AU2020101643A4

AU2020101643A4 - Method for automatically constructing patient-specific anatomic model based on statistical shape model (SSM)

Info

Publication number: AU2020101643A4
Application number: AU2020101643A
Authority: AU
Inventors: Zhaoyang SONG; Wei Wei; Zhuye XU; Xiaoqiang ZHAO
Original assignee: Lanzhou University of Technology
Current assignee: Lanzhou University of Technology
Priority date: 2019-12-29
Filing date: 2020-08-04
Publication date: 2020-09-10
Anticipated expiration: 2028-08-04
Also published as: CN111127488A; CN111127488B

Abstract

The present invention provides a method for automatically constructing a patient-specific anatomic model based on statistical shape model (SSM), which mainly includes four processing phases, namely the phase of splitting of a region of interest, the phase of vertex shrinkage, the phase of response setup, and the phase of generation of the patient-specific anatomic model. According to the invention, firstly, preprocessing of input data and enhancement of a region of interest are carried out, and the region of interest is split from each of CT scan images of specific patients. The split regions of interest are then used as training samples which are subjected to gridding with a triangle, and triangular grids are subjected to iterative shrinkage by a vertex shrinkage strategy. A response is then set up between a template sample and a target sample, and the template sample is aligned to the target sample. Finally, a general expression for anatomic structures of patients is generated. By this method, a patient-specific anatomic model can be effectively constructed to assist medical professionals with quick analysis of a patient's condition, thereby providing a reasonable plan for the patient's follow-up treatment. 1 / 6 0CD a> SampleD/Generation of Image Image Sample Gera en nhancemenD - Registration _, Image -+Expression of E) Respo ) 0 a-= . nse Alignment -*Anatomic U) Setup FIG. 1 FIG. 2

Description

1 /6

0CD a> SampleD/Generation of Image Image Sample Gera ennhancemenD - Registration _, Image -+Expression of E) Respo )0 a-= . nse Alignment -*Anatomic U) Setup

FIG. 1

FIG. 2

I METHOD FOR AUTOMATICALLY CONSTRUCTING PATIENT-SPECIFIC ANATOMIC MODEL BASED ON STATISTICAL SHAPE MODEL (SSM)

TECHNICAL FIELD The present invention belongs to the technical field of medical image processing, and particularly relates to a method for automatically constructing a patient-specific anatomic model. BACKGROUND As new medical image models originated abroad in recent years, statistical shape models (SSMs) have now been widely used for constructing patient-specific anatomic models from shape groups in many domains of medical image processing, such as quantitative and qualitative analysis for computer aided diagnosis, recognition of biomechanical characteristics of the knee joint, virtual reconstruction of facial defects, surgical planning and intraoperative navigation. An effectively constructed patient-specific anatomic model can assist medical professionals with better data visualization and interaction with volumetric data obtained by a three-dimensional imaging modality (e.g., computed tomography (CT) or magnetic resonance imaging (MRI)). Meanwhile, with the aid of the constructed patient-specific anatomic model, it is possible for the medical professionals to judge a patient's condition quickly and accurately at reduced risk of misdiagnosis, thus allowing for effective follow-up treatment. However, it has always been a problem to medical staff on how to effectively construct a patient-specific anatomic model in clinical practice, during which the following major challenges are presented: (1) with regard to most traditional method for constructing patient-specific anatomic models where feature points are manually marked, complicated and time-consuming marking process due to too many feature points, and low feature point marking accuracy resulting from that pseudo-points may be easily marked as feature points; (2) the failure to effectively set up a response between a template sample and a target sample during the construction of a patient-specific anatomic model; and (3) poor performance of the constructed patient-specific anatomic model, leading to the failure to assist the medical professionals with accurate analysis of a patient's condition, thereby affecting the next treatment plan for the patient. SUMMARY In view of the problems in the prior art, an objective of the present invention is to provide a method for automatically constructing a patient-specific anatomic model based on a SSM. By this method, a patient-specific anatomic model can be effectively constructed to assist the medical professionals with quick analysis of a patient's condition, thereby providing a reasonable plan for the patient's follow-up treatment.

To this end, the following technical solution is adopted in the present invention. A method for automatically constructing a patient-specific anatomic model based on a SSM includes the following steps: (1) with regard to a particular diseased site in need of construction of an anatomic model, inputting three-dimensional CT scan images of N different patients with respect to this diseased site as input data, slicing each CT scan image and performing image enhancement on the same, thereby enhancing a region of interest; (2) splitting the region of interest from each CT scan image by a region growing method; (3) generating a grid on the surface of each region of interest by Marching Cubes method while maintaining connectivity between slices, smoothing the surface of the region of interest of each slice image by using a mean filter, and defining the smoothed surfaces of the regions of interest of the slice images as a training data set Q, n {SiS ,.SN; 2

(4) randomly selecting a group of images from the training data set Q, and visualizing the group of images as a template sample Si; (5) subjecting the visualization result of Si to gridding with a triangle, recording the number of surfaces of the triangle and the number of vertexes of the triangle, and then subjecting each group of images in Q to gridding with a triangle; (6) subjecting the triangular grids of Si to iterative shrinkage by a vertex shrinkage strategy; subjecting the remaining samples tS2I3...SN Iin the training data set Q to vertex shrinkage in the same way, and selecting S2 obtained after the vertex shrinkage as a target sample; (7) after all the samples in the training data set Q are subjected to the vertex shrinkage, carrying out registration between Si obtained after the vertex shrinkage and each of { 2,S3...SN obtained after the vertex shrinkage and setting up a response therebetween; and carrying out registration between i and S2 by B-spline free-form deformation method; (8) then aligning S1 to each of other samples in the training data set Q through Procrustes analysis, thereby eliminating effects on the samples caused by rotation and scale variations; (9) simulating the shape change of the patient-specific anatomic model S using normal distribution 0:

s~O(SC) I N = S N ,=,

CN-1 = where S represents a mean shape, C a covariance matrix, and S. a sample image obtained after the alignment in step (8); and (10) generating a general expression of the patient-specific anatomic model S through Principal component analysis, which is represented as the sum of the mean shape and a primary deformation model: M s =:;7+ umern rn=1 where U. and em represent the m th feature value and the m th feature vector of the anatomic model S, respectively; the deformation model is a model at different feature values from the anatomic model S; M is the number of the primary deformation models, which is equal to the number of the feature values; since the feature vectors are essentially sorted in descending order, the following relation is given: U, >U+ upon fast attenuation of U, the final patient-specific anatomic model S is approximated from first G deformation models to: G s= +Xumem rn=1 and selection is made on G according to the following relation: G I Um MI > P

Yu m=1

where P is designed to represent a percentage of all deformation models using first G deformation models. The method provided in the present invention mainly includes four processing phases, namely the phase of splitting of a region of interest, the phase of vertex shrinkage, the phase of response setup, and the phase of generation of the patient-specific anatomic model. According to the invention, firstly, preprocessing of input data and enhancement of a region of interest are carried out, and the region of interest is split from each of CT scan images of specific patients. The split regions of interest are then used as training samples which are subjected to gridding with a triangle, and triangular grids are subjected to iterative shrinkage by a vertex shrinkage strategy, so that the number of vertexes of the triangular grids is effectively reduced while guaranteeing undistorted samples. After the triangular grids are subjected to vertex shrinkage, a response is then set up between a template sample and a target sample by B-spline free-form deformation, and the template sample is aligned to the target sample through Procrustes analysis. Finally, a general expression for anatomic structures of patients is generated through Principal component analysis. The present invention has the following advantages: The present invention proposes a new method for constructing a patient-specific anatomic model on the basis of SSMs, where feature points can be marked automatically and rapidly, and the number of vertexes of triangles is effectively reduced by the vertex shrinkage strategy while guaranteeing undistorted training samples, thereby reducing the time needed by calculation. Meanwhile, the response is set up successfully between the template sample and the target sample, and finally, a good-performance patient-specific anatomic model is constructed, providing a solid foundation for a doctor to analyze a patient's condition. BRIEF DESCRIPTION OF DRAWINGS FIG. 1 is a flow chart of a method according to the present invention. FIG. 2 shows the visualization result of a template sample Si according to an example. FIG. 3 shows the result of gridding of Si according to an example. FIG. 4 is a schematic diagram of vertex shrinkage of 1 according to an example. FIG. 5 shows the result of vertex shrinkage of Si according to an example. FIG. 6 shows registration between a template sample Si and a target sample S 2 according to an example. FIG. 7 is a schematic diagram illustrating a percentage of all deformation models represented using first G deformation models according to an example. FIG. 8 is the visualization result of a final patient-specific anatomic model according to an example. FIG. 9 shows comparison of performance between patient-specific anatomic models different in the number of mark points according to an example. FIG. 10 shows comparison of performance between patient-specific anatomic models obtained by different methods according to an example. DETAILED DESCRIPTION The present invention will be further described in detail below in combination with accompanying drawings and specific examples. Taking left femur for example, a patient-specific anatomic model is constructed by the method provided in the present invention, which specifically includes steps as follows. (1) Three-dimensional CT scan images of 30 different patients with respect to pelvis are input as input data, and each CT scan image is sliced and subjected to image enhancement, thereby enhancing the left femur region of interest. Each CT image has a resolution of 0.9 mm in plane and of 1.5 mm between slices. The information of the 30 patients is shown in Table 1:

Table 1 Anthropometric and Demographic information of 30 Left Femur Images Number of Body Height Surfaces of Image Number of Vertexes No. Age Sex Weight Subjected to of Image Subjected (Kg) (cm) Triangular to Triangular Gridding Gridding 001 25 Male 61 175 102080 51042 002 29 Male 72 180 98744 49370 003 32 Male 70 167 87876 43940 004 26 Male 66 165 109528 54752 005 26 Male 63 170 112800 56410 006 29 Male 80 180 106268 53130 007 30 Male 83 176 156348 78176 008 35 Male 79 175 117744 58876 009 32 Male 75 170 97116 48560 010 32 Male 80 168 76204 38110 011 27 Male 62 166 114452 57224 012 26 Male 60 181 99548 49770 013 30 Male 80 182 95536 47760 014 33 Male 83 178 99512 49746 015 30 Male 85 179 92184 46084 016 42 Female 65 159 95432 47708 017 40 Female 62 158 88288 44140 018 41 Female 62 162 56244 28120 019 36 Female 60 163 80476 40236 020 34 Female 61 172 93272 46640 021 26 Female 58 173 91248 45600 022 24 Female 56 170 107988 53988 023 31 Female 56 160 110216 54922 024 30 Female 57 161 88268 44114 025 25 Female 45 167 8840 4432 026 26 Female 40 168 87296 43648 027 28 Female 46 166 123404 61698 028 30 Female 61 159 86020 43024 029 38 Female 65 158 102688 51344 030 40 Female 60 163 89912 44960 (2) The left femur region of interest is split from each CT scan image by a region growing method. (3) A grid is generated on the surface of each region of interest by Marching Cubes method while maintaining connectivity between slices. The surface of the region of interest of each slice image is smoothed by using a mean filter, and the smoothed surfaces of the regions of interest of the slice images are defined as a training data set , = S 1 iS ,--.SN}, with 2

N=30. (4) A group of femur images 001 is randomly selected from the training data set Q, and visualized as a template sample S 1 , with the result being shown in FIG. 2. (5) The visualization result of Si is subjected to gridding with a triangle (see FIG. 3 for the result), and the number of surfaces of the triangle is recorded as 102080 and the number of vertexes of the triangle as 51042. Then, other sample images in Q are subjected to gridding with a triangle. After the gridding, the number of vertexes of the triangle in each sample femur image ranges from 4432 to 78176, and the number of surfaces of the triangle ranges from 8840 to 156348. (6) As shown in FIG. 4, the triangular grids of Si are subjected to iterative shrinkage by a vertex shrinkage strategy, which specifically includes: (6i) assuming that V" represents a vertex of Si after being subjected to triangular gridding, (V','V) a vertex pair, and (Va,V)9Vo moving vertexes V. and Vb to a new vertex Vq which is connected to V, and deleting V; (6ii) as such, shrinking a series of vertexes ofS I after being subjected to triangular gridding to one vertex: (12, V3, k-) V6, during which one vertex is deleted at each iteration; and (6iii) constraining the vertex pair (a'K') so as to effectively reduce the number of vertexes of the triangular grids while guaranteeing undistorted template sample I, under the following constraint conditions: (a): a being not a special vertex; and (b): (,' ,') being an edge vertex pair. The result of vertex shrinkage of 1 is shown in FIG. 5. Then, the remaining samples 23---SN Iin the training data set Q are subjected to vertex shrinkage in the same way, and S2 obtained after the vertex shrinkage is selected as a target sample. By observation, the number of surfaces of the triangle of SI after being subjected to the vertex shrinkage decreases from 102080 to 13478, and the number of vertexes decreases from 51042 to 6726, thereby effectively reducing the number of vertexes while guaranteeing undistorted femur image. (7) After all the samples in the training data set Q are subjected to the vertex shrinkage, registration is carried out between S obtained after the vertex shrinkage and each of {S2, 5 3..SN obtained after the vertex shrinkage and a response is set up therebetween. As shown in FIG. 6, registration is carried out between S and S2 by B-spline free-form deformation method, which specifically includes: (7i) assuming that each of I and S2 has L mark points, and deforming S1 to S2 by mapping which satisfies the following relation: ie L where - represents the ith mark point, B' the base of B-spline corresponding to , uE [0,1]d a parameter value, and d dimensions of Euclidean space ( d = 2,3 (7ii) carrying out registration of the template sample S1 to the target sample S2 according to the following formula:

Min (deviation + AOsmooth Ulndmarks

where Cdeviation represents the quadratic sum of deviation between the deformed template sample (S1) and the target sample S2

2

Deviation 1) 2,c

#(u)= BB(U) 'j1 v1) where 0 is mapping defined by , I any vertex in vi , and

v(2,c) (1) V1 a vertex in S2 that is closest to V

Cosmooth is a smoothing item having a smoothing coefficient A which is set to decrease gradually from a great initial value; here, A is set to decrease gradually by 10 each time from 100, namely the initial value:

Smooth =2 2

and Olandmarks is a matching error for the mark points to guarantee that each mark point on the template is mapped onto a corresponding target mark point:

L2

Landmarks j(- 2) j=1

where (12 paired )L}Lrepresents mark points between S and

S2 , and the coefficient 9 a weight for the mark points, which is set to 10 herein. (8) S1 is then aligned to each of other samples in the training data set Q through Procrustes analysis, thereby eliminating effects on the samples caused by rotation and scale variations. (9) The shape change of the patient-specific anatomic model S is simulated using normal distribution 0:

s~6(sC) 1 N 2s S" Nn1 iN

C= N -1

where S represents a mean shape, C a covariance matrix, and Sn a sample image obtained after the alignment in step (8). (10) A general expression of the patient-specific anatomic model S is generated through Principal component analysis, which is represented as the sum of the mean shape and a primary deformation model: M S= + u+ e M=1

where um and e- represent the m th feature value and the m th feature vector of the anatomic model S, respectively; the deformation model is a model at different feature values from the anatomic model S; M is the number of the primary deformation models, which is equal to the number of the feature values; since the feature vectors are essentially sorted in descending order, the following relation is given:

um > u+1

upon fast attenuation of U, the final patient-specific anatomic model s is precisely approximated from first G deformation models to: G s = + umem , and the following relation is given:

G

m=1 > M

=1m

where P is designed to represent a percentage of all deformation models using first G deformation models. With regard to 30 left femurs, a decision of P = 98.5% is made, and the number of the corresponding deformation models can be obtained by calculation as G = 7, i.e., 98.5% of deformation models can be represented by first 7 deformation models. By the above method, samples, which have 738, 512 and 356 mark points, respectively, are selected to construct patient-specific anatomic models, respectively. Moreover, using four performance indexes, namely the compactness, specificity, generality and representation of a model as evaluation criteria, comparison between each of Spherical Harmonics Descriptors Method (SPHARM), Minimum Description Length (MDL), Landmark Sliding Method (SLIDE) and the like, and the method provided in the present invention, and results are shown in FIG. 10. The results indicate that the method provided in the present invention is capable of effectively constructing a patient-specific anatomic model which is better in performance than those obtained by such methods as SPHARM, MDL and SLIDE. As described herein, the region growing method, the Marching Cubes method, triangular gridding, Procrustes analysis, Principal component analysis and the like are conventional methods well known in the art, and thus are not explained herein.

Claims

What is claimed is: 1. A method for automatically constructing a patient-specific anatomic model based on a statistical shape model (SSM), comprising the following steps: (1) with regard to a particular diseased site in need of construction of an anatomic model, inputting three-dimensional CT scan images of N different patients with respect to this diseased site as input data, slicing each CT scan image and performing image enhancement on the same, thereby enhancing a region of interest; (2) splitting the region of interest from each CT scan image by a region growing method; (3) generating a grid on the surface of each region of interest by Marching Cubes method while maintaining connectivity between slices, smoothing the surface of the region of interest of each slice image by using a mean filter, and defining the smoothed surfaces of the regions of interest of the slice images as a training data set Q, {Si,2,''SN}; (4) randomly selecting a group of images from the training data set Q, and visualizing the group of images as a template sample Si; (5) subjecting the visualization result of Si to gridding with a triangle, recording the number of surfaces of the triangle and the number of vertexes of the triangle, and then subjecting each group of images in n to gridding with a triangle; (6) subjecting the triangular grids of i to iterative shrinkage by a vertex shrinkage strategy; subjecting the remaining samples tS2,3...SN in the training data set Q to vertex shrinkage in the same way, and selecting S2 obtained after the vertex shrinkage as a target sample; (7) after all the samples in the training data set n are subjected to the vertex shrinkage, carrying out registration between Si obtained after the vertex shrinkage and each of tS2,3SN Iobtained after the vertex shrinkage and setting up a response therebetween; and carrying out registration between S1 and S2 by B-spline free-form deformation method; (8) then aligning Si to each of other samples in the training data set Q through Procrustes analysis, thereby eliminating effects on the samples caused by rotation and scale variations; (9) simulating the shape change of the patient-specific anatomic model S using normal distribution 0:

s~0( SC) 1 N T = -Is, N ,=, iN

C= N -1

wherein S represents a mean shape, C a covariance matrix, and S a sample image obtained after the alignment in step (8); and (10) generating a general expression of the patient-specific anatomic model S through Principal component analysis, which is represented as the sum of the mean shape and a primary deformation model: M S +u e r=1

wherein ur and e represent the m th feature value and the m th feature vector of the anatomic model S, respectively; the deformation model is a model at different feature values from the anatomic model S; M is the number of the primary deformation models, which is equal to the number of the feature values; since the feature vectors are essentially sorted in descending order, the following relation is given: UM > U+

upon fast attenuation of Um, the final patient-specific anatomic model S is approximated from first G deformation models to: G Ss + um,,em rn=1

and selection is made on G according to the following relation: G

YUm

M P

I Um

wherein P is designed to represent a percentage of all deformation models using first G deformation models.
2. The method for automatically constructing a patient-specific anatomic model based on a SSM according to claim 1, wherein the subjecting the triangular grids of Si to iterative shrinkage in step (6) specifically comprises: (6i) assuming that 1 represents a vertex of Si after being subjected to triangular gridding, (VaV) a vertex pair, and ( )a,V)- V moving vertexes Va and V to a new vertex "Y which is connected to V,anddeleting V; (6ii) as such, shrinking a series of vertexes ofS I after being subjected to triangular gridding to one vertex: (Vi,V2,V3,")-+V8, during which one vertex is deleted at each iteration; and (6iii) constraining the vertex pair (Fa'h) so as to effectively reduce the number of vertexes of the triangular grids while guaranteeing undistorted template sample S 1, under the following constraint conditions: (a): V being not a special vertex; and (b): ( ' )being an edge vertex pair.
3. The method for automatically constructing a patient-specific anatomic model based on a SSM according to claim 1, wherein the carrying out registration between S and S 2 by B-spline free-form deformation method in step (7) specifically comprises: (7i) assuming that each of 1 and S2 has L mark points, and deforming Si to S2 by mapping # which satisfies the following relation: kL wherein represents the ith mark point, B the base of B-spline corresponding to uE[0,1]d a parameter value, and d dimensions of Euclidean space (d = 2,3 ; (7ii) carrying out registration of the template sample Si to the target sample S2 according to the following formula:

Min (deviation + AOsmooth Ulndmarks

wherein COdeviation represents the quadratic sum of deviation between the deformed templatesample (Si) and the target sample S2

2 Udeviatio 1) V(2,c)

wherein 0 is mapping defined by #(u)= B,(u) v(i) any vertex in v e, and

v(2,c) a vertex in that is closest to vI

COsmooth is a smoothing item having a smoothing coefficient which is set to decrease gradually from a great initial value:

25 smooth = K a 2 2dx

andClandmarks is a matching error for the mark points to guarantee that each mark point on the template is mapped onto a corresponding target mark point:

L 2 -(1) -(2) 2 landmarks

wherein ( ),g(2)) , V72 (2) (1)2))I represents paired mark points between S

and S2, and the coefficient u a weight for the mark points.
4. The method for automatically constructing a patient-specific anatomic model based on a SSM according to claim 3, wherein in step (7), the deformed template sample (S1 ) is projected to the target sample S2 in a vertex normal direction, thereby setting up a response between S' and S 2 ; and a response is then set up between S1 and each of S3,S4-.SN in the same way; wherein in step (7ii), A is set to decrease gradually by 10 each time from 100, namely the initial value.
5. The method for automatically constructing a patient-specific anatomic model based on a SSM according to claim 1, wherein in step (1), N is an integer ranging from 15 to 60.

1 / 6 04 Aug 2020

#*'

#&"# #&"

#$)) %$ $') %$% #$&*)

#%%) $

' $! # #&"

' $

'), '%+ $ %$ # $'"

# ()') %$ # $$#$) ,&'(( %$% $

" $#$)

$)%# (&%$( %" )*&

FIG. 1 2020101643

FIG. 2

2 /6 04 Aug 2020 2020101643

FIG. 3

Shrinkage

FIG. 4

3 /6 04 Aug 2020 2020101643

FIG. 5

Template sample S1 After gridding S1 After vertex shrinkage S1

Registration result

Target sample S2 After gridding S2 After vertex shrinkage S2

FIG. 6

Proportion of Deformation Models 4 /6

FIG. 8 FIG. 7 Number of Deformation Models

5 /6 04 Aug 2020 Compactnes

Specificity Number of Mark Number of Mark s

Points:738 Number of Mark Points:738 Points:738 Number of Mark Number of Mark Points:738 Number of Mark Points:512 Number of Mark Points:512 Number of Mark Points:512 Points:512 2020101643

Number of Number of Mark Mark Points:356 Number of Mark Points:356 Number of Mark Points:356 Points:356

Number of Deformation Models Number of Deformation Models (a) (b)

Number of Mark Points:738 Number of Mark Points:738 Number of Mark Number of Mark Points:356 Number of Mark Points:512 Points:512 Number of Mark Points:512 Number of Mark Points:356 Number of Mark Number of Mark Points:738 Points:356 Representation Generality

Number of Deformation Models Number of Deformation Models (c) (d) FIG. 9

6 / 6 04 Aug 2020

%#&)$(

& )- ( 2020101643

*#'%%'#) %$%"( *#'%%'#) %$%"(

&'($)) %$ $'" )-

*#'%%'#) %$%"( *#'%%'#) %$%"( FIG. 10