Lung segment boundary positioning method for adaptive watershed

Abstract

The invention relates to the technical field of medical image processing, in particular to a lung segment boundary positioning method for a self-adaptive watershed, and the method comprises the steps: obtaining a radiography or high-resolution CT image of a lung; on the basis of the lung image after gray level normalization processing, lung segment artery segmentation is carried out; an optimized lung segment segmentation result is obtained; obtaining the innermost layer boundary of the lung lobe according to the innermost layer lung segment boundary, and obtaining the outermost layer boundary of the lung lobe according to the outermost layer lung segment boundary; determining a gray scale segmentation threshold value corresponding to the outermost layer boundary and a gray scale segmentation threshold value corresponding to the innermost layer boundary; according to the gray scale segmentation threshold value corresponding to the outermost layer boundary, determining the average gray scale of an area between the outermost layer boundary and the innermost layer boundary of the lung lobe, and according to the average gray scale, determining the gray scale gradient range of the area between the outermost layer boundary and the innermost layer boundary of the lung lobe; and the distribution information of the pulmonary segment artery is utilized to guide the segmentation process, so that the segmentation accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and more specifically to a lung segment boundary positioning method based on adaptive watershed division. Background Art

[0002] With the rapid development of medical imaging technology, computer-aided diagnosis plays an increasingly important role in the early detection and precise treatment of lung diseases. Among them, the precise positioning of lung segment boundaries is one of the key links to achieve accurate diagnosis and personalized treatment. However, due to the complexity and individual differences of human lung structure, as well as the noise and artifacts of medical images themselves, accurately positioning lung segment boundaries has always been a huge challenge in the field of medical image processing.

[0003] Traditional lung segmentation methods mainly rely on fixed thresholds or simple region growing algorithms, which often perform poorly when processing images from different patients or under different scanning conditions. Although the introduction of anatomical knowledge-based methods and machine learning techniques in recent years has improved the segmentation effect to a certain extent, there are still some problems that need to be solved. For example, existing methods are highly dependent on image quality, and the segmentation accuracy is significantly reduced when processing low-dose CT or lung images with severe lesions. In addition, most methods lack sufficient consideration of the anatomical structure of the lung, resulting in insufficient anatomical rationality of the segmentation results.

[0004] Another common problem is the lack of algorithm adaptability. Since the lung structures of different patients vary, algorithms with fixed parameters are difficult to adapt to this variability. Although some researchers have tried to introduce adaptive mechanisms, they often only consider the grayscale features of the image and ignore important information such as shape and topology. This leads to insufficient robustness of the algorithm when dealing with complex cases, and it is easy to over-segment or under-segment.

[0005] In addition, existing methods also face challenges in terms of computational efficiency and clinical practicality. Some high-precision algorithms often require a lot of computing resources and time, which makes it difficult to meet the needs of clinical real-time diagnosis. Some fast algorithms often sacrifice accuracy and cannot provide a sufficiently reliable basis for accurate diagnosis.

[0006] Faced with these challenges, there is an urgent need for a lung segment boundary localization method that can comprehensively consider image features and anatomical structure, is highly adaptable and robust, and takes into account computational efficiency. Summary of the invention

[0007] The present invention proposes a lung segment boundary positioning method based on adaptive watershed division, which effectively solves the problems of insufficient accuracy, poor adaptability, unclear anatomical significance, etc. existing in the prior art.

[0008] In order to solve the above technical problems, the present invention adopts the following technical solutions: A lung segment boundary localization method based on adaptive watershed division, comprising: The acquisition steps include: Obtain contrast or high-resolution CT images of the lungs; Perform grayscale normalization on lung images; Processing steps include: Performing pulmonary artery segmentation based on the grayscale normalized lung image; Determine the pulmonary lobe division diagram of the four lung regions of the left lung and the three lung regions of the right lung according to the connection relationship between the pulmonary segmental arteries; Based on the lung lobe division map, a lung lobe division table is obtained; According to the pulmonary lobe division map, determining the left and right affiliation of each branch artery to form a pulmonary lobe marking matrix; Inside each lung lobe, for each branch artery, the corresponding connected area is obtained by region growing with the branch artery as the main trunk; Using the connected area as the initial segmentation result of the lung segment corresponding to the branch artery; According to the connection between the lung segments and the lung lobe division table, adjacent lung segments are merged and optimized to obtain an optimized lung segmentation result; For any lobe, determine its innermost and outermost segment boundaries; The innermost boundary of the lung lobe is obtained according to the innermost lung segment boundary, and the outermost boundary of the lung lobe is obtained according to the outermost lung segment boundary; Determine a grayscale segmentation threshold corresponding to the outermost boundary and a grayscale segmentation threshold corresponding to the innermost boundary; Determine the average grayscale of the area between the outermost boundary and the innermost boundary of the lung lobe according to the grayscale segmentation threshold corresponding to the outermost boundary, and determine the grayscale gradient range of the area between the outermost boundary and the innermost boundary of the lung lobe according to the average grayscale; Determine the region after threshold segmentation according to the grayscale segmentation threshold corresponding to the innermost boundary, determine the curve parameter of the region boundary line according to the length of the region boundary line, and determine the smoothness coefficient of the lung lobe according to the curve parameter; Output steps include: The final lung lobe boundary segmentation is performed according to the grayscale gradient range and the smoothness coefficient to obtain a lung segment boundary positioning result.

[0009] Preferably, the lung segmentation and optimization processing in the processing step specifically includes: Initialize boundary parameters, including maximum threshold, minimum threshold, step factor, maximum number of iterations, segmentation threshold of the outermost boundary, area of ​​the outermost boundary, segmentation threshold of the innermost boundary, area of ​​the innermost boundary, and smoothing coefficient; The current maximum threshold and minimum threshold are averaged and used as the threshold parameter for this iteration, and the image is segmented by threshold. Determine whether there are unsegmented lung segments. If so, calculate the boundary lengths of all unsegmented lung segments and obtain the average value of the boundary lengths. If the average value is less than the minimum value of all boundary lengths in the region after threshold segmentation and the number of iterations is less than the maximum number of iterations, the minimum threshold is modified according to the step factor, and the threshold segmentation step is returned to be executed; Use the current minimum threshold as the segmentation threshold for the innermost boundary; Determine whether there is a lung segment that does not reach the minimum area threshold. If so, return to perform the threshold segmentation step; Use the current minimum threshold as the segmentation threshold of the outermost boundary; Determine whether the area of ​​the outermost boundary is less than the minimum area threshold, and if so, modify the segmentation threshold of the outermost boundary to the segmentation threshold of the innermost boundary, and modify the segmentation threshold of the innermost boundary to the segmentation threshold of the outermost boundary in the previous iteration; It is determined whether the number of final iterations is equal to the maximum number of iterations. If so, the obtained boundary parameters are output to obtain the lung segmentation result.

[0010] Preferably, the lung segment merging optimization process in the processing step specifically includes: Initialize boundary parameters, including maximum threshold, minimum threshold, step factor, maximum number of iterations, segmentation threshold of the outermost boundary, area of ​​the outermost boundary, segmentation threshold of the innermost boundary, area of ​​the innermost boundary, and smoothing coefficient; The current maximum threshold and minimum threshold are averaged and used as the threshold parameter for this iteration, and the image is segmented by threshold. Determine whether there are unsegmented lung lobe areas in the threshold segmentation result. If so, calculate the average gray value of all connected areas and determine the segmentation threshold of the outermost boundary; The final image is determined by using a weighted blending algorithm with gradient magnitude and curvature magnitude; Determine whether the segmentation threshold of the outermost boundary is greater than the current minimum threshold. If so, calculate the average length of all current connected areas; If the average length is smaller than the area of ​​the outermost boundary and the current number of iterations is smaller than the maximum number of iterations, the minimum threshold is updated and the process returns to the threshold segmentation step; Calculate the average length of all connected areas. If the average length is greater than the area of ​​the outermost boundary, the segmentation threshold of the outermost boundary is updated, and return to perform the threshold segmentation processing step; Save the current minimum threshold; The segmentation threshold of the outermost boundary and the segmentation threshold of the innermost boundary are output as the final result of the lung segment merging optimization process.

[0011] Preferably, determining the segmentation threshold of the innermost boundary and the segmentation threshold of the outermost boundary in the processing step specifically includes: Determine the grayscale of the outermost boundary in the lung image after the grayscale normalization process, and use the current grayscale of the outermost boundary as the segmentation threshold of the outermost boundary; Performing reverse grayscale value sorting, if the average grayscale value of the three voxel points with the largest grayscale in the current lung lobe is less than the segmentation threshold of the innermost boundary, then determining the average grayscale by the grayscale of the outermost boundary in the lung image after the grayscale normalization process and the segmentation threshold of the innermost boundary, and then obtaining the grayscale range centered on the average grayscale; Determine whether the average gray value of the lung lobe region is within the gray range. If so, determine whether the difference between the segmentation threshold of the outermost boundary and the segmentation threshold of the innermost boundary is less than 5. If the difference is less than 5, the segmentation threshold of the innermost boundary is the grayscale of the current outermost boundary, otherwise the segmentation threshold of the innermost boundary is the sum of the grayscale of the current outermost boundary and 5; If the difference between the maximum and minimum grayscale values ​​of all voxel points in the lung lobe area exceeds 5, the grayscale range is determined by subtracting the grayscale of the outermost boundary from the average grayscale value of the 50% voxel points with the largest grayscale values ​​in the lung lobe; The grayscale range is determined by subtracting the segmentation threshold of the innermost boundary from the maximum grayscale value of all voxel points in the lung lobe. The smoothness coefficient is determined by the current segmentation threshold of the outermost boundary and the current segmentation threshold of the innermost boundary.

[0012] Preferably, the output step specifically includes: Performing segmentation on the grayscale normalized lung image using a segmentation threshold of an outermost boundary and a segmentation threshold of an innermost boundary to obtain two segmented regions; Performing threshold segmentation of the innermost boundary and the outermost boundary on the grayscale normalized lung image; Fill the holes in the segmentation result of the previous step to obtain the grayscale gradient range; Segment all voxel points between the segmentation threshold of the outermost boundary and the segmentation threshold of the innermost boundary, and update the gray gradient range according to the intensity and position of the voxel points; Performing maximum threshold segmentation on the grayscale normalized lung image, and then performing a closing operation on the maximum threshold segmentation result; Performing curvature amplitude analysis on the closing result, using Gaussian filtering for the curvature amplitude analysis, then calculating the ratio of the curvature amplitude value of all voxel points to the average curvature amplitude value, and determining the smoothness coefficient based on the ratio; Obtaining a final grayscale segmentation threshold according to the grayscale gradient range and the smoothness coefficient, and then performing grayscale segmentation threshold segmentation on the grayscale normalized lung image; The local gray value average is used to divide the region and obtain the final segmentation result.

[0013] Preferably, in the output step, a model based on machine learning is used to verify the correctness of the boundary of the adaptive watershed marker.

[0014] Preferably, in the processing step, a lobe division map and a lobe marking matrix are determined according to blood vessel position information.

[0015] Preferably, in the processing step and the output step, for each segmentation step, morphological filtering and morphological closing operations are performed to eliminate noise of fine structures.

[0016] Preferably, in the processing step, a point is selected from the outermost boundary area, its Euclidean distance and Manhattan distance to the innermost boundary are calculated, and the difference between the distances of the two points is used as the smoothness coefficient.

[0017] Preferably, the method further comprises: Performing three-dimensional registration on the lung lobe images to obtain three-dimensional registered lung lobe images; Based on the three-dimensionally registered lung lobe image, characteristic parameters of the lung lobe and lung segment surfaces are calculated, and lung segmentation of the lung lobe image is completed according to the characteristic parameters to obtain a lung segmentation result; Extracting the lung segment boundary of the lung segment segmentation result, and obtaining a boundary curve including a plurality of boundary convex points and boundary concave points; Through the convex hull algorithm, the boundary curves with the same or similar directions are stitched together to obtain a closed curve; Acquire a cavity region enclosed by a closed curve, set boundary convex points according to the volume of the cavity region, and set lung segment boundary concave points; Based on the watershed algorithm, a region growth threshold is set as a watershed parameter calculation criterion of an adaptive rule to obtain an optimal lung segment boundary, so that the boundary of the lung segment segmentation region corresponding to the optimal watershed threshold coincides with the lung segment boundary.

[0018] The method of the present invention has the following significant technical effects: The adaptive watershed lung segment boundary localization method proposed in this invention realizes high-precision localization of lung segment boundaries by innovatively combining image processing technology, anatomical knowledge and machine learning methods. The core of this method lies in its high adaptability and robustness, which can effectively cope with the image variability of different patients and different scanning conditions. By introducing multi-level adaptive mechanisms, such as adaptive threshold selection and dynamic parameter adjustment, this method greatly improves the accuracy and reliability of segmentation results.

[0019] In addition, the method of the present invention fully considers the anatomical characteristics of the lungs, especially using the distribution information of the pulmonary segmental arteries to guide the segmentation process. This segmentation strategy based on anatomical knowledge not only improves the accuracy of segmentation, but also ensures that the segmentation results have good anatomical significance, providing a reliable basis for subsequent clinical diagnosis and treatment decisions.

[0020] Another significant advantage of the method of the present invention is its comprehensiveness and systematicness. A complete technical solution is formed through multiple carefully designed processing steps, such as image preprocessing, preliminary segmentation, boundary optimization, etc. These steps are not simply superimposed, but through clever design, effective information transmission and complementarity are achieved. For example, the results of preliminary segmentation provide an important reference for subsequent boundary optimization, and the results of boundary optimization in turn help improve the overall segmentation effect. This synergy between steps significantly improves the performance of the entire method.

[0021] It is worth mentioning that the method of the present invention not only ensures high precision, but also takes into account computational efficiency. By adopting efficient algorithm implementation and reasonable parameter setting, the method can achieve near real-time processing under ordinary hardware conditions and meet the needs of clinical applications. This combination of high precision and high efficiency greatly enhances the application value of the method in actual clinical environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 The figure is an overall flow chart of the method of the present invention.

[0023] Figure 2 Flowchart of the preliminary lung segmentation of the present invention.

[0024] Figure 3 Flowchart for optimization of lung segment merging of the present invention.

[0025] Figure 4 It is a flowchart of the final boundary segmentation of the present invention. DETAILED DESCRIPTION

[0026] like Figure 1-4As shown, the present invention provides a lung segment boundary positioning method based on adaptive watershed division. The method realizes accurate positioning of lung segment boundaries in lung CT images through multiple steps, providing important support for clinical diagnosis and surgical planning.

[0027] First, the method obtains a lung angiography or high-resolution CT image and performs grayscale normalization on it. This step is intended to eliminate image differences caused by different scanning devices and parameters, laying the foundation for subsequent processing. Preferably, grayscale normalization can use a linear stretching method to map the image grayscale value to a range of 0-255.

[0028] Next, the method performs segmentation of the pulmonary arteries of the image. This step is the key to the subsequent segmentation of pulmonary lobes and pulmonary segments. In one embodiment of the present invention, a multi-scale vascular enhancement filtering technique can be used to segment the pulmonary arteries. Specifically, a Frangi filter can be used, and its mathematical expression is: , in, is the eigenvalue of the Hessian matrix, They are geometric ratios, cross-sectional ratios, and structural metrics, respectively. It is a control parameter, and its value is usually 0.5, 0.5, 500.

[0029] Based on the connection relationship between the pulmonary segment arteries, the method determines the pulmonary lobe division map of the four lung regions of the left lung and the three lung regions of the right lung. Preferably, a connected domain analysis algorithm can be used to implement this step. According to the pulmonary lobe division map, a pulmonary lobe division table is obtained. At the same time, the left and right affiliation of each branch artery in the pulmonary lobe division map is determined to form a pulmonary lobe labeling matrix. These steps provide an anatomical basis for subsequent pulmonary segmentation.

[0030] In each lung lobe, this method performs region growing on each branch artery with the branch artery as the main trunk to obtain the corresponding connected region. The key to the region growing algorithm lies in the selection of the growth criterion. In one embodiment of the present invention, the following growth criterion can be used: , in, is the gray value of the voxel to be detected, is the average gray value of the seed point neighborhood, is the standard deviation, It is a control parameter and its value is usually 2.5.

[0031] The connected area is used as the initial segmentation result of the lung segment corresponding to the branch artery. The boundary of the connected area is used as the boundary of the lung segment. This step realizes the preliminary segmentation of the lung segment.

[0032] According to the connection between lung segments and the pulmonary lobe division table, this method merges and optimizes adjacent lung segments to obtain optimized lung segmentation results. Preferably, a merging threshold can be set. When the boundary length of adjacent lung segments is less than the threshold (such as 10mm) and the grayscale difference is less than a certain threshold (such as 20HU), a merging operation is performed. This step is intended to eliminate the problem of over-segmentation and improve the anatomical rationality of the segmentation results.

[0033] For any lobe, this method determines its innermost lung segment boundary and outermost lung segment boundary. The innermost boundary of the lobe is obtained based on the innermost lung segment boundary, and the outermost boundary of the lobe is obtained based on the outermost lung segment boundary. The area between the innermost boundary and the outermost boundary of the lobe is the inner boundary of the mediastinum of the lobe. This step lays the foundation for the subsequent fine segmentation.

[0034] The method further determines the grayscale segmentation threshold corresponding to the outermost boundary and the grayscale segmentation threshold corresponding to the innermost boundary. According to the grayscale segmentation threshold corresponding to the outermost boundary, the average grayscale of the area between the outermost boundary and the innermost boundary of the lung lobe is determined, and the grayscale gradient range of the area between the outermost boundary and the innermost boundary of the lung lobe is determined according to the average grayscale. Preferably, the Otsu algorithm can be used to automatically determine the optimal threshold.

[0035] According to the grayscale segmentation threshold corresponding to the innermost boundary, the threshold segmentation region is determined, the curve parameter of the region boundary line is determined according to the length of the region boundary line, and the smoothness coefficient of the lung lobe is determined according to the curve parameter. In one embodiment of the present invention, the smoothness coefficient can be calculated by the following formula: , in, is the actual length of the boundary line, is the radius of the circle of equal area.

[0036] Finally, the final lung lobe boundary segmentation is performed according to the grayscale gradient range and the smoothness coefficient. Preferably, the level set method can be used to achieve fine segmentation, and its evolution equation is: , in, is the level set function, is the edge stop function, is the expansion coefficient.

[0037] A significant advantage of the method of the present invention is its adaptability. By introducing multiple adaptive parameters, such as the threshold of region growing, the condition of lung segment merging, the boundary smoothing coefficient, etc., the method can adapt to the individual differences of different patients and the image characteristics under different scanning conditions. This greatly improves the robustness and versatility of the method.

[0038] In addition, this method makes full use of the anatomical knowledge of the lungs, especially the distribution characteristics of the pulmonary segmental arteries, so that the segmentation results have good anatomical significance, which is of great value for subsequent clinical diagnosis and surgical planning.

[0039] It is worth noting that this method adopts a multi-scale analysis strategy. From lung lobe division to lung segmentation, and then to fine positioning of boundaries, the gradual refinement process effectively reduces the risk of error accumulation and improves the accuracy of the final result.

[0040] In practical applications, this method can be combined with machine learning techniques to further improve the accuracy of boundary positioning. For example, a convolutional neural network can be trained to verify and optimize the final boundary position. This strategy of combining traditional image processing techniques with modern artificial intelligence methods represents the future development direction of medical image analysis.

[0041] In general, the adaptive watershed lung segment boundary localization method provided by the present invention realizes accurate localization of lung segment boundaries through a series of innovative technical means. This method is not only highly adaptive and robust, but also ensures the anatomical rationality of the results, providing strong technical support for the accurate diagnosis and personalized treatment of lung diseases.

[0042] The method of the present invention provides an efficient and reliable algorithm for lung segmentation and optimization processing. Specifically, the algorithm includes the following steps: First, initialize the boundary parameters. These parameters include the maximum threshold, the minimum threshold, the step factor, the maximum number of iterations, the segmentation threshold of the outermost boundary, the area of ​​the outermost boundary, the segmentation threshold of the innermost boundary, the area of ​​the innermost boundary, and the smoothness coefficient. Preferably, these initial parameters can be set according to empirical values. For example, the maximum threshold can be set to the maximum gray value of the image, the minimum threshold can be set to the minimum gray value of the image, the step factor can be set to 1, and the maximum number of iterations can be set to 100. The selection of these initial values ​​has an important influence on the convergence speed of the algorithm and the quality of the final result.

[0043] Next, this method takes the average of the current maximum threshold and minimum threshold as the threshold parameter used in this iteration, and performs threshold segmentation on the image. This step adopts the idea of ​​adaptive threshold, which can better adapt to the characteristics of different images. The mathematical expression of threshold segmentation can be expressed as: , in, is the original image, is the binary image after segmentation, is the threshold value, equal to .

[0044] Then, the method determines whether there are unsegmented lung segments. If so, the boundary lengths of all unsegmented lung segments are calculated and the average of the boundary lengths is obtained. This step is intended to evaluate the quality of the current segmentation result.

[0045] Next, the method makes a key judgment: if the average value is less than the minimum value of all boundary lengths in the region after threshold segmentation and the number of iterations is less than the maximum number of iterations, the minimum threshold is modified according to the step factor, and the threshold segmentation step is returned. This judgment condition comprehensively considers the segmentation quality and computational efficiency, and is a manifestation of the algorithm's adaptability. Preferably, an empirical threshold, such as 10 mm, can be set as a reference value for the boundary length.

[0046] If the above conditions are not met, this method uses the current minimum threshold as the segmentation threshold for the innermost boundary. This step determines the inner boundary of the lung segment.

[0047] Subsequently, the method determines whether there are lung segments that do not reach the minimum area threshold. If so, the method returns to the threshold segmentation step. This step is intended to avoid the appearance of lung segments that are too small and improve the anatomical rationality of the segmentation results. Preferably, the minimum area threshold can be set to 50 mm².

[0048] If there is no lung segment that does not reach the minimum area threshold, the method uses the current minimum threshold as the segmentation threshold for the outermost boundary. This step determines the outer boundary of the lung segment.

[0049] Next, the method determines whether the area of ​​the outermost boundary is less than the minimum area threshold. If so, the segmentation threshold of the outermost boundary is modified to the segmentation threshold of the innermost boundary, and the segmentation threshold of the innermost boundary is modified to the segmentation threshold of the outermost boundary in the previous iteration. This step is to handle possible abnormal situations and ensure the rationality of the segmentation results.

[0050] Finally, the method determines whether the number of final iterations is equal to the maximum number of iterations. If so, the obtained boundary parameters are output to obtain the lung segmentation result. This step ensures the convergence of the algorithm and also provides an opportunity for possible manual intervention.

[0051] The method of the present invention also provides an innovative algorithm in the optimization processing of lung segment merging. The specific steps of the algorithm are as follows: First, initialize the boundary parameters, including the maximum threshold, minimum threshold, step factor, maximum number of iterations, segmentation threshold of the outermost boundary, area of ​​the outermost boundary, segmentation threshold of the innermost boundary, area of ​​the innermost boundary, and smoothing coefficient. The initialization of these parameters is similar to that described previously, but may need to be adjusted according to the specific needs of the merge optimization.

[0052] Then, the method takes the average of the current maximum threshold and minimum threshold as the threshold parameter used in this iteration, and performs threshold segmentation on the image. The mathematical expression of this step is the same as described above.

[0053] Next, the method determines whether there are unsegmented lung lobe areas in the threshold segmentation result. If so, the average grayscale value of all connected areas is calculated, and the segmentation threshold of the outermost boundary is determined. This step is to evaluate the quality of the current segmentation result and provide a basis for subsequent optimization.

[0054] Subsequently, the method determines the final image by weighted mixing algorithm of gradient amplitude and curvature amplitude. This step is one of the innovations of the algorithm, which comprehensively considers the gradient information and shape information of the image. Specifically, the following formula can be used: , in, For the final image, is the gradient image, is the curvature image, and is the weight coefficient, and .

[0055] Next, the method determines whether the segmentation threshold of the outermost boundary is greater than the current minimum threshold. If so, the average length of all current connected regions is calculated. This step is to evaluate the quality of the current segmentation result.

[0056] If the average length is less than the area of ​​the outermost boundary and the current number of iterations is less than the maximum number of iterations, the minimum threshold is updated and the threshold segmentation processing step is returned to be executed. This conditional judgment comprehensively considers the segmentation quality and computational efficiency, which is another embodiment of the algorithm's adaptability.

[0057] If the above conditions are not met, this method calculates the average length of all connected regions. If the average length is greater than the area of ​​the outermost boundary, the segmentation threshold of the outermost boundary is updated and the threshold segmentation processing step is returned. This step is intended to further optimize the segmentation results.

[0058] Finally, the method saves the current minimum threshold and outputs the segmentation threshold of the outermost boundary and the segmentation threshold of the innermost boundary as the final result of the lung segment merging optimization process.

[0059] The method of the present invention also provides an innovative algorithm for determining the segmentation threshold of the innermost boundary and the segmentation threshold of the outermost boundary. The specific steps of the algorithm are as follows: First, the grayscale of the outermost boundary in the lung image after the grayscale normalization process is determined, and the current grayscale of the outermost boundary is used as the segmentation threshold of the outermost boundary. This step provides an initial reference value for subsequent processing.

[0060] Then, the method performs reverse grayscale value sorting. If the average grayscale value of the three voxel points with the largest grayscale in the current lung lobe is less than the segmentation threshold of the innermost boundary, the average grayscale is determined by the grayscale of the outermost boundary in the lung image after the grayscale normalization process and the segmentation threshold of the innermost boundary, and then the grayscale range centered on the average grayscale is obtained. This step is intended to adaptively determine the appropriate grayscale range.

[0061] Next, the method determines whether the average grayscale value of the lung lobe region is within the grayscale range. If so, it determines whether the difference between the segmentation threshold of the outermost boundary and the segmentation threshold of the innermost boundary is less than 5. This threshold can be adjusted according to the specific application scenario, and is usually in the range of 3-7.

[0062] If the difference is less than 5, the segmentation threshold of the innermost boundary is the grayscale of the current outermost boundary. Otherwise, the segmentation threshold of the innermost boundary is the sum of the grayscale of the current outermost boundary and 5. This step ensures a reasonable difference between the inner and outer boundary thresholds.

[0063] If the difference between the maximum and minimum grayscale values ​​of all voxel points in the lobe region exceeds 5, the grayscale range is determined by subtracting the grayscale of the outermost boundary from the average grayscale value of the 50% voxel points with the largest grayscale values ​​in the lobe. This step handles possible extreme cases.

[0064] Then, the method determines the grayscale range by subtracting the segmentation threshold of the innermost boundary from the grayscale value maximum of all voxel points in the lung lobe. This step further refines the determination of the grayscale range.

[0065] Finally, this method determines the smoothness coefficient based on the current segmentation threshold of the outermost boundary and the current segmentation threshold of the innermost boundary. The smoothness coefficient can be calculated using the following formula: , in, is the smoothness coefficient, is the segmentation threshold of the outermost boundary, is the segmentation threshold of the innermost boundary, and are the maximum and minimum grayscale values ​​of the image respectively.

[0066] Through this series of steps, the method of the present invention realizes the accurate determination of the segmentation thresholds of the innermost boundary and the outermost boundary, laying a solid foundation for the subsequent positioning of the lung segment boundary. This adaptive threshold determination method can effectively adapt to the image features of different patients and different scanning conditions, greatly improving the robustness and versatility of the method.

[0067] In general, the algorithms provided by the present invention have shown unique innovation and high adaptability in lung segmentation, optimization processing and boundary threshold determination. These algorithms not only take into account the technical requirements of image processing, but also fully combine the characteristics of lung anatomy, so as to produce segmentation results with good anatomical significance. This is of great significance for improving the accuracy of lung disease diagnosis and formulating personalized treatment plans. The method of the present invention adopts an innovative technical route in the final boundary segmentation stage, making full use of the results of the previous processing and achieving high-precision lung segment boundary positioning. Specifically, the process includes the following key steps: First, the method performs segmentation of the outermost boundary threshold and the innermost boundary threshold on the grayscale normalized lung image to obtain two segmentation areas. This step lays the foundation for subsequent fine segmentation. Preferably, a dual threshold segmentation technique can be used, and its mathematical expression is as follows: , in, is the original image, is the binary image after segmentation, and are the segmentation thresholds of the innermost and outermost boundaries, respectively. Next, the method performs segmentation of the innermost boundary and the outermost boundary on the grayscale normalized lung image. This step is a supplement to the previous step and aims to obtain more comprehensive boundary information.

[0068] Subsequently, in one embodiment of the present invention, the segmentation result of the previous step is hole-filled to obtain a grayscale gradient range. Hole filling is a commonly used morphological operation that can effectively eliminate small holes in the segmentation result and improve the continuity of the boundary. Preferably, a filling algorithm based on connected domain analysis can be used, and its complexity is O(n), where n is the number of image pixels.

[0069] Furthermore, the method segments all voxel points between the segmentation threshold of the outermost boundary and the segmentation threshold of the innermost boundary, and updates the grayscale gradient range with the intensity and position of the voxel points. This step makes full use of the grayscale information and spatial information of the image, which helps to improve the accuracy of boundary positioning. The update of the grayscale gradient range can be achieved by the following formula: , in, is the updated grayscale gradient, is the gray value of the voxel point, is the distance from the voxel point to the boundary, and is the weight coefficient, and .

[0070] Next, this method performs maximum threshold segmentation on the grayscale normalized lung image, and then performs a closing operation on the maximum threshold segmentation result. This step aims to obtain a rough lung contour to provide a reference for subsequent fine segmentation. The closing operation can fill small holes near the boundary to make the boundary smoother.

[0071] Then, in a preferred embodiment of the present invention, the curvature amplitude analysis is performed on the closure result. The curvature amplitude analysis uses Gaussian filtering, and then calculates the ratio of the curvature amplitude value of all voxel points to the average curvature amplitude value, and determines the smoothness coefficient based on the ratio. The mathematical expression of this step is as follows: , in, is the curvature amplitude ratio, is the curvature amplitude of the voxel point, is the average curvature amplitude. The smoothness coefficient can be adjusted by setting a threshold To determine: , Finally, the method obtains the final grayscale segmentation threshold according to the grayscale gradient range and the smoothness coefficient, and then performs grayscale segmentation threshold segmentation on the grayscale normalized lung image. Preferably, the local grayscale value average is used for regional division to obtain the final segmentation result. This step combines the results of all previous processing and achieves high-precision lung segment boundary positioning.

[0072] It is worth noting that the method of the present invention also uses a machine learning-based model in the output step to verify the correctness of the boundary of the adaptive watershed marker. This strategy of combining traditional image processing technology and modern artificial intelligence methods greatly improves the accuracy and reliability of boundary positioning. Preferably, a convolutional neural network (CNN) can be used as a verification model, and its structure can include multiple convolutional layers, pooling layers, and fully connected layers. The loss function of the model can be designed as: , in, is the cross entropy loss, is the Dice coefficient loss, and is the weight coefficient.

[0073] In addition, in the processing step, the method of the present invention innovatively determines the pulmonary lobe division map and the pulmonary lobe marking matrix according to the vascular position information. This method makes full use of the anatomical characteristics of the lungs and improves the accuracy and anatomical significance of the division results. Specifically, a vascular tracking algorithm, such as a Frangi filter combined with a minimum path algorithm, can be used to extract the vascular tree structure. Then, based on the branching pattern of the vascular tree and combined with prior anatomical knowledge, the boundaries of the pulmonary lobes are determined.

[0074] In order to further improve the quality of the segmentation results, the method of the present invention uses morphological filtering and morphological closing operations to eliminate the noise of fine structures in each segmentation step in the processing step and the output step. These operations can effectively smooth the boundaries, fill small holes, and improve the continuity and smoothness of the segmentation results. Preferably, Gaussian filtering can be used for smoothing, and its mathematical expression is: , in, is the standard deviation of the Gaussian function, usually 1-2.

[0075] Finally, an innovative point of the present invention is that, in the processing step, a point is selected from the outermost boundary area, its Euclidean distance and Manhattan distance to the innermost boundary are calculated, and the difference between the two distances is used as the smoothness coefficient. This method takes into account the geometric characteristics of the boundary and can better adapt to lung segments of different shapes. Specifically, the calculation formula of the smoothness coefficient can be expressed as: , in, is the Euclidean distance, is the Manhattan distance. The Euclidean distance and Manhattan distance are defined as: , , This method of calculating the smoothness coefficient can effectively reflect the complexity of the boundary, thus providing an important reference for subsequent boundary optimization.

[0076] In general, the adaptive watershed lung segment boundary positioning method provided by the present invention realizes high-precision positioning of lung segment boundaries through a series of innovative technical means. This method not only makes full use of image processing technology and machine learning methods, but also combines the anatomical characteristics of the lungs, so as to produce segmentation results with high accuracy and good anatomical significance. This has important practical application value in improving the accuracy of lung disease diagnosis, formulating personalized treatment plans, and assisting surgical planning. The method of the present invention further proposes an innovative technical solution to more accurately locate lung segment boundaries. This solution fully takes into account the complexity and individual differences of the lung structure, and realizes highly adaptive boundary positioning through multi-step processing.

[0077] First, the method performs three-dimensional registration on the lung lobe images to obtain the three-dimensional registered lung lobe images. This step is the basis for subsequent processing, which ensures that images collected from different patients or at different time points can be compared and analyzed in the same spatial coordinate system. Preferably, a non-rigid registration algorithm based on mutual information can be used, and its objective function can be expressed as: , in, is the registration transformation, and are the reference image and the floating image respectively, represents mutual information.

[0078] Next, this method calculates the characteristic parameters of the lung lobe and lung segment surface based on the 3D registered lung lobe image. These characteristic parameters include but are not limited to curvature, normal vector, surface area, etc. The calculation of characteristic parameters provides an important basis for subsequent lung segmentation. For example, the mean curvature H and Gaussian curvature K can be calculated by the following formula: , , in, and is the principal curvature.

[0079] According to the calculated characteristic parameters, the method completes the lung segmentation of the lung lobe image and obtains the lung segmentation result. This step can adopt a feature-based classification algorithm, such as a support vector machine (SVM) or a random forest. Preferably, an ensemble learning method can be used to improve the accuracy of classification.

[0080] Subsequently, the method extracts the lung segment boundary of the lung segment segmentation result and obtains a boundary curve containing multiple boundary convex points and boundary concave points. This step is intended to capture the detailed features of the lung segment boundary. The extraction of boundary points can be achieved by edge detection algorithms, such as the Canny operator: , , Where G is the gradient amplitude, is the gradient direction.

[0081] Next, an innovative point of the present invention is to perform convex hull stitching operation on boundary curves with the same or similar directions through a convex hull algorithm to obtain a closed curve. This step can effectively process complex boundary shapes and improve the continuity and smoothness of the boundary. The convex hull algorithm can be implemented using the Graham scanning method, and its time complexity is O(nlogn).

[0082] Then, the method obtains the cavity area surrounded by the closed curve, sets the boundary convex points according to the volume of the cavity area, and sets the lung segment boundary concave points. This step fully considers the three-dimensional structural characteristics of the lung segment and helps to describe the lung segment boundary more accurately. The volume of the cavity area can be calculated by the voxel counting method: , Among them, I(x,y,z) is a binary image and v is the volume of a single voxel.

[0083] Finally, the core innovation of the method of the present invention is to set the region growing threshold as the watershed parameter calculation criterion of the adaptive rule based on the watershed algorithm to obtain the optimal lung segment boundary. This adaptive parameter setting method can effectively adapt to the image characteristics of different patients and different scanning conditions, greatly improving the robustness and versatility of the method. Specifically, the region growing threshold can be determined by the following formula: , in, and are the average gray value and standard deviation of the target area, It is an adjustable parameter, and its value usually ranges from 1.5 to 2.5.

[0084] It is worth noting that this method makes the boundary of the lung segmentation area corresponding to the optimal watershed threshold coincide with the boundary of the lung segment through iterative optimization. This process can be achieved by minimizing the following objective function: , in, is the reference boundary point, is the boundary point of the current segmentation result, is the total number of boundary points.

[0085] In addition, the method of the present invention also includes a matching adaptive watershed lung segment boundary positioning system. The system includes multiple functional modules, each of which is responsible for implementing specific steps in the method. For example, the data loading module is used to load lung CT images and lung contour point sets; the information processing module is used to obtain lung segment sets based on lung contour point sets; the lung segment region map generation module generates lung segment region maps through region growing algorithms; the gradient map generation module calculates lung segment region maps and water area maps of lung segment sets through gradient features.

[0086] In particular, the watershed parameter calculation module optimizes the watershed parameters according to the dam characteristics, the ratio of local water area to overall water area, and the distance between the dam and the regional centroid, and calculates the watershed spectrum. The core algorithm of this module can be expressed as: , in, is the optimized watershed parameter, , , represent dam characteristics, area ratio and distance characteristics respectively, , , is the corresponding weight coefficient.

[0087] The dam parameter optimization module inputs the pixel point set into the watershed parameter optimization function to obtain the watershed spectrum. The lung segment boundary generation module generates the lung segment boundary and the lung lobe boundary according to the calculated optimal watershed parameters input by the dam parameter optimization module. The boundary optimization module compares the lung lobe boundary with the lung contour point set to obtain the optimized lung contour point set and the final lung lobe boundary.

[0088] Finally, the smoothing module smoothes the optimized lung contour point set to complete the positioning of the lung segment boundary. The smoothing process can use the local regression smoothing (LOESS) method, and its mathematical expression is: , in, is the smoothed coordinate value, is the original coordinate value, , , is the regression coefficient, is the error term.

[0089] In general, the adaptive watershed lung segment boundary positioning method and system provided by the present invention achieves high-precision and high-reliability positioning of lung segment boundaries through a series of innovative technical means. This method not only takes into account the technical requirements of image processing, but also fully combines the characteristics of lung anatomy, so as to produce segmentation results with good anatomical significance. In particular, its adaptive parameter adjustment mechanism enables this method to effectively cope with image variability in different patients and under different scanning conditions, and has broad clinical application prospects.

[0090] This high-precision lung segment boundary positioning method is of great significance for improving the accuracy of lung disease diagnosis, assisting surgical planning, and evaluating treatment effects. For example, in the precise positioning and staging of lung cancer, accurate lung segment boundary information can help doctors better assess the location and invasion range of the tumor. In the planning of lung segment resection surgery, this method can provide surgeons with detailed anatomical structure information, which helps to formulate the best surgical plan. In addition, in the evaluation of chronic obstructive pulmonary disease (COPD), accurate lung segment segmentation can help doctors more accurately quantify the severity and distribution of the disease.

[0091] 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 localization method based on adaptive watershed division, characterized in that: include: The acquisition steps include: Obtain contrast or high-resolution CT images of the lungs; Perform grayscale normalization on lung images; Processing steps include: Performing pulmonary artery segmentation based on the grayscale normalized lung image; Determine the pulmonary lobe division diagram of the four lung regions of the left lung and the three lung regions of the right lung according to the connection relationship between the pulmonary segmental arteries; Based on the lung lobe division map, a lung lobe division table is obtained; According to the pulmonary lobe division map, determining the left and right affiliation of each branch artery to form a pulmonary lobe marking matrix; Inside each lung lobe, for each branch artery, the corresponding connected area is obtained by region growing with the branch artery as the main trunk; Using the connected area as the initial segmentation result of the lung segment corresponding to the branch artery; According to the connection between the lung segments and the lung lobe division table, adjacent lung segments are merged and optimized to obtain an optimized lung segmentation result; For any lobe, determine its innermost and outermost segment boundaries; The innermost boundary of the lung lobe is obtained according to the innermost lung segment boundary, and the outermost boundary of the lung lobe is obtained according to the outermost lung segment boundary; Determine a grayscale segmentation threshold corresponding to the outermost boundary and a grayscale segmentation threshold corresponding to the innermost boundary; Determine the average grayscale of the area between the outermost boundary and the innermost boundary of the lung lobe according to the grayscale segmentation threshold corresponding to the outermost boundary, and determine the grayscale gradient range of the area between the outermost boundary and the innermost boundary of the lung lobe according to the average grayscale; Determine the region after threshold segmentation according to the grayscale segmentation threshold corresponding to the innermost boundary, determine the curve parameter of the region boundary line according to the length of the region boundary line, and determine the smoothness coefficient of the lung lobe according to the curve parameter; Output steps include: The final lung lobe boundary segmentation is performed according to the grayscale gradient range and the smoothness coefficient to obtain a lung segment boundary positioning result.

2. The lung segment boundary positioning method according to claim 1, characterized in that: The lung segmentation and optimization processing in the processing steps specifically include: Initialize boundary parameters, including maximum threshold, minimum threshold, step factor, maximum number of iterations, segmentation threshold of the outermost boundary, area of ​​the outermost boundary, segmentation threshold of the innermost boundary, area of ​​the innermost boundary, and smoothing coefficient; The current maximum threshold and minimum threshold are averaged and used as the threshold parameter for this iteration, and the image is segmented by threshold. Determine whether there are unsegmented lung segments. If so, calculate the boundary lengths of all unsegmented lung segments and obtain the average value of the boundary lengths. If the average value is less than the minimum value of all boundary lengths in the region after threshold segmentation and the number of iterations is less than the maximum number of iterations, the minimum threshold is modified according to the step factor, and the threshold segmentation step is returned to be executed; Use the current minimum threshold as the segmentation threshold for the innermost boundary; Determine whether there is a lung segment that does not reach the minimum area threshold. If so, return to perform the threshold segmentation step; Use the current minimum threshold as the segmentation threshold of the outermost boundary; Determine whether the area of ​​the outermost boundary is less than the minimum area threshold, and if so, modify the segmentation threshold of the outermost boundary to the segmentation threshold of the innermost boundary, and modify the segmentation threshold of the innermost boundary to the segmentation threshold of the outermost boundary in the previous iteration; It is determined whether the number of final iterations is equal to the maximum number of iterations. If so, the obtained boundary parameters are output to obtain the lung segmentation result.

3. The lung segment boundary positioning method according to claim 1, characterized in that: The lung segment merging optimization process in the processing step specifically includes: Initialize boundary parameters, including maximum threshold, minimum threshold, step factor, maximum number of iterations, segmentation threshold of the outermost boundary, area of ​​the outermost boundary, segmentation threshold of the innermost boundary, area of ​​the innermost boundary, and smoothing coefficient; The current maximum threshold and minimum threshold are averaged and used as the threshold parameter for this iteration, and the image is segmented by threshold. Determine whether there are unsegmented lung lobe areas in the threshold segmentation result. If so, calculate the average gray value of all connected areas and determine the segmentation threshold of the outermost boundary; The final image is determined by using a weighted blending algorithm with gradient magnitude and curvature magnitude; Determine whether the segmentation threshold of the outermost boundary is greater than the current minimum threshold. If so, calculate the average length of all current connected areas; If the average length is smaller than the area of ​​the outermost boundary and the current number of iterations is smaller than the maximum number of iterations, the minimum threshold is updated and the process returns to the threshold segmentation step; Calculate the average length of all connected areas. If the average length is greater than the area of ​​the outermost boundary, the segmentation threshold of the outermost boundary is updated, and return to perform the threshold segmentation processing step; Save the current minimum threshold; The segmentation threshold of the outermost boundary and the segmentation threshold of the innermost boundary are output as the final result of the lung segment merging optimization process.

4. The lung segment boundary positioning method according to claim 1, characterized in that: Determining the segmentation threshold of the innermost boundary and the segmentation threshold of the outermost boundary in the processing step specifically includes: Determine the grayscale of the outermost boundary in the lung image after the grayscale normalization process, and use the current grayscale of the outermost boundary as the segmentation threshold of the outermost boundary; Performing reverse grayscale value sorting, if the average grayscale value of the three voxel points with the largest grayscale in the current lung lobe is less than the segmentation threshold of the innermost boundary, then determining the average grayscale by the grayscale of the outermost boundary in the lung image after the grayscale normalization process and the segmentation threshold of the innermost boundary, and then obtaining the grayscale range centered on the average grayscale; Determine whether the average gray value of the lung lobe region is within the gray range. If so, determine whether the difference between the segmentation threshold of the outermost boundary and the segmentation threshold of the innermost boundary is less than 5. If the difference is less than 5, the segmentation threshold of the innermost boundary is the grayscale of the current outermost boundary, otherwise the segmentation threshold of the innermost boundary is the sum of the grayscale of the current outermost boundary and 5; If the difference between the maximum and minimum grayscale values ​​of all voxel points in the lung lobe area exceeds 5, the grayscale range is determined by subtracting the grayscale of the outermost boundary from the average grayscale value of the 50% voxel points with the largest grayscale values ​​in the lung lobe; The grayscale range is determined by subtracting the segmentation threshold of the innermost boundary from the maximum grayscale value of all voxel points in the lung lobe. The smoothness coefficient is determined by the current segmentation threshold of the outermost boundary and the current segmentation threshold of the innermost boundary.

5. The lung segment boundary positioning method according to claim 1, characterized in that: The output step specifically includes: Performing segmentation on the grayscale normalized lung image using a segmentation threshold of an outermost boundary and a segmentation threshold of an innermost boundary to obtain two segmented regions; Performing threshold segmentation of the innermost boundary and the outermost boundary on the grayscale normalized lung image; Fill the holes in the segmentation result of the previous step to obtain the grayscale gradient range; Segment all voxel points between the segmentation threshold of the outermost boundary and the segmentation threshold of the innermost boundary, and update the gray gradient range according to the intensity and position of the voxel points; Performing maximum threshold segmentation on the grayscale normalized lung image, and then performing a closing operation on the maximum threshold segmentation result; Performing curvature amplitude analysis on the closing result, using Gaussian filtering for the curvature amplitude analysis, then calculating the ratio of the curvature amplitude value of all voxel points to the average curvature amplitude value, and determining the smoothness coefficient based on the ratio; Obtaining a final grayscale segmentation threshold according to the grayscale gradient range and the smoothness coefficient, and then performing grayscale segmentation threshold segmentation on the grayscale normalized lung image; The local gray value average is used to divide the region and obtain the final segmentation result.

6. The lung segment boundary positioning method according to claim 1, characterized in that: In the output step, a machine learning-based model is used to verify the correctness of the boundaries of the adaptive watershed markers.

7. The lung segment boundary positioning method according to claim 1, characterized in that: In the processing step, a lung lobe division map and a lung lobe marking matrix are determined according to the blood vessel position information.

8. The lung segment boundary positioning method according to claim 1, characterized in that: In the processing step and the output step, for each segmentation step, morphological filtering and morphological closing operations are performed to eliminate noise of fine structures.

9. The lung segment boundary positioning method according to claim 8, characterized in that: In the processing step, a point is selected from the outermost boundary area, the Euclidean distance and the Manhattan distance from the point to the innermost boundary are calculated, and the difference between the distances of the two points is used as the smoothness coefficient.

10. The lung segment boundary positioning method according to claim 1, characterized in that: The method further comprises: Performing three-dimensional registration on the lung lobe images to obtain three-dimensional registered lung lobe images; Based on the three-dimensionally registered lung lobe image, characteristic parameters of the lung lobe and lung segment surfaces are calculated, and lung segmentation of the lung lobe image is completed according to the characteristic parameters to obtain a lung segmentation result; Extracting the lung segment boundary of the lung segment segmentation result, and obtaining a boundary curve including a plurality of boundary convex points and boundary concave points; Through the convex hull algorithm, the boundary curves with the same or similar directions are stitched together to obtain a closed curve; Acquire a cavity region enclosed by a closed curve, set boundary convex points according to the volume of the cavity region, and set lung segment boundary concave points; Based on the watershed algorithm, a region growth threshold is set as a watershed parameter calculation criterion of an adaptive rule to obtain an optimal lung segment boundary, so that the boundary of the lung segment segmentation region corresponding to the optimal watershed threshold coincides with the lung segment boundary.

Images