Automatic Recognition of Pertrochanteric Bone Fractures in Femur

using Level Sets*

Ling Wang, Hong Cheng, Hai Lan, Yingjie Zheng, Kainan Li

Abstract Femur fracture is one of the common diseases

in the elderly. In this paper, an automatic and efficient
classification method of femoral pertrochanteric fractures is
proposed based on image segmentation techniques. The types
of femoral pertrochanteric fracture are defined firstly according !"#!$%"&&'()
to the difference of fracture parts. To reduce the computational
complexity, only four directions images are used in each
3D femur, i.e. the anterior, the anterolateral, the posterior *+,)"&-.")+"(/,/'$(
and the posterolateral images of a femur. And then those
fracture images are segmented from background using level set
method. Considering the numerical errors and the instability
of evolution in conventional level set formulations, a distance 0"1/2!"&-3"/"%/'$(
regularization term and an external energy constraint term are "!/!$%8,(/"!'%--
used, which are able to maintain a desired shape of the level set 9!,%/2!"-:$;"5
function and speed up the motion of the zero level contour in *+,)"&- ,!/'/'$(
evolution process. After segmentation, the Canny edge detector
is used to extract the fracture edges on femur. The types of
femoral pertrochanteric fracture are classified by comparing
the difference of edges between reference normal images and the 45,&&'6'%,/'$(-7-3',)($&'&
testing images. Except the given four directions of 3D image, the
classification is processed automatically. Experimental results Fig. 1: The framework of pertrochanteric fracture automatic
illustrated the effectiveness of the proposed method. detection.

I. I NTRODUCTION
The incidence of femur fractures increased significantly
among people over 60 years old. Most of these orthopedic
and trauma surgical interventions need careful classification
in order to decide the position and type of implant to be
used intraoperative. Automatic recognition and classification
of femur fractures plays an important role in quantitative
analysis.
For the fracture clinical diagnosis is widely based on
computer-aided images, such as X-ray, CT, ultrasound and
MRI images, many investigations on fracture detection and
diagnosis are based on image processing. Morphological
gradient based image segmentation is used in [1] to detect
radius bone fracture. Fracture detection is performed by a
registered active shape model in [2], where the pelvic CT
images are preprocessed by adaptive windowing, bound-
ary tracing and wavelet transform. Stacked random forests
feature fusion is used in [3] to detect fracture in X-ray
images. After edge and shape features are extracted from
bone, multiple classifiers, such as Back Propagation Neural
Network, K-Nearest Neightbour, Support Vector Machine,
*Research supported by the grant from National Natural Science
Foundation of China (NSFC) ( No. 61273256, No. 61305033 and No.
U1233103F01).
L. Wang and H. Cheng are with Center for Robotics, University of
Electronic Science and Technology of China, Chengdu, China 611731.
(email: eewangling@uestc.edu.cn, hcheng@uestc.edu.cn).
H. Lan and Y. Zheng and K. Li are with Affiliated Hospital of
Chengdu University, Chengdu, China 610106. (email: 13714777@qq.com,
183729860@qq.com, Likainan1961@126.com).
Max/Min Rule, Product Rule, are fused to design as com-
bining classifier [4], [5]. Fracture is detected by computing
the pixels value of each images column or row in [6], where
a threshold is set, there would be a fracture if the value is
less than the threshold.
In the above mentioned investigations, most of them either
only detect the fractures or design a general classifier, such
as traverse fracture, open fracture, simple fracture and com-
minuted fracture. In this paper, a subtle pertrochanteric frac-
ture classifier is proposed, which can distinguish five types
of pertrochanteric fracture. The classification model is shown
in Fig. 1, the original CT image is denoised and smoothed
by Gaussian filter, and then the distance regularization level
set method is used to segment the region of interest (ROI).
After that, the Canny edge detection is used to detect the
construction of femur and then the image is divided into
several parts according to the definition of pertrochanteric
fracture.
For there usually are noises in the segmentation, filters
are used to remove the smaller noise regions and edges
by mistaken segmentation as shown in Fig. 2. A continues
region would be considered as a outliers if its size is less
than 10 10, and then this region would be moved out. A
continues line would be moved out if its length is less than
one forth of the width of each image. The pertrochanteric
fracture is classified by comparing the difference of each
part with reference normal images.

978-1-4577-0220-4/16/$31.00 2016 IEEE 3851

 !" #" $" (a) Two parts fracture (b) Three parts fracture 1

(c) Three parts fracture 2 (d) Four parts fracture

%" '" &"

Fig. 2: Image denoised. (a) is the original image, (b) is

segmented image by level set, (c) is the region of interest.
(d) is the extracted edges by Canny detector, (e) is the edges
in yellow rectangle region of (d), (f) is the lines of interest
(e) Five parts fracture 1 (f) Five parts fracture 2
in (e).

II. P ERTROCHANTERIC F RACTURE M ODEL

For the diagnosis of fracture, the types of femoral
pertrochanteric fractures are defined as five classes [7]: the
two parts, three parts, four parts, five parts and six parts
fractures. In two parts fracture the fracture line is cross the
pertrochanteric, as shown in Fig. 3a, it is between in greater
trochanter and lesser trochanter. The three parts fracture is
defined as follows. There are fractures in greater trochanter
or lesser trochanter go with two parts fracture, as shown
in Fig. 3b and Fig. 3c. In four parts fracture, there are
greater and lesser trochanter fractures go with two parts
fracture, as shown in Fig. 3d. In five parts fracture, there
is anterior or posterior lateral fracture go with four parts
fracture, as shown in Fig. 3e and Fig. 3f. In six parts
fracture, there is anterior and posterior lateral fractures go
with four parts fracture, as shown in Fig. 3g. According to
this definition, only four images of each femur are used in
following discussion, i.e. the anterior, the anterolateral, the
posterior and the posterolateral images of a femur.
III. I MAGE S EGMENTATION
Level set method is a tool for numerical analysis of shapes,
which is used to image segmentation from background
widely. In two dimensions image, the level set method
aims to representing a closed curve by using a level set
auxiliary function . is represented as the zero of by
= (x, y)|(x, y) = 0. If the curve moves in the normal
direction with a speed v, then the level set function satisfies
the level set equation
= v||, (1)
t R() is
where | | is the Euclidean norm, t is time.
In the level set formulations, the level set function maybe
develops irregularities during its evolution, which may cause
numerical errors and eventually destroy the stability of evo-
lution. To overcome this difficulty, in [8], Li et.al proposed
a variation level set formulation to regularize the level
set function during the evolution. The level set evolution
(g) Six parts fracture
Fig. 3: Definition of Pertrochanteric Fracture.
is derived as the gradient flow that minimizes an energy
functional. The energy function is consisted of a distance
regularization term and an external energy that drives the
motion of the zero level set toward desired locations.
Let be a level set function defined in a domain , an
energy function E() is define as
E() = R() + Eext (), (2)
where > 0 is a constant, R() is the level set regularization
term and Eext () is an external energy that depends on the
data of interest.
The distance regularization term R() is defined with a
potential function p(x) as R() = p()dx. It makes
level set evolution has a unique forward-and-backward dif-
fusion effect and maintain a desired shape of the level
set function, particularly a signed distance profile near the
zero level set. The distance regularization effect eliminates
the need for reinitialization and thereby avoids its induced
numerical errors. In order to maintain the signed distance
property = 1 only in a vicinity of the zero level set,
while keeping the as a constant, with = 0, at locations
far away from the zero level set. The p(x) is a double-well
function, which has two minimum values at x = 1 and x = 0.
It is defined as p(x) = (x 1)2 /2. Then the derivative of
R
= div(dp (||)), (3)
here div() is the divergence operator and dp (x) = dp(x)
xdx .
The external energy function
Eext () = gH()dx + g()dx, (4)

3852
here R and > 0 are coefficients, g is an edge indicator
1
function of image I, which defined as g = (1+GI 2 ) and G
is a Gaussian kernel with standard deviation . The H and
are Heaviside function and Dirac delta function, which are
introduced to speed up the motion of the zero level contour
in the evolution process and minimize the energy when the
zero level contour is located at the object boundaries. In
practice, the and H are approximated by the following
smooth functions,
{ 1 [ ]
2 1 + cos( x
) , x6
(x) = (5)
0, x>
[ ]
) , x6
1 x 1
2 1+ + sin( x
H(x) = 1, x> (6)
0, x <
The standard method to minimize the energy function
E() is to find the steady state solution of the gradient flow
equation
E black rectangles are regions of posterior lateral femur. The
= = div(dp (||)) + g() second row is the images of a trochanter fracture, the third
t
+()div(g ). (7)
||
Given an initial 0 (x), the Eqn. (7) can be solved by an iter-
ative narrow-band implementation. The readers are referred
to [8] for more details. The 2D images are segmented from
background by using this level set function. In this paper,
the parameters of the segmentation model are set as follows.
The = 0.2/t, = 10, = 5, = 0.8, in Eqn. (7), (5) number of edges in skeleton, green block, red block, yellow
and (6), where t = 5 is the time step of implementation,
the = 5 is parameter of Gaussian kernel G.
IV. E DGE E XTRACTION
For the fracture diagnosis is based on the information of
edges on bone, the edge detector is used to extract the struc-
tural information from the segmented images. Among the
presented edge detection methods, the Canny edge detection
algorithm [9] is one of the most strictly defined methods that
provides good and reliable detection. Owing to its optimality
to meet with the criteria for edge detection and the simplicity
of process for implementation, it is used to extract the edges
on femoral trochanter fracture. The image is smoothed by
Gaussian filter and two threshold values [t1, t2] are used to
determine potential edges. In the following experiment, the
threshold of Canny edge detector is set as [0.2, 0.35].
A demo results of image segmentation and edge extraction
are shown in Fig. 4, where (a) is the original image, (b) is
the segmented image, (c) is the edges of image, (d) is the
skeletons of image and (e) is the edges without skeleton. It
is noted that the line in (d) is considered as a skeleton if it
is continues with external contour. Fig. 4 illustrated that the
algorithms of segmentation and edge extraction play good
performance.
!" #" $" %" &"
Fig. 4: Image segmentation and edge extraction. (a) The
original image. (b) The segmented image. (c) The edges of
segmented image. (d) The skeletons of image. (e) The edges
in bone.
V. D IAGNOSIS
According to the definition of the five fracture classes,
the images are divided to several parts as intertrochanteric,
greater trochanter section, lesser trochanter section, anterior
and posterior lateral femur section. As shown in Fig. 5, the
first row is the images of a reference normal femur in four
directions, where the green rectangles are regions of greater
trochanter, the red rectangles are regions of lesser trochanter,
the yellow rectangles are regions of anterior lateral femur, the
and forth rows are extracted edges from each part where
the skeletons are subtracted. Then the fracture is classified
by comparing the difference between the test image and
reference image among the mentioned sections.
Let I1 I4 are the four direction images of each patient,
H2 H6 are the hypothesis of two parts fracture to six parts
fracture as defined in section II, F1 F4 are the above men-
tioned four sections fracture, and N, Ng , Nr , Ny , Nk are the
block and black block, respectively. Then the diagnosis is
formulated as hypothesis testing as follows.
P (F1 = 1) = P (Ng > 1|I1 ) P (Ng > 1|I4 )
P (F2 = 1) = P (Nr > 1|I1 ) P (Nr > 1|I2 ) P (Nr > 1|I3 )
P (F3 = 1) = P (Ny > 1|I1 ) P (Ny > 1|I2 ) P (Ny > 1|I4 )
P (F4 = 1) = P (Nk > 1|I3 )
P (H2 = 1) = P (N > 1|I1 ) P (N > 1|I3 )
P (H4 = 1) = P (H2 = 1) P (F1 = 1) P (F2 = 1)
P (H3 = 1) = P (H4 = 0) (P (F1 = 1) P (F2 = 1))
P (H6 = 1) = P (H4 = 1) P (F3 = 1) P (F4 = 1)
P (H5 = 1) = P (H6 = 0) (P (F3 = 1) P (F4 = 1))
VI. E XPERIMENTAL R ESULTS
This section shows the results of the proposed classi-
fication model for real clinical images. 38 patients with
pertrochanteric fracture were collected, where there are
9, 6, 7, 14, 2 people with two parts to six parts fracture,
respectively. Some samples are shown in Fig. 6. The original
reconstructed 3D images from CT are 1024 1024 pixels,
which are down-sampled to 256256 pixels. It is noted that,
the size of segmented images is different from each other for
the difference of viewpoint and difference of patient.
3853
Fig. 5: The partials of images. The first row is the reference images of a normal femur, the second row is the test images
of a trochanter fracture, the third and the forth rows are the extracted edges from each part of the test images.
Fig. 6: Five classes fracture images. From the first image to
the last image, it is two parts to six parts fracture respectively.
TABLE I: The recognition rates
Classes Two Three Four Five Six
Data Number 9 6 7 14 2 pp. 801805.
Recognition Rate 8/9 5/6 5/7 12/14 2/2
The recognition rates of each classes are shown in Table
I. The average recognition rate is 84.2%. There are some
estimation error for the greater or lesser trochanter fracture
are not clearly even diagnosed by human eyes. For the
definition of the five types femoral pertrochanteric fractures
is a novel subject, there is fewer methods to compare.
VII. C ONCLUSION
This paper explore the pertrochanteric fracture classifica-
tion in 3D CT-images. A distance regularization level set
method is used to segment ROI from noisy background,
Canny edge detect operator is used to extract fracture edges,
and then by comparing the difference of edges between
reference normal images and testing images, the fracture is
classified according to the definition of pertrochanteric frac-
ture. Compared with existed works, the proposed method can
classify subtle fractures on pertrochanteric. The experimental
results illustrated the good performance of this method. In
future work, we will discuss the sparse feature representation
and supervised machine learning method on large data sets.
3854
R EFERENCES
[1] B. Swathika, Anandhanarayanan.K, Baskaran.B, and Govindaraj.R,
Radius bone fracture detection using morphological gradient based
image segmentation technique, Int. J. of Comput. Sci. and Inform.
Tech., vol. 6, no. 2, pp. 110, Feb 2015.
[2] J. Wu, P. Davuluri, K. R. Ward, C. Cockrell, R. Hobson, and K. Na-
jarian, Fracture detection in traumatic pelvic CT images, Int. J. of
Biomed. Imag., 2012.
[3] Y. Cao, H. Wang, M. Moradi, P. Prasanna, and T. Syeda-Mahmood,
Fracture detection in x-ray images through stacked random forests
feature fusion, in IEEE Int. Symp. Biomed.l Imag. (ISBI), April 2015,
[4] N. Umadevi and S. Geethalakshmi, Multiple classification system for
fracture detection in human bone x-ray images, in Int. Conf. Comput.
Commun. Net.g Technol. (ICCCNT), July 2012, pp. 18.
[5] V. L. F. Lum, W. K. Leow, Y. Chen, T. S. Howe, and M. A. Png,
Combining classifiers for bone fracture detection in x-ray images, in
IEEE Int. Conf. Image Process., vol. 1, Sept 2005, pp. 11491152.
[6] L. Nascimento and M. Graca Ruano, Computer-aided bone fracture
identification based on ultrasound images, in IEEE Portuguese Meeting
on Bioeng. (ENBENG), Feb 2015, pp. 16.
[7] Z. Chen, K. Li, and Z. Zhang, A finite element analysis of six-segment
classification of femur intertrochanteric fracture, China J. Orthop.
Trauma., vol. 17, no. 5, pp. 433437, May 2015.
[8] C. Li, C. Xu, C. Gui, and M. D. Fox, Distance regularized level
set evolution and its application to image segmentation, IEEE Trans.
Image Process., 2010.
[9] J. Canny, A computational approach to edge detection, IEEE Trans.
Pattern Anal. Mach. Intell., vol. PAMI-8, no. 6, pp. 679698, Nov 1986.