International Journal of Trend in Scientific Research and Development (IJTSRD)

Volume 3 Issue 5, August 2019 Available Online: www.ijtsrd.com e-ISSN: 2456 – 6470

Implementation of Lower Leg Bone

Fracture Detection from X-Ray Images
San Myint, Khaing Thinzar
Technological University, Mandalay, Myanmar

How to cite this paper: San Myint | ABSTRACT

Khaing Thinzar "Implementation of Lower A methodology and various techniques are presented for development of
Leg Bone Fracture Detection from X-Ray fracture detection system using digital image processing. This paper presents
Images" Published in International an implementation of bone fracture detection using medical X-ray image. The
Journal of Trend in goal of this paper is to generate a quick and diagnosis can save time, effort and
Scientific Research cost as well as reduce errors. The main objective of this research is to classify
and Development the lower leg bone fracture using X-ray image because this type of bone are the
(ijtsrd), ISSN: 2456- most commonly occur in bone. This paper classifies the two types of fracture
6470, Volume-3 | (Non-fracture and fracture or Transverse fracture) by using SVM classifier.
Issue-5, August The proposed system has basically five steps, namely, image acquisition, image
2019, pp.2411- IJTSRD27957 pre-processing, image segmentation, feature extraction and image
2415, classification.
https://doi.org/10.31142/ijtsrd27957
KEYWORDS: X-Ray images, Lower Leg bone, Image processing, Fracture type,
Copyright © 2019 by author(s) and SVM Classifier
International Journal of Trend in Scientific 1. INTRODUCTION
Research and Development Journal. This Today, medical image processing is a field of science that is gaining wide
is an Open Access article distributed acceptance in healthcare industry due to its technological advances and
under the terms of software breakthroughs. It plays a vital role in disease diagnosis and improved
the Creative patient care and helps medical practitioners during decision making with
Commons Attribution regard to the type of treatment. Among the various diseases, bone fracture
License (CC BY 4.0) detection and treatment, which affects many people of all ages, is growing
(http://creativecommons.org/licenses/by important in modern society [12Mah].
/4.0)
Computer vision system can help to screen X-ray images for
suspicious cases and alarm the doctors. Depending on the
experts alone for such a critical matter has caused
intolerable errors and hence, the idea of automatic diagnosis
procedure has always been an appealing one.

There are many techniques to detect the bone fracture, such

as Computed Tomography (CT), Magnetic resonance imaging
(MRI), Ultrasound and X-ray which help physicians in
detecting different types of abnormalities. Fracture
diagnosis and treatment, X-ray images are most frequently
used in fracture diagnosis because it is the fastest and easiest
way for the doctors to study the injuries of bones and joints.
Doctors usually use x-ray images to determine whether a
fracture exists. In the recovery process, doctors also use x-
ray images to determine whether the injured bones and
joints have recovered [06Din]. Although CT and MRI images
gives better quality images for body organs than X-ray Figure 1.1: Classification of Bone by Shape [16Wik]
images, X-ray images are faster and cheaper, enjoy wider
availability and are easier to use few limitations. Moreover, 2. Overview of proposed system
the levels of quality of X-ray images are enough for the A. Bone Fracture Detection Using Image Processing in
purpose of bone fracture detection [13Mah]. The system Matlab
proposed in this thesis classifies the existence of fractures in The first paper done by Ms. Snehal Deshmukh, et al. [15Sen],
lower leg bones based only on X-ray images. The human they discussed that several digital image processing
body contains 206 bones with various shapes and structures. techniques applied in edge feature extraction for bone
The largest bones are the femur bones, and the smallest fracture detection. In this paper, firstly, linear filtering is
bones are the auditory ossicles. There are basically five types used to remove noises from the collected images. Secondly,
of bone in human body, such as long, short, flat, irregular and some edge detection operators such as Sobel, Log edge
seasmoid [13mah]. Fig.1 shows the anatomy and shape of detection and Canny edge detection are analyzed and then
the human body. according to the simulation results, the advantages and

@ IJTSRD | Unique Paper ID – IJTSRD27957 | Volume – 3 | Issue – 5 | July - August 2019 Page 2411
 International Journal of Trend in Scientific Research and Development (IJTSRD) @ www.ijtsrd.com eISSN: 2456-6470
disadvantages of these edge detection operators are feature extraction and image classification. The first step, the
compared. It is shown that the Canny operator can obtain input image is obtained from different places and the image
better edge feature and it can also be produced equally good is filtered out by the Gaussian filter. Then, edge detection
edge with the smooth continuous pixels and thin edge. Sobel method is used to detect the bone structure. And then the
edge detection method cannot produce smooth and thin Harris algorithm also used for extracting point features from
edge compared to Canny method. But, the same like other the image. For the classification step, support vector
method, Sobel and Canny methods are very sensitive to the machine (SVM) is used to classify fracture types.
noise pixels. Sometimes, all of the noisy image cannot be
filtered perfectly. A. Image Acquisition
Image acquisition in image processing can be broadly
B. Algorithm and Technique on Various Edge defined as the action of retrieving an image from some
Detection source, usually a hardware-based source, so it can be passed
The next paper, Rashmi,Mukesh Kumar, and through whatever processes need to occur afterward.
RohiniSaxena[13Ras] proposed a survey of various edge Performing image acquisition in image processing is always
detection techniques as Prewitt, Robert, Sobel, Marr Hildrith the first step in the workflow sequence because, without an
and Canny operators. In this paper, the authors have studied image, no processing is possible. In this step, medical images
and evaluate different edge detection techniques. They are given as input and all types of medical images can be
describe that Canny edge detector gives better result as acquired.
compared to other edge detectors on various aspects such as
it is adaptive in nature, performs better for noisy image, give
sharp edges, low probability of detecting false edges, etc. The
authors stated that Canny method is optimal edge detection
technique hence lot of work and improvement on this
algorithm has been done and it can detect edges in color
image without converting in gray image.

C. Edge Detection of Femur Bone in X-Ray Images – A

Comprartive Study of Edge Detectors
Shubhangi D.C, et al. [12Shu] analyzed the performance of
Laplace operator in comparison with other edge detection
methods, namely, Roberts, Sobel, Prewitt, and Canny’s
operators, which are applied to the X-ray images of femur Figure 3.1: Samples of Normal Lower Leg Bone X-Ray
bones. They observed that Robert cross gradient operator is Images
very quick to compute. The resultant image is very similar to
the one obtained by Sobel operator but quality of edge pixels
are found to be degraded due to lot of jerky effect on edges.
The Sobel operator is slower than the Roberts operator and
its performance is poor than the Prewitt operator.From the
experimental results, they observed that the Laplace
operator gives better edge detection results than the other
methods in the investigation of X-ray images of femur bones,
which has significance to medical and forensic experts.

D. Determining the Type of Long Bone Fractures in X-

Ray Images
Mahmoud Al-Ayyoub and Duha Al-Zghool [13Mah] . Figure 3.2: Samples of Fractured Lower Leg Bone X-Ray
considered the problem of determining the fracture type Images
from X-ray images. This paper started with the
preprocessing and noise removal by using histogram B. Image Pre-processing
equalization process. After enhancing the x-ray images and Image pre-processing is the term for operations on images at
removing noise, it then discussed various tools to extract the lowest level of abstraction. These operations do not
useful and distinguishing features based on: edge detection, increase image information content but they decrease it if
corner detection, parallel & fracture lines, texture features entropy is an information measure. Image pre-processing
and peak detection. The next step is image classification uses the redundancy in images. Neighbouring pixels
based on the extracted features to predict fracture types. In corresponding to one real object have the same or similar
this stage, Support Vector Machine (SVM), Decision Tree brightness value. If a distorted pixel can be picked out from
(DT), Navie Bayesian (NB) and Neural Network (NN) were the image, it can be reported as an average value of
tested to classify fracture types. According to the neighbouring pixels. Image pre-processing methods can be
experimental results, SVM classifier was found to be the classified into categories according to the size of the pixel
most accurate with more than 85% accuracy under the 10- neighbourhood that is used for the calculation of new pixel
fold cross validation technique. brightness. The aim of pre-processing is an improvement of
the image data that suppresses undesired distortions or
3. System flow enhances some image features relevant for further
The procedures needed for this system are image processing and analysis task [06Olg].
acquisition, image pre-processing, image segmentation,

@ IJTSRD | Unique Paper ID – IJTSRD27957 | Volume – 3 | Issue – 5 | July - August 2019 Page 2412
 International Journal of Trend in Scientific Research and Development (IJTSRD) @ www.ijtsrd.com eISSN: 2456-6470
In pre-processing step, noise suppression is the most
important task. Noise reduction is an important and basic
part in remote sense image processing. Not only spatial noise
but also spectral noise may exist in the image because of the
influence of natural light, surface topography, mixed pixel,
etc. [08Chi]. Noise is the result of errors in the image
acquisition process that results in pixel values that do not
reflect the true intensities of the real scene. Most commonly
used de-noising techniques are described as follows:
 Median filter
 Average filter
 Gaussian filter
 Winner filter
 Histogram equalization (a) (b)

Among these filters, Gaussian filter is used for noise

reduction in this thesis.

C. Image Segmentation
Image segmentation is a very important step in image
analysis and performance evaluation of processed image
data. The goal of image segmentation is to make simple or
change the representation of an image into something that is
more meaningful and easier to analyze. Image segmentation
is typically used to locate objects and boundaries (lines,
curves, etc.) in image. Image segmentation is the process of
partitioning an image into non-intersecting regions such that
each region is homogeneous and the union of no two
adjacent regions is homogeneous. Segmentation is typically (c) (d)
associated with pattern recognition problems. It is Figure 3.3: Result Images for Edge Detectors
considered the first phase of a pattern recognition process
and in sometimes also referred to as object isolation. D. Feature Extraction
According to the expression of Ashutosh Kumar Chaubey Feature extraction is the main step in various image
[16Ash], segmentation algorithms have been used for a processing applications. A feature is a significant piece of
variety of applications. In this thesis, edge-based information extracted from an image which provides more
segmentation method is used in order to segment region of detail understanding of the image. The commonly used of
bone area from background. feature extraction methods are Gray Level Co-occurrence
Matrix(GLCM), Wavelet transform, Curvelet transform,
Edge-based Method Hough transform and Harris corner detection. Harris corner
Edge-based segmentation exploits spatial information by detection techniques is used in this paper.
detecting the edges in image. Edges correspond to
discontinuities in the homogeneity criterion for segments. Harris Algorithm
Edge-based image segmentation algorithms are sensitive to Harris algorithm detects all corners (in general, intersection
noise and tend to find edges that are irrelevant to the real point) or most true interest points based on the brightness
boundary of the object. Moreover, the edges extracted by of images. In this thesis, Harris corner detection is used to
edge-based algorithms are disjointed and cannot completely detect feature points. Harris features are shape and vector of
represent the boundary of an object. Therefore, the edge- the bone image. These features are used for fracture
linking algorithms are applied to create enclose boundaries. classification using SVM classifier which is a representation
Edge-based segmentation algorithms use edge detectors of the example as points in a feature space. The point which
(Sobel, Prewitt, Robert, Canny, Laplacian) to find edge in the is the direction of the boundary of object changes abruptly.
image. In this research, a comparison of the four types of The Harris detector uses the correlation matrix as the basis
edge detector is tested. Then, Canny edge detection is used of its corner decisions. Classification of image points using
to find discontinuities in bone image. Canny detector eigenvalues of matrix (M). The correlation matrix or second
produces good view of bone structure and gives many moment matrix can be represented as follows:
advantages than other edge detectors. A comparative study
of the four edge detectors are Sobel, Prewitt, Robert and 𝐼 2 𝐼 𝐼
Canny. These result images are shown in the following figure 𝑀 =∑ , 𝑤(𝑥, 𝑦)
𝐼 𝐼 𝐼 2
3.3. In figure 3.3, (a) and (b) are the results of Sobel and
Prewitt edge detectors. Then, figure 3.3(c) and (d) are the
results of Robert edge detector and Canny edge detector. Harris defines the response function (R) which decide the
point is corner or not with the following equation

R = det M – k 𝑡𝑟𝑎𝑐𝑒 M = 𝜆 𝜆 − k 𝜆 + 𝜆 M,

@ IJTSRD | Unique Paper ID – IJTSRD27957 | Volume – 3 | Issue – 5 | July - August 2019 Page 2413
 International Journal of Trend in Scientific Research and Development (IJTSRD) @ www.ijtsrd.com eISSN: 2456-6470
R depends only on eigenvalues (λ λ ) of matrix (M). If hyperplane in more dimensions. A good separation is
λ and λ are large or 𝜆1 is similar 𝜆2, Harris can be defined as achieved by the hyperplane. There can be a lot of separating
Corner. If eigenvalues (𝜆1 and 𝜆2) are small, it is almost hyperplanes and selects as far as possible from data points of
constant in all direction (flat region) and if 𝜆1 very much both classes. The optimal hyper plane is correctly classify the
greater than 𝜆2 or 𝜆2 very much greater than 𝜆1, this can be training data.
defined as edge [13Hua] [03Mar].
4. Experimental Results
E. Image Classification The original X-ray scan of bone fracture image is taken from
Classification is a step of data analysis to study a set of data ‘Mandalay Orthopaedic Hospital’ in Myanmar. The X-ray scan
and categorize them into a number of categories. It takes a of bone image is given as input for this system.
feature vector as an input and responds category to which
the object belongs. It employs two phase of processing, The input X-ray image contains noises which is unwanted
training and testing: in training, characteristic properties of pixels that affect the quality of the image. Therefore, the
typical image features are isolated and in testing, these image pre-processing stage is needed to remove these
feature-space partitions are used to classify image feature. noises. For this system, each X-ray image has a resolution of
200 x 200 pixels in size by resizing. In this thesis, 52 of bone
The most commonly used of classifiers are Bayesian, Neural images are used where 40 images for training and 12 images
Network, K Nearest Neighbour, and Support Vector Machine for testing. Among the training images, 10 images are normal
that are used to classify the sets of data. Bayesian has always or non-fracture image and 30 images are fracture image. The
the minimum error rate but it requires exact knowledge of next step is image pre-processing. The purpose of pre-
class. Bayesian classifier is applied with Gabor Orientation processing is to provide image quality for later process and
(GO), Markov Random Field (MRF), and Intensity Gradient to obtain the best result. After image segmentation is
Direction (IGD) features [14Jar]. performed, the segmented bone image is used for feature
extraction. The extracted feature points are shown in figure
Neural Network classification is fast but its training can be 4.1(a) non-fracture image and (b) fracture image. These
very slow and which requires high processing time for large figures show the feature points from the image by Harris
neural network. algorithm.

K Nearest Neighbour can have excellent performance for

arbitrary class conditional, can be slow for real-time
prediction and a large number of training examples, it is not
robust to noisy data.

Support Vector Machine (SVM), it can affect in high

dimensional spaces. It can avoid under-fitting and over-
fitting. SVM also produces very accurate classifier. If the
number of features is very much greater than the number of
samples, it will give poor performance. In this thesis, support
vector machine (SVM) is used to classify fracture types.

Support Vector Machines

Support Vector Machines (SVM) recently became one of the (a) (b)
most popular classification methods. They have been used in Figure 4.1: Result Images for Harris Corner Detection
a wide variety of applications such as text classification,
facial expression recognition, gene analysis and many others. Harris detects all Corners or intersection points based on the
Support vector machine (SVM) is supervised learning brightness of images. It uses the second moment matrix or
models with associated learning algorithm that analyzes correlation matrix as the basis of its corner decisions. These
data and recognizes patterns. SVM classifies unseen data extracted shape and point feature are used as the input to
based on a set of labeled training data set. The goal of SVM is classify which type of fracture by SVM algorithm.
to find the optimal separating hyperplane which maximizes
the margin of the training data. The maximum margin The SVM algorithm classifies the positive and negative
hyperplane is the one that gives the greatest separation features by training a classifier that uses a linear kernel
between the classes. Among all hyperplanes that separate function to map data into a higher dimensional space. In this
the classes, the maximum margin hyperplane is the one that system, linear kernel detects the features point to point
is as far away as possible from the two convex hulls, each of horizontally. Despite several erroneous matches between
which is formed by connecting the instances of a class. The the two images, SVM find the optimal point from the feature
instances that are closest to the maximum hyperplane are space. This algorithm produces the classification result of
called support vectors. There is at least one support vector match fracture detected for fracture image and no match
for each class, and often there are more. fracture detected for non-fracture image. The output classes
have the two types in this thesis. The classification results of
SVM training data builds the hyper plane or decision SVM algorithm are shown in figure 4.2(a) and (b). In which,
boundary whether a new example falls into one category or (a) is non-fracture and (b) is fracture image.
the other. A hyperplane is a generalization of a plane in one
dimension; a hyperplane is called a point. In two dimensions,
it is a line and a plane in three dimensions. It can call a

@ IJTSRD | Unique Paper ID – IJTSRD27957 | Volume – 3 | Issue – 5 | July - August 2019 Page 2414
 International Journal of Trend in Scientific Research and Development (IJTSRD) @ www.ijtsrd.com eISSN: 2456-6470
fracture images to this proposed system have various shapes
and sizes, so it is hard to get the standard features of these
fracture images. Therefore, this proposed system cannot be
generated the good performance. According to the
experimental results, a combining algorithm can be
produced a good result for fracture detection system.

References
[1] Ms. Senhal Deshmukh, Ms Shivani Zalte, Mr.Shantanu
Vaidia, Mr.Parag Tangade: Bone Fracture Detection
Using Image Processing in Matlab, Internal Journal of
Advent Research in Computer and Electronics, 28
March, (2015).
(a) (b) [2] [Mahmoud Al-Ayyoub, Duha Al-Zghool: Determining the
Figure4.2: Results of SVM classifier for Match Data Images Type of Long Bone Fractures in X- Ray Images, Wseas
Transactions on Information Science and Applications,
Finally, the algorithm produces the output image as the bone Issue 8, Volume 10, August, (2013).
fracture detection result. Therefore, the final results of this
[3] Rashmi, Mukesh Kumar, and Rohini Saxena : Algorithm
thesis are illustrated in figure 4.3(a) and (b) respectively.
and Technique on Various Edge Detection: A Survey, An
International Journal (SIPIJ) Vol.4, No.3, June, (2013).
[4] Shubhangi D. C, Raghavendra, P.S Hiremath : Edge
Detection of Femur Bone in X-Ray Images – A
Comprartive Study of Edge Detectors, International
Journal of Computer Applications, Volume Volume 42-
No.2, March, (2012).
[5] Dmitriy Fradkin and Ilya Muchnik : Support Vector
Machines for Classification, Dimacs Series in Discrete
Mathematics and Theoretical Computer Science,
(2000).
[6] S. K. Mahendran and S. Santhosh Baboo : Ensemble
Systems for Automatic Fracture Detection, IACSIT
International Journal of Engineering and Technology,
(a) (b) Vol. 4, No. 1, February, (2012).
Figure 4.3.Fracture Detection Results
[7] K. E. Bugler, T. O. White, D.B. Thordarson : Focus on
Ankle Fracture, The Journal of Bone and Joint Surgery,
5. Conclusion
(2012).
In this experiment, many X-ray images are collected from
Mandalay Orthopaedic Hospital and various Internet [8] Nathanael .E. Jacob, M. v. Wyawahare: Survey of Bone
websites. In this paper, 52 of bone images are used where 40 Fracture Detection Techniques, International Journal of
images for training and 12 images for testing. Among the Computer Applications, Volume 71-No.17, June, (2013).
training images, 10 images are normal or non-fracture image
[9] Huai Yang Chen & Jinjine Chen: Improvements Based on
and 30 images are fracture image. The algorithm cannot
the Harris Algorithm, Computer and Information
detect correctly in 4 fracture images. According to the test
Science; Vol. 6, (2013).
results, the performance of the detection method affect by
the quality of the image. After that all, this system [10] Redoun Korchiyne, Sidi Mohamed Farssi,
determined whether a fracture exists or not in the image. A Abderrahmane Sbihi, Rajaa Touahni, Mustapha Tahiri
software algorithm capable of providing some theory after Alaoui: A Combine Method of Fractal and GLCM Features
bone fracture detection has been specified, and for MRI and CT Combine Method of Fractal and GLCM
implemented. The proposed system is developed by using Features for MRI and CT Scan Images Classification, An
MATLAB programming language. The input X-ray bone International Journal (SIPIJ) Vol.5, No.4, August,(2014).

@ IJTSRD | Unique Paper ID – IJTSRD27957 | Volume – 3 | Issue – 5 | July - August 2019 Page 2415