WO2013007063A1 - Method for identifying lung fissure on ct image of lungs - Google Patents

Abstract

An automatic detection and segmentation method for efficiently and accurately identifying a lung fissure on a CT image on the basis of three dimensional plane fitting. In the method, a CT image is regarded as a group of three dimensional space point clouds, a lung area is divided into small spherical subdivided bodies, and then a three dimensional plane fitting method is used to look for a lung fissure plane in the small spherical subdivided bodies, so as to convert lung fissure detection having free curve surface characteristics into plane detection, thereby obviously reducing the complexity of the problem. The method has the advantage of being insensitive to noise or abnormal values. Furthermore, in the process of lung fissure detection, identification of lung fissures of different types (that is, an oblique lung fissure and a horizontal lung fissure in left and right lungs) is completed through a simple clustering method. Compared with other methods, the method of the present invention has characteristics of high accuracy, stability, and high efficiency.

Description

Title of Invention: Method for identifying hoof cracks on CT images of the lungs

Technical field:

 The invention relates to the efficient identification of a lung fissure on a chest CT image based on a segmented three-dimensional plane fitting method.

BACKGROUND OF THE INVENTION - Interlobular fissures (referred to as lung fissures) are important markers that reflect the structure of the human lung. A comprehensive understanding of the integrity and structure of the lung fissure has great clinical value in the early detection, classification, disease progression and treatment of lung diseases. Therefore, accurate identification of the lung fissure is important for clinical use. It can be borrowed that high-resolution CT examinations usually contain a large number of images, so it is very time-consuming for experts to manually mark the spatial location of the lung fissure by a picture. The lung fissure usually does not appear on the CT image. It is clear that the accuracy and consistency of the results are often not guaranteed.

 At present, many methods for identifying lung fissures have been proposed in the world, and most of them include two stages, namely initial identification and further optimization. In the specific implementation process ^ initial detection I identification is generally achieved by determining the region of interest (ROlh region -of-interest). For example, Pu et al. used a simple thresholding process to identify the region of interest; Wumker et al. used the Hessum matrix and structural tensor filtering to determine the initial region. In addition, in the literature [5.·8], the initial region is transformed by The anatomical relationship with the blood vessels and the bronchus determines that the spatial position of the lung fissure is specifically determined after the initial region is determined. The method is basically to determine the straight line in the two-dimensional space or in the three-dimensional space. Find a surface or plane to achieve. The two-dimensional line (2D) method includes the growth curve method [9- 1 0], Vanderbrug's linear feature detection [i i], edge detection [8] and Gaussian distribution and/or average curvature analysis [12]. Considering that the three-dimensional surface or plane (3D) is more comprehensive and accurate than the two-dimensional line (2D), Ukil et al. proposed a ridge measurement method with 3D map search function [13]. Finding planes in the region of interest; Pu et al. [1] proposed a computational methodology using Laplacian smoothing and extended Gaussian images (EGI: extended Gaussian image); Rikxoort et al. [3] -4] proposes the use of second-order logic to combine points into planes; Kutaigk [6j et al. proposed an interactive three-dimensional watershed algorithm. Different from these methods, Zhang et al. [14-16] proposed a template-based template matching method to find the original.

Inside

The lung time split has two characteristics on the CT image. The first is the relatively high brightness of the lung tissue compared with the lung tissue around the lung fissure, which is why the human eye can see the original S of the lung fissure; It is the lung fissure that has a higher density in the voxel-based image space. In order to fully exploit this Two features, this invention proposes a method for efficiently and accurately identifying lung fissures on CT images based on plane fitting. Considering that 3D freeform surfaces can be accurately approximated by many small planes, we propose to subdivide the three-dimensional lung region into many small spherical subdivisions (these subdivisions overlap each other), so that the identification of the lung fissure It is converted to look for a series of plane problems in each sphere. This method consists of five basic steps (Figures 1 to 5): (1) segmentation of the lung region (Fig. 1), (2) spatial segmentation of the crotch region (Fig. 2), (3) median filtering (Fig. 3), (4) Identify the lobes in each subdivision using the plane fitting method (Fig. 4), (5) Classification of the spurs (Fig. 5). Filtration utilizes a relatively high lung fissure contrast, and plane fitting takes advantage of the relatively high density of the lung fissure in the voxel space.

 In this method, the CT image is treated as a set of three-dimensional point cloud data. The lung region is subdivided into many small spheres called subdivisions. The knowledge of the lung fissure is through the plane in each subdivision sphere. Fit to find the corresponding plane.

 The invention has many features - first, this method is very versatile, recognizing both straight lines in 2D space and planes in 3D space.

 Second, in terms of stability, this method is not sensitive to noise or outliers.

 Third, in terms of computational efficiency, the entire lung fissure recognition process takes only about s minutes, and the method invented by Pu'er et al. takes about 20 minutes, while the method invented by VAis Rikxoort et al. [ί7] requires approximately 90 minutes.

 Fourth, in terms of accuracy, the newly invented method has higher accuracy.

 Fifth, the implementation of this method is very clean, as long as the plane fitting method is directly applied to the segmentation in the lung region. In addition, the identification of different types of cleavage can be completed simultaneously by a simple clustering method during the lung splitting detection process. To the best of our knowledge, there is currently no way to identify and classify lung fissures at the same time.

DRAWINGS

 Specific embodiments will be described in detail through the following drawings.

 Figure 1 to Figure 5 are the basic steps of the split-splitting method - Figure 1 lung segmentation;

 Figure 2 breakdown of the lung area;

 Figure 3 median filtering:

Figure 4 identifies the lung fissure by plane fitting analysis; Figure 5 Classification of different types of hoof cracks.

 Figure 6 to Figure 说明 illustrate the performance of the plane fitting method by using two-dimensional and three-dimensional examples:

 Figures 6 and 7 show the fit of a line of points assembled from points in a two-dimensional picture. The green line represents the result obtained by the least squares method, and the red line represents the result obtained by the density fitting method described in the present invention;

 Figure 8 shows a cloud map of a sphere that is subdivided from the chest CT:

 Figure 9 uses the inner product of the point vector and the norm vector (that is, the equation (6p visualizes the point cloud in Figure 8;

 Figure! 0 is the crack detection result (indicated in red) obtained by applying the invention;

 Figure 11 shows the function corresponding to the three-dimensional point cloud in Figure 8 («,

 Figures 12 through 14 clearly show the results of lung fissure identification and classification:

 Figure 12 - Sub-CT image;

 Figure 13 identified, classified lung fissures (represented by different colors);

 Figure Μ is the final segmentation and classification map of the final lung fissure obtained after removing the non-lung fissure region indicated by the arrow in Figure 13.

 Figures 15 through ί7 are examples showing the performance of the inventive method in identifying cracks.

 Figures 18 to 20 are examples of lung fissures identified in the disease with severe bronchiectasis (cystic fibrosis).

 Figures 2 through 23 are examples of the method of the invention for identifying lung fissures. The crack in this example is very blurry on the image.

detailed description

 The plane fitting method is a key part of this new invention and will be clearly described below.

 The first step is the choice of the objective function.

Set N points ρ in the Euclidean three-dimensional space _;

- I,..., N, , Our goal is to find a plane F :: { p \ pe R f(p) 0 } to minimize the energy of the following objective function:

Where 00^, /:') is the distance from point _/ to plane F, w _: > 0 (equal to ! in this study) is the weight of a particular point (equal to ί , which will be all points in this study) All right.) From Equation 1, it can be deduced that the distance function plays a key role in minimizing. It is worth noting that in this study, the scattered points are located in a predefined sphere (in the subdivision). The versatility of the algorithm, because for a set of known points in the open space, there is always a 'surrounding ball'. In the Euclidean king's space, the degree of freedom of the plane is 3, which can be determined by three parameters. The plane F can be defined as f(p) = cos(<3⁄4) sin(^)x + sin( ) sin(^) - + cos( z - p (2) where (",έ?,ρ) is defined The three parameters of plane F. Although the formula (i) can be directly solved by the least squares method, where D{P:,F) =! (^)μ, although the efficiency is high, the result solved by the least square method is It is very susceptible to noise or outliers (Figure 6-7). To this end, we choose to use the distance function D defined by equation (3), which has a relatively accurate solution and is not sensitive to outliers [18-19]:

D{p F) ' (}}

Among them, "is a scalar. In this study, "there is a close relationship with the cracking thickness of the 3 cities, and it is set to a value of about half of the thickness of the lung fissure described (for example, plexus. 0 to L5 mm). Here, we can re-form the formula (!) as

Bringing equations (1) and (3) into equations. By finding a set of optimization parameters (ie (cr^p)), the plane fitting process plexus energy minimizes CF) (ie, equation (1)) Convert to energy maximization (ie, equation (5)).

 N

Ε' (α, θ, ρ) Ε, (F) ^3⁄4 ² ^ w _f E(F)

=∑w _; (u ² ~D( _P; ,F) ² )

= w<(·"' - f(P: ) ² ) + . w;- ( ² - ·"'― )

As can be seen in the cluster (5), only when the distance from the point to the plane F is less than "these points affect the objective function ^Tf ." Since the density of the points in the open space is used to identify the plane of the lung fissure in the subdivision (if any), we propose to use the function of equation (5) to describe the plane F(«, i?, ρ). The cross-sectional area is 3. Thus, the problem to be solved by the density-based plane fitting is to find the optimization parameter (",6*,ρ) by maximizing equation (6),

among them'? 'Expressing the inner product, ) = r ² — · ο — ) ² ) is the cross-sectional area relative to the plane of the sphere ( , i?, ), o and The center and radius of the sphere are respectively divided by S to ensure that the objective function S is independent of S. Therefore, the objective function S can only be determined by the dot density near the plane. Since the objective function (6) is not a convex function, the gradient descent method cannot be used to optimize the parameters. Although exhaustive search is also feasible, the computational complexity is quite high. For example, for grid system - the computational complexity of time 为 is ^^^^ ^. It is difficult to accept in actual operation. Therefore, an effective optimization algorithm is needed. Next, the integral method is used to optimize the parameter ρ. The hypothesis parameter (", ) is fixed. Let " = (cos(a) sin(^), sin(a) sln(^) , cos(^)) represents the norm q of the plane F _; ::: ρ _ί . η:::: cos( ) ύη(θ)χ _; + sin(a) sin(i?) v, + cos( s (7) Bring equation (7) into equation (6) and rewrite it as w _t . (u ¹ - jp _; ) ² ) W; {u ¹ - (q _i - p) ² )

 /(Λ) <«

 ε(ρ\α, θ)

π τ -- (η . ο - ρ) " ) π " -- (η · ο -- ρ) ¹ ) π τ^ -- (η - ο - ρΥ)

(8) Let a variable and a unit vector Μ, when two miri, max q, - < m + 2r .

Ci II - ^w

 (a)) where N = 2r/ λ, which is the lattice spacing of discrete p. Here, we will set a small value (say, (mm), A, B, C is the integral of a, b, c in equation (10) =

(30)

Cj ^:::: L ^c i ^::::C , 〗10 p(^m + jl), i 目标 The objective function formula (8) is: Where l = (p~m~u)/A, = — w + )/ . Here p.- and it are integral multiples of A; therefore, Z and are integers. Using the integral method, we can get the optimized parameters and reduce the computational complexity) λ 0(Ν χ Ν _ρ ) to 0(N + N _p ). The final step is to use the step-by-step refinement strategy to optimize the ^( ,έ ?) denotes maxf (ρ β,6 , and can obtain the optimal parameter pair by solving the equation (12) (" _.pi ,6 _pi )

-gmax ε ( J) . (12) Although the objective function (cr, i?) is not a convex function related to (", έ?), but in the optimal parameter pair (". _ρ near the region ( Ω ) function The value is as large as the function value in the region (described in Figure 1!). We propose to use a step-by-step refinement strategy to efficiently search for the optimal parameter pair (α,6 . We first use a relatively large error ( that is, the pitch search "to search in the space :) in the formula (12) is near optimal parameters (ί ^,), then the predefined error is small then 8 _refine (e.g. ::: 0.01) Search for the final optimization parameter pair in the subdivision space defined by i\a _c ~05δ,α _ε +0.5^]and[^ _c -~0,5ό,θ ₍ , +0.5 ]) (". _ρί , 6^). Considering the characteristics of the lung fissure, in order to avoid eye damage in the local minimum, δ = 0.1 is sufficient for the initial search. For example, in Figure 8-11, the point taken from the chest CT examination A spherical subdivision of clouds, Ω = 2 {(α, Θ) I ''(« - a _op十. (Θ - θ _ορ :† < 0.3} , Figure 11 shows the corresponding 3D point cloud target in Figure 8. function(",? ), where S(«,i?) = maxs(p|«, ). Search for ρ using the thinning strategy

(« ) time reduced from ^ -- to +

, ( ) actual search time from the original

The approximate 10 ^{5 of} Srefir.e ^ efine ' drops to 10 ³ . In order to improve efficiency and ensure the desired accuracy, homogenization (that is, uniform sampling) corresponds to the surface of a half unit sphere (", candidate norm vector "=(cos(«)sin(6,sin(«) Sin(^,cos( )). Otherwise, when the value is small (say, =0, even if set to a different value, "::::(0,0,1) is the same.), there will be many similarities Direction search is repeated in the present invention, for each generation (", and Fan 2

The number vector η, specifying Θ = 0: _: π and _α 0: /sin{6): The time to search for the optimization parameter pair ( ί?) is reduced from ₊ - _Γ · one to

 S ' S r'ej -me

2π , δ ^Δ

δ ² ' ό, r ² ej.m. e . Figures 6-11 verify the performance of the fitting algorithm by two-dimensional and three-dimensional examples. Although there are many outliers, the lines and surfaces (planes) in the picture are still recognized by the quasi-硗. Figures 6 and 7 show the straight line fit of the set of points in a two-dimensional picture. The green line represents the result from the least squares method, while the red line represents the result from the invented density fitting method. Figure 8 shows a pair of balls drawn from the chest CT 3D point cloud image of the body. Figure 9 visualizes the point cloud in Figure 8 using the inner product of the point vector and the norm vector (i.e., equation (6)). Figure iO shows the fit plane (indicated in red) by applying the method of the invention. Figure U shows the function («,?) corresponding to the three-dimensional point cloud in Figure 8.

 Basic steps of splitting the lung fissure: as shown in Figures 1 to 5

 * Pulmonary segmentation: The main purpose of segmentation of the lungs (Figure 2) is to narrow the search for cleft palate recognition, thereby improving computational efficiency and eliminating the possibility of penile crack recognition errors (false) outside the lung region. .

* Hoof area subdivision: Further subdivision of the segmented lung area. First, create a subdivision grid system for each month of the month in the CT with the axis-aligned bounding box AABB, each cell size is 5x5x5 mm3. The center of the subdivision sphere is set at the apex of the grid system, with radius r (see Figure 3). Since the lung fissure generally has a relatively low curvature, we will set the r value to 10 mm. The reason why I use a spherical surface is not a cube because of the rotation invariance of the sphere and the efficient calculation of the cross-sectional area of the sphere in any direction. The spacing is 5 mm. In the Euclidean three-dimensional space, any point G to the nearest vertex 0 is less than ^3 x 5 ² / 2 mm or 4.4 mm.

 * Point filtering: In order to narrow the search for lung fissures, we use two filters to process the hoof area. First, the luminance value from the city crack is generally from -900 HU (Hausfield unit) to -300 HU, so that voxels outside this luminance value range are eliminated as non-fractured voxels. Second, since most of the fissure regions have higher brightness than the surrounding area (voxels), some voxels with local values smaller than the intermediate intensity values can be proposed. The size of the local area of interest is defined as the size of the sphere in the subdivision of the lung region.

* Pulmonary split recognition: Identify each other by plane fitting in each subdivision! Suppose a sphere with a radius of 10 mm centered at ₀ (X., ), and the set of points to be selected in the sphere after the filtering operation ^^; "Apply the objective function of (6) to fit The optimal plane f after parameterization of ( , θ, ρ). Since the voxel brightness values representing the lung fissures are different in different detections and in different regions in the same detection. Equation (6) The weighting values of all the points in the point are set to the same value (that is, ^ 1), that is, only the plane fitting density distribution is considered. In order to accurately identify the planar structure in each subdivision (ie, the lung fissure), only the The plane of the sphere (0) is less than 5 mm in plane _FO, p), which is ^/ ^ /(^ ^5!7^, where (0) is defined in equation (2). This ensures the sphere; The cross-sectional area of F in r(r ² -d ² ) is larger than ? r(10 ² -5 ² ) « 236 mm The distance from any point G to the nearest lattice vertex is less than 5 mm, then there is interest from the center of the sphere to the Point 细分 Subdivision smaller than 5 mm (center is at the nearest vertex;) Therefore, all points on the mesh are considered in the fitting process. Assume the recognized ball Optimal plane ^ (,), At the same time, ^^ and ( - Oj + ( ? . · ο respectively indicate the norm vector and the (circle) cross-sectional center associated with the spherical volume, if the energy function £fF is greater than the predetermined threshold 7, the identified circular cross-section (plane It is considered to be (assuming) a part of the lung fissure. In this invention, an adaptive threshold of 7" is applied (say, Γ

Applying the density-based fitting algorithm to all subdivisions will identify the lung fissure (as shown in Figures 12-14). Figure 12 - Figure 14 shows the results of lung fissure recognition and classification: Figure 12 - Sub-CT image: Figure 13 identified, classified lung fissures (covered in different colors), Figure 14 is removed by the arrow in Figure 13 Final segmentation and classification map of the final lung fissure obtained after the non-lung fissure

* Classification of lung fissures: In addition to identifying lung fissures, this new invention can simultaneously classify lung fissures by assuming that two lung fissure patches f are identified in two adjacent spheres _; and if one of the following conditions is met, Suppose the two belong to the same aggregation (that is, ~),

/ is the 26-connected neighbor of Fj, and ί ·" ί ^{> (} 3⁄4 and G ÷

< e ₂ .

{2) 3k, sA., E ~ F _k and _; ~ F _k .

Identified and assumed plane block parameters in the adjacent sphere <¾ to allow or curvature continuity between the control and the distance between the corner, and for controlling parameters of _62. Here, by experience, it will be set to 0.98 mm, and the e ₂ will be set to 2 mm. The aggregation interval is arranged in descending order of the number of cracks contained in each aggregate. The first (largest) aggregate of the left lung is used to represent the left oblique split. The first and second (maximum) aggregates of the right lung are used to represent the right oblique and transverse lung fissures, respectively (Figures 12-14). All other aggregates were removed as non-lung fissures

 Result

 To assess the performance of the invented method, we collected 20 Q cases of chronic obstructive pulmonary disease (COPD) cranial CT data and used this method to detect lung fissures in these data, using the same data as the previous method. Compare. The recognition performance is measured by Hausdorff distance and the cumulative error distance distribution (CEDD). The lung fissures identified by these two methods are about 95% within 2 mm, and the average error is about 0.6 mm. The average crack is the largest. The error is 16 mm. In terms of computing time, this new method takes about 8 minutes on a normal computer (configured as an Intel Core 13, 3.2 Hz central processor, 6.0G memory desktop) for lung fissure on CT images.

 In order to visually demonstrate the performance of the new invention method in the detection of lung fissures, I have provided examples in Figures 15-17. In the example, the identified crack is color-coded.

The examples provided in Figures 18 through 20 show lung fissure recognition in an abnormality test with severe bronchiectasis (cystic fibrosis). 21 to 23 are examples in which it is difficult to identify a lung fissure by the new invention method. It is very difficult to clearly see the location of the lungs. Applying the previous method [1] to this test is completely invalid.

 Here, a fully automated method for identifying and classifying lung fissures is described, by exploiting the characteristics of the density distribution of the lung fissures in the open space of the image. This method converts complex free surface recognition into a series of plane fittings, which is characterized by versatility and efficiency.

While the invention has been described by reference to the specific embodiments of the present invention, it will be understood that many modifications, modifications and/or changes may be made without departing from the spirit and scope of the invention. The invention is intended to be defined by the scope of the claims and their equivalents.

references

[1] J. Pu _f J. Κ. Leader, Β. Zheng, F. noi!mann, C. Fuhrman, FC Sciurba, and D, Gur, "A computational geometry approach to automated puinionary fissure segmentation in ct examinations/' IEEE Transactions on Medical Imaging, voi. 28, no. S _f pp 710-719, 2009.

[2] R. Waemker, T, Blow, and T, Blaffert, "Unsupervised extraction of the puimonary interlobar fissures from hagh resolution thoracic ct data '" In Proc. 13th Int. Congress ExhlbitionComputer Assist. RadioL Surg. (CARS), Berlin , Germany, 2005 ₍ pp. 1121-1126.

 [3] E. . van Rikxoor B, van Ginneken, iVL Kl k, and iVL Prokop, "Supervised enhancement filters: Apphcation to fissure detection in chest ct scans/' iEEE Transactions on Medical imaging, vol. 27, no. 1, Pp. 1-10, Jan. 2008,

[4] E iVL vn Rikxoor B de Hoop, S van de Vorst, M, Prokop, and 8, van Ginneken, "Automate segmentation of pulmonary segments from voiumetric chest ct scans/' IEEE Transactions on Medical Imaging, vol. 28, no , 4 _; pp. 621-630, 2009

[5] S, Saiia, M. Kubo,. Y. Kawata, N. Niki, H. Ohmaisu, and N. Moriyama, "An algorithm for the extraction of puimonary Assures from low-dose muitisHce ct image " System and Computers in Japan, vol 37, no, 9, pp _t 63-76, 2006.

[6] J. -M, Kuhnigk, HK Hahn, M. Hmdennach, V Dicken, S, rass, and H.-O. Peitgen, "Lung iobe segmentation by anatomy-gLfided 3d watershed transform/' in SP\E _f 2003, pp. 1482—1490.

 [7] S UkH and J, IVL Reiniwdt "Anatomy-guided lung iobe segmentation m x-ray ct images;" IEEE Transactions on Medical Imaging, vol. 28, no. 2, pp. 202-214, 2009.

 [8] X. Zhou, T Hayashi, T, Mara,, H Fuji and R, Yokoyama, "Automatic recognition of lung lobes and Assures from muiis-slice ct images, in SPIE,. 2004, pp. 1629-1633.

 [9]丄Wang,, M, Betkel, and』, Ko, "Shape-based curvature growing model and adaptive regu arization for pulmonary fissure segmentation in ct/'' in MiCCAI, LNCS 3216, 2004, pp, 541-548.

[10] J. Wang, M. Betke, and and, o,, "Pulmonary fissure segmentation on CT" Medical Image Analysis, voL 10, pp 530-547, 2006.

 [11] M, Kubo, N. Niki, S. Nakagawa. . Eguchi, . Kaneko, N oriyama, H. Omatsu, R. \ R, and N. Yamaguchi, "Extraction algorithm of pulmonary fissures from thin-section ci images Based on linear feature detector method " ΕΕΕ on on Nuclear Science, vol 46, no. 6, pp. 2128-2133, 1999.

[12] . Kubo., Y. avvata, . Niki, K. Eguchi, H, Omatsu, R, K. R _; M. aneko, , Kusumoto, N, oriyama., K, Mori, and H Nishiyama, "Automatic Extraction of pulmonary fissures from multidetector-rovv ct images," In Proceedings of the IEEE international Conference on Image Processing (iCIPOl), 2001, pp. 1091-1094,

[13] . Li, X. Wu, D Chen, and iVI Sonka, "Optima! surface segmentation in volumetric mages- a graph - theoretic approach/' iEEE transactions on Pattern Analysis and Machine InieHsgence, VoL IS, no. 1 _; pp 119-134, 2006.

[14] L 2hang and J, Reinhardt, "Detection of pulmonary lobar Assures using fuzzy logic, in Proc. of SPIE 1999, pp. 1SS --199.

 [15] L Zhang, Hoffman, and 』. Reinhardt "Pulmonary iobe segmentation by graph search with 3 -d shape constraints/' in SPIE, 2001, pp. 204-215.

 [16] L Zhang, EA Hoffman, and J. IVL Remhardt, "Atlas-driven lung iobe segmentation in vo umelnc x-ray cl images/' iEEE Transactions on Medical imaging, vol. 25, no. 1, . 1-16 , 2006.

 [17] E van Rikxoort, M. Prokop, B. de Hoop,, M, V iergever, J. Pluim, and B. van Ginneken, "Automatic segmentation of pulmonary iobes robust against incomplete fissures," IEEE Transactions on Medical imaging, Vol 29, no, 6, pp. 1286—1296, 2010,

 [IS] P. To rr and D. Murray, "The development and comparison of robust methods for estimating the fundamental matrix" !nternaisonai journal of Computer Vision, vol 24, no. pp. 271-300, 1997.

 [19] S. Choi, T. Kim, and W. Yu, "Performance evaluation of RAN SAC family/' in Proc. of the 20th Brstssh Machine Vision Conference, 2009; pp. 1. 12

[20]". Pu, DS Pa k, X. eng, J. E, Roos, and GD Rubin, "Shape break-and-repair strategy and its application to computerized medical image segmentation/' IEEE Transactions on Visualization & Computer Graphics , vol 17, no. 1,

ZZ6.0/llOZN3/X3d C90.00/CT0Z OAV

Claims

 Claim

The rights include:

1. An automatic calculation method for lung fissure recognition based on an analytical plane fitting method, the method comprising

Select the objective function; optimize the parameters

2. The method of claim 1 wherein the selection of the objective function comprises:

Set to plane F is f(p) = cos(a) sin(i?)x + sin(a) sin( ' + ο%(θ)ζ - ρ;

Ifip) ² if f(P;f <u

Define a distance function that is insensitive to noise or outliers ∑) D( D

(α,θ, p):::: E'' (F):::: u ² w; - The function '(«, ί?, ρ) in E(F) divided by the plane F(«, p )of

= ^w i ( ^{u 2} - f(P; ) ² ) ÷ v

The cross-sectional area S is to ensure that the objective function f is independent of S, therefore, the objective function is determined by the dot density near the plane

3. The method of claim j, wherein optimizing the parameter further comprises: optimizing the parameter p by means of integration.

The optimization parameter pair ^^ also includes: using a multi-scale or a step-by-step refinement method to effectively search for parameter pairs

5. The method according to claim 4, further comprising: ^ a relatively large error (search pitch Θ) in (", searching for an approximate optimization parameter pair in space];

Use a predefined smaller error. 0.01) to search within the subdivision space defined by (|a _f . - 0.5δ, a _c + 0.5c?] and [θ _(: - 0.5 θ _:: + (15δ]) Finally optimize the parameter pair ( _a.t , θ ₀ λ.

6. Subdivision methods for the lung area, including the use of axis-aligned bounding boxes 建 Create a subdivision system for each CT test.:

7. ffl Two filters handle the segmented lung area.

8. Identify the planar fitting method of the segmentation in the candidate segment.

9. In the process of lung fissure detection, in order to simultaneously identify different types of lung fissure clustering methods.