JOURNAL OF MAGNETIC RESONANCE IMAGING 39:1457–1467 (2014)
Original Research
Semiautomated Analysis of Carotid Artery Wall Thickness in MRI Luca Saba, MD,1* Hao Gao, PhD,2 Eytan Raz, MD,3 S. Vinitha Sree, PhD,4 Lorenzo Mannelli, MD,5 Niranjan Tallapally, PhD,6 Filippo Molinari, PhD,7 Pier Paolo Bassareo, MD,8 U. Rajendra Acharya, PhD,9 Holger Poppert, MD,10 and Jasjit S. Suri, PhD11 Purpose: To develop a semiautomatic method based on level set method (LSM) for carotid arterial wall thickness (CAWT) measurement.
observers is less than 10%, suggesting LSM can segment arterial wall well compared with manual tracings. The Jaccard Similarity (Js) analysis showed a good agreement for the segmentation results between proposed method and GT (Js 0.71 6 0.08), the mean curve distance for lumen boundary is 0.34 6 0.2 mm between the proposed method and GT, and 0.47 6 0.2 mm for outer wall boundary.
Materials and Methods: Magnetic resonance imaging (MRI) of diseased carotid arteries was acquired from 10 patients. Ground truth (GT) data for arterial wall segmentation was collected from three experienced vascular clinicians. The semiautomatic variational LSM was employed to segment lumen and arterial wall outer boundaries on 102 MR images. Two computer-based measurements, arterial wall thickness (WT) and arterial wall area (AWA), were computed and compared with GT. Linear regression, Bland–Altman, and bias correlation analysis on WT and AWA were applied for evaluating the performance of the semiautomatic method.
Conclusion: The proposed LSM can generate reasonably accurate lumen and outer wall boundaries compared to manual segmentation, and can work similar to a human reader. However, it tends to overestimate CAWT and AWA compared to the manual segmentation for arterial wall with small area. Key Words: carotid arterial wall thickness; plaque; level set method; segmentation J. Magn. Reson. Imaging 2014;39:1457–1467. C 2013 Wiley Periodicals, Inc. V
Results: Arterial wall thickness measured by radial distance measure (RDM) and polyline distance measure (PDM) correlated well between GT and variational LSM (r ¼ 0.83 for RDM and r ¼ 0.64 for PDM, P < 0.05). The absolute arterial wall area bias between LSM and three
1
Department of Radiology, Azienda Ospedaliero Universitaria (A.O.U.), di Cagliari – Polo di Monserrato, Cagliari, Italy. 2 Centre for Excellence in Signal and Image Processing, Department of Electronic and Electrical, University of Strathclyde, Strathclyde, UK. 3 Department of Radiology, New York University School of Medicine, New York, New York, USA. 4 Visiting Scientist, Global Biomedical Technologies, Roseville, California, USA. 5 University of Washington, Seattle, Washington, USA. 6 27558 Kingsgate Way, Farmington Hills, Michigan, USA. 7 Biolab, Department of Electronics, Politecnico di Torino, Torino, Italy. 8 Department of Cardiology, Azienda Ospedaliero Universitaria (A.O.U.), di Cagliari – Polo di Monserrato, Cagliari, Italy. 9 Department of ECE, Ngee Ann Polytechnic, Singapore. 10 Neurologische Klinik und Poliklinik Technische Universit€ at M€ unchen, M€ unchen, Germany. 11 Diagnostic and Monitoring Division, AtheroPoint LLC, Roseville, California, and Department of Biomedical Engineering, Idaho State University (Aff.), Pocatello, Idaho, USA. *Address reprint requests to: L.S. Department of Radiology, Azienda Ospedaliero Universitaria (A.O.U.), di Cagliari – Polo di Monserrato s.s. 554 Monserrato (Cagliari) 09045, Italy. E-mail: [email protected]. Received November 17, 2012; Accepted June 18, 2013. DOI 10.1002/jmri.24307 View this article online at wileyonlinelibrary.com. C 2013 Wiley Periodicals, Inc. V
ATHEROSCLEROSIS is the third leading cause of death in the world (1). Early detection of atherosclerosis and proper treatment and lifestyle changes would prevent the onset of cardiovascular diseases like stroke and cardiac failure. With advancements in medical imaging technology, methods are being developed for studying plaque risk stratification, characterization of plaque burdens, and real-time degree of stenosis estimation, lipid size, fibrous cap thickness, and arterial wall thickness. Magnetic resonance imaging (MRI) has been widely used for imaging diseased vascular wall and it assists in accurate characterization of plaque (2). One of the key factors of its success is the ability to identify the adventitia boundary in transverse images of the vessel wall. Research has demonstrated that MRI is able to detect the boundaries of arterial walls with good reproducibility, and the mean arterial wall thickness derived from those detected boundaries has been used for measuring plaque burden (3). Studies (4) have demonstrated that carotid arterial wall thickness (CAWT) correlates well with atherosclerosis burden and is reliable in predicting clinical events and
1457
1458
Saba et al.
treatment outcome. However, the usefulness of CAWT depends on the accuracy and precision of arterial wall boundary segmentation followed by wall thickness calculation using methods such as calculating the average of several radial thickness measurements (4). Methods from manual to semiautomatic and automatic segmentation such as level set methods (LSM) (5), active contour models (ACM) (6), and statistical methods such as active shape model (7) have been widely used for diseased arterial wall detection on MR images (8–14). In this study we present a semiautomatic method based on level set method for segmenting carotid plaque lumen and outer wall boundaries. Carotid wall thickness from segmented boundaries were calculated and compared with manual segmentation carried out by trained clinicians to study the accuracy of quantification, and we also present the performance comparison with ACMs.
X
X
k>0 and y is a constant, d is a univariate Dirac function, H is the Heaviside function, g is the edge indicator function defined as: g5
1 11jrGr Ij2
[6]
Gr is the Gaussian kernel with given s, I is the given image matrix. The first term in Eq. (5) computes the length of zero level curve of u and the second term is the speed up function in the area of V w ¼ fðx; yÞj wðx; yÞ < 0g. Finally, Eq. (2) can be expanded as: du ru ru 5l ru2div 1kdðuÞdiv g 1tgdðuÞ [7] dt jruj jruj The above equation can be easily implemented by simple finite difference scheme.
MATERIALS AND METHODS Level Set Method (LSM) LSM has been widely used for segmentation in the image-processing field (12). The basic idea is to represent the contours as the zero level set of an implicit function, which is defined in a higher dimension, usually denoted as level set function. Then the level set function evolves according to certain partial differential equations (PDEs). In this study, variational level set method without reinitialization (15) was used for arterial wall segmentation. The evolution equation of the traditional level set formulation of the moving front C can be written in the following general form: @u 1F jruj50 [1] @t F is the speed function, which depends on the imaging data and level set function u. C can be represented by zero level set as C(t) ¼ {(x,y)ju(t,x,y) ¼ 0}. In this study we limit our problem domain in the 2D image space. For traditional LSM, u is required to be kept close to a signed distance function during the evolution, therefore reinitialization is required constantly, even though the reinitialization can be very complicated and can have great side effects. In order to overcome these difficulties, the evolution equation is redefined as: @u @E 1 50 [2] @t @u EðuÞ5lP ðuÞ1eðuÞ
motion of the zero level curve of u. In the segmentation procedure, it can be defined as: ð ð eðuÞ5k gdðuÞjrujdxdy1t gH ð2uÞdxdy [5]
[3]
where P(u) is the measurement of the distance of how close a function u is to a signed distance function. P(u) is defined as: ð 1 [4] PðuÞ5 ðjruj21Þ2 dxdy 2 X
P(u) will help eliminate the reinitialization of u during level set evolution. l > 0, which controls the effect of penalizing the deviation of u from a signed distance function. e(u) is the energy term which will drive the
GVF Method Based on ACM, gradient vector flow (GVF) was developed by Xu and Prince (16) by introducing a new static external force field to overcome the disadvantages in traditional ACM, such as sensitivity to initial position and difficulties in penetrating into concavities. The proposed new external force field V(x,y) 5 [u(x,y) v(x,y)], named the GVF field, is solved with the following Euler equations: lr2 u2ðu2fx Þ fx2 1fy2 50 [8] [9] lr2 v2ðv2fx Þ fx2 1fy2 50 where l is the regularization parameter (l>0), r2 is the Laplacian operator, f is the edge map defined as: f ðx; yÞ5jrðGd ðx; yÞ I ðx; yÞÞ2 j
[10]
MRI Protocol and Data Collection MRI Ten patients were selected from consecutive patients from the stroke unit (17) in the January 2012. The mean age of patients (seven males and three females) in this study was 67 6 6 years. The internal carotid artery stenosis was diagnosed by Doppler and duplex sonography, and without showing heavy calcification. In our institution the inclusion criteria for performing MRA of carotid arteries are the following: presence of an ultrasound analysis that showed a carotid stenosis >50% and evidence of plaque alteration (an irregular surface, intraplaque hemorrhage, ulceration). Multispectral MRI scans were performed on the subjects to obtain plaque morphology. The protocol was approved by the local Ethics Committee, and written informed consent was obtained from each patient before the study.
Semiautomated Analysis of CAWT in MRI
1459
software implemented in MatLab (MathWorks, Natick, MA). The procedure of LSM segmentation was done on a slice-by-slice basis as follows: 1) The boundary inside lumen was initialized by a radial expansion from a specific point; 2) The initial boundary will evolve according to Eq. (7), and stop at the luminal boundary. Usually the iteration is chosen to be 100 to ensure the initial boundary will evolve towards luminal boundary as close as possible; 3) The segmented lumen boundary grows outward by two pixels for the initialization of outer wall boundary segmentation; 4) From step 3, the newly initialized boundary will evolve similar as in step 2 according to Eq. (7). Due to poorer contrast between outer wall and surrounding tissues, usually 30 iteration steps were chosen in order to avoid leaking problem in the proposed variational LSM. Figure 2 shows the diagram for the proposed variational LSM method. Data Evaluation Figure 1. Manual segmentation for GT (obtained with ImgTracer) (a): MRI slice of carotid plaque; (b) lumen boundary tracing; (c) outer wall boundary tracing superimposed with lumen boundary; (d) ROI of the carotid wall.
MRI was performed using a 1.5T scanner (Siemens Medical Systems, Erlangen, Germany) with bilateral phased-array surface coils (PACC-SS15, Machnet, Netherlands). The MRI scan was centered on the carotid bifurcation to cover a large range for proper stenosis imaging. T1-weighted, T2-weighted and proton-densityweighted (PDW) studies were performed for each patient. In this study only PDW images were used for plaque segmentation. The parameters for PDW sequence were: flip angle 180 , field of view (FOV) 16 16 cm, TR 700 msec, TE 10 msec, number of excitations 2, slice thickness 2 mm, matrix size 640 640. The longitudinal coverage for each carotid artery was 72 mm. No further process was applied to assess MR image quality. Figure 1a shows a typical MR image of carotid plaque. The MRI scan for one patient was 18 minutes.
The GT and LSM CAWT were computed for each slice. Two kinds of methods were used to obtain CAWT in this study: radial distance measurement (RDM), and polyline distance measurement (PDM). The procedure for RDM is as follows: 1) the center of the lumen was defined after the segmentation according to the lumen boundary; 2) ray lines were constructed from the lumen center. RDM was defined along the same ray lines across the lumen and outer wall boundaries, as shown in Fig. 3a. PDM was defined as the closest distance from the estimated lumen boundary to the outer wall boundary, and PDM is not sensitive to the number of points constituting the profiles. More details can be found in a previously published work (18). Figure 3b presents the RDM and PDM obtained for the plaque MR image shown in Fig. 3a. Since RDM may not always measure the distance along the perpendicular direction between contours, RDM tends to be slightly larger than PDM, as evident in Fig. 3b.
Ground Truth (GT) Data Collection ImgTracer was used for GT data collection. It is a userfriendly software tool designed by Global Biomedical Technologies (Naples, FL) mainly for radiologists for tracing different kinds of anatomy in 2D and 3D medical images. Lumen boundary can be traced by clicking 10– 15 points in a clockwise or anticlockwise direction as shown in Fig. 1b. Similarly, the outer wall can be traced by clicking 10–15 points in an orderly fashion, shown in Fig. 1c. These boundaries are saved as GT data or manual boundaries. Furthermore, the ImgTracer allows computing the region of interest (ROI) of the carotid wall to study the components of the wall, such as plaque area (Fig. 1d). Three experienced clinicians segmented the carotid plaque MR images by using ImgTracer and these were considered as manual tracings. Image Processing and Segmentation Method The variational LSM method was used for segmentation, and it was carried out using the in-house
Figure 2. Diagram for the proposed LSM. [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]
1460
Saba et al.
Figure 3. Definition of RDM (a) and comparison of RDM and PDM for the same plaque (b).
The agreement analysis between LSM delineated and manual delineated carotid arterial wall was done in terms of spatial overlap by using Jaccard Similarity Factor (Js), defined as: Js 5
pixelsðPLSB \ PGT Þ pixelsðPLSB [ PGT Þ
[11]
where PLSB stands for the set of pixels identified by the semiautomatic delineation, and PGT indicates the set of pixels identified in manual segmentation as GT. Therefore, Js is a ratio of number of pixels in the intersection of these sets over the union of these sets. If the two sets of pixels are perfectly in the same spatial position, then Js will reach the maximum value of 1. Furthermore, the curve distance, defined by the mean perpendicular distance for each pair of lumen/ outer wall boundaries from LSM and manual methods, was used to measure the accuracy of LSM delineated arterial wall boundaries.
RESULTS Results Using LSM Figure 4 shows the typical segmentation results obtained using LSM. Initial boundary was defined as in Fig. 4a, which is not close enough to the real luminal boundary. Figure 4b shows the detected luminal boundary using LSM. Figure 4c shows the outer wall boundary combined with luminal boundary and it is evident that the luminal wall has been well segmented. Figure 4d is the overlay of manual and LSM segmentation results. The detected boundaries are very close to the manual segmentation boundaries. However, LSM in Fig. 4 overestimates the arterial wall area by 10%. In total, 102 slices of MR images of carotid plaques were segmented by the proposed LSM and also manually by three trained clinicians. The CAWT values obtained using RDM and PDM are summarized in Table 1 and expressed as mean 6 standard deviation. Unlike RDM, PDM metric overestimates the CAWT thickness. As expected and observed, the average PDM is smaller than RDM in both manual segmentations and our LSM method.
Bland–Altman and Regression Analysis for the CAWT Bland–Altman analysis and regression analysis were performed between the three observers of the GT and the LSM method, summarized in Figs. 5 and 6. In order to test the interobserver variability, Bland–Altman plots between the three observers were analyzed. The limit of agreement between the GT was from 29.8% to þ25.9% in the case of best concordance Table 1 Thickness Comparison Between RDM and PDM
Figure 4. Segmentation results from LSM. (a) Initial contour for lumen boundary; (b) lumen boundary from LSM; (c) lumen/outer wall boundary from LSM; (d) Overlay with GT (line with star stands for GT).
LSM Obser1 Obser2 Obser3
RDM (mm)
PDM (mm)
P
2.34 2.24 2.06 2.28
2.2 6 0.55 1.93 6 0.56 1.64 6 0.44 1.97 6 0.57
Original Research
Semiautomated Analysis of Carotid Artery Wall Thickness in MRI Luca Saba, MD,1* Hao Gao, PhD,2 Eytan Raz, MD,3 S. Vinitha Sree, PhD,4 Lorenzo Mannelli, MD,5 Niranjan Tallapally, PhD,6 Filippo Molinari, PhD,7 Pier Paolo Bassareo, MD,8 U. Rajendra Acharya, PhD,9 Holger Poppert, MD,10 and Jasjit S. Suri, PhD11 Purpose: To develop a semiautomatic method based on level set method (LSM) for carotid arterial wall thickness (CAWT) measurement.
observers is less than 10%, suggesting LSM can segment arterial wall well compared with manual tracings. The Jaccard Similarity (Js) analysis showed a good agreement for the segmentation results between proposed method and GT (Js 0.71 6 0.08), the mean curve distance for lumen boundary is 0.34 6 0.2 mm between the proposed method and GT, and 0.47 6 0.2 mm for outer wall boundary.
Materials and Methods: Magnetic resonance imaging (MRI) of diseased carotid arteries was acquired from 10 patients. Ground truth (GT) data for arterial wall segmentation was collected from three experienced vascular clinicians. The semiautomatic variational LSM was employed to segment lumen and arterial wall outer boundaries on 102 MR images. Two computer-based measurements, arterial wall thickness (WT) and arterial wall area (AWA), were computed and compared with GT. Linear regression, Bland–Altman, and bias correlation analysis on WT and AWA were applied for evaluating the performance of the semiautomatic method.
Conclusion: The proposed LSM can generate reasonably accurate lumen and outer wall boundaries compared to manual segmentation, and can work similar to a human reader. However, it tends to overestimate CAWT and AWA compared to the manual segmentation for arterial wall with small area. Key Words: carotid arterial wall thickness; plaque; level set method; segmentation J. Magn. Reson. Imaging 2014;39:1457–1467. C 2013 Wiley Periodicals, Inc. V
Results: Arterial wall thickness measured by radial distance measure (RDM) and polyline distance measure (PDM) correlated well between GT and variational LSM (r ¼ 0.83 for RDM and r ¼ 0.64 for PDM, P < 0.05). The absolute arterial wall area bias between LSM and three
1
Department of Radiology, Azienda Ospedaliero Universitaria (A.O.U.), di Cagliari – Polo di Monserrato, Cagliari, Italy. 2 Centre for Excellence in Signal and Image Processing, Department of Electronic and Electrical, University of Strathclyde, Strathclyde, UK. 3 Department of Radiology, New York University School of Medicine, New York, New York, USA. 4 Visiting Scientist, Global Biomedical Technologies, Roseville, California, USA. 5 University of Washington, Seattle, Washington, USA. 6 27558 Kingsgate Way, Farmington Hills, Michigan, USA. 7 Biolab, Department of Electronics, Politecnico di Torino, Torino, Italy. 8 Department of Cardiology, Azienda Ospedaliero Universitaria (A.O.U.), di Cagliari – Polo di Monserrato, Cagliari, Italy. 9 Department of ECE, Ngee Ann Polytechnic, Singapore. 10 Neurologische Klinik und Poliklinik Technische Universit€ at M€ unchen, M€ unchen, Germany. 11 Diagnostic and Monitoring Division, AtheroPoint LLC, Roseville, California, and Department of Biomedical Engineering, Idaho State University (Aff.), Pocatello, Idaho, USA. *Address reprint requests to: L.S. Department of Radiology, Azienda Ospedaliero Universitaria (A.O.U.), di Cagliari – Polo di Monserrato s.s. 554 Monserrato (Cagliari) 09045, Italy. E-mail: [email protected]. Received November 17, 2012; Accepted June 18, 2013. DOI 10.1002/jmri.24307 View this article online at wileyonlinelibrary.com. C 2013 Wiley Periodicals, Inc. V
ATHEROSCLEROSIS is the third leading cause of death in the world (1). Early detection of atherosclerosis and proper treatment and lifestyle changes would prevent the onset of cardiovascular diseases like stroke and cardiac failure. With advancements in medical imaging technology, methods are being developed for studying plaque risk stratification, characterization of plaque burdens, and real-time degree of stenosis estimation, lipid size, fibrous cap thickness, and arterial wall thickness. Magnetic resonance imaging (MRI) has been widely used for imaging diseased vascular wall and it assists in accurate characterization of plaque (2). One of the key factors of its success is the ability to identify the adventitia boundary in transverse images of the vessel wall. Research has demonstrated that MRI is able to detect the boundaries of arterial walls with good reproducibility, and the mean arterial wall thickness derived from those detected boundaries has been used for measuring plaque burden (3). Studies (4) have demonstrated that carotid arterial wall thickness (CAWT) correlates well with atherosclerosis burden and is reliable in predicting clinical events and
1457
1458
Saba et al.
treatment outcome. However, the usefulness of CAWT depends on the accuracy and precision of arterial wall boundary segmentation followed by wall thickness calculation using methods such as calculating the average of several radial thickness measurements (4). Methods from manual to semiautomatic and automatic segmentation such as level set methods (LSM) (5), active contour models (ACM) (6), and statistical methods such as active shape model (7) have been widely used for diseased arterial wall detection on MR images (8–14). In this study we present a semiautomatic method based on level set method for segmenting carotid plaque lumen and outer wall boundaries. Carotid wall thickness from segmented boundaries were calculated and compared with manual segmentation carried out by trained clinicians to study the accuracy of quantification, and we also present the performance comparison with ACMs.
X
X
k>0 and y is a constant, d is a univariate Dirac function, H is the Heaviside function, g is the edge indicator function defined as: g5
1 11jrGr Ij2
[6]
Gr is the Gaussian kernel with given s, I is the given image matrix. The first term in Eq. (5) computes the length of zero level curve of u and the second term is the speed up function in the area of V w ¼ fðx; yÞj wðx; yÞ < 0g. Finally, Eq. (2) can be expanded as: du ru ru 5l ru2div 1kdðuÞdiv g 1tgdðuÞ [7] dt jruj jruj The above equation can be easily implemented by simple finite difference scheme.
MATERIALS AND METHODS Level Set Method (LSM) LSM has been widely used for segmentation in the image-processing field (12). The basic idea is to represent the contours as the zero level set of an implicit function, which is defined in a higher dimension, usually denoted as level set function. Then the level set function evolves according to certain partial differential equations (PDEs). In this study, variational level set method without reinitialization (15) was used for arterial wall segmentation. The evolution equation of the traditional level set formulation of the moving front C can be written in the following general form: @u 1F jruj50 [1] @t F is the speed function, which depends on the imaging data and level set function u. C can be represented by zero level set as C(t) ¼ {(x,y)ju(t,x,y) ¼ 0}. In this study we limit our problem domain in the 2D image space. For traditional LSM, u is required to be kept close to a signed distance function during the evolution, therefore reinitialization is required constantly, even though the reinitialization can be very complicated and can have great side effects. In order to overcome these difficulties, the evolution equation is redefined as: @u @E 1 50 [2] @t @u EðuÞ5lP ðuÞ1eðuÞ
motion of the zero level curve of u. In the segmentation procedure, it can be defined as: ð ð eðuÞ5k gdðuÞjrujdxdy1t gH ð2uÞdxdy [5]
[3]
where P(u) is the measurement of the distance of how close a function u is to a signed distance function. P(u) is defined as: ð 1 [4] PðuÞ5 ðjruj21Þ2 dxdy 2 X
P(u) will help eliminate the reinitialization of u during level set evolution. l > 0, which controls the effect of penalizing the deviation of u from a signed distance function. e(u) is the energy term which will drive the
GVF Method Based on ACM, gradient vector flow (GVF) was developed by Xu and Prince (16) by introducing a new static external force field to overcome the disadvantages in traditional ACM, such as sensitivity to initial position and difficulties in penetrating into concavities. The proposed new external force field V(x,y) 5 [u(x,y) v(x,y)], named the GVF field, is solved with the following Euler equations: lr2 u2ðu2fx Þ fx2 1fy2 50 [8] [9] lr2 v2ðv2fx Þ fx2 1fy2 50 where l is the regularization parameter (l>0), r2 is the Laplacian operator, f is the edge map defined as: f ðx; yÞ5jrðGd ðx; yÞ I ðx; yÞÞ2 j
[10]
MRI Protocol and Data Collection MRI Ten patients were selected from consecutive patients from the stroke unit (17) in the January 2012. The mean age of patients (seven males and three females) in this study was 67 6 6 years. The internal carotid artery stenosis was diagnosed by Doppler and duplex sonography, and without showing heavy calcification. In our institution the inclusion criteria for performing MRA of carotid arteries are the following: presence of an ultrasound analysis that showed a carotid stenosis >50% and evidence of plaque alteration (an irregular surface, intraplaque hemorrhage, ulceration). Multispectral MRI scans were performed on the subjects to obtain plaque morphology. The protocol was approved by the local Ethics Committee, and written informed consent was obtained from each patient before the study.
Semiautomated Analysis of CAWT in MRI
1459
software implemented in MatLab (MathWorks, Natick, MA). The procedure of LSM segmentation was done on a slice-by-slice basis as follows: 1) The boundary inside lumen was initialized by a radial expansion from a specific point; 2) The initial boundary will evolve according to Eq. (7), and stop at the luminal boundary. Usually the iteration is chosen to be 100 to ensure the initial boundary will evolve towards luminal boundary as close as possible; 3) The segmented lumen boundary grows outward by two pixels for the initialization of outer wall boundary segmentation; 4) From step 3, the newly initialized boundary will evolve similar as in step 2 according to Eq. (7). Due to poorer contrast between outer wall and surrounding tissues, usually 30 iteration steps were chosen in order to avoid leaking problem in the proposed variational LSM. Figure 2 shows the diagram for the proposed variational LSM method. Data Evaluation Figure 1. Manual segmentation for GT (obtained with ImgTracer) (a): MRI slice of carotid plaque; (b) lumen boundary tracing; (c) outer wall boundary tracing superimposed with lumen boundary; (d) ROI of the carotid wall.
MRI was performed using a 1.5T scanner (Siemens Medical Systems, Erlangen, Germany) with bilateral phased-array surface coils (PACC-SS15, Machnet, Netherlands). The MRI scan was centered on the carotid bifurcation to cover a large range for proper stenosis imaging. T1-weighted, T2-weighted and proton-densityweighted (PDW) studies were performed for each patient. In this study only PDW images were used for plaque segmentation. The parameters for PDW sequence were: flip angle 180 , field of view (FOV) 16 16 cm, TR 700 msec, TE 10 msec, number of excitations 2, slice thickness 2 mm, matrix size 640 640. The longitudinal coverage for each carotid artery was 72 mm. No further process was applied to assess MR image quality. Figure 1a shows a typical MR image of carotid plaque. The MRI scan for one patient was 18 minutes.
The GT and LSM CAWT were computed for each slice. Two kinds of methods were used to obtain CAWT in this study: radial distance measurement (RDM), and polyline distance measurement (PDM). The procedure for RDM is as follows: 1) the center of the lumen was defined after the segmentation according to the lumen boundary; 2) ray lines were constructed from the lumen center. RDM was defined along the same ray lines across the lumen and outer wall boundaries, as shown in Fig. 3a. PDM was defined as the closest distance from the estimated lumen boundary to the outer wall boundary, and PDM is not sensitive to the number of points constituting the profiles. More details can be found in a previously published work (18). Figure 3b presents the RDM and PDM obtained for the plaque MR image shown in Fig. 3a. Since RDM may not always measure the distance along the perpendicular direction between contours, RDM tends to be slightly larger than PDM, as evident in Fig. 3b.
Ground Truth (GT) Data Collection ImgTracer was used for GT data collection. It is a userfriendly software tool designed by Global Biomedical Technologies (Naples, FL) mainly for radiologists for tracing different kinds of anatomy in 2D and 3D medical images. Lumen boundary can be traced by clicking 10– 15 points in a clockwise or anticlockwise direction as shown in Fig. 1b. Similarly, the outer wall can be traced by clicking 10–15 points in an orderly fashion, shown in Fig. 1c. These boundaries are saved as GT data or manual boundaries. Furthermore, the ImgTracer allows computing the region of interest (ROI) of the carotid wall to study the components of the wall, such as plaque area (Fig. 1d). Three experienced clinicians segmented the carotid plaque MR images by using ImgTracer and these were considered as manual tracings. Image Processing and Segmentation Method The variational LSM method was used for segmentation, and it was carried out using the in-house
Figure 2. Diagram for the proposed LSM. [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]
1460
Saba et al.
Figure 3. Definition of RDM (a) and comparison of RDM and PDM for the same plaque (b).
The agreement analysis between LSM delineated and manual delineated carotid arterial wall was done in terms of spatial overlap by using Jaccard Similarity Factor (Js), defined as: Js 5
pixelsðPLSB \ PGT Þ pixelsðPLSB [ PGT Þ
[11]
where PLSB stands for the set of pixels identified by the semiautomatic delineation, and PGT indicates the set of pixels identified in manual segmentation as GT. Therefore, Js is a ratio of number of pixels in the intersection of these sets over the union of these sets. If the two sets of pixels are perfectly in the same spatial position, then Js will reach the maximum value of 1. Furthermore, the curve distance, defined by the mean perpendicular distance for each pair of lumen/ outer wall boundaries from LSM and manual methods, was used to measure the accuracy of LSM delineated arterial wall boundaries.
RESULTS Results Using LSM Figure 4 shows the typical segmentation results obtained using LSM. Initial boundary was defined as in Fig. 4a, which is not close enough to the real luminal boundary. Figure 4b shows the detected luminal boundary using LSM. Figure 4c shows the outer wall boundary combined with luminal boundary and it is evident that the luminal wall has been well segmented. Figure 4d is the overlay of manual and LSM segmentation results. The detected boundaries are very close to the manual segmentation boundaries. However, LSM in Fig. 4 overestimates the arterial wall area by 10%. In total, 102 slices of MR images of carotid plaques were segmented by the proposed LSM and also manually by three trained clinicians. The CAWT values obtained using RDM and PDM are summarized in Table 1 and expressed as mean 6 standard deviation. Unlike RDM, PDM metric overestimates the CAWT thickness. As expected and observed, the average PDM is smaller than RDM in both manual segmentations and our LSM method.
Bland–Altman and Regression Analysis for the CAWT Bland–Altman analysis and regression analysis were performed between the three observers of the GT and the LSM method, summarized in Figs. 5 and 6. In order to test the interobserver variability, Bland–Altman plots between the three observers were analyzed. The limit of agreement between the GT was from 29.8% to þ25.9% in the case of best concordance Table 1 Thickness Comparison Between RDM and PDM
Figure 4. Segmentation results from LSM. (a) Initial contour for lumen boundary; (b) lumen boundary from LSM; (c) lumen/outer wall boundary from LSM; (d) Overlay with GT (line with star stands for GT).
LSM Obser1 Obser2 Obser3
RDM (mm)
PDM (mm)
P
2.34 2.24 2.06 2.28
2.2 6 0.55 1.93 6 0.56 1.64 6 0.44 1.97 6 0.57
