A method and system for evaluating the consistency of multiple / subsequent vessel segmentation results

Abstract

This invention provides a method and system for evaluating the consistency of multiple / subsequent vessel segmentation results. It involves reading the original data of a vessel dataset and multiple / subsequent vessel segmentation result matrices, then randomly selecting two vessel segmentation result matrices to calculate their union, intersection, and symmetric difference. Any point in the symmetric difference is selected, and iterative growth is performed using a convolution kernel to simulate a path diffusion process. Under the constraints of the union range and a given geodesic distance, a spherical set is obtained. The intersection of this spherical set with the union of the vessel segmentation result matrices and the intersection of this intersection with the two / subsequent vessel segmentation result matrices are calculated. The above calculations and statistical analysis are performed on each point to obtain statistical results, which are then mapped back to the original data and the two / subsequent vessel segmentation result matrices. This yields the statistical results for any two / subsequent vessel segmentation result matrices, and the visualized results are output. This invention can be used to measure the performance of vessel segmentation algorithms or compare their differences, accelerating the development and verification process of vessel segmentation algorithms.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a method and system for evaluating the consistency of multiple / subsequent blood vessel segmentation results. Background Technology

[0002] Vascular morphological changes are associated with various pathological features, making vascular segmentation crucial for biomedical image analysis. However, due to the complexity of vascular structures, anatomical differences among patients, the scarcity of annotated public datasets, and the variability in image quality, achieving accurate vascular segmentation remains challenging. From the perspective of the development path of fully automated vascular segmentation methods, developing high-quality evaluation methods for vascular segmentation results is important and urgently needed, as judging the quality of annotation / segmentation results is an unavoidable sub-problem in the vascular segmentation process. Whether using manual annotation or other developed vascular segmentation methods, objective comparison, verification, and evaluation of their annotation / segmentation results are necessary to promote the development of vascular segmentation methods.

[0003] A classic approach is to compare the blood vessel segmentation results predicted by a certain algorithm with the "gold standard" (ground truth) dataset annotated by experts to measure the consistency / discrepancy, thereby evaluating the performance of this blood vessel segmentation algorithm. Current objective evaluation methods for annotation / segmentation results still primarily focus on characterizing and comparing the consistency or inconsistency of multiple / instance segmentation results as a whole. Typical examples of this type of global evaluation method / metric include: Dice similarity coefficient (DSC), intersection-over-union ratio (IoU), Hausdorff distance (HD), and F1 score. Taking the intersection-union ratio (IoU) as an example, set X is the first segmentation result obtained from a certain blood vessel data set, in which 120 voxels are labeled as blood vessels. Set Y is the second segmentation result obtained using the same or different segmentation methods, in which 100 voxels are labeled as blood vessels. There are 80 overlapping voxels in the regions labeled as blood vessels in sets X and Y. Therefore, |X∪Y|=140, |X∩Y|=80, and IoU=|X∩Y| / |X∪Y|=80 / 140=0.571. Here, IoU=0.571 characterizes the overall overlap between sets X and Y, that is, the global consistency / inconsistency between the two blood vessel segmentation results. Taking Hausdorff distance (HD) as an example, this metric reflects the maximum Euclidean distance between the "gold standard" set and the segmentation algorithm's prediction set in the entire 3D vessel segmentation results. This means that while this metric can tell us the maximum deviation between the prediction result and the gold standard result, it does not provide the specific local spatial location or spatial distribution of this deviation. This type of global evaluation method is intuitive, but it is more susceptible to subtle errors in thicker vessel branches, while being insensitive to obvious errors in thinner vessel branches. In fact, in vessel segmentation tasks, accurately segmenting thinner vessel branches is more difficult and important than segmenting thicker ones.

[0004] While the need to understand the spatial distribution of vessel segmentation errors has always existed, only a few methods can partially meet this requirement. For example, a novel function measuring the geometric topological loss of segmentation results calculates the radial distance of voxels based on the tubular centerline and Euclidean distance, assigning greater weight to voxels closer to the centerline, thus achieving finer-grained segmentation error characterization / control. Further improvements to this method calculate the vessel cross-sectional size and voxel radial distance based on the vessel boundary and Euclidean distance, assigning greater weight to finer vessel branches and voxels closer to the vessel boundary, thereby achieving better segmentation even for small vessel branches. However, accurately determining the vessel centerline and vessel boundary in complex vascular networks is inherently challenging, especially in medium and small vessels (where the vessel cross-section has fewer voxels). Furthermore, the calculation of the centerline and edge in the aforementioned methods all rely on the referenced "gold standard" dataset, assuming the "gold standard" dataset is accurate. However, in reality, differences in annotators and annotation methods make the "gold standard" dataset highly susceptible to change, further limiting the practicality of such methods. If there are significant differences between multiple / subsequent segmentation results, the center line and boundary will be significantly different, which may also make this type of method unsuitable.

[0005] Current blood vessel segmentation algorithms still have some practical shortcomings and deficiencies. First, the impact of errors varies across blood vessels of different sizes. For example, when errors exist simultaneously in 10 voxels across blood vessels of different sizes, the errors in small and medium-sized vessels have a greater impact on the accuracy of the segmentation results compared to large vessels; that is, the weight of the error's impact is unbalanced across blood vessels of different sizes. Second, while many blood vessel segmentation algorithms have been developed in the field of image processing, horizontal comparisons between different algorithms are mainly based on the aforementioned method of globally / holistically evaluating the consistency of multiple / instance segmentation results. This leads users to understand that errors exist between multiple / instance segmentation results, but they are not clear about how these errors are distributed with blood vessel size. Finally, fully characterizing the distribution of blood vessel segmentation errors with blood vessel size is a crucial foundation for blood vessel image processing. However, current methods that only characterize a small number of segmentation errors locally are still insufficient to meet the requirements and rely on technical solutions based on the centerline and blood vessel boundaries of the "gold standard" dataset, resulting in high complexity. Summary of the Invention

[0006] Purpose of the invention: The first purpose of this invention is to provide a method for evaluating the consistency of multiple / instance vessel segmentation results that can quantify the differences between multiple / instance vessel segmentation results of the same vessel dataset; the second purpose is to provide an evaluation system for evaluating the consistency of multiple / instance vessel segmentation results corresponding to the above evaluation method.

[0007] Technical solution: A method for evaluating the consistency of multiple / subsequent vessel segmentation results, comprising the following steps:

[0008] (1) Read the original data of the same blood vessel dataset and multiple blood vessel segmentation result matrices, and randomly select two blood vessel segmentation result matrices s1 and s2 from them;

[0009] (2) Calculate the union V, intersection S, and symmetric difference Δs of s1 and s2;

[0010] (3) Select any point p∈Δs and perform iterative growth by simulating the path diffusion process using a convolution kernel. Under the geometric constraints of the two / times blood vessel segmentation result matrices s1 and s2 and the given geodesic distance r, calculate the spherical set B. The center of the spherical set B is point p, and the radius of the sphere is the geodesic distance r. Calculate the intersection B of the spherical set B with the union V of s1 and s2. V And calculate B respectively V The intersection B with s1 and s2 s1 B s2 ;

[0011] (4) Repeat step (3) to output each point p and its corresponding B. V B s1 B s2 Statistical analysis is performed to obtain the proportion of blood vessels of different sizes and the statistical results of the differences between the two / subsequent blood vessel segmentation result matrices s1 and s2. The statistical results are output and mapped back to the original data and the two / subsequent blood vessel segmentation result matrices s1 and s2.

[0012] (5) Repeat the above steps to obtain the statistical results of any two blood vessel segmentation result matrices in the multiple blood vessel segmentation result matrices. Output the visualization processing results for the statistical results of any two blood vessel segmentation result matrices.

[0013] Specifically, the blood vessel segmentation result matrix is ​​a two-dimensional or three-dimensional binary image matrix. In the binary image matrix, element 1 indicates that the voxel point belongs to the blood vessel region, and element 0 indicates that the voxel point does not belong to the blood vessel region.

[0014] Specifically, in step (3), the iterative growth process, which simulates the path diffusion process using convolutional kernels, includes: iteratively performing two-dimensional or three-dimensional convolution operations, applying the convolutional kernels to the union V of the blood vessel segmentation result matrix, and using the convolutional kernels as weight filters to simulate the diffusion of the blood vessel region in each iteration. The convolution iteration continues until a set sphere radius r is reached. The aforementioned convolutional kernel is a kernel function whose weights decrease as the sphere radius r increases.

[0015] Specifically, in step (3), the geometric constraints of the two / times blood vessel segmentation result matrices and the geodesic distance include: the growable region of the convolution kernel iteration is restricted to the union V of the two / times blood vessel segmentation result matrices s1 and s2, and the sphere radius r is the shortest path length of the blood vessel region from one point to another in the blood vessel segmentation result, which is constrained according to the geodesic distance.

[0016] Specifically, in step (4), the statistical analysis includes: calculating the set B corresponding to each point p. V B s1 B s2 The modulus, and the statistical values ​​of B for different values. V B s1 B s2 The number of points p corresponding to the modulus, set B V The modulus is the blood vessel size at the corresponding point p, set B s1 B s2 The modulus is the difference between the two / subsequent blood vessel segmentation result matrices s1 and s2 at point p.

[0017] Specifically, visualization processing includes: using set B V Using the modulus and the corresponding number of points p as the horizontal and vertical axes, a graph showing the proportion of blood vessels of different sizes is obtained; using set B... V The number of points p after weighting by the modulus and the corresponding given weighting factor is used as the horizontal and vertical axes to plot the distribution of the difference between the two / secondary vessel segmentation result matrices with respect to vessel size. The aforementioned weighting factor is B. s1 B s2 The absolute value of the difference in modulus divided by the corresponding B V The numerical values ​​obtained from the modulus; based on the coordinates of each point p, the differences between the two / subsequent blood vessel segmentation result matrices are traced and marked, and then visualized in a two-dimensional image or three-dimensional model.

[0018] This invention also provides an evaluation system for the consistency of multiple / subsequent vessel segmentation results, comprising:

[0019] Data input module: used to read the raw data of the same blood vessel dataset and multiple / subsequent blood vessel segmentation result matrices, and to randomly select two / subsequent blood vessel segmentation result matrices s1 and s2 from the multiple / subsequent blood vessel segmentation result matrices;

[0020] Matrix processing module: used to calculate the union V, intersection S, and symmetric difference Δs of s1 and s2;

[0021] Convolution Iteration Module: Select any point p∈Δs, and perform iterative growth by simulating the path diffusion process using a convolution kernel. Under the geometric constraints of two / times blood vessel segmentation result matrices s1 and s2 and a given geodesic distance r, calculate a spherical set B. The center of this spherical set B is point p, and the radius of the sphere is the geodesic distance r. Calculate the intersection B of the spherical set B with the union V of s1 and s2. V And calculate B respectively V The intersection B with s1 and s2 s1 B s2 ;

[0022] Statistical analysis module: Used to repeatedly run the convolution iteration module, outputting each point p and its corresponding B. V B s1 B s2 Statistical analysis is performed to obtain the proportion of blood vessels of different sizes and the statistical results of the differences between the two / subsequent blood vessel segmentation result matrices s1 and s2. The statistical results are output and mapped back to the original data and the two / subsequent blood vessel segmentation result matrices s1 and s2.

[0023] Visualization output module: This module is used to repeatedly run the above sub-modules to obtain the statistical results of any two blood vessel segmentation result matrices from multiple blood vessel segmentation result matrices. It then outputs the visualization results of the statistical results of any two blood vessel segmentation result matrices.

[0024] Beneficial effects: Compared with the prior art, the significant effects of the present invention are:

[0025] (1) The evaluation method provided by this invention can assess the variability of multiple / segmentation results with finer granularity based on local blood vessel area / volume, characterizing the spatial distribution of blood vessel segmentation error with varying blood vessel size. This provides a complementary scheme / theory for global evaluation methods such as Dice similarity coefficient (DSC), intersection-over-union ratio (IoU), Hausdorff distance (HD), and F1 score. Firstly, the statistical results of this method can reflect the proportion of blood vessels of different sizes in a given blood vessel dataset. This is beneficial for developers of blood vessel segmentation algorithms or users in the clinical medical field to clearly understand the blood vessel size proportion characteristics of the segmented blood vessel network, such as recognizing the relative proportions of large, medium, and small blood vessels. Secondly, this method can clearly define the magnitude of the variability in blood vessel segments of different sizes in multiple segmentation results, such as identifying which part of the large, medium, and small blood vessels has the largest size error during segmentation. Providing this novel statistical data has significant theoretical value for accurately evaluating algorithm performance and subsequent algorithm development.

[0026] (2) The evaluation method provided by this invention has significant practical value in vascular image post-processing and radiological practice. This method can be used to evaluate the quality of manual annotation of vascular images. Since each annotator's experience and habits are different, it is necessary to evaluate the reliability of vascular annotation quality. By evaluating the consistency of multiple manual vascular annotation results, a higher-quality gold standard dataset can be further improved and created. Characterizing the spatial distribution of vascular segmentation error with vascular thickness is particularly important. This method can also trace the source of error between the segmented image and the original vascular image, allowing users to intuitively understand the spatial distribution of differences between multiple segmentation results. Similarly, this method is also needed to compare the results obtained by different vascular segmentation algorithms, thereby clarifying the errors between algorithms or between the segmentation results of a certain algorithm and the gold standard, facilitating more targeted research and development. For example, in this method, higher loss function weights can be set for vascular locations with worse consistency with the gold standard dataset to improve the accuracy of the segmentation algorithm.

[0027] (3) The method for calculating blood vessel thickness in this invention is based on constraints of local blood vessel area / volume and geodesic distance, rather than on blood vessel centerline / boundary and Euclidean distance. Furthermore, the characterization of local errors utilizes a weighted calculation of blood vessel thickness based on local blood vessel volume, thus addressing the issue that errors of the same number of voxels have a greater negative impact on smaller blood vessels. Constraints based on local blood vessel area / volume and geodesic distance ensure that the calculation of blood vessel thickness is based on multiple / instance results simultaneously, rather than relying on the centerline or blood vessel boundary of a single result, avoiding related calculation errors. This also means that this method does not require one of the two / instance blood vessel segmentation results being compared to be a gold standard dataset or restrict the proximity / similarity of multiple / instance segmentation results; even if there are significant differences between two / instance segmentation results, this method can accurately and intuitively reflect this situation. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating the method for evaluating the consistency of multiple / segmented blood vessel segmentation results in Embodiment 1 of the present invention.

[0029] Figure 2 This is a schematic diagram illustrating the principle of the method for evaluating the consistency of multiple / segmented blood vessel segmentation results in Embodiment 1 of the present invention.

[0030] Figure 3 This is a visualization result of the evaluation method for the consistency of multiple / segmentation results of blood vessel segmentation in Embodiment 2 of the present invention. Detailed Implementation

[0031] The present invention will be further described below with reference to the accompanying drawings.

[0032] Example 1

[0033] Please see Figure 1 As shown, this embodiment uses two / multiple vessel segmentation results as an example to illustrate a method for evaluating the consistency of multiple / multiple vessel segmentation results, including the following steps:

[0034] (1) Read the original data of the same blood vessel dataset and two / secondary blood vessel segmentation result matrices s1 and s2.

[0035] The blood vessel segmentation result matrix is ​​a two-dimensional or three-dimensional binary image matrix, i.e., a 0-1 matrix. In the binary image matrix, an element 1 indicates that the voxel point belongs to the blood vessel region, and an element 0 indicates that the voxel point does not belong to the blood vessel region.

[0036] (2) Calculate the union of s1 and s2, V = s1∪s2, and the intersection of s1 and s2. interest =s1∩s2 and the symmetry difference Δs=V∩(s interest )^c=(s1∪s2)\(s1∩s2).

[0037] (3) Select any point p∈Δs, and perform two-dimensional or three-dimensional convolution operations iteratively. Apply the convolution kernel to the union V = s1∪s2 of the blood vessel segmentation result matrix. In each iteration, use the convolution kernel as a weight filter to simulate the diffusion of the blood vessel region. When the value of some regions exceeds 1 during the convolution iteration, these values ​​will be reset to 1 to ensure that the result remains within the valid binary range. The convolution iteration stops when the set sphere radius r is reached. Define and calculate the spherical set B(p,r), where the center of the spherical set B is point p and the sphere radius is the geodesic distance r. Calculate the intersection B of the spherical set B and the union V of s1 and s2. V (p,r), and calculate B V The intersection of (p,r) with s1 and s2

[0038] The growable region of the above convolutional kernel iteration is limited to the union V = s1∪s2 of the two / times vessel segmentation result matrices s1 and s2. The sphere radius r is the shortest path length from one point to another in the vessel segmentation result. This shortest path length is constrained by geodesic distance, not Euclidean distance. Due to the high vessel density, using Euclidean distance constraints can easily lead to erroneous growth by "penetrating" one vessel from another during the iterative growth process. By performing iterative growth under the geometric constraints of vessel segmentation results and geodesic distance, and simulating the path diffusion process with convolutional kernels, the entire vascular network can be gradually "explored" while maintaining the continuity of the vascular structure.

[0039] In this embodiment, the convolution kernel is a kernel function whose weight decreases as the radius r of the sphere increases. It has a "bell-shaped" characteristic that varies with distance. It can be 2D or 3D and is suitable for two-dimensional or three-dimensional blood vessel segmentation results / matrices s1 and s2, respectively.

[0040] (4) Repeat step (3) to output each point p and its corresponding B. V (p,r), B s1 (p,r), B s2 For (p,r), perform statistical analysis and calculate B corresponding to each point p. V (p,r), B s1 (p,r), B s2 The modulus of (p,r) ||B V (p,r)||、||B s1 (p,r)||、||B s2 (p,r)||, and count the values ​​of ||B for different values. V (p,r)||、||B s1 (p,r)||、||B s2 (p,r)||The number of points p corresponding to ||B V (p,r)|| represents the blood vessel size at the corresponding point p, ||B s1 (p,r)||、||B s2 (p,r)|| represents the difference between the two / subsequent vessel segmentation result matrices s1 and s2 at point p. Output the above statistical results and map them back to the original data and the two / subsequent vessel segmentation result matrices s1 and s2, and output the visualization results.

[0041] The visualization results can be divided into two levels. The first level is based on ||B V (p,r)|| (arranged in ascending or descending order) and the corresponding number of points p are plotted on the x and y axes. This subplot reflects the number of vessels of different sizes in the union V = s1∪s2 of the two / subsequent vessel segmentation results being examined, thus obtaining a proportion map of vessels of different sizes in the vascular bundle; with ||B V The number of points p after weighting (p,r)|| with the corresponding given weight factor is plotted on the x and y axes to obtain the distribution of the difference between the two / secondary vessel segmentation result matrices with respect to vessel size. Here, the weight factor is ||B s1 (p,r)||、||B s2 The absolute value of the difference between (p,r)|| divided by the corresponding ||B V The value obtained by (p,r)|| is |(||B s1 (p,r)||-||B s2 (p,r)||)| / ||B V(p,r)||, this weighting factor characterizes the magnitude of the difference between the two / segmentation results and the local blood vessel size at point p.

[0042] The second level involves tracing and marking the differences between the two / subsequent blood vessel segmentation result matrices s1 and s2 based on the coordinates of each point p. This facilitates a more intuitive analysis of the differences between the two / subsequent segmentation results, which are then visualized in a two-dimensional image or three-dimensional model.

[0043] The following is combined with Figure 2 To provide a more intuitive explanation, Figure 2 Each sub-figure uses a two-dimensional cross-sectional diagram. Please refer to sub-figure a. For the original data and two blood vessel segmentation results s1 and s2 of the same blood vessel dataset (red represents the original data, green represents the blood vessel segmentation result s1, and blue represents the blood vessel segmentation result s2), sub-figure a is magnified locally to obtain sub-figure b. It can be seen that the difference between the two blood vessel segmentation results lies in the incomplete matching of the segmentation result space. Define and calculate the following sets: the union of s1 and s2 V = s1∪s2 (the blue part of sub-figure c), and the intersection s interest =s1∩s2 (the gray part of subfigure d) and the symmetry difference Δs = V∩(s interest )^c=(s1∪s2)\(s1∩s2) (the yellow part of subgraph e), under the geometric constraints of V=s1∪s2 and geodesic distance, traverse each point p of the symmetric difference Δs in the form of convolution iteration, calculate the corresponding spherical set B(p,r), and the intersection B of the spherical set and the union V=s1∪s2 of the two blood vessel segmentation results. V (p,r), then ||B V (p,r)|| delineates the size of the blood vessel (gray area in subplot f), while subplot g provides a more intuitive ||B V (p,r)||An example of characterizing blood vessel size, obtaining the corresponding ||B at different points p1, p2. V (p1,r)||、||B V (p2,r)|| delineates the different blood vessel sizes at these two points.

[0044] This embodiment describes the consistency analysis method for two / subsequent vessel segmentation result matrices. For two or more subsequent vessel segmentation result matrices, by repeating the above steps, the statistical results and visualization results of any two subsequent vessel segmentation result matrices can be obtained.

[0045] Example 2

[0046] The following example illustrates a method for evaluating the consistency of multiple / segmented vessel segmentation results provided in Example 1.

[0047] In this embodiment, the raw data of the vascular dataset is a 3D magnetic resonance liver vascular imaging dataset. Please refer to [link to relevant documentation]. Figure 3 As shown, sub-image a is the original data, using a two-dimensional screenshot, showing a local area of ​​the blood vessel segmentation results s1 and s2 superimposed on a certain layer of the original image. Sub-image b is the blood vessel segmentation result s1, and sub-image c is the blood vessel segmentation result s2. Both are displayed in three dimensions, showing the 3D blood vessel structure of the two result matrices s1 and s2, which are manually annotated by two independent individuals. In this embodiment, the difference between the two result matrices s1 and s2 is intentional, in order to demonstrate the recognition performance of this method for blood vessel segments with obvious segmentation differences.

[0048] With a sphere radius r = 10 voxels, statistical and visualization results were obtained using the method described in Example 1. Please refer to [link to example]. Figure 3 The subgraph d, with ||B V (p,r)|| is the x-coordinate (for all ||B) V (p,r)||arranged in ascending order), and two visualizations are displayed simultaneously: one is based on ||B||taking different values. V (p,r)||The number of points p at time point is the ordinate. At this point, the height of each bar reflects the proportion of blood vessels of different thicknesses in the vascular tree, i.e., the dark histogram in subgraph d. From this, we can see that small blood vessels account for the largest proportion (dark red), while medium-sized blood vessels (dark yellow) and large blood vessels (dark green) account for a smaller proportion. The so-called division of large, medium, and small blood vessels can be arbitrarily defined according to specific needs. Secondly, it reflects the difference in the size of blood vessel segments between the two segmentation results. At this point, the ordinate is the number of points p within a certain interval multiplied by |(||B s1 (p,r)||-||B s2 (p,r)||)| / ||B V (p,r)||, which is the light-colored histogram in the figure. In this embodiment, the error is greatest in the small blood vessel segment, which is represented by the highest light red bar. This corresponds to the fact that s1 in this embodiment segmented significantly more small blood vessels than s2. When such a significantly high value appears, it indicates that there is a significant difference between the two segmentation results of the blood vessel segment in this size range. The height of the light yellow and light green bars is almost 0, which indicates that the difference between the large and medium-sized blood vessel segments in these two segmentation results is very small.

[0049] The differences between the two segmentation results are traced and labeled to the corresponding s1, s2, and original data, providing a visual display of the differences between the two segmentation results. Please refer to [link / reference]. Figure 3As shown in sub-image e, a two-dimensional screenshot reveals a local layer of the original image where the differences between the two segmentation result matrices are superimposed on a specific part of the image. The red area represents the differences in small-sized blood vessels, the yellow area represents the differences in medium-sized blood vessels, and the green area represents the differences in large-sized blood vessels. Please refer to [link to relevant documentation]. Figure 3 Subfigure f, displayed in three dimensions, shows the difference distribution of two manually labeled results / matrices s1 and s2 for two independent individuals. Similar to subfigure e, the red area represents the difference in small-sized vessels, the yellow area represents the difference in medium-sized vessels, and the green area represents the difference in large-sized vessels. From subfigures e and f, it can be seen that the differences are most pronounced among small-sized vessels, which is consistent with... Figure 3 The results of the quantized data of the subgraph d, which reflect the largest error in the small blood vessel part in this embodiment, are consistent.

[0050] Example 3

[0051] This embodiment provides an evaluation system for the consistency of multiple / subsequent vessel segmentation results, corresponding to the evaluation method for the consistency of multiple / subsequent vessel segmentation results described in Embodiment 1, including:

[0052] Data input module: used to read the raw data of the same blood vessel dataset and multiple / subsequent blood vessel segmentation result matrices, and to randomly select two / subsequent blood vessel segmentation result matrices s1 and s2 from the multiple / subsequent blood vessel segmentation result matrices;

[0053] Matrix processing module: used to calculate the union V, intersection S, and symmetric difference Δs of s1 and s2;

[0054] Convolution Iteration Module: Select any point p∈Δs, and perform iterative growth by simulating the path diffusion process using a convolution kernel. Under the geometric constraints of two / times blood vessel segmentation result matrices s1 and s2 and a given geodesic distance r, calculate a spherical set B. The center of this spherical set B is point p, and the radius of the sphere is the geodesic distance r. Calculate the intersection B of the spherical set B with the union V of s1 and s2. V And calculate B respectively V The intersection B with s1 and s2 s1 B s2 Statistical analysis module: used to repeatedly run the convolution iteration module, outputting each point p and its corresponding B. V B s1 B s2 Statistical analysis is performed to obtain the proportion of blood vessels of different sizes and the statistical results of the differences between the two / subsequent blood vessel segmentation result matrices s1 and s2. The statistical results are output and mapped back to the original data and the two / subsequent blood vessel segmentation result matrices s1 and s2.

[0055] Visualization output module: This module is used to repeatedly run the above sub-modules to obtain the statistical results of any two blood vessel segmentation result matrices from multiple blood vessel segmentation result matrices. It then outputs the visualization results of the statistical results of any two blood vessel segmentation result matrices.

Claims

1. A method for evaluating the consistency of multiple blood vessel segmentation results, characterized in that, Includes the following steps: (1) Read the original data of the same blood vessel dataset and multiple blood vessel segmentation result matrices, and randomly select two blood vessel segmentation result matrices from the multiple blood vessel segmentation result matrices. , ; (2) Calculation , union Intersection and symmetry difference ; (3) Select any point By iteratively performing two-dimensional or three-dimensional convolution operations, the convolution kernel is applied to the union of the blood vessel segmentation result matrix. In each iteration, the convolution kernel is used as a weight filter to simulate the diffusion of the blood vessel region. The convolution iteration continues until the set sphere radius is reached. The growth region of the convolutional kernel iteration is limited to two / subsequent vessel segmentation result matrices. , union Inside, the radius of the sphere The shortest path length from one point to another within the vascular segmentation results is calculated using geodesic distance constraints to obtain a spherical set. The spherical set The center is Point, the radius of the sphere is the geodesic distance Calculate the spherical set Union intersection and calculate respectively and , intersection , ; (4) Repeat step (3) to output each point. and corresponding , , Statistical analysis was performed to obtain the proportion of blood vessels of different sizes and the matrix of two blood vessel segmentation results. , The statistical results of the differences are output and mapped back to the original data and the two vessel segmentation result matrices. , middle; (5) Repeat the above steps to obtain the statistical results of any two blood vessel segmentation result matrices in multiple blood vessel segmentation result matrices. Output the visualization processing results for the statistical results of any two blood vessel segmentation result matrices.

2. The method for evaluating the consistency of multiple blood vessel segmentation results according to claim 1, characterized in that: The blood vessel segmentation result matrix is ​​a two-dimensional or three-dimensional binary image matrix. In the binary image matrix, element 1 indicates that the element point belongs to the blood vessel region, and element 0 indicates that the element point does not belong to the blood vessel region.

3. The method for evaluating the consistency of multiple blood vessel segmentation results according to claim 1, characterized in that: The convolution kernel is determined by the radius of the sphere. Increasing the weight of the kernel function results in a decrease in its weight.

4. The method for evaluating the consistency of multiple blood vessel segmentation results according to claim 1, characterized in that: In step (4), the statistical analysis includes: calculating the value of each point. corresponding set , , The modulus, and the statistical analysis of different values. , , The point corresponding to the modulus Number, set The modulus is the corresponding point. The size of the blood vessels at the site, collection , The modulus is a point The two blood vessel segmentation result matrices at the location , The differences between them.

5. The method for evaluating the consistency of multiple blood vessel segmentation results according to claim 4, characterized in that: The visualization process includes: using a set The modulus and the corresponding points Plotting the number of vessels as the horizontal and vertical axes yields a diagram showing the proportion of vessels of different sizes; using a set... The points are obtained by weighting the modulus and the corresponding given weight factors. Plotting the number of vessels as the horizontal and vertical axes yields a distribution map of the difference between the two vessel segmentation result matrices with respect to vessel size.

6. The method for evaluating the consistency of multiple blood vessel segmentation results according to claim 5, characterized in that: The weighting factor is , The absolute value of the difference of the modulus divided by the corresponding The value obtained by modulo.

7. The method for evaluating the consistency of multiple blood vessel segmentation results according to claim 4, characterized in that: The visualization process includes: based on each point The coordinates are traced and the parts of the two blood vessel segmentation result matrices that differ from each other are visualized in the two-dimensional image or three-dimensional model.

8. A system for evaluating the consistency of multiple blood vessel segmentation results, characterized in that, include: Data input module: Used to read the raw data of the same blood vessel dataset and multiple blood vessel segmentation result matrices, and to randomly select two blood vessel segmentation result matrices from the multiple matrices. , ; Matrix processing module: used for calculation , union Intersection and symmetry difference ; Convolution Iteration Module: Select any point By iteratively performing two-dimensional or three-dimensional convolution operations, the convolution kernel is applied to the union of the blood vessel segmentation result matrix. In each iteration, the convolution kernel is used as a weight filter to simulate the diffusion of the blood vessel region. The convolution iteration continues until the set sphere radius is reached. The growth region of the convolutional kernel iteration is limited to two / subsequent vessel segmentation result matrices. , union Inside, the radius of the sphere The shortest path length from one point to another within the vascular segmentation results is calculated using geodesic distance constraints to obtain a spherical set. The spherical set The center is Point, the radius of the sphere is the geodesic distance Calculate the spherical set Union intersection and calculate respectively and , intersection , ; Statistical Analysis Module: Used to repeatedly run the convolution iteration module and output each point. and corresponding , , Statistical analysis was performed to obtain the proportion of blood vessels of different sizes and the matrix of two blood vessel segmentation results. , The statistical results of the differences are output and mapped back to the original data and the two vessel segmentation result matrices. , middle; The visualization output module is used to repeatedly run the data input module, matrix processing module, convolution iteration module, and statistical analysis module in sequence to obtain the statistical results of any two blood vessel segmentation result matrices from multiple blood vessel segmentation result matrices. For the statistical results of any two blood vessel segmentation result matrices, the visualization processing results are output.

Images