SlideShare a Scribd company logo
Mestiri Makram & Hamrouni Kamel
International Journal of Image Processing (IJIP), Volume (8) : Issue (2) : 2014 17
Reeb Graph for Automatic 3D Cephalometry
Mestiri Makram mmestiri@gmail.com
Enit/Electrical/Image Processing
Enit
Elnasr 2, 2037, Tunisia
Hamrouni Kamel Hamrouni.kamel@lsts.com
Enit/Electrical/Image Processing
Enit
Elnasr 2, 2037, Tunisia
Abstract
The purpose of this study is to present a method of three-dimensional computed tomographic
(3D-CT) cephalometrics and its use to study cranio/maxilla-facial malformations. We propose a
system for automatic localization of cephalometric landmarks using reeb graphs. Volumetric
images of a patient were reconstructed into 3D mesh. The proposed method is carried out in
three steps: we begin by applying 3d mesh skull simplification, this mesh was reconstructed from
a head volumetric medical image, and then we extract a reeb graph. Reeb graph mesh extraction
represents a skeleton composed in a number of nodes and arcs. We are interested in the node
position; we noted that some reeb nodes could be considered as cephalometric landmarks under
specific conditions. The third step is to identify these nodes automatically by using elastic mesh
registration using “thin plate” transformation and clustering. Preliminary results show a landmarks
recognition rate of more than 90%, very close to the manually provided landmarks positions made
by a medical stuff.
Keywords: Reeb Graph, Cephalometry Landmarks, Thin Plate, LCP.
1. INTRODUCTION
Radiographic cephalometry has been one of the most important diagnostic tools in maxillofacial
diseases, since its introduction in the early 1930. It deals with the morphological scientific study of
the dimensions of all the structures present in a human head, usually through the use of
standardized lateral head radiographs. Generations of doctors have relied on the interpretation of
these images for their diagnosis and treatment planning as well as for the long-term follow-up of
growth and treatment results. Also in the planning for surgical orthodontic corrections of jaw
discrepancies, lateral/antero-posterior cephalograms have been valuable tools.
Parameters measurement is based on a set of agreed upon feature point’s landmarks. The
detection of the landmarks plays an essential role in diagnosis and treatment planning by doctors.
Head-growing analyzing with anatomical landmarks was first proposed by Arc Thomson, in 1917
[1]. His approach was based on a deforming grid. Changes in the landmarks position resulted in
deformations of the grid. His method was applied not only to quantify the effects of growth, but
also to relate different individuals and even different species. A malocclusion is a misalignment of
teeth or incorrect relation between the teeth of the two dental arches, Broadbent [2] and later
Brodie [3] applied a method based in landmarks to quantify malocclusions and study their effects.
Mestiri Makram & Hamrouni Kamel
International Journal of Image Processing (IJIP), Volume (8) : Issue (2) : 2014 18
They used radio-graph image to define the Landmarks, head structure like bony or soft tissue,
had to be identified. This approach was used for measurement specification at an individual of
known age, sex and race to quantify head anatomy differences for diagnosis, such as the one
shown in Figure 1.
FIGURE 1: Manual Extraction of Cephalometry Landmarks in X-ray Image.
Downs [4] proposed in 1948 the first cephalometric analysis method. His approach was based on
measurements of 10 angles from a lateral radiographs from a group of selected individuals; he
calculates the average values and gives them a clinical significance. His approach was the basis
for most methods used at present.
Ricketts [5] and Steiner [6] proposed traced lines between significant landmarks, so their length
and angles can be measured and compared with standard values shown in Figure 2.
FIGURE 2: Line Traced Between Significant Landmarks [6].
Locating landmarks depends on medical expertise to locate the landmarks manually. This can
take an experienced orthodontist up to 30 min. The process is time consuming, subject to human
error and tedious.
An automated system would help to eliminate the above-mentioned problems and hence produce
repeatable results.
2. RELATED WORKS
All the automatic cephalometric landmarks recognition methods use 2D image processing, the
various methods used by researchers in this field found in the literature can be divided into three
different approaches: knowledge-based feature detection, pattern matching and neural networks
with fuzzy inference system.
Mestiri Makram & Hamrouni Kamel
International Journal of Image Processing (IJIP), Volume (8) : Issue (2) : 2014 19
Levy [7] proposed in 1986 the first automatic extraction of cephalometric landmarks. His method
was based on knowledge based line extraction technique. He begins with applying median filter
and histogram equalization to enhance the X-rays images, then he applies a Mero-Vassy
operator [8] to extract relevant edges using knowledge-based line tracker. The landmarks are
then located according to their geometric definition online crossing. This method requires good X-
rays image (cannot be guaranteed in practice). Parthasaraty [9] proposed an enhancement of
levy approach by introducing a multi resolution pyramidal analysis of X-rays images to reduce the
processing time.
Yen [10], Contereras [11], Jackson [12], Cohen [13] and Davis and Forsyth [14] presented similar
knowledge-based edge tracking methods. These methods depend highly on X-ray image quality
and can be used only for landmarks located on an edge.
The second approache used mathematical or statistical models to narrow down the search area
for each landmark, and then shape-matching techniques are used to locate the exact location of
each landmark. Cardillo and Sid-Ahmed [15] proposed a method based on a combination of
mathematical modeling methods, like affine transformation to remove the shifts, rotations and
size differences to reduce the size of the search area. Then, they applied a shape recognition
algorithm based on gray-scale mathematical morphology to locate the landmarks. The method
locates 76% of the landmarks within 2 mm. Grauetal[16], adds a line detection module to select
the most significant lines in Sid-Ahmed[15] approach. Then, he applies mathematical morphology
transformations for shape recognition. The disadvantage of theses methods is that it’s very
sensitive to noise present in X-rays images. Desvignes [17] proposed a statistical method based
on estimation of landmarks locations using adaptive coordinate space where locations are
registered. The method was tested on a set of 58 X-rays, of which 28 were used for training. They
obtain 99.5% of the landmarks but with 4:5 mm mean error between the real position and the
estimated position, far from the authorized range (±2 mm). Hutton [18] used active shape models
for cephalometric landmarking. They established a template of possible deformations from a
training set of hand-annotated image. The result was 35% of 16 landmarks within an acceptable
range of ±2 mm. Implementation could be used as a first estimation location of the landmarks.
The third category of researchers used neural networks and fuzzy inference systems to locate the
landmarks. Uchino [19] proposed a fuzzy learning machine that could learn the relation between
the gray levels of the image and the location of the landmarks. They begin by dividing the image
into nine blocks, each one is an input to the fuzzy learning machine. The weights were adjusted
by learning the coordinates of the landmark. The block containing the landmark was divided in
nine separate blocks until landmark location is obtained. This method produced an average error
of 2:3 mm. this method depends highly on scale, rotation and shift of X-rays images. Innes [20]
highlights regions containing craniofacial features using Pulse Coupled Neural Networks (PCNN)
from X-rays images. They get 36.7% rate for the region containing sella landmark, 88.1% for the
region containing the chin landamark, and 93.6% for the region containing the nose landmark.
The most disadvantage of PCNN’s method is that they require a considerable manual contribution
to set the required parameters.
In this paper, we are interested in the extension of 2D X-rays to 3D image cephalometry analysis.
Some work has been carried out to study the potential advantages of 3D imaging methods, such
as computed tomography for cephalometric analysis. Kragskov [21] made a comparative result,
using human dry skulls. Ferrario [22] studied 3D facial morphometry using infrared cameras, but
only as a supplement for classic cephalometry.
We propose a novel method based on the use of reeb graph for Automatic localization of
craniofacial landmarks. The task is a difficult one due to the complexity and variability of
cephalometry landmarks.
Mestiri Makram & Hamrouni Kamel
International Journal of Image Processing (IJIP), Volume (8) : Issue (2) : 2014 20
3. PROPOSED METHOD
As seen in previous works, many researchers have attempted automatic cephalometric
landmarks detections. However, a major drawback of the existing techniques is that they use a
2D representation of a 3D structure. In this paper we are interested in a 3D cephalometric
landmarks. The development of spiral CT and cone beam CT has revolutionized the medical
image techniques, the former providing outstanding resolution and the latter, with its low cost,
allowing unique accessibility, as a result 3-D imaging has become an essential tool in planning
and managing the treatment of facial deformity.
3.1 3D Mesh Database
The image data-base in composed of Dicom images stored in a series of 2D grey level images
from a Spiral CT scan (with 512 x 512 matrixes, 110 kV, and 80 mAs) performed at 1 mm slice
thickness. The extraction of hard and soft head tissue (bones and skin) is done by using a
threshold and region growing technique. Each bones and skin parts are separately constructed
using marching cube algorithm as shown in Figure 3.
FIGURE 3: 3D Mesh Extraction of Hard and Soft Tissues.
The results of this extraction method are 3 meshes: interior/exterior skull bones, and mandibular
bone Figure-4.
FIGURE 4: 3D Mesh Extraction from Volumetric CT-Scan .
3.2 3D Cephalometric Landmarks
After getting the mesh database, 3D cephalometric landmarks must be localized in hard tissue. In
the literature we found 20 hard landmarks, some of them are presented in table-1.
CT-Scan Image
Threshold and
region growing
Interior
Cranium bone
Exterior
Cranium bone
Mandibule Head Skin
3D mesh reconstruction
(Marching Cube)
Volumic grey level images
3D mesh images
Mestiri Makram & Hamrouni Kamel
International Journal of Image Processing (IJIP), Volume (8) : Issue (2) : 2014 21
No Abbreviation Hard Tissue
1 N Nasion is the midpoint of the frontonasal suture
2 S Sella is the centre of the hypophyseal fossa (sella turcica).
3 Po Porion is the most superior point of each external acoustic meatus.
4 Or Orbitale is the most inferior point of each infra- orbital rim.
5 ANS Anterior Nasal Spine is the most anterior midpoint of the anterior nasal spine of the maxilla.
6 PNS Posterior Nasal Spine is the most posterior midpoint of the posterior nasal spine of the palatine bone.
7 PMP Posterior Maxillary Point is the point of maximum concavity of the posterior border of the palatine bone in the horizontal plane at both sides.
8 UI Upper Incisor is the most mesial point of the tip of the crown of each upper central incisor.
9 LI Lower Incisor is the most mesial point of the tip of the crown of each lower central incisor.
10 UMcusp Upper Molar cusp is the most inferior point of the mesial cusp of the crown of each first upper molar in the profile plane.
TABLE 1: 10 of the 20 Cephalometric Landmarks Proposed by Swennen [23].
Figure 5 shows some sample of CT-scan images with located landmarks [23].
FIGURE 5: A: Right hard cephalometric landmarks, B: Facial hard cephalometric landmarks, C: Soft
cephalometric landmarks.
3.3 Automatic Localization of Craniofacial Landmarks
Figure 6 presented the bloc diagram of the proposed method for automatic localization of 3D
cephalometric landmarks.
After transforming the volumetric medical images to a 3D surface mesh, we need to make a
reference shape model containing all cephalometric landmarks information’s. This is made with
what is called “cephalometric head atlas”.
FIGURE 6: Automatic localization of 3D cephalometric landmarks using reeb graphs.
Cephalometric
Head Atlas
Patient Head Mesh
Mesh Simplification
(Edge Collapse)
Reeb Graph
Extraction
Elastic Registration
(Thin Plate Spline)
Automatic Landmarks
localisation
(NNB algorithm)
Rigid registration
(ICP)
A B C
Mestiri Makram & Hamrouni Kamel
International Journal of Image Processing (IJIP), Volume (8) : Issue (2) : 2014 22
To automatically identify 3D landmarks, we need to have prior knowledge information’s. We
create a cephalometric head atlas based on medical stuff help. This atlas is composed of skull
and skin meshes with 3D cephalometric landmarks as shown in Figure 7.
FIGURE 7: Cephalometric atlas of skull. A: initial clean skull mesh, B: Cephalometric landmarks and C:
Manual localization of cephalometric landmarks.
In our method we choose to automatically localize cephalometric landmarks using reeb graph.
Below we will expose the reeb graph extraction.
The Reeb graph is a chart of connectivity on a surface between its critical points. The main
advantage of the reeb graph is that it makes it possible to represent 3D anatomy in a simple
topology way. It is composed of nodes and arcs; each whole of level associates a node. The
graph of Reeb is obtained starting from the computation of a µ function introduced by the theory
of Morse [24] defined on the closed surface of an object in its critical points [25].
We define µ: S ! R defined on surface S of a 3D object. The Reeb graph is a space quotient of
the graph of µ in S, defined by the relation of following equivalence between X є S and Y є S:
· µ(X)=µ(Y) (1)
X~Y " · X et Y are in the same Related component of
µ-1(µ(X)), and the image µ(Y) is S.
The aspect of the graph is entirely related to the choice of the µ function, which determine the
properties of stability and invariance of the resulting graphs. Several functions of application were
proposed in the state of the art for the construction of Reeb graphs [26]. One example is the
function (µ(ν(x,y,z)) = z / ν є S), adapted well for the models whose points are mainly distributed
in the Z-axis direction.
FIGURE 8: Reeb Graph with a z Projection µ Function.
A B C
Mestiri Makram & Hamrouni Kamel
International Journal of Image Processing (IJIP), Volume (8) : Issue (2) : 2014 23
The function suggested in [27] as «outdistances geodesic extrema of curve » is based on growth
of areas starting from local Gaussian curves on tops (germs). The results depend on the position
of the germs and require no disturbed data; it’s a precise calculation of the local curves. Hilaga
[27] defines a function by computing the distance from a point of the surface to the center of mass
G of the object:
(µ(ν) = d(G, ν) / ν є S et d : Euclidian distance) (2)
The function defined by Thierny[28] at the characteristic points and the geodesic distances is
defined by the following equation:
µμ v = 1 − δ v, v! (3)
With δ v, v! = min!"!!!δ(v, v!!
)
Where δ is the function of the geodesic distances and v! are characteristic points. The function
suggested in [29] is the integral of the geodetic distances g(v, p) of v to the other point’s p of the
surface (4).
µμ v = g v, p dS!!!
. (4)
In our approach we choose this µμ function and tested it with different mesh simplification. Figure-9
shows an example of reeb graph extraction.
FIGURE 9: A: Mandibule Patient Mesh, B: Reeb Graph Extraction.
Surface mesh simplification is the process of reducing the faces number used in the mesh
surface, while keeping preserved as much as possible the overall shape, volume and boundaries.
Cignoni [29] proposed a comparison of mesh simplification algorithms, and divided these
methods into 7 approaches: coplanar facets merging, controlled vertex/edge/face decimation, re-
tiling, energy function optimization, vertex clustering, wavelet-based approaches, and
simplification via intermediate hierarchical representation.
We test many simplification methods with reeb graph transformation, and we choose to use the
edge collapse method. These techniques reduce a model’s complexity by repeated use of the
simple edge collapse operation. Researchers have proposed various methods of determining the
“minimal cost edge” to collapse at each step (figure-10).
FIGURE 10: Edge Collapse Algorithm [30].
A C
Mestiri Makram & Hamrouni Kamel
International Journal of Image Processing (IJIP), Volume (8) : Issue (2) : 2014 24
We try to define the effect of the edge collapse simplification process on the reeb
graph extraction. The result in shown in figure-11
FIGURE 11: Patient Skull Simplification. A: 332.932 vertices and 662.564 faces, B: 20.047 vertices and
38.958 faces and C: 3.887 vertices and 7320 faces.
We notice that, after mesh simplification, the number of nodes in the reeb graph decreased and
they became closed to cephalometric landmarks. And, of course, the computation time is reduced
Figure-12.
FIGURE 12: Decrease of reeb graph nodes and reconciliation with cephalometric lanmarks.
Mesh registration is a technique used to find a transformation for mapping a source mesh known
as a reference mesh and a target mesh. This technique associated to a head cephalometric atlas
should help us to get closer to cephalometric landmarks. Every elastic registration process should
begin with a rigid one. In this study, we targeted the rigid registration between two different
meshes using the ICP algorithm (Iterative Closest Point).
The principle of ICP is to iterate between a step of mapping data and another step of optimization
of rigid transformation until convergence. The transformation used for registration is composed of
a 3D rotation and translation. At each iteration, the Algorithm provides a list of matched points
and an estimate of the transformation. The algorithm converges when the error in distance
between matched points is less than a given threshold. The ICP algorithm is presented in the
following [30] where Np is the number of points CP1 and CP2 and k is the index of current
iteration. Figure-13 shows a example of rigid registration between mesh patient and
cephalometric mesh atlas.
FIGURE 13: Rigid registration between patient and the cephalometric atlas
A B C
Mestiri Makram & Hamrouni Kamel
International Journal of Image Processing (IJIP), Volume (8) : Issue (2) : 2014 25
Mesh elastic registration is a mesh deformation process; one of the transformations that are able
to represent elastic deformations is the thin-plate spline (TPS). Thin plate splines were introduced
by Bookstein[31] for geometric design. In 2D images, the TPS model describes the transformed
coordinates (xT,yT) both independently as a function(7) of the original coordinates (x, y):
(xT,yT)=(fx(x,y),fy(x,y)) (7)
The algorithm begins with a given displacements of a number of landmark points, the TPS model
interpolates those points, while maintaining maximal smoothness. For each landmark point (x,y),
the displacement is represented by an additional z-coordinate, and, for each point, the thin plate
is fixed at position (x, y, z). The strain energy is calculated by integrating the second derivative
over the entire surface that can be minimized by solving a set of linear equations.
dxdy
y
z
yx
z
x
z
R
∫∫ ⎟
⎟
⎠
⎞
⎜
⎜
⎝
⎛
⎟⎟
⎠
⎞
⎜⎜
⎝
⎛
∂
∂
+⎟⎟
⎠
⎞
⎜⎜
⎝
⎛
∂∂
∂
+⎟⎟
⎠
⎞
⎜⎜
⎝
⎛
∂
∂
2
2
2
2222
2
2
.2 (8)
The TPS model for one of the transformed coordinates is given by parameter vectors a and D (9):
F(xT,yT)= a1+a2x+a3y+…+
( )( )∑=
−
n
i ii yxLFD1
,
(9)
Where F(r)= r2log(r) is the basis function, a = [a1 a2 a3 a4]T defines the affine part of the
transformation, D gives an additional non-linear deformation, and the Li are the landmarks that
the TPS interpolates figure-4.
FIGURE 14: A template configuration (left) and a target configuration (right) of five landmarks each.
The deformation right grid illustrates the thin-plate spline function between these configurations
as applied to the left regular grid.
The method interpolates some of the points using smoother transformation controlled by a
parameter µ, which weights the optimization of landmark distance and smoothness. For µ = 0,
there is full interpolation, while for very large µ, there is only an affine transformation. In our
methods the landmarks used in the TPS algorithm are the reeb graph nodes as shown in figure-
15.
FIGURE 15: Elastic Registration using Thin Plate.
Mestiri Makram & Hamrouni Kamel
International Journal of Image Processing (IJIP), Volume (8) : Issue (2) : 2014 26
As result of the elastic registration, the patient reeb graph nodes become closer to the atlas
cephalometric head landmarks. Since, some reeb graph nodes are not cephalometric landmarks,
we have to select the real ones.
Automatic localization of cephalometry landmarks
This step starts with the creation of possible landmark localization in cephalometric atlas. Then, it
identifies the reeb graph nodes that exist in the mesh patient. For each node, it searches the
closest landmarks in the cephalometric atlas through the calculation of Euclidean distance
between an acceptable range of ±2 mm. Finally, as a result all patient reeb graph nodes are
associated to a nearest cephalometric landmarks[32].
FIGURE 16: Localization of Cephalometric Landmarks in Patient Skull.
4. TEST AND RESULTS
To validate our approach we asked a doctor to conduct manually a cephalometric analysis on the
mesh patient. We then calculate the error between its points and cephalometric landmarks
automatically localized by our method. The result is showed in the following table:
No Abbreviation Error in mm Recognized
1 N 0.5 Yes
2 S 2.6 No
3 Po 1.3 Yes
4 Or 0.4 Yes
5 ANS 0.7 Yes
6 PNS 2.8 No
7 PMP 2 Yes
8 UI 1.3 Yes
9 LI 1.7 Yes
10 UMcusp 1.8 Yes
11 LMcusp 0.9 Yes
12 Men 0.8 Yes
13 Go 1.5 Yes
14 Fz 1.7 Yes
15 Zy 0.3 Yes
16 A 0.4 Yes
17 B 0.5 Yes
18 Pog 0.8 Yes
19 Ba 0.6 Yes
20 Co 1.4 Yes
TABLE 2: Result of cephalometric landmarks detection
Mestiri Makram & Hamrouni Kamel
International Journal of Image Processing (IJIP), Volume (8) : Issue (2) : 2014 27
Our method localized 18 landmarks on 20. This mean a percentage of 90% landmarks in patient
mesh was recognized.
FIGURE 17: Localization of A- Sella and B- PNS Landmarks.
The Sella landmark (S) is very difficult to be localized. It is not localized in the surface mesh, and
the doctors need to insert it taking into account the interior skull surface. The Posterior Nasal
Spine Landmark (PNS) is defined on the exo-cranial skull base view of the 3-D hard tissue
surface representation. This landmark depends highly on the accuracy of the mesh
reconstruction.
5. CONCLUSION
In this paper, we have proposed a novel method for automatic 3D localization of cephalometric
landmarks. Our approach is based on reeb graph nodes. In particular, we have presented results
for localized 90% of cephalometric landmarks in CT-Scan medical images. We initiated a new
approach for automatic three-dimensional cephalometric analysis. The proposed method needs
to be validated on a larger database.
6. REFERENCES
[1] A.Thomson « On growth and form ». Cambridge, UK: Cambridge Univ. Press, 1992.
[2] Broadbent HB. « A new X-ray technique and its application to orthodontia ». Angle Orthod
1931;
[3] Brodie AG. « On the growth pattern of the human head from the third month to the eighth
year of life ». Am J Anat 1941; 68:209–62.
[4] Downs WB. « Variations in facial relationships: their significance in treatment and
prognosis ».Am J Orthod 1948; 34:812–40.
[5] Ricketts RM. « Cephalometric Analysis and Synthesis ». Angle Or- thod, 1961; 31:141–56.
[6] Steiner C. « Cephalometrics in clinical practice ». Angle Orthod 1959; 29:8–29.
[7] A. Levy-Mandel, A. Venetsanopoulos, J. Tsotsos, « Knowledge-based landmarking of
cephalograms », Comput. Biomed. Res. 19 (3) (1986) 282–309.
[8] L. Mero, Z. Vassy, « A simpliffed and fast version of the Hueckel operator for finding optimal
edges in pictures », Proceedings of the Fourth International Joint Conference on Artificial
Intelligence, Tbilissi, Georgia, USSR, September 1975, pp. 650–655.
[9] S. Parthasaraty, S. Nugent, P.G. Gregson, D.F. Fay, « Automatic landmarking of
cephalograms », Comput. Biomed. Res. 22 (1989) 248–269.
A C
Mestiri Makram & Hamrouni Kamel
International Journal of Image Processing (IJIP), Volume (8) : Issue (2) : 2014 28
[10] C.K. Yan, A. Venetsanopoulos, E. Filleray, « An expert system for landmarking of
cephalograms », Proceedings of the Sixth International Workshop on Expert Systems and
Applications, 1986, pp. 337–356.
[11] J.L. Contereras-Vidal, J. Garza-Garza, « Knowledge-based system for image processing and
interpolation of cephalograms », Proceedings of the Canadian Conference on Electrical and
Computer Engineering, 1990, pp. 75.11–75.14.
[12] P.H. Jackson, G.C. Dickson, D.J. Birnie, Digital « image processing for cephalometric
radiographic: a preliminary report », Br. J. Orthod. 12 (1985) 122
[13] A.M. Cohen, A.D. Linney, « A low cost system for computer-based cephalometric analysis »,
Br. J. Orthod. 13 (1986) 105–108.
[14] Forsyth D. D.Davis, « Knowledge-based cephalometric analysis: a comparison with
clinicians using interactive computer methods», Comput. Biomed. Res. 27 (1994) 10–28.
[15] M. Sid-Ahmed, J. Cardillo, « An image processing system for the automatic extraction of
craniofacial landmarks », IEEE Trans. Med. Imaging 13 (2) (1994) 275–289.
[16] V. Grautel, M.C. Juan, C. Monserrat, C. Knoll, « Automatic localization of cephalometric
landmarks », J. Biomed. Inf. 34 (2001) 146–156.
[17] M. Desvignes, B. Romaniuk, R. Demoment, M. Revenu, M.J. Deshayes, « Computer
assisted landmarking of cephalometric radiographs », Proceedings of the Fourth IEEE
Southwest Symposium on Image Analysis and Interpretation, 2000, pp. 296–300.
[18] T.J. Hutton, S. Cunningham, P. Hammond, « An evaluation of active shape models for the
automatic identification of cephalometric landmarks », Eur. J. Orthodont. 22 (5) (2000) 499–
508.
[19] E. Uchino, T. Yamakawa, « High speed fuzzy learning machine with guarantee of global
minimum and its application to chaotic system identification and medical image processing »,
Proceedings of the Seventh International Conference on Tools with Artificial Intelligence,
1995, pp. 242–249.
[20] A. Innes, V. Ciesielski, J. Mamutil, S. John, « Landmark detection for cephalometric radiology
images using pulse coupled neural networks », International Conference in Computing in
Communications, June 2002, pp. 391–396.
[21] Kragskov J, Bosch C, Gyldensted « Comparison of the reliability of craniofacial anatomic
landmarks based on cephalo-metric radiographs and 3-dimensional CT scans », Cleft Palate
Crani- ofac J 1997; 34:111–6.
[22] Ferrario VF, Sforza C, Poggio CE, « Facial 3-dimensional morphometry ». Am J Orthod
Dentofac Orthop 1996; 109:86–93.
[23] Gwen R. J. Swennen,Filip Schutyser,Jarg-Erich Hausamen « Three-dimensional
cephalometry: a color atlas and manual » Amazon Book.
[24] Milnor (j.w), « Morse theory », Princeton University Press, Princeton, NJ, (1963)
[25] Reeb, G. « Sur les points singuliers d'une forme complètement intégrable ou d'une fonction
numérique ». Livre stringer (1946).
Mestiri Makram & Hamrouni Kamel
International Journal of Image Processing (IJIP), Volume (8) : Issue (2) : 2014 29
[26] Tony Tong Thése « Indexation 3D de bases de données d’objets par graphes de Reeb
améliorés » (2005)
[27] Hilaga, Shigawa, « Topology Matching for fully automatic Similarity Estimation of 3D
Shape », ACM SIGGRAPH, Los Angeles, CA, USA (2001).
[28] J. Tierny, J. Vandeborne et M. Daouadi, « Graphe de Reeb de haut niveau de maillages
polygonaux ». Journées de l’association Francophone d’informatique graphique, bordeaux
(2006).
[29] P. Cignoni, C. Montani, R. Scopigno : « A comparison of mesh simplification algorithms »,
Int. Symp. on 3D Data Processing, Visualization, and Transmission.
[30] Besl P.J., McKay N.D., “A Method for Registration of 3D Shapes", IEEE Trans. On Pattern
Analysis and Machine Intelligence (TPAMI), Vol.14(2), pp. 239-255,1992.
[31] F. L. Bookstein, “Principal Warps: Thin-Plate Splines and the Decomposition of
Deformations” IEEE Transaction on pattern analysis and machine intelligence. Vol 11, No 6,
June 1989.
[32] Makram Mestiri, Sami Bourouis and Kamel Hamrouni, “Automatic Mesh Segmentation using
atlas projection and thin plate spline: Application for a Segmentation of Skull Ossicles”,
VISAPP 2011.

More Related Content

PDF
Proposition of local automatic algorithm for landmark detection in 3D cephalo...
PDF
Brain Tumor Extraction from T1- Weighted MRI using Co-clustering and Level Se...
PDF
Iris Localization - a Biometric Approach Referring Daugman's Algorithm
PDF
ROLE OF HYBRID LEVEL SET IN FETAL CONTOUR EXTRACTION
PDF
Az03303230327
PDF
AN AUTOMATIC SCREENING METHOD TO DETECT OPTIC DISC IN THE RETINA
PDF
Detection of Cranial- Facial Malformations: Towards an Automatic Method
PDF
Literature survey for 3 d reconstruction of brain mri
Proposition of local automatic algorithm for landmark detection in 3D cephalo...
Brain Tumor Extraction from T1- Weighted MRI using Co-clustering and Level Se...
Iris Localization - a Biometric Approach Referring Daugman's Algorithm
ROLE OF HYBRID LEVEL SET IN FETAL CONTOUR EXTRACTION
Az03303230327
AN AUTOMATIC SCREENING METHOD TO DETECT OPTIC DISC IN THE RETINA
Detection of Cranial- Facial Malformations: Towards an Automatic Method
Literature survey for 3 d reconstruction of brain mri

What's hot (19)

PDF
Literature survey for 3 d reconstruction of brain mri images
PDF
Ijetcas14 315
PDF
Intelligent computing techniques on medical image segmentation and analysis a...
PDF
A robust iris recognition method on adverse conditions
PDF
Computer Aided Diagnosis using Margin and Posterior Acoustic Featuresfor Brea...
PDF
International Journal of Computer Science, Engineering and Information Techno...
PDF
Automatic detection of optic disc and blood vessels from retinal images using...
PDF
A new hybrid method for the segmentation of the brain mris
PDF
Detection of uveal melanoma using fuzzy and neural networks classifiers
PDF
Automatic whole-body bone scan image segmentation based on constrained local ...
PDF
Ijarcet vol-2-issue-7-2262-2267
PDF
A SIMPLE IMAGE PROCESSING APPROACH TO ABNORMAL SLICES DETECTION FROM MRI TUMO...
PDF
Mri image registration based segmentation framework for whole heart
PDF
Hybrid Speckle Noise Reduction Method for Abdominal Circumference Segmentatio...
PDF
EV-SIFT - An Extended Scale Invariant Face Recognition for Plastic Surgery Fa...
PDF
Ijcse13 05-01-001
PDF
G43043540
PDF
International Journal of Engineering Research and Development (IJERD)
PDF
JBSC_online
Literature survey for 3 d reconstruction of brain mri images
Ijetcas14 315
Intelligent computing techniques on medical image segmentation and analysis a...
A robust iris recognition method on adverse conditions
Computer Aided Diagnosis using Margin and Posterior Acoustic Featuresfor Brea...
International Journal of Computer Science, Engineering and Information Techno...
Automatic detection of optic disc and blood vessels from retinal images using...
A new hybrid method for the segmentation of the brain mris
Detection of uveal melanoma using fuzzy and neural networks classifiers
Automatic whole-body bone scan image segmentation based on constrained local ...
Ijarcet vol-2-issue-7-2262-2267
A SIMPLE IMAGE PROCESSING APPROACH TO ABNORMAL SLICES DETECTION FROM MRI TUMO...
Mri image registration based segmentation framework for whole heart
Hybrid Speckle Noise Reduction Method for Abdominal Circumference Segmentatio...
EV-SIFT - An Extended Scale Invariant Face Recognition for Plastic Surgery Fa...
Ijcse13 05-01-001
G43043540
International Journal of Engineering Research and Development (IJERD)
JBSC_online
Ad

Viewers also liked (8)

PDF
Discover all Bretagne has to offer at in-Cosmetics 2016
PDF
Foreign Assistance Newsletter - June/July 2016 Edition
DOCX
Sara Abo Shady 2014
PDF
Journal du Salon de l'Entrepreneuriat 2016
PDF
Rapport 2015 de l'Institut Pasteur de Tunis
PPT
transformation 3d
PDF
Notes 2D-Transformation Unit 2 Computer graphics
Discover all Bretagne has to offer at in-Cosmetics 2016
Foreign Assistance Newsletter - June/July 2016 Edition
Sara Abo Shady 2014
Journal du Salon de l'Entrepreneuriat 2016
Rapport 2015 de l'Institut Pasteur de Tunis
transformation 3d
Notes 2D-Transformation Unit 2 Computer graphics
Ad

Similar to Reeb Graph for Automatic 3D Cephalometry (20)

PDF
Powerful processing to three-dimensional facial recognition using triple info...
PPTX
Cephalometrics
PPTX
Cephalometrics analysis -ih
PPT
Cephalometric in orthodontics/prosthodontic courses
PPTX
Cephalometrics analysis in orthodontics.pptx
DOCX
cephalometry in orthodontic.docx
PPT
Cephalometrics in orthodontics/prosthodontic courses
PPTX
HARD TISSUE CEPHALOMETRIC ANALYSIS.pptx
PPTX
Cephalometric Radiograph : Orthodontic applications.pptx
PPT
Limitations of cephalometrics of ceph
DOC
Cephalometrics history, evolution, and land marks
DOC
Cephalometrics history, evolution, and land marks/orthodontic courses by indi...
PPT
Cephalometrics for orthodontics
PPTX
lateral cephalometry in orthodontics
PPT
3D Cephalometrics and morphometrics /certified fixed orthodontic courses by ...
PPT
Cephlometric history, evolution & landmarks
PPT
CEPHALOMETRICS.ppt A Scientific approach to the scrutiny of human craniofacia...
PPTX
Cephalometrics (hard and soft tissue ) - in detail
PPTX
Lateral cephalometric analysis of hard tissues
PPT
3 d cephalometrics and morphometrics /certified fixed orthodontic courses
Powerful processing to three-dimensional facial recognition using triple info...
Cephalometrics
Cephalometrics analysis -ih
Cephalometric in orthodontics/prosthodontic courses
Cephalometrics analysis in orthodontics.pptx
cephalometry in orthodontic.docx
Cephalometrics in orthodontics/prosthodontic courses
HARD TISSUE CEPHALOMETRIC ANALYSIS.pptx
Cephalometric Radiograph : Orthodontic applications.pptx
Limitations of cephalometrics of ceph
Cephalometrics history, evolution, and land marks
Cephalometrics history, evolution, and land marks/orthodontic courses by indi...
Cephalometrics for orthodontics
lateral cephalometry in orthodontics
3D Cephalometrics and morphometrics /certified fixed orthodontic courses by ...
Cephlometric history, evolution & landmarks
CEPHALOMETRICS.ppt A Scientific approach to the scrutiny of human craniofacia...
Cephalometrics (hard and soft tissue ) - in detail
Lateral cephalometric analysis of hard tissues
3 d cephalometrics and morphometrics /certified fixed orthodontic courses

Recently uploaded (20)

PPTX
BOWEL ELIMINATION FACTORS AFFECTING AND TYPES
PDF
Supply Chain Operations Speaking Notes -ICLT Program
PDF
Complications of Minimal Access Surgery at WLH
PPTX
Renaissance Architecture: A Journey from Faith to Humanism
PPTX
Introduction_to_Human_Anatomy_and_Physiology_for_B.Pharm.pptx
PDF
Basic Mud Logging Guide for educational purpose
PPTX
GDM (1) (1).pptx small presentation for students
PDF
Sports Quiz easy sports quiz sports quiz
PDF
FourierSeries-QuestionsWithAnswers(Part-A).pdf
PDF
2.FourierTransform-ShortQuestionswithAnswers.pdf
PDF
Abdominal Access Techniques with Prof. Dr. R K Mishra
PPTX
Institutional Correction lecture only . . .
PPTX
PPH.pptx obstetrics and gynecology in nursing
PDF
RMMM.pdf make it easy to upload and study
PPTX
Lesson notes of climatology university.
PPTX
Microbial diseases, their pathogenesis and prophylaxis
PDF
3rd Neelam Sanjeevareddy Memorial Lecture.pdf
PDF
TR - Agricultural Crops Production NC III.pdf
PDF
Insiders guide to clinical Medicine.pdf
PDF
Microbial disease of the cardiovascular and lymphatic systems
BOWEL ELIMINATION FACTORS AFFECTING AND TYPES
Supply Chain Operations Speaking Notes -ICLT Program
Complications of Minimal Access Surgery at WLH
Renaissance Architecture: A Journey from Faith to Humanism
Introduction_to_Human_Anatomy_and_Physiology_for_B.Pharm.pptx
Basic Mud Logging Guide for educational purpose
GDM (1) (1).pptx small presentation for students
Sports Quiz easy sports quiz sports quiz
FourierSeries-QuestionsWithAnswers(Part-A).pdf
2.FourierTransform-ShortQuestionswithAnswers.pdf
Abdominal Access Techniques with Prof. Dr. R K Mishra
Institutional Correction lecture only . . .
PPH.pptx obstetrics and gynecology in nursing
RMMM.pdf make it easy to upload and study
Lesson notes of climatology university.
Microbial diseases, their pathogenesis and prophylaxis
3rd Neelam Sanjeevareddy Memorial Lecture.pdf
TR - Agricultural Crops Production NC III.pdf
Insiders guide to clinical Medicine.pdf
Microbial disease of the cardiovascular and lymphatic systems

Reeb Graph for Automatic 3D Cephalometry

  • 1. Mestiri Makram & Hamrouni Kamel International Journal of Image Processing (IJIP), Volume (8) : Issue (2) : 2014 17 Reeb Graph for Automatic 3D Cephalometry Mestiri Makram mmestiri@gmail.com Enit/Electrical/Image Processing Enit Elnasr 2, 2037, Tunisia Hamrouni Kamel Hamrouni.kamel@lsts.com Enit/Electrical/Image Processing Enit Elnasr 2, 2037, Tunisia Abstract The purpose of this study is to present a method of three-dimensional computed tomographic (3D-CT) cephalometrics and its use to study cranio/maxilla-facial malformations. We propose a system for automatic localization of cephalometric landmarks using reeb graphs. Volumetric images of a patient were reconstructed into 3D mesh. The proposed method is carried out in three steps: we begin by applying 3d mesh skull simplification, this mesh was reconstructed from a head volumetric medical image, and then we extract a reeb graph. Reeb graph mesh extraction represents a skeleton composed in a number of nodes and arcs. We are interested in the node position; we noted that some reeb nodes could be considered as cephalometric landmarks under specific conditions. The third step is to identify these nodes automatically by using elastic mesh registration using “thin plate” transformation and clustering. Preliminary results show a landmarks recognition rate of more than 90%, very close to the manually provided landmarks positions made by a medical stuff. Keywords: Reeb Graph, Cephalometry Landmarks, Thin Plate, LCP. 1. INTRODUCTION Radiographic cephalometry has been one of the most important diagnostic tools in maxillofacial diseases, since its introduction in the early 1930. It deals with the morphological scientific study of the dimensions of all the structures present in a human head, usually through the use of standardized lateral head radiographs. Generations of doctors have relied on the interpretation of these images for their diagnosis and treatment planning as well as for the long-term follow-up of growth and treatment results. Also in the planning for surgical orthodontic corrections of jaw discrepancies, lateral/antero-posterior cephalograms have been valuable tools. Parameters measurement is based on a set of agreed upon feature point’s landmarks. The detection of the landmarks plays an essential role in diagnosis and treatment planning by doctors. Head-growing analyzing with anatomical landmarks was first proposed by Arc Thomson, in 1917 [1]. His approach was based on a deforming grid. Changes in the landmarks position resulted in deformations of the grid. His method was applied not only to quantify the effects of growth, but also to relate different individuals and even different species. A malocclusion is a misalignment of teeth or incorrect relation between the teeth of the two dental arches, Broadbent [2] and later Brodie [3] applied a method based in landmarks to quantify malocclusions and study their effects.
  • 2. Mestiri Makram & Hamrouni Kamel International Journal of Image Processing (IJIP), Volume (8) : Issue (2) : 2014 18 They used radio-graph image to define the Landmarks, head structure like bony or soft tissue, had to be identified. This approach was used for measurement specification at an individual of known age, sex and race to quantify head anatomy differences for diagnosis, such as the one shown in Figure 1. FIGURE 1: Manual Extraction of Cephalometry Landmarks in X-ray Image. Downs [4] proposed in 1948 the first cephalometric analysis method. His approach was based on measurements of 10 angles from a lateral radiographs from a group of selected individuals; he calculates the average values and gives them a clinical significance. His approach was the basis for most methods used at present. Ricketts [5] and Steiner [6] proposed traced lines between significant landmarks, so their length and angles can be measured and compared with standard values shown in Figure 2. FIGURE 2: Line Traced Between Significant Landmarks [6]. Locating landmarks depends on medical expertise to locate the landmarks manually. This can take an experienced orthodontist up to 30 min. The process is time consuming, subject to human error and tedious. An automated system would help to eliminate the above-mentioned problems and hence produce repeatable results. 2. RELATED WORKS All the automatic cephalometric landmarks recognition methods use 2D image processing, the various methods used by researchers in this field found in the literature can be divided into three different approaches: knowledge-based feature detection, pattern matching and neural networks with fuzzy inference system.
  • 3. Mestiri Makram & Hamrouni Kamel International Journal of Image Processing (IJIP), Volume (8) : Issue (2) : 2014 19 Levy [7] proposed in 1986 the first automatic extraction of cephalometric landmarks. His method was based on knowledge based line extraction technique. He begins with applying median filter and histogram equalization to enhance the X-rays images, then he applies a Mero-Vassy operator [8] to extract relevant edges using knowledge-based line tracker. The landmarks are then located according to their geometric definition online crossing. This method requires good X- rays image (cannot be guaranteed in practice). Parthasaraty [9] proposed an enhancement of levy approach by introducing a multi resolution pyramidal analysis of X-rays images to reduce the processing time. Yen [10], Contereras [11], Jackson [12], Cohen [13] and Davis and Forsyth [14] presented similar knowledge-based edge tracking methods. These methods depend highly on X-ray image quality and can be used only for landmarks located on an edge. The second approache used mathematical or statistical models to narrow down the search area for each landmark, and then shape-matching techniques are used to locate the exact location of each landmark. Cardillo and Sid-Ahmed [15] proposed a method based on a combination of mathematical modeling methods, like affine transformation to remove the shifts, rotations and size differences to reduce the size of the search area. Then, they applied a shape recognition algorithm based on gray-scale mathematical morphology to locate the landmarks. The method locates 76% of the landmarks within 2 mm. Grauetal[16], adds a line detection module to select the most significant lines in Sid-Ahmed[15] approach. Then, he applies mathematical morphology transformations for shape recognition. The disadvantage of theses methods is that it’s very sensitive to noise present in X-rays images. Desvignes [17] proposed a statistical method based on estimation of landmarks locations using adaptive coordinate space where locations are registered. The method was tested on a set of 58 X-rays, of which 28 were used for training. They obtain 99.5% of the landmarks but with 4:5 mm mean error between the real position and the estimated position, far from the authorized range (±2 mm). Hutton [18] used active shape models for cephalometric landmarking. They established a template of possible deformations from a training set of hand-annotated image. The result was 35% of 16 landmarks within an acceptable range of ±2 mm. Implementation could be used as a first estimation location of the landmarks. The third category of researchers used neural networks and fuzzy inference systems to locate the landmarks. Uchino [19] proposed a fuzzy learning machine that could learn the relation between the gray levels of the image and the location of the landmarks. They begin by dividing the image into nine blocks, each one is an input to the fuzzy learning machine. The weights were adjusted by learning the coordinates of the landmark. The block containing the landmark was divided in nine separate blocks until landmark location is obtained. This method produced an average error of 2:3 mm. this method depends highly on scale, rotation and shift of X-rays images. Innes [20] highlights regions containing craniofacial features using Pulse Coupled Neural Networks (PCNN) from X-rays images. They get 36.7% rate for the region containing sella landmark, 88.1% for the region containing the chin landamark, and 93.6% for the region containing the nose landmark. The most disadvantage of PCNN’s method is that they require a considerable manual contribution to set the required parameters. In this paper, we are interested in the extension of 2D X-rays to 3D image cephalometry analysis. Some work has been carried out to study the potential advantages of 3D imaging methods, such as computed tomography for cephalometric analysis. Kragskov [21] made a comparative result, using human dry skulls. Ferrario [22] studied 3D facial morphometry using infrared cameras, but only as a supplement for classic cephalometry. We propose a novel method based on the use of reeb graph for Automatic localization of craniofacial landmarks. The task is a difficult one due to the complexity and variability of cephalometry landmarks.
  • 4. Mestiri Makram & Hamrouni Kamel International Journal of Image Processing (IJIP), Volume (8) : Issue (2) : 2014 20 3. PROPOSED METHOD As seen in previous works, many researchers have attempted automatic cephalometric landmarks detections. However, a major drawback of the existing techniques is that they use a 2D representation of a 3D structure. In this paper we are interested in a 3D cephalometric landmarks. The development of spiral CT and cone beam CT has revolutionized the medical image techniques, the former providing outstanding resolution and the latter, with its low cost, allowing unique accessibility, as a result 3-D imaging has become an essential tool in planning and managing the treatment of facial deformity. 3.1 3D Mesh Database The image data-base in composed of Dicom images stored in a series of 2D grey level images from a Spiral CT scan (with 512 x 512 matrixes, 110 kV, and 80 mAs) performed at 1 mm slice thickness. The extraction of hard and soft head tissue (bones and skin) is done by using a threshold and region growing technique. Each bones and skin parts are separately constructed using marching cube algorithm as shown in Figure 3. FIGURE 3: 3D Mesh Extraction of Hard and Soft Tissues. The results of this extraction method are 3 meshes: interior/exterior skull bones, and mandibular bone Figure-4. FIGURE 4: 3D Mesh Extraction from Volumetric CT-Scan . 3.2 3D Cephalometric Landmarks After getting the mesh database, 3D cephalometric landmarks must be localized in hard tissue. In the literature we found 20 hard landmarks, some of them are presented in table-1. CT-Scan Image Threshold and region growing Interior Cranium bone Exterior Cranium bone Mandibule Head Skin 3D mesh reconstruction (Marching Cube) Volumic grey level images 3D mesh images
  • 5. Mestiri Makram & Hamrouni Kamel International Journal of Image Processing (IJIP), Volume (8) : Issue (2) : 2014 21 No Abbreviation Hard Tissue 1 N Nasion is the midpoint of the frontonasal suture 2 S Sella is the centre of the hypophyseal fossa (sella turcica). 3 Po Porion is the most superior point of each external acoustic meatus. 4 Or Orbitale is the most inferior point of each infra- orbital rim. 5 ANS Anterior Nasal Spine is the most anterior midpoint of the anterior nasal spine of the maxilla. 6 PNS Posterior Nasal Spine is the most posterior midpoint of the posterior nasal spine of the palatine bone. 7 PMP Posterior Maxillary Point is the point of maximum concavity of the posterior border of the palatine bone in the horizontal plane at both sides. 8 UI Upper Incisor is the most mesial point of the tip of the crown of each upper central incisor. 9 LI Lower Incisor is the most mesial point of the tip of the crown of each lower central incisor. 10 UMcusp Upper Molar cusp is the most inferior point of the mesial cusp of the crown of each first upper molar in the profile plane. TABLE 1: 10 of the 20 Cephalometric Landmarks Proposed by Swennen [23]. Figure 5 shows some sample of CT-scan images with located landmarks [23]. FIGURE 5: A: Right hard cephalometric landmarks, B: Facial hard cephalometric landmarks, C: Soft cephalometric landmarks. 3.3 Automatic Localization of Craniofacial Landmarks Figure 6 presented the bloc diagram of the proposed method for automatic localization of 3D cephalometric landmarks. After transforming the volumetric medical images to a 3D surface mesh, we need to make a reference shape model containing all cephalometric landmarks information’s. This is made with what is called “cephalometric head atlas”. FIGURE 6: Automatic localization of 3D cephalometric landmarks using reeb graphs. Cephalometric Head Atlas Patient Head Mesh Mesh Simplification (Edge Collapse) Reeb Graph Extraction Elastic Registration (Thin Plate Spline) Automatic Landmarks localisation (NNB algorithm) Rigid registration (ICP) A B C
  • 6. Mestiri Makram & Hamrouni Kamel International Journal of Image Processing (IJIP), Volume (8) : Issue (2) : 2014 22 To automatically identify 3D landmarks, we need to have prior knowledge information’s. We create a cephalometric head atlas based on medical stuff help. This atlas is composed of skull and skin meshes with 3D cephalometric landmarks as shown in Figure 7. FIGURE 7: Cephalometric atlas of skull. A: initial clean skull mesh, B: Cephalometric landmarks and C: Manual localization of cephalometric landmarks. In our method we choose to automatically localize cephalometric landmarks using reeb graph. Below we will expose the reeb graph extraction. The Reeb graph is a chart of connectivity on a surface between its critical points. The main advantage of the reeb graph is that it makes it possible to represent 3D anatomy in a simple topology way. It is composed of nodes and arcs; each whole of level associates a node. The graph of Reeb is obtained starting from the computation of a µ function introduced by the theory of Morse [24] defined on the closed surface of an object in its critical points [25]. We define µ: S ! R defined on surface S of a 3D object. The Reeb graph is a space quotient of the graph of µ in S, defined by the relation of following equivalence between X є S and Y є S: · µ(X)=µ(Y) (1) X~Y " · X et Y are in the same Related component of µ-1(µ(X)), and the image µ(Y) is S. The aspect of the graph is entirely related to the choice of the µ function, which determine the properties of stability and invariance of the resulting graphs. Several functions of application were proposed in the state of the art for the construction of Reeb graphs [26]. One example is the function (µ(ν(x,y,z)) = z / ν є S), adapted well for the models whose points are mainly distributed in the Z-axis direction. FIGURE 8: Reeb Graph with a z Projection µ Function. A B C
  • 7. Mestiri Makram & Hamrouni Kamel International Journal of Image Processing (IJIP), Volume (8) : Issue (2) : 2014 23 The function suggested in [27] as «outdistances geodesic extrema of curve » is based on growth of areas starting from local Gaussian curves on tops (germs). The results depend on the position of the germs and require no disturbed data; it’s a precise calculation of the local curves. Hilaga [27] defines a function by computing the distance from a point of the surface to the center of mass G of the object: (µ(ν) = d(G, ν) / ν є S et d : Euclidian distance) (2) The function defined by Thierny[28] at the characteristic points and the geodesic distances is defined by the following equation: µμ v = 1 − δ v, v! (3) With δ v, v! = min!"!!!δ(v, v!! ) Where δ is the function of the geodesic distances and v! are characteristic points. The function suggested in [29] is the integral of the geodetic distances g(v, p) of v to the other point’s p of the surface (4). µμ v = g v, p dS!!! . (4) In our approach we choose this µμ function and tested it with different mesh simplification. Figure-9 shows an example of reeb graph extraction. FIGURE 9: A: Mandibule Patient Mesh, B: Reeb Graph Extraction. Surface mesh simplification is the process of reducing the faces number used in the mesh surface, while keeping preserved as much as possible the overall shape, volume and boundaries. Cignoni [29] proposed a comparison of mesh simplification algorithms, and divided these methods into 7 approaches: coplanar facets merging, controlled vertex/edge/face decimation, re- tiling, energy function optimization, vertex clustering, wavelet-based approaches, and simplification via intermediate hierarchical representation. We test many simplification methods with reeb graph transformation, and we choose to use the edge collapse method. These techniques reduce a model’s complexity by repeated use of the simple edge collapse operation. Researchers have proposed various methods of determining the “minimal cost edge” to collapse at each step (figure-10). FIGURE 10: Edge Collapse Algorithm [30]. A C
  • 8. Mestiri Makram & Hamrouni Kamel International Journal of Image Processing (IJIP), Volume (8) : Issue (2) : 2014 24 We try to define the effect of the edge collapse simplification process on the reeb graph extraction. The result in shown in figure-11 FIGURE 11: Patient Skull Simplification. A: 332.932 vertices and 662.564 faces, B: 20.047 vertices and 38.958 faces and C: 3.887 vertices and 7320 faces. We notice that, after mesh simplification, the number of nodes in the reeb graph decreased and they became closed to cephalometric landmarks. And, of course, the computation time is reduced Figure-12. FIGURE 12: Decrease of reeb graph nodes and reconciliation with cephalometric lanmarks. Mesh registration is a technique used to find a transformation for mapping a source mesh known as a reference mesh and a target mesh. This technique associated to a head cephalometric atlas should help us to get closer to cephalometric landmarks. Every elastic registration process should begin with a rigid one. In this study, we targeted the rigid registration between two different meshes using the ICP algorithm (Iterative Closest Point). The principle of ICP is to iterate between a step of mapping data and another step of optimization of rigid transformation until convergence. The transformation used for registration is composed of a 3D rotation and translation. At each iteration, the Algorithm provides a list of matched points and an estimate of the transformation. The algorithm converges when the error in distance between matched points is less than a given threshold. The ICP algorithm is presented in the following [30] where Np is the number of points CP1 and CP2 and k is the index of current iteration. Figure-13 shows a example of rigid registration between mesh patient and cephalometric mesh atlas. FIGURE 13: Rigid registration between patient and the cephalometric atlas A B C
  • 9. Mestiri Makram & Hamrouni Kamel International Journal of Image Processing (IJIP), Volume (8) : Issue (2) : 2014 25 Mesh elastic registration is a mesh deformation process; one of the transformations that are able to represent elastic deformations is the thin-plate spline (TPS). Thin plate splines were introduced by Bookstein[31] for geometric design. In 2D images, the TPS model describes the transformed coordinates (xT,yT) both independently as a function(7) of the original coordinates (x, y): (xT,yT)=(fx(x,y),fy(x,y)) (7) The algorithm begins with a given displacements of a number of landmark points, the TPS model interpolates those points, while maintaining maximal smoothness. For each landmark point (x,y), the displacement is represented by an additional z-coordinate, and, for each point, the thin plate is fixed at position (x, y, z). The strain energy is calculated by integrating the second derivative over the entire surface that can be minimized by solving a set of linear equations. dxdy y z yx z x z R ∫∫ ⎟ ⎟ ⎠ ⎞ ⎜ ⎜ ⎝ ⎛ ⎟⎟ ⎠ ⎞ ⎜⎜ ⎝ ⎛ ∂ ∂ +⎟⎟ ⎠ ⎞ ⎜⎜ ⎝ ⎛ ∂∂ ∂ +⎟⎟ ⎠ ⎞ ⎜⎜ ⎝ ⎛ ∂ ∂ 2 2 2 2222 2 2 .2 (8) The TPS model for one of the transformed coordinates is given by parameter vectors a and D (9): F(xT,yT)= a1+a2x+a3y+…+ ( )( )∑= − n i ii yxLFD1 , (9) Where F(r)= r2log(r) is the basis function, a = [a1 a2 a3 a4]T defines the affine part of the transformation, D gives an additional non-linear deformation, and the Li are the landmarks that the TPS interpolates figure-4. FIGURE 14: A template configuration (left) and a target configuration (right) of five landmarks each. The deformation right grid illustrates the thin-plate spline function between these configurations as applied to the left regular grid. The method interpolates some of the points using smoother transformation controlled by a parameter µ, which weights the optimization of landmark distance and smoothness. For µ = 0, there is full interpolation, while for very large µ, there is only an affine transformation. In our methods the landmarks used in the TPS algorithm are the reeb graph nodes as shown in figure- 15. FIGURE 15: Elastic Registration using Thin Plate.
  • 10. Mestiri Makram & Hamrouni Kamel International Journal of Image Processing (IJIP), Volume (8) : Issue (2) : 2014 26 As result of the elastic registration, the patient reeb graph nodes become closer to the atlas cephalometric head landmarks. Since, some reeb graph nodes are not cephalometric landmarks, we have to select the real ones. Automatic localization of cephalometry landmarks This step starts with the creation of possible landmark localization in cephalometric atlas. Then, it identifies the reeb graph nodes that exist in the mesh patient. For each node, it searches the closest landmarks in the cephalometric atlas through the calculation of Euclidean distance between an acceptable range of ±2 mm. Finally, as a result all patient reeb graph nodes are associated to a nearest cephalometric landmarks[32]. FIGURE 16: Localization of Cephalometric Landmarks in Patient Skull. 4. TEST AND RESULTS To validate our approach we asked a doctor to conduct manually a cephalometric analysis on the mesh patient. We then calculate the error between its points and cephalometric landmarks automatically localized by our method. The result is showed in the following table: No Abbreviation Error in mm Recognized 1 N 0.5 Yes 2 S 2.6 No 3 Po 1.3 Yes 4 Or 0.4 Yes 5 ANS 0.7 Yes 6 PNS 2.8 No 7 PMP 2 Yes 8 UI 1.3 Yes 9 LI 1.7 Yes 10 UMcusp 1.8 Yes 11 LMcusp 0.9 Yes 12 Men 0.8 Yes 13 Go 1.5 Yes 14 Fz 1.7 Yes 15 Zy 0.3 Yes 16 A 0.4 Yes 17 B 0.5 Yes 18 Pog 0.8 Yes 19 Ba 0.6 Yes 20 Co 1.4 Yes TABLE 2: Result of cephalometric landmarks detection
  • 11. Mestiri Makram & Hamrouni Kamel International Journal of Image Processing (IJIP), Volume (8) : Issue (2) : 2014 27 Our method localized 18 landmarks on 20. This mean a percentage of 90% landmarks in patient mesh was recognized. FIGURE 17: Localization of A- Sella and B- PNS Landmarks. The Sella landmark (S) is very difficult to be localized. It is not localized in the surface mesh, and the doctors need to insert it taking into account the interior skull surface. The Posterior Nasal Spine Landmark (PNS) is defined on the exo-cranial skull base view of the 3-D hard tissue surface representation. This landmark depends highly on the accuracy of the mesh reconstruction. 5. CONCLUSION In this paper, we have proposed a novel method for automatic 3D localization of cephalometric landmarks. Our approach is based on reeb graph nodes. In particular, we have presented results for localized 90% of cephalometric landmarks in CT-Scan medical images. We initiated a new approach for automatic three-dimensional cephalometric analysis. The proposed method needs to be validated on a larger database. 6. REFERENCES [1] A.Thomson « On growth and form ». Cambridge, UK: Cambridge Univ. Press, 1992. [2] Broadbent HB. « A new X-ray technique and its application to orthodontia ». Angle Orthod 1931; [3] Brodie AG. « On the growth pattern of the human head from the third month to the eighth year of life ». Am J Anat 1941; 68:209–62. [4] Downs WB. « Variations in facial relationships: their significance in treatment and prognosis ».Am J Orthod 1948; 34:812–40. [5] Ricketts RM. « Cephalometric Analysis and Synthesis ». Angle Or- thod, 1961; 31:141–56. [6] Steiner C. « Cephalometrics in clinical practice ». Angle Orthod 1959; 29:8–29. [7] A. Levy-Mandel, A. Venetsanopoulos, J. Tsotsos, « Knowledge-based landmarking of cephalograms », Comput. Biomed. Res. 19 (3) (1986) 282–309. [8] L. Mero, Z. Vassy, « A simpliffed and fast version of the Hueckel operator for finding optimal edges in pictures », Proceedings of the Fourth International Joint Conference on Artificial Intelligence, Tbilissi, Georgia, USSR, September 1975, pp. 650–655. [9] S. Parthasaraty, S. Nugent, P.G. Gregson, D.F. Fay, « Automatic landmarking of cephalograms », Comput. Biomed. Res. 22 (1989) 248–269. A C
  • 12. Mestiri Makram & Hamrouni Kamel International Journal of Image Processing (IJIP), Volume (8) : Issue (2) : 2014 28 [10] C.K. Yan, A. Venetsanopoulos, E. Filleray, « An expert system for landmarking of cephalograms », Proceedings of the Sixth International Workshop on Expert Systems and Applications, 1986, pp. 337–356. [11] J.L. Contereras-Vidal, J. Garza-Garza, « Knowledge-based system for image processing and interpolation of cephalograms », Proceedings of the Canadian Conference on Electrical and Computer Engineering, 1990, pp. 75.11–75.14. [12] P.H. Jackson, G.C. Dickson, D.J. Birnie, Digital « image processing for cephalometric radiographic: a preliminary report », Br. J. Orthod. 12 (1985) 122 [13] A.M. Cohen, A.D. Linney, « A low cost system for computer-based cephalometric analysis », Br. J. Orthod. 13 (1986) 105–108. [14] Forsyth D. D.Davis, « Knowledge-based cephalometric analysis: a comparison with clinicians using interactive computer methods», Comput. Biomed. Res. 27 (1994) 10–28. [15] M. Sid-Ahmed, J. Cardillo, « An image processing system for the automatic extraction of craniofacial landmarks », IEEE Trans. Med. Imaging 13 (2) (1994) 275–289. [16] V. Grautel, M.C. Juan, C. Monserrat, C. Knoll, « Automatic localization of cephalometric landmarks », J. Biomed. Inf. 34 (2001) 146–156. [17] M. Desvignes, B. Romaniuk, R. Demoment, M. Revenu, M.J. Deshayes, « Computer assisted landmarking of cephalometric radiographs », Proceedings of the Fourth IEEE Southwest Symposium on Image Analysis and Interpretation, 2000, pp. 296–300. [18] T.J. Hutton, S. Cunningham, P. Hammond, « An evaluation of active shape models for the automatic identification of cephalometric landmarks », Eur. J. Orthodont. 22 (5) (2000) 499– 508. [19] E. Uchino, T. Yamakawa, « High speed fuzzy learning machine with guarantee of global minimum and its application to chaotic system identification and medical image processing », Proceedings of the Seventh International Conference on Tools with Artificial Intelligence, 1995, pp. 242–249. [20] A. Innes, V. Ciesielski, J. Mamutil, S. John, « Landmark detection for cephalometric radiology images using pulse coupled neural networks », International Conference in Computing in Communications, June 2002, pp. 391–396. [21] Kragskov J, Bosch C, Gyldensted « Comparison of the reliability of craniofacial anatomic landmarks based on cephalo-metric radiographs and 3-dimensional CT scans », Cleft Palate Crani- ofac J 1997; 34:111–6. [22] Ferrario VF, Sforza C, Poggio CE, « Facial 3-dimensional morphometry ». Am J Orthod Dentofac Orthop 1996; 109:86–93. [23] Gwen R. J. Swennen,Filip Schutyser,Jarg-Erich Hausamen « Three-dimensional cephalometry: a color atlas and manual » Amazon Book. [24] Milnor (j.w), « Morse theory », Princeton University Press, Princeton, NJ, (1963) [25] Reeb, G. « Sur les points singuliers d'une forme complètement intégrable ou d'une fonction numérique ». Livre stringer (1946).
  • 13. Mestiri Makram & Hamrouni Kamel International Journal of Image Processing (IJIP), Volume (8) : Issue (2) : 2014 29 [26] Tony Tong Thése « Indexation 3D de bases de données d’objets par graphes de Reeb améliorés » (2005) [27] Hilaga, Shigawa, « Topology Matching for fully automatic Similarity Estimation of 3D Shape », ACM SIGGRAPH, Los Angeles, CA, USA (2001). [28] J. Tierny, J. Vandeborne et M. Daouadi, « Graphe de Reeb de haut niveau de maillages polygonaux ». Journées de l’association Francophone d’informatique graphique, bordeaux (2006). [29] P. Cignoni, C. Montani, R. Scopigno : « A comparison of mesh simplification algorithms », Int. Symp. on 3D Data Processing, Visualization, and Transmission. [30] Besl P.J., McKay N.D., “A Method for Registration of 3D Shapes", IEEE Trans. On Pattern Analysis and Machine Intelligence (TPAMI), Vol.14(2), pp. 239-255,1992. [31] F. L. Bookstein, “Principal Warps: Thin-Plate Splines and the Decomposition of Deformations” IEEE Transaction on pattern analysis and machine intelligence. Vol 11, No 6, June 1989. [32] Makram Mestiri, Sami Bourouis and Kamel Hamrouni, “Automatic Mesh Segmentation using atlas projection and thin plate spline: Application for a Segmentation of Skull Ossicles”, VISAPP 2011.