See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/269309071

An analysis of x-ray image enhancement methods for vertebral bone

segmentation

Conference Paper · March 2014

DOI: 10.1109/CSPA.2014.6805749

CITATIONS READS

31 3,240

5 authors, including:

Aini Hussain Noorita Tahir

Universiti Kebangsaan Malaysia Universiti Teknologi MARA
552 PUBLICATIONS   6,541 CITATIONS    174 PUBLICATIONS   1,082 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Squat exercise abnormality detection by analyzing joint angle for knee osteoarthritis rehabilitation View project

Intelligent Controller View project

All content following this page was uploaded by Aouache Mustapha on 22 December 2014.

The user has requested enhancement of the downloaded file.

 2014 IEEE 10th International Colloquium on Signal Processing & its Applications (CSPA2014), 7 - 9 Mac. 2014, Kuala Lumpur, Malaysia

An Analysis of X-Ray Image Enhancement Methods

for Vertebral Bone Segmentation

Ili Ayuni Mohd Ikhsan1, Aini Hussain2, Mohd Asyraf Zulkifley3, Nooritawati Md. Tahir4, Aouache Mustapha5
EESE Dept., Faculty of Engineering and Built Environment, Universiti Kebangsaan Malaysia, 43600, Bangi, Malaysia
*Faculty of Electrical Engineering, Univeriti Teknologi Mara, 40450 Shah Alam, Malaysia
[email protected], [email protected], [email protected],
[email protected]*, [email protected]

Abstract— Image enhancement is a critical component in getting indicator for osteoporosis disease. Osteoporotic facture
a good segmentation, especially for x-ray images. Magnification of normally occurs to elderly woman due to aging factor [8].
the contrast and sharpness of the image will increase the accuracy Usually, for a small facture, it might go undetected by the
of the subsequent modules for an autonomous disease diagnosis physician or doctor, thus leads to a wrong diagnosis[9].
system. In this paper, we analyze various methods of pre-
Autonomous diagnosis system can assist the doctor in
processing techniques for vertebral bone segmentation. Three
methods are considered which are histogram equalization (HE), examining the bone health, which acts as a second opinion
gamma correction (GC) and contrast limited adaptive histogram [6][9]. The common characteristics of an x-ray image are
equalizer (CLAHE). This work aims to compare and quantify the grayscale color space, high noise, low intensity, poor contrast
precision and accuracy of the techniques that are used to enhance and weak boundaries representation, which will normally affect
the image quality. Experimental results of the system yield the information of the image [10]. An example of grayscale x-
favorable results where the most accurate technique is CLAHE, ray image of cervical vertebra is shown in Figure 1. Grayscale
followed by GC and HE. image represents the intensity information of the image pixel.
The pixel intensity representation is from 0 (black) to 255
Keywords—I, Histogram technique, Gamma Correction
technique, Contrast Limited Adaptive Histogram Equalizer
(white)[11]. With limited information and low quality image,
segmentation and feature extraction of x-ray image is a
challenging task.
I. INTRODUCTION
Image processing techniques have been used in a wide
variety of applications nowadays to enhance the quality of raw
image data[1]. As for example, in medical imaging, it has been
accepted as a diagnostic tool to facilitate the doctors for fast
checkup [2]. The system provides information and details of the
patient’s health where an expert can analyze the symptom just
by looking at the medical images. The importance of good image
processing techniques in medical imaging are to produce better
image quality, to sharpen the details of the image and to measure
the features and structures of the images[3]. Moreover, the Fig. 1: X-ray image of the cervical vertebra
technologies nowadays enable a rapid diagnosis assessment
compared to the manual method [4]. Medical image processing The quality of the image can be enhanced by applying pre-
allows the doctors to see fine details of the images that are processing technique, which is the fundamental step in medical
difficult to detect or distinguish just by using naked eyes [5]. image processing, especially for segmentation and feature
There are several medical image modalities, for instance, x-ray extraction [11]. The primary aim of this paper is to compare and
images, computed tomography image (CT scan), single photon quantify the results of image enhancement of cervical x-ray
emission computed tomography (SPECT), magnetic resonance image using different pre-processing techniques.
image (MRI), position emission tomography (PET) and digital In previous study[12], various methods have been proposed
radiography [5][6]. for the pre-processing module to enhance the quality of the
X-ray image is one of the oldest photographic films that is image. Two common domains used in image contrast are spatial
mostly used in medical diagnosis and treatment [6][7]. X-ray (SD) and frequency domain (FD). The FD utilizes Fourier
image is a very useful modality for the physicians and doctors transform technique while SD focuses on modifying the pixel
to determine and analyze the bone fracture which is an values based on neighbouring pixels with a unique frequency
important symptom used for diagnosis[6]. It is also a good range [13][11]. According to Mustapha et al., saturation

978-1-4799-3091-3/14/$31.00 ©2014 IEEE 208

 2014 IEEE 10th International Colloquium on Signal Processing & its Applications (CSPA2014), 7 - 9 Mac. 2014, Kuala Lumpur, Malaysia

occurred because of the negative aspect of image contrast National Centre for Biomedical Communication that provides
enhancement [13]. Thus, they have proposed the GC approach a database of more than 17000 spine X-ray images [15][16].
and a non-linear contrast image enhancement process which are The three image enhancement methods are as follows:
independent to neighbourhood pixels and can be performed a. Histogram Equalization (HE)
directly to each individual pixel. This approach is applied to b. Gamma Correction (GC)
improve the localization of the cervical vertebrae and its c. Contrast Limited Adaptive Histogram Equalization
features [12]. Besides, HE has also been applied for x-ray image (CLAHE)
enhancement [7]. The information from the HE not only can be
used in image enhancement, it can be applied to image A. Image Histogram
segmentation and compression [11]. Apart from that, CLAHE
is a popular image enhancement technique that modifies the Image histogram is a graphical representation of intensity
grey-level value based on some characteristics to improve the distribution [7][14]. The y-axis represents the tonal variation
contrast and minimize the noise [7]. while x-axis of the graph represents the pixels of the image[14].
The structure of the paper is presented as follows; Section Figure 3 (a) and (b) shows the original image and histogram of
II focuses on the conceptual framework and overview of the the image respectively. The image contains more pixels with
pre-processing techniques. In Section III, the experimental high intensity and therefore, the histogram skew towards the
results are presented and finally, general conclusion is given in right side which is closed to 255.
Section IV.

II. METHOD
Pre-processing is a fundamental step in image processing.
This module takes the raw image, which have the lowest level
of abstraction to suppresses the irrelevant data as well as
enhances the important information for further image
processing steps[14]. This paper will compare three methods
or techniques of pre-processing. The conceptual framework for
this paper is shown in Figure 2.
Fig. 3: (a) Original Image, (b) Histogram of the image

Test Image Original Image

Image B. Histogram Equalizer (HE)

Acquisition
HE works by adjusting the contrast of the image. This
technique improves the image appearance by scaling out the
intensity range of the image [14]. Through re-assignment of
Image Enhancement pixel value, the distribution on the histogram is stretched out to
Edge detection
produce a more uniformly distribution [7]. An example of each
HE image and its histogram are shown in Figure 4(a) and (b)
Edge detection
respectively.
Comparison
TP, TN, FP, FN

Matching
Measure accuracy &
sensitivity

Fig. 2: The conceptual framework

The first step in medical image processing is good image

acquisition as the subsequent steps rely heavily on a good
quality input. The x-ray images were taken from Web-based
Medical Information retrieval System (Web-MIRS) which was
collected by the second National Health and Nutrition
Fig. 4 (a) Histogram Equalization (b) Histogram of the image
Examination Survey (NHANES II). The system is developed
by The Communications Engineering Branch of Lister Hill

209
 2014 IEEE 10th International Colloquium on Signal Processing & its Applications (CSPA2014), 7 - 9 Mac. 2014, Kuala Lumpur, Malaysia

C. Gamma Correction (GC) heavily on edge detection, the performance is quantified based
GC is a non-linear contrast enhancement technique that is on the detected edge before and after image enhancement. The
applied directly to each pixel and independently changes the output results are based on the measurement of sensitivity
dynamic range of all pixels in the image. The intensity value will (TPR), specificity (SPC) and accuracy of each method. There
be shifted and modified due to the change in gamma value [12]. are four parameters needed to compute these metrics, which are
Figure 5(a) and (b) show the gamma correction and its histogram true positive (TP), true negative (TN), false positive (FP) and
respectively. false negative (FP). The comparison is based on the event
classification as stated in Table I. Sensitivity metric emphasizes
more on positive edge detected while specificity emphasizes on
negative detection rate of the edge, whereas accuracy determines
the true detection of the pixel regardless of positive or negative
detection. Sensitivity is a ratio of positive experimental edge
detected over total expected number of positive pixels.
Meanwhile, specificity is the ratio of negative experimental edge
detected over expected negative detection. The ground truth is
traced out manually, which the dataset was validated by peer
review among the lab members. TP classification is based on
positive ground truth and positive test image, whereas FN
involves positive ground truth and negative test image. FP
involves negative ground truth and positive test image, while TN
Fig. 5: (a) Gamma correction, (b) Histogram of the image
involves negative ground truth and negative test image.

Table I: Event Classification of edge detection

D. Contrast Limited Adaptive Histogram Equalization
(CLAHE) Ground truth High High Low Low
Test image High Low High Low
A good contrast through modification and adjustment of the
gray-level intensity can be achieved through contrast limited Classification Output TP FP FN TN
adaptive histogram equalization technique. This technique is
inspired by HE where several histograms will be constructed,
one each to distinct part of the image. CLAHE utilizes a small ்௉
Sensitivity (TPR) ൌ
region of the image (tiles), where it contrast is enhanced through ்௉ାிே
adaptive histogram equalization [7]. The result of CLAHE (1)
்ே
technique is shown in Figure 6 (a) and its histogram is shown in Specificity (SPC) ൌ
Figure 6 (b). ி௉ା்ே
(2)
்௉ା்ே
Accuracy (ACC) ൌ ்௉ା்ேାி௉ାிே
(3)
III. RESULTS & DISCUSSION
In this paper, three pre-processing techniques are compared
based on the output of edge detection before and after image
enhancement. The results of these technique are shown in Table
II.

Table II: Segmentation results of histogram equalization, gamma correction

and CLAHE.
Fig. 6: (a) CLAHE image, (b) Histogram of the image .
TP FP FN TN
E. Evaluation Metrics HE 30.3% 35.7% 32.3% 33.4%
GC 36.5% 30.2% 35.3% 33.2%
For the assessment, three evaluation metrics are computed CLAHE 32.2% 34.1% 32.4% 33.3%
by comparing the original image with the image that has
undergone preprocessing technique. For preliminary result, to Table III summarizes the performance of the system in terms
validate our techniques, there are five different images with of accuracy, sensitivity and specificity. The results show that
variety of image qualities are selected. Since bone fracture relies the best pre-processing technique is gamma correction as
compared to the histogram equalization and CLAHE methods.

210
 2014 IEEE 10th International Colloquium on Signal Processing & its Applications (CSPA2014), 7 - 9 Mac. 2014, Kuala Lumpur, Malaysia

Table III Accuracy, Sensitivity and Specificity result

Accuracy Sensitivity Specificity

HE GC CLAHE HE GC CLAHE HE GC CLAHE
Image 1 0.9731 0.9738 0.9754 0.0220 0.0198 0.0224 0.9959 0.9959 0.9959
Image 2 0.9816 0.9825 0.9848 0.0444 0.0393 0.0416 0.9965 0.9965 0.9966
Image 3 0.9831 0.9857 0.9848 0.0418 0.0384 0.0401 0.9971 0.9970 0.9972
Image 4 0.9970 0.9823 0.9838 0.0235 0.0247 0.0186 0.9972 0.9973 0.9973
Image 5 0.9865 0.9902 0.9864 0.0234 0.0343 0.0264 0.9970 0.9970 0.9970
Average 0.98026 0.98290 0.98304 0.03102 0.03130 0.02982 0.99674 0.99674 0.99680

From both tables, we can summarize that CLAHE obtained classification fusion techniques in x-ray images,” Glob. J.
a more stable result for all three evaluation metrics and woks Comput. Sci. Technol., vol. 11, no. 14, 2011, pp. 22–28.
the best for x-ray images of cervical vertebra. [7] N. R. S. Parveen and M. M. Sathik, “Enhancement of bone
fracture images by equalization methods,” in 2009
IV. CONCLUSION International Conference on Computer Technology and
Development, 2009, vol. 2, pp. 391–394
In conclusion, three methods for pre-processing have been [8] S. Wesarg, M. Erdt, K. Kafchitsas, and M. F. Khan, “CAD
compares, which are HE, GC and CLAHE. The result shows of osteoporosis in vertebrae using dual-energy CT,” in
that GC achieved the best sensitivity while CLAHE obtains Computer-Based Medical Systems (CBMS), 2010 IEEE
the best accuracy. Since accuracy measures true detection, 23rd International Symposium on, 2010, pp. 358–363.
either positive or negative, it is the most suitable performance [9] A. Mustapha, A. Hussain, and S. A. Samad, “Automatic
metric for our intended application. Thus, the most suitable Vertebral Fracture Assessment System ( AVFAS ) for
algorithm for image enhancement is CLAHE. In future, the spinal pathologies diagnosis based on radiograph x-ray
images,” Vis. Informatics Bridg. Res. Pract., vol. 5857,
research will focus on improving these problems through
2009, pp. 122–135.
better segmentation and extraction scheme. A color [10] R. Ramani, N. S. Vanitha, and S. Valarmathy, “The pre-
constancy approach can be explored to further improve the processing techniques for breast cancer detection in
contrast of the image[17]. mammography images,” Int. J. Image, Graph. Signal
Process., vol. 5, no. 5, 2013, pp. 47–54.
ACKNOWLEDGEMENT [11] C. Chaudhary and M. K. Patil, “Review of image
This research is supported by Universiti Kebangsaan enhancement techniques using histogram,” Int. J. Appl. or
Malaysia (Geran Universiti Penyelidikan UKM: GUP-2013- Innov. Eng. Manag., vol. 2, no. 5,2013, pp. 343–349.
035) and Ministry of Science and Technology Malaysia [12] A. Mustapha, A. Hussain, and S. A. Samad, “A new
approach for noise reduction in spine radiograph images
(ScienceFund: 06-01-02-SF1018).
using a non-linear contrast adjustment scheme based
REFERENCES adaptive factor,” Sci. Res. Essays, vol. 6, no. 20, 2011, pp.
4246–4258.
[1] K. M. M. Rao, “Overview of image processing,” in [13] Ada. and R. Kaur, “Feature extraction and principal
Proceedings of a workshop on image processing and component analysis for lung cancer detection in ct scan
pattern recognition, 2004, pp. 1–7. images,” Int. J. Adv. Res. Comput. Sci. Softw. Eng., vol. 3,
[2] H. Zhu, W. Sun, M. Wu, G. Guan, and Y. Guan, “Pre- no. 3, 2013, pp. 187–190.
processing of x-ray medical image based on improved [14] V. Agarwal, “Analysis of Histogram Equalization in image
temporal recursive self-adaptive filter,” 2008 9th Int. Conf. preprocessing,” BIOINFO Human- Comput. Interact., vol.
Young Comput. Sci.,2008, pp. 758–763. 1, no. 1, 2011, pp. 4–7.
[3] M. S. V Sokashe, “Computer assisted method for cervical [15] X. Zhiyun, L. R. Long, S. Antani, and G. R. Thoma,
vertebrae segmentation from x-ray images,” Int. J. Adv. “Pathology-based vertebral image retrieval,” in Biomedical
Res. Comput. Commun. Eng., vol. 2, no. 11, 2013, pp. Imaging: From Nano to Macro, 2011 IEEE International
4387–4389. Symposium on, 2011, pp. 1893–1896.
[4] N. Khayati, W. Lejouad-Chaari, and S. Sevestre-Ghalila, [16] C. Yuchou, S. Antani, L. Dah-Jye, K. Gledhill, L. R. Long,
“A distributed image processing support system and P. Christensen, “CBIR of spine x-ray images on inter-
application to medical imaging,” in Imaging Systems and vertebral disc space and shape profiles,” in Computer-
Techniques, 2008. IST 2008. IEEE International Workshop Based Medical Systems, 2008. CBMS ’08. 21st IEEE
on, 2008, pp. 261–264. International Symposium on, 2008, pp. 224–229.
[5] M. Aouache, A. Hussain, S. A. Samad, A. H. Hamzaini, [17] M. A. Zulkifley, B. Moran, and D. Rawlinson, “Robust
and A. K. Ariffin, “Active shape modeling of medical foreground detection: a fusion of masked grey world,
images for vertebral fracture computer assisted assessment probabilistic gradient information and extended
system,” in Research and Development, 2007. SCOReD conditional random field approach.,” Sensors (Basel)., vol.
2007. 5th Student Conference on, 2007, pp. 1–6. 12, no. 5, 2011, pp. 5623-4
[6] S. K. Mahendran and S. S. Baboo, “An enhanced tibia
fracture detection tool using image processing and

211

View publication stats