Computerized Medical Imaging and Graphics 33 (2009) 359368

Contents lists available at ScienceDirect

Computerized Medical Imaging and Graphics

journal homepage: www.elsevier.com/locate/compmedimag

Automated identication of anatomical landmarks on 3D bone models

reconstructed from CT scan images
K. Subburaj a , B. Ravi a, , Manish Agarwal b
a
OrthoCAD Network Research Centre, Department of Mechanical Engineering, Indian Institute of Technology Bombay, Powai, Mumbai 400076, India
b
Department of Surgical Oncology, Tata Memorial Hospital, Parel, Mumbai 400012, India

a r t i c l e i n f o a b s t r a c t

Article history: Identication of anatomical landmarks on skeletal tissue reconstructed from CT/MR images is indis-
Received 5 February 2009 pensable in patient-specic preoperative planning (tumour referencing, deformity evaluation, resection
Accepted 2 March 2009 planning, and implant alignment and anchoring) as well as intra-operative navigation (bone registra-
tion and instruments referencing). Interactive localisation of landmarks on patient-specic anatomical
Keywords: models is time-consuming and may lack in repeatability and accuracy. We present a computer graphics-
Anatomical landmarks
based method for automatic localisation and identication (labelling) of anatomical landmarks on a 3D
Curvature and its derivatives
model of bone reconstructed from CT images of a patient. The model surface is segmented into different
Knee
Virtual surgery planning
landmark regions (peak, ridge, pit and ravine) based on surface curvature. These regions are labelled auto-
matically by an iterative process using a spatial adjacency relationship matrix between the landmarks.
The methodology has been implemented in a software program and its results (automatically identied
landmarks) are compared with those manually palpated by three experienced orthopaedic surgeons, on
three 3D reconstructed bone models. The variability in location of landmarks was found to be in the range
of 2.155.98 mm by manual method (inter surgeon) and 1.924.88 mm by our program. Both methods
performed well in identifying sharp features. Overall, the performance of the automated methodology
was better or similar to the manual method and its results were reproducible. It is expected to have a
variety of applications in surgery planning and intra-operative navigation.
2009 Elsevier Ltd. All rights reserved.

1. Introduction
Anatomical landmarks are distinct regions or points on skeletal
tissues with uniqueness in shape characteristics in their vicinity.
Most of the landmarks used during surgical procedures are manu-
ally palpable and geometrically recognizable. For example, in knee
joint major landmarks include: on distal femur bonemost distal
point of medial and lateral condyles (MC and LC), medial and lat-
eral epicondyle eminences (ME and LE), peak on medial and lateral
side of anterior ridge (MP and LP) of prominent patellar groove, and
adductor magnus attachment tubercle (AT); on proximal tibiatibial
tuberosity (TT), Gerdys tubercle (GT), apex of Fibula (HF), most
medial and lateral point of tibial plateau (MP and LP), and medial
and lateral intercondylar tubercle (MIT and LIT). These are shown
in Fig. 1.
Accurate localisation of 3D landmarks on skeletal tissue is
indispensable in biomechanical studies and computer integrated
surgery, including custom implant design and positioning [14],
intra-operative navigation, joint kinematics study [57], deformi-
Corresponding author. Tel.: +91 22 2576 7510; fax: +91 22 2572 6875.
E-mail addresses: b.ravi@iitb.ac.in, prof.b.ravi@gmail.com (B. Ravi).
ties assessment [8], tumour resection [9], referencing [10], and
registration in shape models [11] (Lorenz and Krahnstover [29]). In
situations involving tumours and highly deformed bones, patient-
specic 3D anatomical models reconstructed from CT or MR images
are used for visualisation and surgical planning (tumour local-
isation, resection planning, and prosthesis positioning). Various
methods used in current practice for localising anatomical land-
marks on 3D bone models are given in Table 1 along with the
issues associated with them. Variability associated with identify-
ing and locating anatomical landmarks (for example, on the knee)
has the potential to affect the surgical outcome. Semi-automatic
methods for landmark identication with interactive control offer
the possibility of overcoming the above problems. However, exist-
ing computational approaches to landmark detection often suffer
from a signicant number of false detections [12].
Several studies have been reported on interactive localisation
of landmarks on dry bone models, laser digitized data [8,13] or
medical images [3]. Van Sint Jan [14] presented a comprehensive
list of bony landmarks and described the procedure for locating
them on a patient by palpation. Della Croce et al. [15] carried out
experimental studies on manual location of lower limb landmarks
on dry bone models with a group of surgeons and concluded that
the variations are in the range 625 mm. Liu et al. [8] manually

0895-6111/$ see front matter 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compmedimag.2009.03.001
360 K. Subburaj et al. / Computerized Medical Imaging and Graphics 33 (2009) 359368

Fig. 1. Anatomical landmarks (a) on distal femur (b) on proximal tibia.

identied landmarks on a scanned surface of foot model, guided
by curvature values, for evaluating tibial torsion. Reproducibility of
the results appears to be dependent on users knowledge of land-
marks. The effect of scanning artefacts is not very clear. Yang et al.
[16] studied the relationships between bony and soft tissue land-
marks using cephalometric radiographs to diagnose facial growth
abnormalities prior to treatment. Maudgil et al. [17] manually iden-
tied anatomical landmarks on a 3D model reconstructed from MRI
images for morphometric analysis. In general, manual localisation
of landmarks and related measurements consume time, require a
high level of expertise, are tedious and time-consuming, and may
lack in accuracy/repeatability.
Advances in medical imaging, image analysis, and computa-
tional capabilities offer an opportunity to exploring shape property
for localising anatomical landmarks. Invariant geometric measures
such as curvatures, extreme points, and higher order derivatives
can be used for classifying surface regions [18] based on their
shapes. Surface curvature is independent of position and orienta-
tion of patients and it can help in reliable detection of geometric
features, from which landmarks can be derived. Surfaces can be
classied into various regions such as ridges, umbilicus and singular
lines [19,20]. Drerup and Hierholzer [27] identied lumbar dim-
ples based on surface curvature, and used it to study spine shape
and muscle movement over pelvis. Ehrhardt et al. [28] used pre-
dened templates to locate landmarks on 3D hip models for surgery
Table 1
Various methods used for localising landmarks in surgical procedures.
Method Issues
Manual identication Driven by expertise of the examiner
Inter examiner variability
Time-consuming
Lack of repeatability
planning. Accurate morphing and registering the atlas model with
the patient model is however necessary before transferring the
landmarks. Frangi et al. [11] generated dense landmark nodes (dec-
imating dense triangular data of the boundary) on 3D shapes and
used for constructing active shape models. This method does not
actually identify or localise individual landmarks. Rohr [21] and
Worz and Rohr [12] showed the use of invariant principles (rst
order derivatives) in identifying landmarks from 2D CT images.
These image based landmarks are used for multi-modal image
registration. This could be extended to localising 3D skeletal land-
marks from reconstructed three-dimensional anatomical models,
but does not appear to have been attempted so far.
There appears to be very little reported work on fully or
semi-automated identication of anatomical landmarks on skeletal
tissue. One reason could be that these landmarks are inuenced by
bone morphology, and hence prior knowledge of anatomical land-
marks is necessary to locate and identify them precisely [9]. Also
positional uncertainty of anatomical landmarks is caused by the
fact that they are not clearly identiable discrete points, but are
rather relatively large and curved areas [15]. The ideal situation
would be the ability to palpitate or compute the same reference
point on bony prominence in a repeated manner. Characterizing
specic landmarks for the desired application based on their invari-
ant geometric characteristics and identifying them would be an
effective way to automate the procedure, increase robustness and
repeatability, and reduce false identications.
In this work, we have developed a systematic approach to
identify anatomical landmarks driven by their general geomet-
ric characteristics (curvature analysis) and adjacency relationship
between the landmarks (rules), and demonstrated the same for a
knee joint. The methodology is described rst, followed by the
results of its implementation. The last section summarizes the work
and concludes with directions for future research.

Recognizable markers on skin Errors due to skin movement artefacts 2. Methodology

before scanning
Anatomical atlas (transferring
pre-dened landmarks from
generic model into patients
model)
Inserting metallic pin instead of
bony landmarks for registration
Inaccurate or impossible in obese
patients
Accurate registration of 3D models
before transfer
Unconsidered anatomical deformities
Unique bone morphology
Cannot be performed on patients
If performed, then associated trauma
Extreme risk in osteoporosis patients
The overall methodology for identifying anatomical landmarks
is shown in Fig. 2. First a 3D surface model is reconstructed from
a set of CT images. It is assumed that the reconstructed model is
perfectly segmented (bone density thresholding) to represent the
true anatomical morphology of the bone. Then principal curvatures
and their derivatives are computed on every vertex of the triangu-
lated surface. These vertices are segregated based on the curvature
descriptors into various regions. Thereafter anatomical landmarks
are identied and labelled from these regions using pre-dened
rules (spatial adjacency). The methodology is described in detail
below.
 K. Subburaj et al. / Computerized Medical Imaging and Graphics 33 (2009) 359368
Fig. 2. Overview of the methodology for identifying landmarks.
2.1. Notations
A surface mesh M consists of a set of vertices V = {vi }i IR3 ,
which are connected by a set of edges E = {ej = (vj1 , vj2 )}j and a
set of triangles T = {tk = (vk1 , vk2 , vk3 )}k . Let v V be a vertex of
a triangle mesh M and let v1 , . . . , vn be the ordered neighbouring
vertices of v as shown in Fig. 3.
The edges ei = vi v.
The angles between two successive edges are i = (ei , ei+1 ).
Fig. 4. Reference planes and model orientation (a) coronal, (b) sagittal, (c) axial.
361
Fig. 3. Notations used in triangular mesh representation.
The triangle between ei and ei+1 is named as ti = (v, vi , vi+1 ).
Face normal n i = ei ei+1 /||ei ei+1 ||.
The dihedral angle at an edge ei is i = (n  i1 , n
 i ) (the angle
between the normals of the adjacent triangles).
All position measurements (for landmarks) are taken with
respect to the CT coordinates (Fig. 4). The horizontal left-to-
right arrow represents medio-lateral (ML), horizontal front-to-rear
represents anterio-posterior (AP), and vertical bottom-to-top rep-
resents distal-proximal (DPx) directions. There are three reference
planes. A coronal plane (YZ) is perpendicular to the ground, and
divides the model into anterior and posterior portions (Fig. 4a). A
sagittal plane (XZ) is also perpendicular to the ground, but divides
the model into left and right (medial and lateral) portions (Fig. 4b).
An axial or transverse plane (XY) is parallel to the ground, and
divides the model into inferior and superior (distal and proximal)
portions (Fig. 4c).
2.2. Shape characteristics of landmark regions
Anatomical landmarks are distinct regions or points on bones
with uniqueness in shape characteristics in their vicinity, and can
be classied on the basis of their shape characteristics, described
here. Geometrically invariant measures such as curvatures, concav-
ity, and convexity are useful in identifying anatomical landmarks
on a 3D anatomical model. A set of common geometric shapes such
362 K. Subburaj et al. / Computerized Medical Imaging and Graphics 33 (2009) 359368
Fig. 5. Shape of landmark regions: peak, pit, ridge, and ravine.
as peak, ridge, pit, and ravine is observed in regard to the anatom-
ical landmarks on skeleton and used in orthopaedic surgery, for
example, in knee replacement or reconstruction (Fig. 5).
2.2.1. Peak and ridge
Peak and ridge are n 2 dimensional elements of an n-
dimensional polytope. A peak can be described as a point with local
maximum real-valued function and high gradient with respect to
surroundings. A ridge is a narrow, raised strip of line formed at the
junction of two sloping surfaces diverging towards the ground. In
its broadest sense, the notion of peak and ridge generalizes the idea
of a local maximum of a real-valued function such as curvature and
its derivatives.
2.2.2. Pit and ravine
Pit and ravine have a similar characteristic that is they represent
a local minimum of an intrinsic real-valued function. Pit is a point
surrounded by a high walled locally depressed region on surface.
A ravine can be dened as a line formed by intersection of two
concave surfaces.
2.3. Estimation of discrete curvature
The curvature is computed on each vertex of the triangulated
surface data generated by surface tting on volumetric data. Let p be
a point on the surface S. Consider all curves Ci on S passing through
the point p on the surface. Every such Ci has an associated curvature
ci given at p. Of those curvatures ci , at least one is characterized as
maximal k1 and one as minimal k2 . These two curvatures k1 and
k2 are known as the principal curvatures of the surface. Gaussian
curvature K is the product of k1 and k2 . Mean curvature H is the
average of k1 and k2 . We also use integral Gaussian curvature K
and the integral mean curvature H with respect to the area A. The
well-known GaussBonnet theorem for polygonal cases [22] gives:
 
n all of its three vertices are in the same group.
K = KdA = 2 i In the second stage, probable landmark regions are formed by
A the edge-connected triangles within a group. Every seed triangle
i=1
 region (Rijk ) is created with a seed triangle from a seed list (Sij ).
1
n
 
H = HdA = ei  i 
A
4 ciating edge-connected adjacent triangles. This continues until no
i=1
Gaussian curvature is an intrinsic measure of curvature, i.e., its
value depends only on how distances are measured on the surface,
not on the way it is embedded in space. On the other hand, mean
curvature is an extrinsic measure of curvature, i.e., the curvature
of an embedded surface in some ambient space such as Euclidean
Table 2
Different surface types based on curvature.
K<0 K=0
H<0 Ridge (1) Ridge (2)
H=0 Flat (5)
H>0 Valley (6) Valley (7)
space. To derive the curvatures at the vertex v from these inte-
gral values, we assume the curvatures to be uniformly distributed
around the vertex, and normalized by the area [23]:
n
K 2
i=1 i
Gaussian curvature : K= =
A 1/3A
n   
ei  i 
H i=1
Mean curvature : H= =
A 1/3A
where A is the sum of the areas of neighbouring faces around a
vertex v. The area used is the barycentric area SB which is one third
of the area of the triangles adjacent to v, and can be constructed by
connecting the edge midpoints with the barycentres of the adjacent
triangles.
2.4. Extraction of landmark regions
As introduced earlier in Section 2.2, landmark regions are
localised or extracted by grouping vertices and triangles based on
their curvature values (K and H). This is independent of spatial
location of vertices and is carried out in two stages. First, vertices
are grouped according to their H value. Two clusters of vertices
(Ci v , {i = h, l}) are obtained from this grouping (maximum
threshold = CTMAX and minimum threshold = CTMIN ). All anatomi-
cal landmarks shown in Fig. 1 are located at high or low curvature
(H) regions of model surface (Table 2). The vertices in Ch lead to
landmark regions such as peaks and ridges, while vertices in Cl
lead to pits and valleys. Variations of specic landmark regions are
segregated from these clusters (Ch and Cl ) based on their K val-
ues (negative K, zero K, and positive K). This second level grouping
leads to six groups (Gi Ch , {i = 1, 2, 3} and Gi Cl , {i = 4, 5, 6} of
vertices. Landmark regions are extracted from these groups in the
second stage of this algorithm. A set of seed-triangles list (Sij Gi ,
{i=1,2, . . ., 6; j=1,2, . . . m}) is formed from corresponding groups Gi
for region growing. A triangle is added to a seed list of the group if
must be a part of a region and the corresponding group. A new
A recursive process is started for the current region (Rijk ) by asso-
edge-connected triangle within the region is left in the seed list.
Once, a triangle is associated with a region, it is removed from the
seed list. Likewise, new regions are created until no triangle is left in
the seed list. The same procedure is repeated in all seed lists (S1S6)
to form regions. After all landmark regions ((Rijk Sij , {i = 1,2, . . ., 6;
j = 1,2, . . ., m; k = 1,2, . . ., n}) are formed, we proceed to eliminate
or lter unwanted regions. For this purpose, a set of rules has been
Table 3
Anatomical directions according to the sign in spatial adjacency matrix.
K>0
Sign ML AP DPx
Peak (3)
ve Medial Anterior Distal
Pit (8) +ve Lateral Posterior Proximal
 K. Subburaj et al. / Computerized Medical Imaging and Graphics 33 (2009) 359368 363

Table 4
Spatial adjacency matrix of distal femur landmarks.

Node Shape AT ME LE MP LP MC LC
Shape Peak Peak Peak Peak Peak Peak Peak

AT Peak L:Px M:P:Px M:P M:P:D M:Px M:P:Px

ME Peak M:D M:A:D M:P M:P:D M:Px M:Px
LE Peak L:A:D L:P:Px L:P L:P:D L:Px L:Px
MP Peak L:A:D L:A M:A M:P:D L:A:Px M:A:Px
LP Peak L:A:Px L:A:Px M:A:Px L:A:Px L:A:Px M:A:Px
MC Peak L:D L:D M:D M:P:D M:P:D M:P:D
LC Peak L:A:D L:D M:D L:P:D L:P:D L:A:Px

designed based on the number of triangles associated with a region,
surface area of the region, and the location of the region. The steps
of the algorithm for localising landmark regions are given below:
Algorithm. Localisation of landmark regions
Input: List of vertices ( ) with curvature values.
Output: Landmark regions R
Procedure:
Step 1: set CTMax and CTMin.
Step 2: get clusters Ci v , {i = h, l} by thresholding.
Step 3: form groups Gi Ch , {i = 1, 2, 3} and Gi Cl , {i = 4, 5, 6}
from corresponding clusters using K.
Step 5: construct triangle seed lists Si j Gi , {i = 1,2, . . ., 6; j = 1,2, . . .,
M}.
Step 6: get landmark regions Rijk Sij , {i = 1,2, . . ., 6; j = 1,2, . . ., M;
k = 1,2, . . ., N} by segmenting homogenous triangles in particular
seed list.
According to differential geometry, local surface shape is
uniquely determined by the rst and second fundamental forms
(Do Carmo [22]). Gaussian and mean curvature combine these rst
and second fundamental forms in two different ways to obtain
scalar surface features that are invariant to rotations, translations
and changes in parameterization [24]. There are eight fundamen-
tal viewpoint-independent surface types that can be characterized
using only the sign of the mean curvature (H) and Gaussian curva-
ture (K). In this work, we focus on only ve surface types, which
are typically observed in anatomical models and used as landmark
regions (Table 2). A unique geometric identier is given to each
landmark region before identifying them anatomically.
The geometric identier for a landmark region LRi is computed
as follows:
LRi = 1 + 3(1 + sgn(H, )) + (1 sgn(K, ))
where
 
+1 : x>
sgn(x, ) = 0 : |x|
1 : x<
The geometric identier of a landmark region is used to iden-
tify the geometric shape according to Table 1. These regions are
ltered out by surface area, geometric characteristics of the asso-
ciated triangles, and spatial location of the region with respect to
the required landmarks. The extracted landmark regions have to
be identied as a set of specic anatomical landmarks of the bone
being focussed. We have used spatial conguration of the land-
marks for this purpose, which is explained in the next section.
2.5. Spatial adjacency matrix
Most of the landmarks used in surgery are palpable (in geomet-
ric terms: high and low curvature regions) and vary only in size
and spatial location depending on patients age, gender, ethnic-
ity, etc. These anatomical landmarks are usually constrained in a
spatial relationship that is the same for all models. We therefore
propose an additional characteristic for landmark identication:
relative position (adjacency) of a landmark with respect to other
known landmarks. This is captured in a network diagram, where
a node represents a landmark and an arc represents the relation-
ship between two landmarks. We characterize an arc by the spatial
adjacency of a landmark relative to its neighbouring landmarks.
This is encoded in terms of anatomical directions and sequence
M/L:A/P:D/Px (Table 3). After all landmarks are localised using the
methods described in Sections 2.2 and 2.4, the network diagram
is formed. The network diagram is shown as a spatial adjacency
relationship matrix for computational purpose. This matrix is rst
lled with vector distances between pairs of landmarks before con-
verting them into a sequential code given as above. This vector
distance carries the directional information indirectly in terms of
the sign (). Table 3 shows the interpretation of the signed dis-
tance values in terms of anatomical directions. Tables 4 and 5 show
the spatial-adjacency relations of the anatomical landmarks of the
knee (Fig. 1) as a matrix. These landmarks were selected based
on instructions given in [14] and in consultation with orthopaedic
surgeons.
The spatial adjacency relationship of the landmarks given in
the above tables is used to label the anatomical landmarks after
localisation. We have used simplistic way of representing spatial
adjacency relationship, since incorporating average distance and
standard deviation between landmarks (used in active shape mod-
els) would be superuous for identifying anatomical landmarks.
The relations between localised landmark regions are mapped one-
to-one with the spatial network diagram. Implementation of the

Table 5
Spatial adjacency matrix of proximal tibia and bula landmarks.

Node Shape TT MP LP GT HF LIT MIT

Shape Peak Peak Peak Peak Peak Peak Peak

TT Peak L:A:D M:A:D M:A:D M:A:D A:D A:D

MP Peak M:P:Px M M:P:Px M:A:Px M M
LP Peak L:P:Px L L:P:Px L:A:Px L L
GT Peak L:P:Px L:A:D M:A:D M:A L:A:D L:A:D
HF Peak L:P:Px L:P:D M:P:D L:P L:P:D L:P:D
LIT Peak P:Px L M M:P:Px M:A:Px L
MIT Peak P:Px L M M:P:Px M:A:Px M
364 K. Subburaj et al. / Computerized Medical Imaging and Graphics 33 (2009) 359368

Fig. 6. Reconstructed knee bone model from CT images.

above mentioned methodology and a case study is presented in the
next section.
3. Results and discussion
A set of CT images were obtained for the left knee of a 26-year-
old male (format: DICOM, resolution: 512 512 pixels; thickness:
0.67 mm; slices: 668). A software program developed in our lab
was employed for reconstructing its 3D model [25]. The CT images
were smoothed and processed with a gradient operator to highlight
high intensity change. Images were thresholded (range of intensity
values) by analyzing image intensity histograms to eliminate soft
tissue for more effective extraction of bone. The processed images
were stacked and converted into a surface model (Fig. 6) by sur-
face tiling. The triangulated surface model data is represented by
winged edge data structure with explicit topological information
for easy data accessibility, necessary in curvature computation and
geometric landmark localisation. The geometric errors created dur-
ing conversion of triangle surface data from volumetric data such
as degeneracy, holes at a vertex, gaps along a line of intersection,
and collapsed edges of small triangles [26], are considered and cor-
rected by a set of designed mesh lters. After 3D reconstruction,
the curvature values are computed on the bone surface model using
the algorithm given in Section 2.3. Fig. 7 shows normalised curva-
ture values on each vertex mapped with colour values. Red and blue
indicate regions with high and low curvature, and yellow and green
indicate smooth transition of curvature from high to low.
After computing curvature values on each vertex, landmark
regions (Rijk ), which contain probable landmark locations, are seg-
regated as described in Section 2.4. The curvature threshold (CTMAX
and CTMIN ) values used in this investigation are the upper and lower
25 percentile of spread of the curvature in curvature histogram.
Clustering, grouping, and region-growing based on curvature sim-
ilarity among neighbouring triangles are performed to segregate
landmark regions. Fig. 8 shows the landmark regions on the knee
model.
Among the segregated landmark regions, there is a possibil-
ity of generating false hints in landmarks identication. Hence
unwanted regions are eliminated as per steps given in Section 2.4

Table 6
Parameters of the distal femur and proximal tibia landmarks.

Distal femur Proximal tibia

LM Location (mm) LM Location (mm)

F1 45.77, 30.05, 69.31 T1 80.81, 64.2, 62.0

F2 71.27, 23.71, 76.33 T2 86.62, 63.5, 62.8
F3 21.61, 63.21, 69.22 T3 23.68, 62.96, 61.53
F4 100.53, 64.08, 68.06 T4 96.73, 61.13, 60.41
F5 26.35, 61.89, 76.28 T5 92.95, 85.03, 46.52
F6 35.18, 66.96, 44.57 T6 82.04, 26.68, 22.31
F7 76.57, 65.78, 45.56 T7 90.79, 38.85, 44.45
Fig. 7. Curvature map of 3D anatomical model of knee.
 K. Subburaj et al. / Computerized Medical Imaging and Graphics 33 (2009) 359368 365

Fig. 8. Landmark regions on knee model.

Table 7
Computed adjacency relationship matrix between distal femur landmarks.

Shape F1 F2 F3 F4 F5 F6 F7
Shape Peak Peak Peak Peak Peak Peak Peak

F1 Peak 0:0:0 25.5:6.3:7 24.2:33.2:0.1 54.8:34:1.3 19.4:31.8:7 10.6:36.9:24.7 30.8:35.7:23.8

F2 Peak 0:0:0 49.7:39.5:7.1 29.3:40.4:8.3 44.9:38.2:0 36.1:43.3:31.8 5.3:41.6:30.8
F3 Peak 0:0:0 78.9:0.9:1.2 4.7:1.3:7.1 13.6:3.8:24.7 55:2.1:23.7
F4 Peak 0:0:0 74.2:2.2:8.2 65.4:2.9:23.5 24:1.2:22.5
F5 Peak 0:0:0 8.8:5.1:31.7 50.2:3.9:30.7
F6 Peak 0,0,0 41.4:1.2:1
F7 Peak 0:0:0

to reduce false identication of landmarks. Since the relative loca-
tion of selected landmarks shown in Fig. 1 are known in terms
of anatomical directions, identifying probable landmarks among
these landmark regions is carried out with only modest computa-
tion. The location of a landmark is given by the mean location of the
points in the particular landmark region, and is used as its reference
point. Surface area of a landmark gives an approximate indication
of its spread. The regions are stored temporarily as separate objects
(Table 6) and characterized as peak, ridge, pit, and ravine.
A relationship matrix is formed between the landmarks based
on their relative locations (Tables 7 and 8) with respect to other
regions. This is evaluated by the landmark identication rules
using geometric characteristics and adjacency relations between
landmarks. For example, take landmark F1, with respect to land-
mark F2 (code = M:P:D), F1 is located medial (ML < 0), posterior
(AP < 0), and distal (DPx < 0). Similarly, it can be seen that F1 is
located lateral (ML < 0) and anterior (AP < 0) with respect to F3
(code = L:A). The distal-proximal relation is omitted due to a small
difference (<1 mm) in value.
Given that F1 and F2 have an adjacency relation dened by the
code M:P:D, this code is looked up in Table 4. It appears 6 times. Thus
indicates that F1 should be among these six landmarks AT, ME, MP,
LP, MC, and LC. Now, the second code L:A is searched in this set of
six. The code is found with the landmarks MP and ME. This implies
that, F1 should be MP and F3 should be ME. There is a possibility that
no code in lookup table matches with the code we are looking for.
In such a case, the lowest value among three directions (ML, AP, and
DPx) is dropped and the process is repeated. If a wrong landmark
had been identied, it will be rectied in a later stage, when we
identify other landmarks (same landmark may be identied more
than once). Now, the identied anatomical landmark MP leads to a
set of temporary candidates for the remaining unknown landmarks.

Table 8
Computed adjacency relationship matrix between proximal tibia landmarks.

Shape T1 T2 T3 T4 T5 T6 T7
Shape Peak Peak Peak Peak Peak Peak Peak

T1 Peak 0:0:0 1.7:0.7:0.8 62.9:0.5:1.3 10.1:2.4:2.4 6.3:21.5:14.3 10.6:36.8:40.5 4.2:24.7:22.4

T2 Peak 0:0:0 61.2:1.3:0.5 11.9:3.1:1.6 8.1:20.8:13.5 8.8:37.5:39.7 5.9:25.4:21.6
T3 Peak 0:0:0 73.1:1.8:1.1 69.3:22.1:13 52.4:36.3:39.2 67.1:24.1:21.1
T4 Peak 0:0:0 3.8:23.9:11.9 20.7:34.5:38.1 5.9:22.3:20
T5 Peak 0:0:0 16.9:58.4:26.2 2.2:46.2:8.1
T6 Peak 0:0:0 14.8:12.2:18.1
T7 Peak 0:0:0
366 K. Subburaj et al. / Computerized Medical Imaging and Graphics 33 (2009) 359368

Table 9
Location and inter examiner variability of landmark MP in manual method.

Direction Model no. Gender Age Examiner Mean position (mm) Std. deviation (mm) Mean deviation (mm)

A (mm) B (mm) C (mm)

1 M 17 80.99 84.83 88.24 84.69 3.63 3.46

ML 2 M 20 92.89 90.01 96.46 93.12 3.23
3 M 26 94.99 98.37 102.2 98.46 3.52

1 M 17 39.92 45.03 50.04 45.00 5.06 5.10

AP 2 M 20 59.08 62.07 69.23 63.46 5.22
3 M 26 67.33 60.98 70.90 66.40 5.02

1 M 17 70.29 74.36 79.02 74.56 4.37 4.38

DPx 2 M 20 91.47 82.73 87.01 87.07 4.37
3 M 26 92.43 88.09 96.92 92.48 4.42

Table 10
Inter examiner variability in location of all landmarks in manual method.

Distal femur landmarks Proximal tibia landmarks

ML (mm) AP (mm) DPx (mm) Mean (mm) ML (mm) AP (mm) DPx (mm) Mean (mm)

MP 3.46 5.10 4.38 4.31 TT 3.23 1.69 5.42 3.45

LP 2.98 3.51 3.91 3.47 MP 6.14 6.38 5.42 5.98
ME 3.67 6.59 4.18 4.81 LP 5.14 5.07 5.91 5.37
LE 2.67 4.13 4.91 3.90 GT 4.72 3.43 3.17 3.77
AM 5.61 2.94 5.92 4.82 HF 2.37 3.83 2.54 2.91
MC 4.07 4.59 3.08 3.91 LIT 2.37 2.01 2.68 2.35
LC 2.34 3.30 4.91 3.52 MIT 2.32 2.19 1.95 2.15

Table 11
Deviation in localising landmark MP by automatic method with respect to mean location from manual method.

Direction Model no. Gender Age Location (automatic) (mm) Location (manual) (mm) Deviation (mm) Mean (mm)

1 M 17 82.65 84.69 2.04 2.42

ML 2 M 20 95.81 93.12 2.69
3 M 26 100.98 98.46 2.52

1 M 17 49.31 45.00 4.31 4.15

AP 2 M 20 58.67 63.46 4.79
3 M 26 63.05 66.40 3.35

1 M 17 77.39 74.56 2.83 3.04

DPx 2 M 20 83.96 87.07 3.11
3 M 26 89.31 92.48 3.17

This is a recursive process and carried out on all landmarks to ensure
the correctness of identication. Any anatomical deformation in the
bone may lead to errors in spatial location of identied landmarks,
which is reduced by recursive identication. The best set is selected
based on voting (a vote is assigned to each landmark in the list
of probable landmarks). This prevents incorrect identication of
landmarks caused by oating point and geometric errors.
Using the above procedure, on femur bone, the identied land-
marks are medial peak on anterior ridge (F1:MP), lateral peak on
anterior ridge (F2:LP), medial epicondyle eminence (F3:ME), lateral
epicondyle eminence (F4:LE), adductor magnus attachment tuber-
cle (F5:AT), most distal point of medial condyle (F6:MC), and most
distal point of lateral condyle (F7:LC). On tibia bone, the identied
anatomical landmarks are medial intercondylar tubercle (T1:MIT),
lateral intercondylar tubercle (T2:LIT), medial most point of tibial
plateau (T3:MP), lateral most point of tibial plateau (T4:LP), bula
head point (T5:HF), tibial tubercle (T6:TT), and Gerdys tubercle
(T7:GT).
Three 3D knee models reconstructed from CT scan images were
used to evaluate the effectiveness of the proposed methodology
with respect to manual method. In manual method, localisation
of anatomical landmarks was carried out by three experienced
orthopaedic surgeons. The average position of the landmarks of all
surgeons has been used as the gold standard. As an example, man-
ually located landmark MP (medial peak on anterior patella tracking
ridges) and calculated deviation from its mean position are given

Table 12
Deviation in location of all landmarks in automatic method with respect to mean location from manual method.

Distal femur Proximal tibia

ML (mm) AP (mm) DPx (mm) Mean (mm) ML (mm) AP (mm) DPx (mm) Mean (mm)

MP 2.42 4.15 3.04 3.34 TT 3.21 1.82 4.27 3.1

LP 1.74 3.04 3.15 2.64 MP 5.01 5.12 4.50 4.88
ME 3.81 6.18 3.23 4.41 LP 5.28 3.76 4.83 4.63
LE 2.35 4.02 4.53 3.63 GT 3.94 2.61 3.81 3.45
AM 3.55 2.55 4.08 3.93 HF 2.13 3.99 2.59 2.90
MC 3.79 4.58 3.61 2.99 LIT 2.04 1.93 1.84 1.94
LC 2.13 3.33 4.40 2.29 MIT 1.97 1.72 2.08 1.92
 K. Subburaj et al. / Computerized Medical Imaging and Graphics 33 (2009) 359368
Fig. 9. Comparison of variations in manual and automatic method of locating land-
marks.
in Table 9. In the same way, all other landmarks are also manually
located and deviations were calculated (Table 10). Localisation of
anatomical landmark MP by the proposed automated methodology
and its deviation from the mean position calculated from man-
ual method are given in Table 11. The deviation in location of all
landmarks by automated method is given in Table 12.
In manual localisation method, the mean deviation in location of
landmarks was found to be in the range of 2.155.98 mm (Table 10).
In automatic localisation method, the mean deviation of location
of landmarks was in the range of 1.924.88 mm (Table 12). The
comparative results are shown in Fig. 9. It is clear that the auto-
matic identication consistently performed better than manual
method for all landmarks (except in one, where both have the same
deviation). The relative performance was better on the femur side
compared to tibial side, owing to more spread-out landmarks on
tibia. It is also seen that both methods perform well for sharp fea-
tures, like tibial intercondylar tubercles (MIT and LIT). For fuzzy
features like MP and LP of tibia and ME and AM of femur, which
are not clearly identiable, the deviation is low for both methods.
In terms of variability in localising anatomical landmarks, the inter
observer variability is higher than the automated one. This indi-
cates that the automatic method localise the anatomical landmarks
closer to the gold standard coordinates.
The 3D anatomical model reconstructed from CT axial images
may have eroded surfaces and islands due to the presence of noise
and defective tissue (e.g., tumours). To reduce the error caused by
noise present in CT images and enhance the boundary pixels, a
set of 2D image lters was incorporated. This also improved the
effectiveness of segmentation and edge extraction. Surface curva-
ture estimations are sensitive to mesh quality. To increase the mesh
quality, geometric errors created during conversion of triangle sur-
face data from volumetric data are considered and corrected by a
set of designed mesh lters. Even then, reliable estimates of sur-
face curvature sign can be computed. Hence the sign of the surface
curvature is a powerful representation for categorizing surface type
without noise sensitivity. During landmarks identication, a num-
ber of probable landmark regions are segregated or ltered out by
a set of geometric criteria (surface area of the region, geometric
characteristics of the associated triangles, and spatial location). This
improves the robustness of identifying the nal location of anatom-
ical landmarks and reduces the identication of false landmarks.
4. Conclusion
We have described an automated approach for localising and
identifying landmarks on 3D anatomical models. This has been
achieved by combining surface characteristics (curvature) and
spatial adjacency of the landmarks on a particular bone. False locali-
sation of anatomical landmarks is minimized by the use of recursive
367
search and voting algorithm and spatial-adjacency relation. The
quantitative results show that they are reproducible and generally
superior or close to manual method of localisation. The auto-
mated methodology reduces the planning and registration time,
and increases repeatability. It also reduces dependency on (error-
prone) manual procedure and the risk associated with the insertion
of pin or markers for bone model registration.
The directions for further work include: (i) recognition of
anatomical surfaces for surface based registration in addition to
the anatomical landmarks, which will increase the accuracy of
registering 3D virtual models with patients bone during surgical
procedure; (ii) incorporating active shape models (manual or auto-
mated) along with the geometric reasoning algorithms to localise
and identify landmarks on patient-specic 3D anatomical models;
and (iii) including the requirement of the landmarks in the pro-
cess to minimize the landmark regions need to be segregated like
tumour resection planning, positioning and orienting of tumour
knee prosthesis after resection, etc.
Disclosure
None.
Acknowledgements
This work is a part of an ongoing project in OrthoCAD Net-
work Research Centre at Indian Institute of Technology Bombay
for developing a computer aided orthopaedic implant design and
surgery planning system in collaboration with Tata Memorial Hospi-
tal, Mumbai. It is supported by the Ofce of the Principal Scientic
Adviser to the Government of India, New Delhi.
References
[1] Wong KC, Kumta SM, Chiu KH, Cheung KW, Leung KS, Unwin P, et al. Com-
puter assisted pelvic tumour resection and reconstruction with a custom-made
prosthesis using an innovative adaptation and its validation. Computer Aided
Surgery 2007;12(4):22532.
[2] Viceconti M, Testi D, Simeoni M, Zannoni C. An automated method to position
prosthetic components within multiple anatomical spaces. Computer Methods
and Programs in Biomedicine 2003;70(2):1217.
[3] Grifn FM, Math K, Scuderi GR, Insal JN, Poilvache PL. Anatomy of the
epicondyles of the distal femur MRI analysis of normal knees. Journal of Arthro-
plasty 2000;15(3):3549.
[4] Robinson M, Eckhoff DG, Reinig KD, Bagur MM, Bach JM. Variability of landmark
identication in total knee arthroplasty. Clinical Orthopaedic Related Research
2006;442:5762.
[5] Morton NA, Maletsky LP, Pal S, Laz PJ. Effect of variability in anatomical land-
mark location on knee kinematic description. Journal of Orthopaedic Research
2007;25(9):122130.
[6] Stagni R, Fantozzi S, Cappello A. Propagation of anatomical landmark misplace-
ment to knee kinematics: performance of single and double calibration. Gait
and Posture 2006;24(2):13741.
[7] Cappozzo A, Catani F, Della Croce U, Leardinis A. Position and orientation in
space of bones during movement: anatomical frame denition and determina-
tion. Clinical Biomechanics 1995;10(4):1718.
[8] Liu X, Kim W, Drerup B. Foot 3D characterization and localization of anatomical
landmarks of the foot by FASTscan. Real-Time Imaging 2004;10(4):21728.
[9] Agarwal MG, Anchan C, Shah M, Puri A, Pai S. Limb salvage surgery for osteosar-
coma: effective low-cost treatment. Clinical Orthopaedics and Related Research
2007;459:8291.
[10] Apaydin N, Uz A, Bozkurt M, Elhan A. The anatomical relationships of the aux-
iliary nerve and surgical landmarks for its localization from the anterior aspect
of the shoulder. Clinical Anatomy 2007;20(3):2737.
[11] Frangi AF, Rueckert D, Schnabel JA, Niessen WJ. Automatic construction of
multiple-object three-dimensional statistical shape models: application to car-
diac modelling. IEEE Transactions on Medical Imaging 2002;21(9):115166.
[12] Worz S, Rohr K. Localization of anatomical point landmarks in 3D medical
images by tting 3D parametric intensity models. Medical Image Analysis
2006;10(1):4158.
[13] Yahara H, Higuma N, Fukui Y, Nishihara S, Mochimaru M, Kouchi M. Estimation
of anatomical landmark positions from model of 3-dimensional foot by the FFD
method. Systems and Computers in Japan 2005;36(6):2638.
[14] Van Sint Jan S. Colour atlas of skeletal landmark denitions: guidelines for
reproducible manual and virtual palpations. Churchill Livingstone/Elsevier;
2007.
368 K. Subburaj et al. / Computerized Medical Imaging and Graphics 33 (2009) 359368
[15] Della Croce U, Cappozzo A, Kerrigan DC. Pelvis and lower limb anatomical
landmark calibration precision and its propagation to bone geometry and
joint angles. Medical and Biological Engineering and Computing 1999;37(1):
15561.
[16] Yang J, Ling X, Lu Y, Wei M, Ding G. Cephalometric image analysis and mea-
surement for orthognathic surgery. Medical and Biological Engineering and
Computing 2001;39(3):27984.
[17] Maudgil DD, Free SL, Sisodiya SM, Lemieux L, Woermann FG, Fish DR, et al.
Identifying homologous anatomical landmarks on reconstructed magnetic res-
onance images of the human cerebral cortical surface. Journal of Anatomy
1998;193(4):55971.
[18] Koenderink J, Dooen AV. Surface shape and curvature scales. Image and Vision
Computing 1992;10(8):55765.
[19] Maekawa T, Wolter EE, Patrikalakis NM. Umbilics and lines of curvature for
shape interrogation. Computer Aided Geometric Design 1996;13(2):13361.
[20] Page DL, Sun Y, Koschan AF, Paik J, Abidi MA. Normal vector voting: crease
detection and curvature estimation on large, noisy meshes. Graphical Models
2002;64(3-4):199229.
[21] Rohr K. Extraction of 3D anatomical point landmarks based on invariance prin-
ciples. Pattern Recognition 1999;32(1):315.
[22] Do Carmo MP. Differential geometry of curves and surfaces. New Jersey: Pren-
tice Hall; 1976.
[23] Meyer M, Desbrun M, Schrder P, Barr AH. Discrete differential-geometry oper-
ators for triangulated 2-manifolds. Visualization and Mathematics 2002:3557.
[24] Besl PJ, Jain RC. Segmentation through variable-order surface tting. IEEE Pat-
tern Analysis and Machine Intelligence 1988;10(2):16792.
[25] Subburaj K, Ravi B. High resolution medical models and geometric reasoning
starting from CT/MR images. In: Wang G, Li H, Zha H, Zhou B, editors. Pro-
ceedings of 10th IEEE International Conference on Computer Aided Design and
Computer Graphics. Beijing: IEEE Computer Society; 2007. p. 4414.
[26] Nagy MS, Matyasi GY. Analysis of STL les. Mathematical and Computer Mod-
elling 2003;38:94560.
[27] Drerup B, Hierholzer E. Automatic localization of anatomical landmarks on the
back surface and construction of a body-xed coordinate system. Journal of
Biomechanics 1987;20(10):96170.
[28] Ehrhardt J, Handels H, Plotz W, Poppl SJ. Atlas-based recognition of anatomi-
cal structures and landmarks and the automatic computation of orthopaedic
parameters to support the virtual planning of hip operations. Methods of Infor-
mation in Medicine 2004;43(4):3917.
[29] Lorenz C, Krahnstover N. Generation of point-based 3D statistical shape
models for anatomical objects. Computer Vision and Image Understanding
2000;77(2):17591.
K. Subburaj is a PhD scholar in the Department of Mechanical Engineering, Indian
Institute of Technology Bombay, India and attached to the OrthoCAD Network
Research Centre. He received his M.Eng from the M.S. University of Baroda, India in
2005. His current research interests are BioCAD, 3D Geometric Reasoning, Surgery
Planning, and Rapid Prototyping.
Dr. B. Ravi is a professor of Mechanical Engineering at IIT Bombay and Founder-
coordinator of OrthoCAD Network Research Centre. He completed his engineering
degree from National Institute of Technology, Rourkela in 1986, followed by mas-
ters and PhD from Indian Institute of Science, Bangalore in 1992. His areas of
research include casting design and simulation, product lifecycle engineering, and
Bio-CAD/CAM. He has published about 160 technical papers, and given over 100
invited talks. He is a reviewer for ASME, IEEE, IJAMS, and IJPR. He is a member
of ASME Bio Manufacturing Committee, and Fellow of the Institution of Engineers
(India).
Dr. Manish Agarwal is an orthopaedic oncologist and associate professor at Tata
Memorial Hospital, Mumbai. He completed his MBBS and MS (Ortho) from Uni-
versity of Bombay in 1987 and 1992 respectively. His main interest is limb salvage
surgery for bone and soft tissue tumours. He is a member of the musculoskeletal
tumor society of North America and has several publications in peer reviewed jour-
nals. His current research includes development of an indigenous tumor prosthetic
system as well as clinical trials with herbs like Curcumin and ashwagandha for bone
sarcomas.