Int J CARS (2006) 1:137–148 DOI 10.1007/s11548-006-0046-4
O R I G I NA L A RT I C L E
Muscles of mastication model-based MR image segmentation H. P. Ng · S. H. Ong · Q. Hu · K. W. C. Foong · P. S. Goh · W. L. Nowinski
Published online: 11 October 2006 © CARS 2006
Abstract Objective The muscles of mastication play a major role in the orodigestive system as the principal motive force for the mandible. An algorithm for segmenting these muscles from magnetic resonance (MR) images was developed and tested. Materials and methods Anatomical information about the muscles of mastication in MR images is used to obtain the spatial relationships relating the muscle region of interest (ROI) and head ROI. A model-based technique that involves the spatial relationships between head and muscle ROIs as well as muscle templates is developed. In the segmentation stage, the muscle ROI is derived from the model. Within the muscle ROI, anisotropic diffusion is applied to smooth the texture, H. P. Ng (B) · K. W. C. Foong NUS Graduate School for Integrative Sciences and Engineering, Singapore, Singapore e-mail:
[email protected] H. P. Ng · Q. Hu · W. L. Nowinski Biomedical Imaging Lab, Agency for Science Technology and Research, Singapore, Singapore S. H. Ong Department of Electrical and Computer Engineering, National University of Singapore, Singapore, Singapore S. H. Ong Division of Bioengineering, National University of Singapore, Singapore, Singapore K. W. C. Foong Department of Preventive Dentistry, National University of Singapore, Singapore, Singapore P. S. Goh Department of Diagnostic Radiology, National University of Singapore, Singapore, Singapore
followed by thresholding to exclude bone and fat. The muscle template and morphological operators are employed to obtain an initial estimate of the muscle boundary, which then serves as the input contour to the gradient vector flow snake that iterates to the final segmentation. Results The method was applied to segmentation of the masseter, lateral pterygoid and medial pterygoid in 75 images. The overlap indices (κ) achieved are 91.4, 92.1 and 91.2%, respectively. Conclusion A model-based method for segmenting the muscles of mastication from MR images was developed and tested. The results show good agreement between manual and automatic segmentations. Keywords Masseter · Pterygoids · Muscles · MRI · Model-based segmentation
Introduction The human face plays a key role in interpersonal relationships and has a direct influence on the quality of life. Hence, facial surgeries are carried out daily to help patients who suffer from craniofacial problems such as hemifacial microsomia and misshapen jaws. Simulation systems for facial surgery have been developed to assist clinicians in pre-surgery planning. A variety of facial models have been proposed [1–3]. A physics-based model has been developed for static soft tissue prediction and muscle simulation [4]. This model assumes that different tissue groups have similar properties and uses the linear elastic modeling approach to simplify the highly complicated biomechanical behavior of different tissue types. In the facial model for
138
pre-surgery simulation [5], a laser range scan provides a photorealistic 3D model of the patient’s face, whereas a CT scan gives a 3D representation of the skull. The CT and laser data sets are registered via selected cephalometric landmarks. The basic components of the soft-tissue model are mass points and, connecting them, springs with elasto-mechanical properties. The main drawback with existing models is the assumption that different muscles have similar properties whereas they are different in reality. Another shortcoming is that they do not take into consideration the actual location, shape and size of the muscles of mastication. The availability of such information would be extremely useful for planning maxillofacial surgeries. A good understanding of the mastication system is essential for maxillofacial surgeons in the surgical procedures concerning the jaws. The muscles of mastication play a major role in mastication. The large masseter muscle is the strongest jaw muscle and acts to raise the jaw and clench the teeth. The pterygoid muscles, used in various combinations, can elevate, depress, or protract the mandible, or slide it from side to side. Although the muscles may be seen in MR scans of the patient, clinicians would like to visualize them in 3D and measure important quantities such as their volume and surface area. Compared to manual contour tracing, computerized techniques would result in greater consistency and savings in clinicians’ time. Image segmentation is a key component of medical image analysis. A variety of image processing techniques have been developed for the segmentation of MR images. The traditional active contour model [6], which matches a deformable model to an image, has been constantly improved and used extensively in MR image segmentation [7, 8]. It is an energy minimizing spline whose energy depends on its shape and location within the image. Thresholding is another method that has been widely used in the segmentation of medical images. It is a simple but often effective means for obtaining segmentation of images in which different structures have contrasting intensities. Examples of connectivity-based thresholding, which finds a boundary between two regions using the path connection algorithm and changing the threshold adaptively, can be found in [9, 10]. A major limitation of thresholding is that it does not take into account the spatial characteristics of an image and thus is sensitive to the noise, artifacts and intensity inhomogeneities that can occur in MR images. A recent approach, supervised range-constrained thresholding, which confines the analysis in the histogram of a region of interest (ROI), is able to provide a good solution when the image quality is poor [11].
Int J CARS (2006) 1:137–148
Besides thresholding, there is also an increasing use of model-based techniques for segmentation in MR images, such as the work described in [12], which incorporates prior knowledge for segmenting the corpus callosum from MR images with little human intervention. Another example is found in [13], where segmentation is carried out via matching of probability distributions belonging to photometric variables that incorporate learned shape and appearance models for the objects of interest. The absence of a computerized procedure for the segmentation of the muscles of mastication may be due to the fact that, firstly, the muscles and their surrounding tissues have similar gray levels and, secondly, distinct boundaries between them are lacking. We solve this problem by a two-stage approach. In the training stage, we use reference MR images in which the targeted muscles have been manually segmented to determine the spatial relationships between the head and the muscles and to construct muscle templates. These muscle models are employed in the segmentation stage to specify the muscle ROIs in the study image and also to discard unwanted background pixels. An initial segmentation is obtained by using morphological operators and this serves as the input contour to the gradient vector flow (GVF) [14] snake which iterates to obtain the final segmentation result. The use of the GVF snake is preferred more than the conventional active contour model as it is able to converge to concave boundaries of the muscles. The proposed method is used to segment the masseter, lateral pterygoid and medial pterygoid, shown in Fig. 1, from 2D MR images. The following section describes the materials and proposed method. The subsequent sections present the results and discussion, respectively. The last section concludes the paper.
Materials and methods Data acquisition The image data were obtained using a 1.5T MRI scanner. The imaging protocols that were explored include fluid attenuated inversion recovery (FLAIR), fast spin echo (FSE), gradient echo (GRE), spoiled gradient recall (SPGR) and fast low angle shot (FLASH). Figure 1 shows an image acquired using the FLASH sequence and Fig. 2 the images acquired using the FSE, GRE and SPGR sequences. The T1 FLASH images are clearly superior in displaying the anatomy of the facial muscles. Hence, in this work, we acquire the image data using T1 FLASH with the following specifications: 1 mm slice thickness, 512 × 512 matrix, 240 mm FOV, TR = 9.93 ms
Int J CARS (2006) 1:137–148
139
Fig. 1 Muscles of mastication in MR T1 FLASH image
Fig. 2 MR images acquired using a T1 weighted FSE, b T2* weighted GRE, c T1 weighted SPGR
and TE = 4.86 ms. We acquired a total of ten data sets from ten different subjects, of which five are for training (training data) and the other five for validation (study data).
Overview of method The proposed method is a two-stage process (Fig. 3). In the first stage, we determine the spatial relationship between the targeted muscle ROI and the head ROI in images from the training data sets. The targeted muscle and head ROIs are the bounding boxes of the muscle and head regions, respectively, in a 2D image (Fig. 5a). The spatial relationship serves as the prior knowledge for training the system to identify the targeted muscle ROI in the study image. In the second stage, we apply the segmentation algorithm to the images from the study data sets to segment the targeted muscle.
Acquisition of prior knowledge The muscles of mastication in the five training data sets are first manually segmented by medical experts. The first 2D axial slice in each data set is in the region of the orbit whereas the last slice is in the neck region. We assign a start index Ii to the first axial slice in which the targeted muscle appears and an end index If to the last slice in the data set. For each of the training data sets, we compute the relative location of the slice Il with the largest surface area of the targeted muscle: Ir =
I l − Ii If − Ii
The average values of Ir for the masseter, lateral pterygoid and medial pterygoid are 0.487, 0.383 and 0.423, respectively. The actual Ir in the training data sets ranges from 0.477 to 0.494, 0.368 to 0.395 and 0.405 to 0.440 for
140
Int J CARS (2006) 1:137–148
Acquisition of prior knowledge from reference data sets
Segmentation process on test data sets Input Ii and If for muscle in each training data set
Input Ii and If for muscle in each training data set
Locate reference image
Locate reference image
Automatic detection of head ROI
Manual contour tracing of the muscle for each 2-D reference image
Automatic detection of muscle ROI by the system using prior knowledge
Automatic detection of head ROI and muscle ROI for each image
Apply anisotropic diffusion to muscle ROI
Derive a relative relationship between head ROI and muscle ROI for each image
Obtain an average of this relationship for all these reference images
Thresholding to remove fat and bone after FCM clustering Check for overlap with template of targeted muscle for initial segmentation Perform GVF snake to refine initial segmentation
Fig. 3 Overview of methodology
the masseter, lateral pterygoid and medial pterygoid, respectively. Having selected the reference slice from each data set for each of the muscles, we now automatically detect the head ROI (Fig. 4a) from the projections of the image in the x (horizontal) and y (vertical) directions (Fig. 4b). The muscle ROI is defined to be the bounding box of the manually segmented muscle in the reference image. The muscles of mastication are paired muscles with symmetrical origins and attachments on each side of the midline. The masseter has a lateral origin and attachment, and is centered laterally to the ramus of the mandible. The medial and lateral pterygoid muscles have medial origins on the pterygoid processes and lateral attachments to the ramus and neck of the mandible, respectively. We make use of this anatomical information, together with our observations from the five training data sets to specify a spatial relationship between the
ROI of each muscle of mastication and the head ROI in the reference slice. We take one reference slice of the masseter from each of the five training data sets. The spatial relationship between the head and masseter ROIs (Fig. 5a) is specified in terms of the distance between the boundaries of the head ROI and the origin of the masseter ROI. For a reference slice of the masseter, the distances l1 , w1 , j1 , k1 , m1 , n1 (Fig. 6a) are measured, and the relative distances calculated as follows: m1,r =
m1 , l1
n1,r =
n1 , w1
k1,r =
k1 , l1
j1,r =
j1 w1
To obtain a good estimate of the spatial relationship, we use the mean values of m1,r , n1,r , k1,r , j1,r obtained from the five reference slices. The model of the muscle requires a template of the muscle region in addition to the spatial relationship. This is obtained first by manually tracing the masseter boundary on each of the five reference slices. The boundary (or contour) is represented as a linked list of points whose coordinates are normalized to the length and breadth of the head ROI to allow for different head sizes. The five masseter contours are aligned using their centroids, and the mean contour obtained by averaging the radial distances for each unit polar angle from 0 to 360◦ . The mean contour is the template of the masseter. The spatial relationships between the head and the lateral and medial pterygoid ROIs (Fig. 5b, c) are specified in terms of the distance from the origin of the head ROI to the origin of the muscle ROIs. Following the procedure used for the masseter, we measure the various spatial distances l2 , w2 , j2 , k2 , m2 , n2 for the lateral pterygoid (Fig. 6b) and calculate the relative distances m2,r , n2,r , k2,r , j2,r of each of the five reference slices from the five training data sets. This procedure is also used for the medial pterygoid where we first measure l3 , w3 , j3 , k3 , m3 , n3 followed by calculation of m3,r , n3,r , k3,r , j3,r . The process of constructing the templates of the lateral and medial pterygoids follows that for the masseter. Segmentation Given an image from the study data set, the system first automatically determines the head ROI based on the vertical and horizontal projections (as described in the previous section). The system then makes use of the spatial relationship between the targeted muscle ROI and the head ROI to identify the targeted muscle ROI in the study image. A combination of image processing techniques, which includes anisotropic diffusion, template
Int J CARS (2006) 1:137–148
141
Fig. 4 a Head ROI, b projections in vertical and horizontal directions
(a)
(b) 4
x 104
x 104
3.5
3.5
3
3
2.5
2.5 2
Start 2
End
point 5
1.5
point
1
1
0.5
0.5
0
0 0
100
200
300
400
500
600
0
vertical
100
200
300
400
500
600
horizontal
correlation and GVF snake, is used to arrive at the final segmentation.
of an appropriate value of s is further discussed in the Results section.
Determination of muscle ROI
Image smoothing
Given that the width and length of the head ROI in the study image are ai and bi , respectively, where i = 1, 2, 3 denotes the masseter, lateral pterygoid and medial pterygoid, respectively; we derive equations for the spatial parameters ci , di , ei , fi (Fig. 7)
The muscles consist of bundles of multi-nucleated cells known as muscle fibers, which give rise to a highly textured appearance in the MR image. An image of the masseter with its muscle fibers is shown in Fig. 8. Inaccurate results will be produced when we apply active contour and GVF techniques to such textured regions, thus necessitating prior image smoothing with anisotropic diffusion [15] (number of iterations = 20, timestep = 0.2, one smoothing per iteration).
ci = ai ji,r s,
di = bi ki,r s,
ei = ai ni,r s,
fi = bi mi,r s
Parameter s is a scaling factor introduced to allow for the slight variations in location, shape and size of the muscles between different subjects. In our work, the minimum and maximum values of s are found to be 1.3 and 2.5, respectively. When s is 2.5, our method is applicable to images that have been rotated 15◦ clockwise or anti-clockwise from the upright position. The selection
Coarse segmentation The muscle ROI comprises soft tissue, fat and bone. Compared to soft tissue, bone has relatively low and fat relatively high intensity values. Using this knowledge, we are able to approximately separate the soft
142
Int J CARS (2006) 1:137–148
Length l1
(a)
distance n1 width j1
distance m1
width w1 length k1
Masseter ROI
Head ROI
(b)
length li length ki width ji Pterygoid ROI
width w i
Head ROI
distance ni
distance mi
i = 2 for lateral pterygoid i = 3 for medial pterygoid Fig. 6 a Spatial relationship between head ROI and masseter ROI in reference image, b spatial relationship between head ROI and lateral pterygoid ROI in reference image
Initial segmentation of muscles
Fig. 5 a Spatial relationship between boundary of head ROI and origin of masseter ROI, b spatial relationship between origin of head ROI and origin of lateral pterygoid ROI, c spatial relationship between origin of head ROI and origin of medial pterygoid ROI
tissue from the bone and the fat. We make use of the fuzzy c-means (FCM) algorithm [16] with three clusters to divide the muscle ROI into three different clusters. Thresholding is then carried out, with the lower and upper thresholds being the minimum and maximum intensity values of the second cluster, to form the binary image of the ROI. We discuss the effect of using four clusters in the Results section.
The templates of the masseter, lateral and medial pterygoids are shown in Fig. 9. The targeted muscle template (T1 × T2 pixels) is moved from left to right, top to bottom across the targeted muscle ROI (R1 × R2 pixels). We measure the quantity of overlap between the ROI and the template at each position and this is done (R1 − T1 )(R2 − T2 ) times. If the regions in the targeted muscle ROI and the template overlap by 75% or more, we note the positions of the overlapping pixels. We discuss the effect of changing the threshold value to 65 and 85% in the Results section. After all the checking has been done, a binary image is formed with the pixels noted earlier taking a value of ‘1’. Morphological opening is then applied to the binary image to open the connections between the muscle region and its surrounding soft tissue. Connected component labeling [17] is employed to find the largest
Int J CARS (2006) 1:137–148
(a)
143
Length b1
[18] and used to initialize the GVF snake [14] for refining the initial segmentation of the muscle. distance d1
width e1
distance c1
width a1 length f1
Masseter ROI
Head ROI
(b)
length bi length fi width ei Pterygoid ROI
width ai
Head ROI
distance di
distance ci
i = 2 for lateral pterygoid i = 3 for medial pterygoid Fig. 7 a Spatial relationship between head ROI and masseter ROI in study image, b spatial relationship between head ROI and pterygoid ROI in study image
Evaluation In this study, the manual segmentation of the muscles of mastication by an expert radiologist serves as the ground truth in the evaluation of our proposed method. The κ index [19] is used to evaluate the agreement between the computerized segmentation and manual segmentation: N(M ∩ S) × 100% κ =2× N(M) + N(S) where M and S denote the regions obtained by manual segmentations and the proposed method, respectively, M ∩ S the intersection between M and S, and N(·) the number of pixels in a region. The value of κ ranges from 0 (no overlap) to 1 (exact overlap). In addition, we make use of false positive rate (FPR) to measure the probability of the method incorrectly giving a positive result: N(Fp ) × 100% FPR = N(M) where N Fp denotes the number of pixels which the method incorrectly determines as positive. The FNR gives the probability of a negative result N(Fn ) × 100% FNR = N(M) where N(Fn ) denotes the number of pixels which the method incorrectly determines as negative. Results Our proposed method was applied to five MR study data sets. The user is required to input Ii and If for each of the targeted muscles in each data set. The reference slice will then be automatically derived. For each muscle, we performed 2D segmentation on the reference slice of the muscle, as well as on the two slices superior and two slices inferior to the reference slice. Hence, a total of 75 segmentation results (25 for each muscle) were obtained. Accuracy
Fig. 8 Masseter with its muscle fibers
connected component. Morphological closing is used to fill up the holes in the resulting image. The initial segmentation of the muscle is thus obtained. The edge map of the initial segmentation is derived using Canny filter
We perform 2D segmentation of the masseter on a total of 25 MR images from the five study data sets. In Fig. 10, we show a set of masseter results obtained after each stage. These results are obtained using s = 1.3. Numerical validations are performed on the 25 segmentation
144
Int J CARS (2006) 1:137–148
Fig. 9 Template of a masseter, b lateral pterygoid, c medial pterygoid
results by comparison with the ground truths, with the results tabulated in Table 1. The average κ, FPR and FNR are 91.4, 7.8 and 9.4%, respectively. The ROIs of the lateral pterygoid and medial pterygoid were identified by our system. They are obtained using s = 1.3. The average κ, FPR and FNR are 92.1, 5.9, 9.9, respectively, for the lateral pterygoid (Table 2), and 91.2, 8.6, 9.1, respectively, for the medial pterygoid (Tables 3).
Sensitivity to scaling factor s and rotation of MR image In Segmentation section we introduce a scaling factor s to enlarge the muscle ROI in the study image. A larger ROI may be needed to ensure that the muscle is fully enclosed in it. Experimental results indicate that the ROI will fail to fully enclose the muscle for s below 1.2. We have also tested our method for its robustness to situations where the image is rotated by 15◦ to the left and right, or 15◦ up and down from the upright position. To handle such cases, s needs to be as large as 2.5. It was observed through experimental results that for s = 2.5, the ROI will contain more spurious components compared to s = 1.3. However, by checking for the overlap between the ROI and the template, together with morphological operators, the method is able to provide good initializations to the GVF snake; the average κ of the segmentation results is greater than 90%. For values of s greater than 2.5, it was observed that the ROI contains more soft tissue that has relatively similar intensity values to the muscle, and hence the results may not be as good. The validation shows that our proposed method is tolerant to a 15◦ rotation of the head. When dealing with data sets where the head was rotated by more than 15◦ , a solution will be to locate the midsagittal plane (MSP)
in the image [20] and rotate the image till the MSP is in an upright position. Justification of parameters Throughout our work presented here, we have made use of FCM with three clusters. We also obtained segmentation results through FCM with four clusters. When we change the number of clusters from three to four, the average interval between the lower and upper thresholds is reduced from 100 to 70. The size of the remaining region after thresholding is smaller, and the probability of forming an overlap with the template at 75% or more decreases. Hence, the initialization of the GVF snake falls within the boundaries, and the GVF snake may be trapped in local minima as it propagates outwards. Compared to using FCM with three clusters, κ decreases and FNR increases. It is also possible that no region in the ROI overlaps with the template at 75%. The objective of checking for the overlap between template and ROI is to reduce the quantity of unwanted soft tissue in the ROI. It is mentioned in Segmentation section that we set the threshold for the overlap between template and muscle ROI to be 75%. Instead of 75%, we have also experimented with 65 and 85%. The experimental results indicate that, when the overlap is 65%, we are still able to obtain good segmentations of the masseter. However, it is also possible that there is minimal removal of unwanted soft tissue. The initialization of the GVF snake falls in the unwanted soft tissue and the snake does not converge to the muscle boundary. Compared to setting the overlap to 75%, κ decreases and FPR increases. In contrast, when we set the amount of overlap to be 85%, there may be excessive removal of soft tissue. Hence the initializations of the GVF snake tend to lie within the actual boundaries of the muscle.
Int J CARS (2006) 1:137–148
145
Fig. 10 Results at each stage for segmentation of masseter
Table 1 Validation results on segmentation of masseter Image index 1 2 3 4 5 6 7 8 9 10 11 12 13
κ (%) 93.0 93.5 92.8 91.3 92.8 90.6 89.5 91.5 90.2 88.6 91.2 92.4 93.0
FPR (%) 2.6 3.8 4.5 6.3 7.7 11.6 10.4 5.8 9.8 9.2 10.6 10.9 7.5
FNR (%) 11.4 9.2 9.9 11.1 6.7 7.2 10.6 11.2 9.8 13.6 7 4.3 6.5
In such situations, as the snake grows outward, it may be trapped in local minima and not converge to the actual boundaries. When the overlap threshold is set
Image index 14 15 16 17 18 19 20 21 22 23 24 25 Mean
κ (%) 91.8 90.2 87.8 90.8 92.0 91.5 90.3 91.4 91.8 93.3 92.1 91.6 91.4
FPR (%) 10.3 9.5 9.7 7.3 9.2 7.4 8.2 6.3 7.1 5.2 5.8 7.9 7.8
FNR (%) 6.1 10.1 14.7 11.1 6.8 9.6 11.2 10.9 9.3 8.2 10 8.9 9.4
within (65, 85%), the segmentation results are reasonable (κ > 90%), so for all the experiments we set the parameter as 75%.
146
Int J CARS (2006) 1:137–148
Table 2 Validation results on segmentation of lateral pterygoid Image index 1 2 3 4 5 6 7 8 9 10 11 12 13
κ (%) 93.6 93.8 95.2 92.4 91.8 90.8 91.2 92.0 92.2 91.5 89.8 91.0 92.8
FPR (%) 5.7 4.3 3.2 5.8 6.0 5.5 7.8 5.7 6.2 9.4 5.6 4.3 3.8
FNR (%) 7.1 8.1 6.4 9.4 10.4 12.9 9.8 10.3 9.4 7.6 14.8 13.7 10.6
Image index 14 15 16 17 18 19 20 21 22 23 24 25 Mean
κ (%) 91.3 91.0 88.9 91.7 93.0 92.4 90.5 91.9 92.8 95.3 93.1 92.6 92.1
FPR (%) 6.1 5.5 7.7 7.2 4.5 7.5 9 6.3 5.9 3.2 5.1 6.5 5.9
FNR (%) 11.3 12.5 14.5 9.4 9.5 7.7 10 9.9 8.5 6.2 8.7 8.3 9.9
Image index 14 15 16 17 18 19 20 21 22 23 24 25 Mean
κ (%) 90.5 88.5 90.8 91.5 92.1 90.9 89.5 90.4 90.8 92.3 91.8 87.1 91.2
FPR (%) 9.7 8.5 7.2 8.1 9.2 10.1 9.3 7.3 8.1 9.2 9.7 8.8 8.6
FNR (%) 9.3 14.5 11.2 8.9 6.6 8.1 11.7 11.9 10.3 6.2 6.7 17 9.1
Table 3 Validation results on segmentation of medial pterygoid Image index 1 2 3 4 5 6 7 8 9 10 11 12 13
κ (%) 91.2 92.5 92.8 92.1 91.8 89.6 92.0 93.3 91.1 89.6 91.9 91.5 93.7
FPR (%) 8.7 7.1 5.5 7.3 9.2 10.6 9.2 8.5 9.2 9.6 8.8 7.5 7.8
FNR (%) 8.9 7.9 8.9 8.5 7.2 10.2 6.8 4.9 8.6 11.2 7.4 9.5 4.8
Discussion Segmenting the muscles of mastication from MR images is difficult because a muscle and its surrounding tissue have similar gray levels, often with no distinct boundaries between them. Despite this, our proposed method, which involves prior knowledge of the location and a template of the muscle, is able to segment with average κ greater than 90%. We have applied the method to the segmentation of the masseter, lateral and medial pterygoids. By training the system using the spatial relationship between the head ROI and muscle ROI, the system is able to automatically determine the muscle ROI in a study image. In Fig. 10, the output at each stage of the method for segmenting the masseter is displayed. The soft tissue is approximately separated from the fat and bone through FCM with three clusters. In the ROI after thresholding, soft tissue is present together with muscles. By checking for the overlap between the template and the regions in the ROI, we are able to remove the majority of the unwanted soft tissue. After checking
for overlap with the template, small unwanted regions remaining in the muscle ROI are then removed through the use of connected components labeling. This gives us an initial segmentation of the muscle. Even though the model is 2D in nature, it is tolerant to a rotation of 15◦ and can be applied to two neighboring slices. This is an encouraging sign of good model although we may need to validate the model against more data sets as well as data from different imaging centers. The initialization process for the GVF snake is automatic. The initial segmentations serve as good initializations for the GVF snake; since they are close to the actual boundaries of the muscles, relatively fewer iterations are required before the GVF snake converges. Furthermore, accurate segmentations are, to a certain extent, dependent on the initial state. Our method to initialize the GVF snake proves to be good as we are able to obtain κ better than 90%. The presence of clear and distinct boundaries affects the accuracy of our segmentation results. A better accuracy will be achieved when there are clear boundaries
Int J CARS (2006) 1:137–148
between the masseter and its surrounding soft tissue (κ = 93.5%). On the contrary, the lack of a distinctive boundary may result in a less accurate segmentation. The human factor could play a part in the quantum of error as we have only one medical expert to aid us in manually segmenting the ground truth in this work. We also have limited number of data sets currently. In the long run, we plan to acquire a larger number of data sets, as well as obtain the assistance of more medical experts in our work.
Conclusion We developed a computerized method for segmenting the muscles of mastication from MR images, which to our best knowledge, is currently unavailable. This method uses prior knowledge on the location of the muscle with respect to the head, obtained from the ground truths by medical experts. These manual contour tracings also provide us with the template used in obtaining the initial segmentations which are used to initialize the GVF snake to refine the segmentations of the facial muscle. This initialization proves to be good as we achieved an average κ of greater than 90%. This initialization also results in fewer iterations for the GVF snake to converge as the initial segmentations are near the actual boundaries. There is an increasing emphasis on 3D medical image segmentation and many established techniques, such as the level set approach [21], graph cuts [22, 23] and the watersnakes [24], are applied to 3D image segmentation. Currently, our proposed method is targeted for 2D segmentation only and an extension from 2D to 3D is on the way. Practical reasons, such as computational complexity and how the targeted structures have been defined, are taken into consideration in the design of our technique. The work presented in [25, 26] are examples where 2D methods are applied sequentially to the slices of a 3D image. For the segmentation of muscles of mastication, repeatedly applying our 2D method to all the 2D slices in the data set may not achieve good segmentation results. This is because in a 3D MR data set, there are slices where no clear boundary exists between the muscle and the surrounding tissue. As such, we will need to make use of the neighboring slices which provide additional information. Other problems pertaining to extension from 2D to 3D include intensity inhomogeneity between slices and the variation in the size of the muscles in different slices which may result in higher error rate.
147 Acknowledgments This project is funded by NUS R-222-000011-112 from the Faculty of Dentistry, National University of Singapore. The first author is grateful to Agency for Science, Technology and Research (A*Star), Singapore for funding his PhD studies. The authors thank Mr Christopher Au, Principal Radiographer at National University Hospital, Singapore for his assistance in data acquisition.
References 1. Noguchi N, Goto M (2003) Computer simulation system for orthognathic surgery. J Orthod Craniofac Res 6(1):176–178 2. Xia J, Samman N, Yeung RW, Wang D, Shen SG, Ip HH, Tideman H (2000) Computer-assisted three-dimensional surgical planning and simulation. 3D soft tissue planning and prediction. Int J Oral Maxillofac Surg 29(4):250–258 3. Wong TY, Fang JJ, Chung CH, Huang JS, Lee JW (2005) Comparison of 2 methods of making surgical models for correction of facial asymmetry. Int J Oral Maxillofac Surg 63(2):200–208 4. Gladilin E, Zachow S, Deuflhard P, Hege HC (2004) Anatomy and physics-based facial animation for craniofacial surgery simulations. Med Biol Eng Comput 42(2):167–170 5. Meehan M, Teschner M, Girod S (2003) Three-dimensional simulation and prediction of craniofacial surgery. J Orthod Craniofac Res 6(1):103–107 6. Kass M, Witkin A, Terzopoulos D (1987) Snakes: active contour models. Int J Comput Vis 1(4):321–331 7. Ray N, Acton ST, Altes T, Lange EE, Brookeman JR (2003) Merging parametric active contours within homogeneous image regions for MRI-based lung segmentation. IEEE Trans Med Imaging 22(2):189–199 8. Pluempitiwiriyawej C, Moura JMF, Wu YJ, Ho C (2005) STACS: new active contour scheme for cardiac MR image segmentation. IEEE Trans Med Imaging 24(5):593–603 9. Lee C, Huh S, Ketter TA, Unser M, (1998) Unsupervised connectivity-based thresholding segmentation of midsagittal brain MR images. Comput Biol Med 28(3):309–338 10. Kim DY, Park JW (2005) Connectivity-based local adaptive thresholding for carotid artery segmentation using MRA images. Image Vis Comput 23(14):1277–1287 11. Hu Q, Hou Z, Nowinski WL (2006) Supervised rangeconstrained thresholding. IEEE Trans Image Process 15(1): 228–240 12. Lundervold A, Duta N, Taxt T, Jain AK (1999) Model-guided segmentation of corpus callosum in MR images. IEEE Conf Comput Vis Pattern Recognit 1:231–237 13. Freedman D, Radke RJ, Zhang T, Jeong Y, Lovelock DM, Chen GT (2005) Model-based segmentation of medical imagery by matching distributions. IEEE Trans Med Imaging 24(3):281–292 14. Xu C, Prince JL (1998) Snakes, shapes, and gradient vector flow. IEEE Trans Image Process 7(3):359–369 15. Perona P, Malik J (1990) Scale-space and edge detection using anisotropic diffusion. IEEE Trans Pattern Anal Mach Intell 12(7):629–639 16. Bezdek JC (1981) Pattern recognition with fuzzy objective function algorithm. Plenum, New York 17. Hu Q, Qian G, Nowinski WL (2005) Fast connectedcomponent labelling in three-dimensional binary images based on iterative recursion. Comput Vis Image Underst 99(3):414–434 18. Canny J, (1986) A computational approach to edge detection. IEEE Trans Pattern Anal Mach Intell 8(4):679–698
148 19. Leemput VK, Maes F, Vandermeulen D, Suetens P (1999) Automated model-based tissue classification of MR images of the brain. IEEE Trans Med Imaging 18(10):897–908 20. Hu Q, Nowinski WL (2003) A rapid algorithm for robust and automatic extraction of the midsagittal plane of the human cerebrum from neuroimages based on local symmetry and outlier removal. NeuroImage 20(4):2153–2165 21. Malladi R, Sethian JA, Vermuri BC (1995) Shape modelling with front propagation: a level set approach. IEEE Trans Pattern Anal Mach Intell 17(2):158–174 22. Kolmogorov V, Zabin R (2004) What energy functions can be minimized via graph cuts?. IEEE Trans Pattern Analy Mach Intell 26(2):147–159
Int J CARS (2006) 1:137–148 23. Lombaert H, Sun Y, Grady L, Xu C (2005) A multilevel banded graph cuts method for fast image segmentation. IEEE Int Conf Comput Vis 1:259–265 24. Nguyen HT, Worring M, Boomgaard RVD (2003) Watersnakes: energy-driven watershed segmentation. IEEE Trans Pattern Anal Mach Intell 25(3):330–342 25. Pham DL, Prince JL, Dagher AP Xu C (1997) An automated technique for statistical characterization of brain tissues in magnetic resonance image. Int J Pattern Recognit Artif Intell 11(8):1189–1211 26. Ge Y, Fitzpatrick JM, Dawant BM, Bao J, Kessler RM, Margolin RA (1996) Accurate localization of cortical convolutions in MR brain images. IEEE Trans Med Imaging 15(4):418–428
