the vascular tree;

[0052] Generate a probabilistic model of the vascular tree.

[0053] Preferably, the vascular tree probability model is expressed as:

[0054] ,

[0055] in, represents the vascular tree, Represents image data, Indicates the diameter, Indicates shape, Indicates the distribution.

[0056] The method of the present invention has the following significant technical effects:

[0057] First, by introducing a multi-scale analysis strategy, this method can simultaneously capture both large-scale structure and small-scale details of the lung. This multi-scale strategy not only improves segmentation accuracy but also enhances the robustness of the algorithm. When dealing with complex lung structures, this method can maintain the integrity of large-scale structures while accurately identifying small-scale details, thereby achieving precise segmentation of overlapping lung segments.

[0058] Secondly, the improved region growing algorithm proposed in this paper combines grayscale information, gradient features, and shape constraints to significantly improve segmentation accuracy and stability. This multi-factor integrated region growing strategy can effectively avoid over- and under-growing issues, particularly when dealing with overlapping lung segments, resulting in more accurate segmentation results.

[0059] Furthermore, the probabilistic vascular tree model introduced in this paper provides powerful prior knowledge for the segmentation process. This probabilistic model, based on anatomical features, not only improves the algorithm's accuracy but also enhances its adaptability to individual differences and pathological conditions. This characteristic is particularly important when dealing with abnormal or diseased lungs, significantly improving the reliability of segmentation results.

[0060] Furthermore, the boundary optimization technology employed in this invention achieves fine-tuning of the segmentation results by comprehensively considering regional similarity, degree of fusion, and boundary smoothness. This not only improves the accuracy of the segmentation boundaries but also enhances the visual quality and clinical usability of the results.

[0061] Finally, significant synergy exists between the various components of our method. Multiscale analysis provides reliable seed points for region growing, the probabilistic model guides the growing process, and boundary optimization further improves the quality of the final result. This synergy enables our method to excel in handling complex lung structures, particularly in regions with overlapping lung segments, achieving results far superior to existing techniques.

[0062] In summary, the multi-scale flow domain analysis method presented in this paper provides a comprehensive and efficient solution for accurately processing overlapping lung segments. This method not only overcomes the limitations of existing technologies but also achieves significant improvements in accuracy, robustness, and adaptability. This has significant implications for improving the diagnostic accuracy of lung diseases, optimizing surgical planning, and enhancing treatment assessment, bringing new possibilities to related clinical practice. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 The figure is an overall flow chart of the method of the present invention.

[0064] Figure 2 This is a pretreatment flow chart of the present invention.

[0065] Figure 3 The flowchart of the sampling distance setting of the present invention is shown in FIG.

[0066] Figure 4 This is a flow chart of region growing of the present invention.

[0067] Figure 5 This is a flowchart of the multi-scale data fusion of the present invention.

[0068] Figure 6 This is a flow chart of boundary optimization of the present invention. DETAILED DESCRIPTION

[0069] like Figure 1-6 As shown in the figure, the present invention provides a method for processing overlapping lung segments using multi-scale watershed analysis. This method aims to address the problem of insufficient accuracy of traditional lung segmentation techniques when processing overlapping regions. By combining multi-scale analysis with a watershed algorithm, it achieves accurate segmentation of complex lung structures.

[0070] First, the method of the present invention involves obtaining a three-dimensional lung CT image. In the field of medical imaging, CT (computed tomography) technology is an important means of obtaining three-dimensional information about the human body's internal structures. The present invention preferably utilizes spiral CT scanning, which can produce high-resolution, low-noise raw image data. For example, a 64-slice or 128-slice multislice spiral CT machine can be used, with scanning parameters set to: tube voltage 120 kV, tube current 200 mA, slice thickness 1 mm, and interval 0.8 mm. This parameter setting ensures image quality while keeping radiation dose within a reasonable range.

[0071] Next, the present invention preprocesses the three-dimensional lung CT image to obtain the initial shape of the lung segment model. Preprocessing is a key step in image analysis and can significantly improve the accuracy and efficiency of subsequent processing. In one embodiment of the present invention, preprocessing includes the following steps:

[0072] 1. Image denoising: A 3D Gaussian filter is used to denoise the original CT image. The standard deviation σ of the Gaussian filter can be dynamically adjusted based on the image noise level, typically ranging from 0.5 to 2.

[0073] 2. Grayscale normalization: Convert the CT value (Hounsfield Unit, HU) to the range of [-1000, 1000] to facilitate subsequent threshold processing. The conversion formula is:

[0074] ,

[0075] in is the normalized image, is the original image.

[0076] 3. Lung parenchyma segmentation: The improved maximum inter-class variance method (OTSU algorithm) is used to perform preliminary segmentation of the lung parenchyma. The objective function of the OTSU algorithm can be expressed as:

[0077] ,

[0078] in, is the between-class variance, is the threshold, In the present invention, considering the characteristics of lung CT images, the initial threshold of the OTSU algorithm is set to -500HU, which can converge to the optimal threshold more quickly.

[0079] 4. Morphological processing: Open and close the segmentation results to remove small noise areas and fill small holes. The size of the structural element is usually selected as a 3×3×3 cube.

[0080] Through the above preprocessing steps, the initial shape of the lung segment model is obtained. This initial shape is the basis for subsequent multi-scale analysis, and it contains the general outline and main structural information of the lung.

[0081] After obtaining the initial morphology, the method of the present invention sets multiple sampling distances based on the initial morphology of the lung segment model. The selection of sampling distances directly affects the accuracy and computational efficiency of subsequent analysis. The present invention preferably adopts an adaptive multi-scale sampling strategy. Specifically, three sampling distances are set according to different levels of lung structure:

[0082] 1. Large-scale sampling: used to capture structures at the lobe level, with the sampling distance set to 1 / 10 of the maximum lung diameter.

[0083] 2. Medium-scale sampling: used to capture the structure at the lung segment level, and the sampling distance is set to 1 / 3 of the large-scale sampling distance.

[0084] 3. Small-scale sampling: used to capture small bronchi and blood vessels, with the sampling distance set to 1 / 2 of the medium-scale sampling distance.

[0085] This multi-scale sampling strategy can take into account both large-scale structures and local details, which is conducive to subsequent accurate segmentation.

[0086] Based on the multiple sampling distances, the present invention extracts sampling points on the initial morphology of the lung segment model. The sampling points are extracted using an improved Poisson disk sampling algorithm, which can ensure that the sampling points are evenly distributed while avoiding excessive density of sampling points. The main steps of the algorithm are as follows:

[0087] 1. Initialization: Randomly select a point on the model surface as the starting point.

[0088] 2. Generate candidate points: With the current point as the center, generate k candidate points uniformly on the sphere (k is usually 30).

[0089] 3. Verify candidate points: Check whether each candidate point meets the minimum distance constraint.

[0090] 4. Select a valid point: Randomly select one of the candidate points that meet the constraints as a new sampling point.

[0091] 5. Iteration: Repeat steps 2-4 until no new sampling points can be added.

[0092] Through this method, a series of evenly distributed sampling points can be obtained at different scales, which will serve as seed points for subsequent region growth.

[0093] Next, the present invention performs region growing based on the sampling points to form a sub-watershed of the lung segment model. Region growing is a classic image segmentation method, and the present invention improves it to make it more suitable for the characteristics of lung CT images. The improved region growing algorithm includes the following steps:

[0094] 1. Initialization: Set the sampling point as the seed point and initialize an empty growth area.

[0095] 2. Neighborhood check: For the 26 neighboring pixels of the seed point, calculate their similarity with the seed point. The similarity function is defined as:

[0096] ,

[0097] in, is the seed point, is the neighborhood pixel, and are their grayscale values, It is an adjustable parameter, usually set to 10-20.

[0098] 3. Growth judgment: If the similarity is greater than the threshold T (usually 0.7-0.9), the neighborhood pixels are added to the growth area.

[0099] 4. Update: Use the newly added pixels as new seed points.

[0100] 5. Iteration: Repeat steps 2-4 until no new pixels can be added.

[0101] Through this improved region growing algorithm, a series of sub-watersheds can be obtained, which represent the lung segment structures at different scales.

[0102] Finally, the present invention performs multi-scale data fusion based on the sub-watersheds obtained at different sampling distances. The fusion process adopts a weighted voting method, and the weight of each scale is dynamically adjusted according to its corresponding sampling distance. The fusion function can be expressed as:

[0103] ,

[0104] in, is the result after fusion, is the sub-basin of the i-th scale, is the corresponding weight, satisfying The weight calculation formula is:

[0105] ,

[0106] in, is the sampling distance of the i-th scale, To adjust the parameter, the value is usually taken as the average value of the sampling distance.

[0107] This multi-scale fusion method can produce a final segmentation result that integrates information from different scales. This result not only preserves large-scale structural information but also includes small-scale detail features, thereby achieving accurate processing of overlapping lung segments.

[0108] In a preferred embodiment of the present invention, the preprocessing step of 3D lung CT images is further refined and optimized. This preprocessing step is crucial for subsequent lung segment analysis because it directly affects the accuracy of segmentation and the overall performance of the algorithm.

[0109] First, the method filters and smoothes the three-dimensional lung CT image. The purpose of this step is to reduce image noise and improve the stability of subsequent processing. Preferably, a three-dimensional Gaussian filter can be used for processing. The standard deviation σ of the Gaussian filter can be dynamically adjusted according to the noise level of the image, and is usually between 0.5 and 2. For example, for images with high noise, a larger σ value, such as 1.5, can be selected; while for high-quality images with low noise, a smaller σ value, such as 0.8, can be selected. This adaptive filtering strategy can effectively suppress noise while preserving image details.

[0110] Next, this method uses the maximum inter-class variance method to binarize the 3D lung CT image based on the CT value to obtain the lung tissue mask. The maximum inter-class variance method, also known as the OTSU algorithm, is a classic image segmentation method. In this invention, the OTSU algorithm is improved to make it more suitable for the characteristics of lung CT images. Specifically, the objective function of the algorithm can be expressed as:

[0111] ,

[0112] in, is the inter-class variance, t is the threshold, and L is the number of grayscale levels. Considering the characteristics of lung CT images, this method sets the initial threshold of the OTSU algorithm to -500 HU. This initial value, selected based on statistical analysis of extensive clinical data, allows the algorithm to converge to the optimal threshold more quickly. This improvement allows the method to more accurately segment lung tissue regions, laying the foundation for subsequent analysis.

[0113] Finally, this method extracts blood vessels from the 3D lung CT image to obtain a preliminary vascular cast. Blood vessel extraction is a complex process, and this method employs a multiscale vessel enhancement filter based on the Hessian matrix. This method effectively enhances vascular structure while suppressing interference from other tissues. Specifically, for each voxel in the image, its Hessian matrix is calculated:

[0114] ,

[0115] in, 、 etc. represent the second-order partial derivatives of the image in the corresponding direction. By analyzing the eigenvalues of the Hessian matrix, we can determine whether the point belongs to the vascular structure. The vascular enhancement function can be defined as:

[0116] ,

[0117] in, 、 、 is the eigenvalue of the Hessian matrix ( ), 、 and is the shape index calculated based on these eigenvalues, and is an adjustable parameter. By applying this filter at multiple scales, vascular structures of different sizes can be enhanced.

[0118] Through the above preprocessing steps, this method can obtain high-quality lung tissue masks and preliminary vascular castings, providing a reliable basis for subsequent detailed analysis.

[0119] In the process of setting multiple sampling distances, the method of the present invention adopts an adaptive strategy. Specifically, the method sets a first sampling ratio and a second sampling ratio. The setting of these two sampling ratios directly affects the accuracy and efficiency of subsequent analysis.

[0120] Preferably, the first sampling ratio can be set to a larger value to capture the large-scale structure of the lung. For example, the first sampling ratio can be set to 1 / 10 of the maximum diameter of the lung. This ratio is selected based on statistical analysis of a large amount of lung CT data and can effectively capture the contours of the lung lobes and major lung segments.

[0121] The second sampling ratio is set to a smaller value to capture more detailed structures. Preferably, the second sampling ratio can be set to 1 / 3 to 1 / 2 of the first sampling ratio. This ratio can effectively capture small lung segments and fine bronchial structures.

[0122] An innovative aspect of this method is that the first and second sampling ratios are dynamically adjusted based on the average radius of the different bronchi in the initial morphology of the lung segment model. This adaptive adjustment strategy can better adapt to the individual differences of different patients. Specifically, the adjustment process can be expressed as:

[0123] ,

[0124] ,

[0125] and are the adjusted first and second sampling ratios, is the mean radius of the bronchus, and is the adjustment coefficient. Usually, The value can be between 8 and 12. The value can be between 3 and 5. The specific values of these parameters can be determined by statistical analysis of a large amount of clinical data.

[0126] Through this adaptive sampling strategy, this method can achieve good sampling effects in different lung structures, providing a reliable basis for subsequent region growing.

[0127] During the region growing process, the method of the present invention adopts an improved strategy to better adapt to the characteristics of lung CT images. First, the method uses each sampling point as a seed point. These seed points are obtained based on the multi-scale sampling strategy described above, so their distribution reflects both the large-scale structure of the lung and the local details.

[0128] Next, this method performs region growing on these seed points based on a preset growth rule. This preset growth rule is a key innovation of this method, which comprehensively considers multiple factors such as CT value, gradient information, and shape constraints. Specifically, the growth rule can be expressed as a cost function:

[0129] ,

[0130] in, Represents pixels The difference between the CT value of the seed point and the CT value of the express The gradient size of the point, is a shape constraint term used to maintain the smoothness of the growing region. 、 and are weight coefficients, and their sum is 1. These weights can be adjusted according to the specific application scenario. For example, for normal lungs with clear boundaries, the weight of the gradient term can be increased; while for lungs with lesions, the weight of the shape constraint term may need to be increased.

[0131] In each iteration, the method calculates the cost function value of all pixels in the neighborhood of the seed point and adds the pixel with the lowest cost to the growing area. This process continues until no new pixels can be added or the preset maximum number of iterations is reached.

[0132] Through this improved region growing strategy, this method can form sub-watersheds of the lung segment model including the seed points. These sub-watersheds accurately reflect the local structure of the lung segment and lay the foundation for subsequent multi-scale fusion.

[0133] In the specific implementation of the preset growth rules, the method of the present invention adopts a multi-step strategy to ensure the accuracy and stability of the growth process. First, the method expands the sampling points according to different radii to obtain growth points. The purpose of this step is to explore the local area around the sampling points and provide more information for subsequent judgment.

[0134] Specifically, the expansion process can be expressed as:

[0135] ,

[0136] in, are the coordinates of the sampling points, is the expansion radius. This method usually uses multiple different radii, for example, you can set mm, mm and mm. This multi-radius strategy can capture local structural information at different scales.

[0137] Next, the method determines whether the acquired growth point is included in the lung segment model. This determination process is based on the cost function described above. , calculate its cost function value .if Less than the preset threshold , then the point is considered to be included in the lung segment model. Threshold The choice of is very important, it directly affects the range and accuracy of growth.

[0138] Preferably, It can be set to 1.2 to 1.5 times the cost function value of the seed point. If the judgment result shows that the growth point is not included in the lung segment model, this method will continue to expand according to different radii. This process can be expressed as:

[0139] ,

[0140] in, is the expansion factor, which is usually between 1.2 and 1.5. This progressive expansion strategy ensures that no region that may belong to a lung segment is missed.

[0141] If the judgment result shows that the growth point is included in the lung segment model, the method considers that a sub-watershed of the lung segment model has been obtained, which accurately reflects the local lung segment structure centered at the sampling point.

[0142] By using this meticulously predefined growth rule, the method can effectively control computational complexity while ensuring accuracy. For a typical 512×512×300 CT dataset, the entire region growing process can typically be completed within 30 seconds, making it possible for practical clinical applications.

[0143] In summary, the multi-scale flow domain analysis method proposed in this paper provides a new and efficient solution for accurately processing overlapping lung segments. Through rational preprocessing, adaptive sampling strategies, and an improved region growing algorithm, this method can accurately identify and segment complex lung structures, providing a reliable imaging basis for subsequent clinical diagnosis and treatment planning.

[0144] In another preferred embodiment of the present invention, the multi-scale data fusion process is further optimized and refined. This fusion process is one of the core steps of the method and directly affects the accuracy and reliability of the final lung segment overlap region processing.

[0145] Specifically, this method first obtains a set of sub-watersheds based on a first sampling ratio and a set of sub-watersheds based on a second sampling ratio. These two sets represent lung segment structural information at different scales. The sub-watersheds corresponding to the first sampling ratio generally reflect larger-scale structural features, such as the outlines of lung lobes and major lung segments; while the sub-watersheds corresponding to the second sampling ratio capture more detailed structural information, such as small lung segments and delicate bronchial branches.

[0146] Next, this method performs an intersection operation on these two sub-watershed sets. The purpose of this step is to find areas that are recognized as the same structure at different scales, thereby improving the reliability of segmentation. The intersection operation can be expressed as:

[0147] ,

[0148] Among them, R is the final fusion result, and are the sub-basin sets at the first and second sampling scales, respectively.

[0149] However, simply taking the intersection may lead to the loss of some important structures. To solve this problem, this method introduces a weighted fusion strategy. Specifically, a fusion function can be defined:

[0150] ,

[0151] in, Respectively represent the degree of belonging of point x at the first and second sampling ratios (0 or 1), and is the weight coefficient, and + = 1. The weight selection can be adjusted according to the specific application scenario. For example, for applications that require accurate recognition of small structures, you can increase For applications that require more stable large-scale segmentation, the value of value.

[0152] Through this weighted fusion strategy, our method can obtain the intersection of sub-watersheds under the combined effect of the two sampling ratios. This result not only preserves large-scale structural information but also includes small-scale details, thereby achieving accurate processing of overlapping lung segments.

[0153] In practical applications, this method can further introduce multiple sampling ratios to form a multi-scale pyramid structure. This structure can be expressed as:

[0154] ,

[0155] Where n is the number of sampling ratios, is the sub-basin set under the i-th sampling ratio, This multi-scale fusion strategy can capture the characteristics of lung structure more comprehensively and further improve the accuracy and robustness of segmentation.

[0156] A key innovation of the present invention is to optimize the boundaries of the multi-scale data fusion results. This step aims to obtain smoother and more natural lung segment boundaries, improving the visual quality and clinical usability of the segmentation results.

[0157] To achieve this goal, this method constructs a comprehensive energy function. This energy function takes into account multiple factors, including region similarity, fusion degree, and region boundary smoothness. Specifically, the energy function can be expressed as:

[0158] ,

[0159] in, represents the region edge, is the regional similarity term, is the boundary smoothness term.

[0160] Regional similarity item Used to ensure the consistency of the segmentation result with the original image, it can be defined as:

[0161] ,

[0162] Among them, I (x) is the gray value of the image at point x, and are the average grayscale values inside and outside the region, respectively.

[0163] The regional fusion degree term F(C) is used to measure the consistency between the boundary and the multi-scale fusion result, which can be defined as:

[0164] ,

[0165] Where H is the Heaviside function, is the level set function, and F(x) is the multi-scale fusion result.

[0166] The boundary smoothness term S(C) is used to control the smoothness of the boundary and can be defined as:

[0167] ,

[0168] That is, the length of the boundary.

[0169] Based on this energy function, this method uses the variational level set method for boundary optimization. Specifically, the level set function can be updated by solving the following partial differential equation: : ,

[0170] in, is the Dirac function, is the curvature.

[0171] By iteratively solving this equation, this method can obtain the optimal regional boundary. This boundary not only accurately reflects the structure of the lung segment but also has good smoothness, which is beneficial for subsequent clinical applications.

[0172] In practical applications, the control factor 、 and The choice of is very important. Preferably, these parameters can be dynamically adjusted according to the specific application scenario. For example, for lesions of the lung with fuzzy boundaries, the value of \beta can be appropriately increased to enhance the impact of the multi-scale fusion result; for applications that need to preserve fine structures, the value of \beta can be reduced. , reducing the smoothness constraint.

[0173] The method of the present invention also includes the step of constructing a probabilistic vascular tree model, which is a key innovation. Based on probability functions and Boolean algebraic set theory, this step transforms the pulmonary vascular tree into a probabilistic space, providing powerful prior knowledge for subsequent analysis.

[0174] Specifically, this method first performs a probability distribution analysis on the diameter, shape, and distribution of the vascular tree. This analysis can be based on a large amount of clinical data and use statistical methods to obtain a probability distribution model for each feature. For example, the distribution of vascular diameter may conform to a lognormal distribution:

[0175] ,

[0176] Where D is the diameter of the blood vessel, and is the distribution parameter.

[0177] Similarly, the shape and distribution of blood vessels can also be described by appropriate probability models. For example, the branching angle of blood vessels may follow a Beta distribution, while the blood vessel density may follow a Poisson distribution.

[0178] Based on these probability models, this method generates a comprehensive vascular tree probability model. This model can be expressed as:

[0179] ,

[0180] Where T represents the vascular tree, I represents the image data, D represents the diameter, S represents the shape, and L represents the distribution. P(T) is the prior probability of the vascular tree, which can be set based on anatomical knowledge.

[0181] An important application of this probabilistic model is providing guidance during region growing. For example, when determining whether a voxel belongs to a blood vessel, in addition to considering its grayscale value and gradient information, the probability that it conforms to the vascular tree model can also be calculated. This effectively reduces missegmentation and improves the accuracy and robustness of the algorithm.

[0182] In general, the multi-scale watershed analysis method provided by the present invention combines an improved region growing algorithm, a multi-scale fusion strategy, boundary optimization technology, and a probabilistic vascular tree model to provide a comprehensive and efficient solution for the precise processing of overlapping lung segments. This method can not only accurately identify and segment complex lung structures, but also has good adaptability and robustness, and can cope with various clinical scenarios. Through this method, a more reliable imaging basis can be provided for the diagnosis, surgical planning, and treatment evaluation of lung diseases, thereby improving the accuracy and efficiency of related clinical decisions.

[0183] The above description is only a preferred specific embodiment of the present invention; however, the protection scope of the present invention is not limited thereto; any person familiar with the art who makes equivalent replacements or changes based on the scheme and improved concepts of the present invention within the scope disclosed by the present invention shall be covered by the protection scope of the present invention.

Claims

1. A lung segment overlapping region processing method for multi-scale watershed analysis, characterized in that: include: The acquisition steps include: Acquire three-dimensional lung CT images; Preprocessing the three-dimensional lung CT image to obtain an initial shape of a lung segment model; Processing steps include: Setting a plurality of sampling distances based on the initial shape of the lung segment model; extracting sampling points on the initial form of the lung segment model according to the multiple sampling distances; Based on the sampling points, region growing is performed to form a sub-watershed of the lung segment model; Multi-scale data fusion is performed based on the sub-watersheds obtained at different sampling distances; Output steps include: performing boundary optimization on the result of the multi-scale data fusion; Output the optimized lung segment overlapping area processing results; The setting of multiple sampling distances specifically includes: Set the first sampling ratio and the second sampling ratio; adjusting the first sampling ratio and the second sampling ratio according to the average radius of different bronchi in the initial form of the lung segment model; The multi-scale data fusion specifically includes: Acquire a sub-watershed set based on a first sampling ratio and a sub-watershed set based on a second sampling ratio; Performing an intersection operation on the sub-watershed set of the first sampling ratio and the sub-watershed set of the second sampling ratio; The intersection of sub-basins under the joint action of the two sampling ratios is obtained.

2. The method according to claim 1, characterized in that Preprocessing the three-dimensional lung CT image specifically includes: performing filtering and smoothing processing on the three-dimensional lung CT image; Based on the CT value, the three-dimensional lung CT image is binarized using the maximum inter-class variance method to obtain a lung tissue mask; Blood vessels are extracted from the three-dimensional lung CT image to obtain a preliminary blood vessel casting model.

3. The method according to claim 1, characterized in that The region growing specifically includes: Each sampling point is used as a seed point; Based on a preset growth rule, performing region growing on the seed point; A sub-flow domain of the lung segment model including the seed point is formed.

4. The method according to claim 3, characterized in that The preset growth rules include: Expand the sampling points according to different radii to obtain growth points; determining whether the growth point is included in the lung segment model; If it is not included in the lung segment model, it will continue to be expanded according to different radii; If included in the lung segment model, the sub-watershed of the lung segment model is obtained.

5. The method according to claim 1, wherein Boundary optimization of the result of the multi-scale data fusion specifically includes: Construct an energy function that includes parameters of region similarity, fusion degree, and region boundary smoothness; Based on the energy function, performing boundary optimization calculation; Get the optimal region boundary.

6. The method according to claim 5, characterized in that The energy function is defined as: ， Among them, C is the region boundary, R(C) is the region similarity, F(C) is the degree of regional fusion, S(C) is the boundary smoothness, and α, β, and γ are control factors.

7. The method according to claim 1, characterized in that It also includes the steps of building a probabilistic vascular tree model: Based on probability functions and Boolean algebra set theory, the pulmonary vascular tree is transformed into a probability space; Conduct probability distribution analysis on the diameter, shape and distribution of the vascular tree; Generate a probabilistic vascular tree model.

8. The method according to claim 7, characterized in that The probabilistic vascular tree model is expressed as: ， in, represents the vascular tree, represents image data, Indicates the diameter, Indicates shape, Indicates the distribution.

Images