# INTRODUCTION edical image processing is a field of science that is gaining wide acceptance in healthcare industry due to its technological advances and software breakthroughs. It plays a vital role in disease diagnosis and improved patient care and helps medical practitioners during decision making with regard to the type of treatment. Several state-of-the-art equipments produce human organs in digital form. Examples of such devices include X-Ray-based devices, Computed Tomography (CT), Magnetic Resonance Imaging (MRI), Ultrasound (US), Positron Emission Tomography (PET) and Single Photon Emission Computed Tomography (SPECT). Out of these, X-Ray is one the oldest and frequently used devices, as they are non-invasive, painless and economical. A bone x-ray makes images of any bone in the body, including the hand, wrist, arm, elbow, shoulder, foot, ankle, leg (shin), knee, thigh, hip, pelvis or spine (Tamisiea, 2008). A typical bone ailment is the fracture, which occurs when bone cannot withstand outside force like direct blows, twisting injuries and falls. Fractures are cracks in bones and are defined as a medical condition in which there is a break in the continuity of the bone. Detection and correct treatment of fractures are considered important, as a wrong diagnosis often lead to ineffective patient management, increased dissatisfaction and expensive litigation. The importance of fracture detection comes from the fact that in clinical practice, a tired radiologist has been found to miss fracture cases after looking through many images containing healthy bones. Computer detection of fractures can assist the doctors by flagging suspicious cases for closer examinations and thus improve the timeliness and accuracy of their diagnosis. Moreover, tibia fractures are the subject of ongoing controversy and discussion. Despite newer innovations, automatic detection of tibial fractures essentially remains unresolved as these injuries are different and variable in presentation and their outcomes are unpredictable. The main focus of this work is to automatically detect fractures in long bones and in particular, leg bone (often referred as Tibia), from plain diagnostic xrays using a series of sequential steps. The region of interest in tibia is the Diaphysis region. # II. # PREVIOUS STUDY Since Wilhelm Roentgen discovered the existence of x-rays in 1895, medical imaging has advanced at a tremendous rate and has become the fundamental diagnostic tool in modern healthcare. This section reviews some works that has focused on the various steps during fracture detection. Preprocessing X-ray images are frequently degraded by Poisson noise, which degrades the visual quality of the image and obscures important information required for accurate diagnosis. Including a denoising step in automatic fracture detection process is used by several researchers (Sakata and Ogawa, 2009). For this purpose, a Wiener filter is recommended by Tsukahara et al. (1998). The advantage of using this filter is that it is more effective at edge preservation but the result if often blurred. Wavelets have been used by Niijima (2000) where a decomposition step identifies noise dominant regions and different shrinkage thresholding techniques M are used to reduce noise. Several proposal for a) denoising use wavelets (Kaur and Singh, 2010;Federico and Kaufmann, 2007).They produce very effective with Gaussian noise but its performance slightly degrades when used with Poisson noise. To answer this problem, Luisier et al. (2010), proposed a fast interscale wavelet denoising method for Poisson-corrupted images. This model is referred to as 'PURESHRINK', which used Haar wavelet transformation with soft thresholding to reduce Poisson noise. This method showed improved noise removal result but in some cases introduced 'staircase' artifacts. This can be enhanced by replacing haar with a decimated biorthogonal Haar (Bi-Haar) transform (Zhang et al., 2008) and then use Independent Component Analysis (ICA) (Marusic et al., 2005) to remove dependencies between the data streams associated with each wavelet decomposition. Use of anisotropic diffusion for denoising image has recently received a great deal of attention because of its impressive performance in preserving edge sharpness and suppressing noise (Tsukahara et al., 1998). They work well with both Gaussian and Poisson noise. However, while considering x-ray images, consideration of edges and local details between heterogeneous portions can provide much enhanced performance. This work considers this by developing an edge enhanced adaptive anisotropic diffusion filter. # b) Segmentation The solutions proposed for bone image segmentation can be categorized into two main groups, namely, gray level feature-based methods and texture feature-based methods (Sharma and Aggarwal, 2010). Gray level feature-based methods analyze the gray level or color features of an image to segment an image. Histogram-based methods, Edge-based methods, region-based methods all belong to this category. Textural features of an image are considered important from both segmentation and classification point of view. The aim here is to divide images into regions with similar texture properties. All these methods use random patterns/textures but work well for segmenting medical images. Apart from these, model based segmentation and Atlas based segmentation techniques have also been proposed. Model based methods involve active shape and appearance model during training. They work by determining and analyzing the statistical influence between the model and image during segmentation. These methods often require manual interaction and exhibits poor convergence to concave boundaries. All these methods work on a common objective, that is, to provide solution for efficient automatic medical image segmentation. Recently the use of morphological transformation and wavelet transformation are gaining more attention for segmenting medical images. In wavelet based techniques the result of wavelet transform is used as features during segmentation. Morphological transformation combines geometrical features and edge features to segment the image. Use of morphological-based segmentation helps to avoid the problem of over and under segmentation (Najman and Schmitt, 1996). The advantage of using wavelet transform is that it provides a precise and unifying framework for the analysis and characterization of an image at different scales. This advantage can be fully exploited for efficient segmentation of X-Ray images. Further, the use of wavelets increases the speed of the segmentation and reduces the number of computations performed by selecting only the prominent pixels within an image. Hough transforms have long been used as computationally efficient methods for detecting particular shapes in images. The use of hough transformation in identifying fractures have also been proved advantages. The Hough transform (Duda and Hart, 1972) is a feature extraction technique in image analysis, computer vision, and digital image processing. It is concerned with the identification of straight lines, position of arbitrary shapes, most circles or ellipses. The important case of Hough transform is the linear transform for detecting straight lines. Compared with other algorithms that detect straight lines, Hough transform can be used to find and link segments in an image. A line in the image space is mapped to a point in the parameter space. Similarly, each pixel of the image space is transformed to a parameterized curve of the parameter space. Each transformed point in the parameter space is considered as a candidate for being a line and accumulated in the corresponding cell of an accumulator. Finally, a cell with a local maximum of scores is selected, and its parameter coordinates are used to represent a line segment in the image space. The main advantage of the Hough transform technique is that it is tolerant of gaps in feature boundary description and is relatively unaffected by image noise. # Global Journal of Computer # METHODOLOGY The methodology consists of the following steps during the identification of fractures in Tibia bone from x ray images. (i) Preprocessing -Procedure that contains steps that enhance the x-ray input image in a way that its result improves the fracture detection process. (ii) Segmentation -Procedure that has two steps. The first step separates the bone structure from the x-ray image and the second step, identifies the diaphysis region from the segmented bone structure. (iii) Fracture Detection -A classifier fusion method is used, where results from different classifiers are combined to detect fractures. # a) Preprocessing This stage consists of procedures that enhance the features of an input x-ray image so that the output image improves the performance of the subsequent stages of the proposed system. This stage proposes a new technique, called, SACEN (Simultaneous Automatic Contrast adjustment, Edge Enhancement and Noise Removal) algorithm, which can adjust contrast, enhance edges and remove noise from an x-ray image simultaneously. The SACEN has the following steps : # b) Segmentation This stage of the study has two algorithms. The first algorithm segments the bone image from the x-ray image and the second algorithm describes a technique that identifies the diaphysis region of a bone image. The two algorithms are independent and are used in different places during fracture detection. The algorithm segments bone region from background uses a combination of wavelets (1-level Haar) and morphological transformation (dilation, erosion, opening and closing). Usage of mathematical operations for segmentation eliminates small objects and enhances the connectivity of objects, thus generating an image with areas that have elements with only connected regions. The resultant image after segmentation is referred to as "Bone Region Image (BRI)' in this paper. The BRI has diaphysis region and epiphyses region along with fleshy region. An activecontour based segmentation is first performed to remove the fleshy region, followed by the application of diaphysis segmentation algorithm. The algorithm used for this purpose is proposed by Donnelley (2008). This image is termed as 'Diaphysis Region Image (DRI)' in this paper. The usage of these two segmentation techniques narrows down the search region during fracture detection and thus aids in accurate detection of fractures in a time efficient manner. # c) Fracture Detection The last stage of the proposed system is the actual detection of fracture. It has been proved that the performance of fracture detectors can be improved both in terms of accuracy and sensitivity by combining The basic fusion rule used is that if more than 2 classifiers report fracture then the image is said to have fracture. A modified Hough transform with gradient analysis is used to identify the location of fracture. IV. # EXPERIMENTAL RESULTS To analyze the performance of the proposed classifiers, several experiments were conducted using 1000 test images. Out of these images, 600 images The performance of the proposed system is analyzed using two performance metrics, namely, detection rate and speed of detection. The software was developed using MATLAB 2009a and all the experiments were conducted on Pentium IV machine with 4GB RAM. # a) Detection Rate Detection rate is the number of correctly detected fractures to the total number of fractured/normal images and the results from the experiments is shown in Table 4.1. From the table, it is evident that the proposed fusion classifier (FC) that combines three classifiers BPNN, SVM and NB produce better detection results. # b) Speed of Detection The time taken in seconds for the proposed algorithm to report the detection of fracture or nonfracture is presented in Table 4 From the table, it is clear that the proposed FC is slow when compared with its base classifiers. The reason behind this is that the number of computations involved during classification is high because of the fusion process. However, it can still be considered negligible, as the average difference between FC and other classifiers is less than 1.89 (fractured images) and less than 1.35 seconds (for non-fractured images). # c) Visual Results ![c) Fracture DetectionThe fracture detection techniques proposed can be loosely categorized into classification-based and transform-based. The first published work on the detection of fractures in x-ray images is that ofTian et al. (2003). The method detected femur fractures by computing the angle between the neck axis and shaft axis. Subsequently, Gabor, MRSAR, and gradient intensity were used for fracture detection and a simple voting scheme was used to combine the individual classifiers that work on single features (Lim et al., 2004; Chen et al., 2004). Since the individual classifiers tend to complement each other, the combined method improves both the accuracy and sensitivity significantly. A similar approach of combining classifiers was also proposed by Lim et al. (2005) who combined probabilistic combination methods for segmentation.](image-2.png "") ![Science and Technology Volume XI Issue XIV Version which in turn slows the feature extraction and fracture detection process. III.](image-3.png "") ![classifiers (Lim et al., 2005; Chen et al., 2004). Motivated by this fact, the present work uses a novel classifierfusion algorithm to identify and detect fractures from DRD. This paper uses the texture features for fracture detection. The texture features used are GLCM Mean, GLCM Variance, Energy, Entropy, Homogeneity, Gabor Orientation, Markov Random Field (MRF) and Intensity Gradient Direction (IGD). A fusion-based classifier for fracture detection is proposed in this paper for fracture detection. All the selected classifiers are modeled to work as a binary classifier, thus reporting whether a fracture is detected or not detected. If detected, the position of the fracture is highlighted. Three classifiers are considered, namely, Feed Forward Back Propagation Neural Networks (BPNN), Support Vector Machine (SVM) classifier and a Naïve Bayes (NB) classifier. The fusion classifiers build from base classifiers are 1. Texture features with BPNN (Named as TBPNN Model) 2. Texture features with SVM (Named as TSVM Model) 3. Texture features with NB (named as TNB Model) Texture features with BPNN + SVM + NB (named FC model)](image-4.png "") 4![Global Journal of Computer Science and Technology Volume XI Issue XIV Version I 2011 August were images with fractures and rest were normal x-ray images. Four hundred fractured x-ray images and 250 from normal x-ray images were used during training.](image-5.png "4 .") ![Figure ---shows sample images from the dataset, where F1-F3 are images with fracture, while N1-N3 are normal images.](image-6.png "Figure") 41![Figure 4.1 : Sample Images](image-7.png "Figure 4 . 1 :") ![Figure ---shows the visual results of the proposed fracture detection system. Figures ---a and bshows correctly identified fractures. It is evident from the visual results that the performance of the proposed system is greatly improved and can identify more than one fracture (if present) in the image. This is evident from the results of Figure---cwhere multiple fractures in different places are detected.](image-8.png "Figure") 42![Figure 4.2 : Visual Results Several experiments were conducted to analyze the performance the proposed fracture detection systems based on texture features and classifiers with a common aim of finding the best combination of features and best fusion of classifiers to identify fractures in x-ray bone images. Experimental results showed that the](image-9.png "Figure 4 . 2 :") 41 : Detection RateBone typeTBPNN TSVM TNBFCFractured83.1285.71 86.08 91.27Non-Fractured92.9694.55 95.91 98.43 42Bone typeTBPNN TSVM TNBFCFractured8.57.69.310.36Non-Fractured8.77.99.410.02 © 2011 Global Journals Inc. (US) August ## August V. ## CONCLUSION Bone fractures are a common affliction and even in most developed countries the number of fractures associated with age-related bone loss and accidental fractures are increasing rapidly. From both orthopaedic and radiologic point of view, the fully automatic detection and classification of fractures in long-bones is an important but difficult problem. The present work focuses on providing a solution to the automatic discovery of bone fracture in leg long bones. For this purpose several image processing techniques (for preprocessing, segmentation and feature extraction) were used. The extracted features were then given as input to a fusion-based classification system to detect the presence / absence of fracture(s) in an image. Several experiments were conducted to analyze the performance of the proposed fusion classifier-based detection system with respect to its efficiency in terms of correct detection and speed of the algorithm. The performance was compared with its traditional single classification system. Experimental results proved that the proposed amalgamation of techniques showed improved results in terms of accuracy in detecting fractures and the speed of detection also. In future, other features like shape are to be considered and its affect on detection rate is to be analyzed. Moreover, its applicability to other long bones, like hand, back bone can also be analyzed. ## BIBLIOGRAPHY * Detecting femur fractures by texture analysis of trabeculae YChen WHYap WKLeow TSHowe MAPng Proc. Int. Conf. on Pattern Recognition Int. Conf. on Pattern Recognition 2004 * Detecting femur fractures by texture analysis of trabeculae YChen WHYap WKLeow TSHowe MAPng Proc. Int. Conf. on Pattern Recognition Int. Conf. on Pattern Recognition 2004 * Computer aided long-bone segmentation and fracture detection MDonnelley Computer and Information Science 2008 Ph.D Thesis * Use of the hough transformation to detect lines and curves in pictures RDuda PHart Comm. ACM 15 1972 * Denoising in digital speckle pattern interferometry using wave atoms AFederico GHKaufmann Opt. Lett 32 2007 * An Enhanced Tibia Fracture Detection Tool Using Image Processing and Classification Fusion Techniques in X-Ray Images * AKaur KSingh Speckle noise reduction by using wavelets, National Conference on Computational Instrumentation CSIO NCCI 2010 INDIA 2010 * Detection of femur and radius fractures in x-ray images SELim YXing YChen WKLeow TSHowe MAPng Proc. 2nd Int. Conf. on Advances in Medical Signal and Info. Proc 2nd Int. Conf. on Advances in Medical Signal and Info 2004 * Combining classifiers for bone fracture detection in X-ray images VL FLim LWLeow TChen TSHowe MAPng IEEE International Conference on Image Processing 2005. 2005 1 * Combining classifiers for bone fracture detection in X-ray images VL FLim LWLeow TChen TSHowe MAPng IEEE International Conference on Image Processing 2005. 2005 1 * Fast interscale wavelet denoising of Poissoncorrupted images, signal Processing FLuisier CVonesch TBlu MUnser 2010 Elsevier * Image Denoising Using Over-complete Wavelet Representations SMarusic GDeng DB HTay Proceedings of the European Signal Processing Conference the European Signal Processing Conference 2005. 2005 * Geodesic saliency of watershed contours and hierarchical segmentation, Pattern Analysis and Machine Intelligence LNajman MSchmitt IEEE Transactions on Pattern Analysis and Machine Intelligence 18 12 1996 * Wavelet image analysis MNiijima 2000 Kagaku Gijutsu Publishing in Japanese * Contrast limited adaptive histogram equalization image processing to improve the detection of simulated speculations in dense mammograms EDPisano SZong BMHemminger MDeluca REJohnston KMuller PMBraeuning SMPizer Journal of Digital Imaging 11 4 1998 * Image Denoising using Wavelet Embedded Anisotropic Diffusion (WEAD) JRajan MRKaimal Proceedings of IET International Conference on Visual Information Engineering IET International Conference on Visual Information Engineering 2006 * Noise reduction and contrast enhancement for small-dose X-Ray images in wavelet domain MSakata KOgawa IEEE Nuclear Science Symposium Conference Record (NSS/MIC) Orlando, FL 2009 * Automated medical image segmentation techniques NSharma LMAggarwal J Med Phys 35 2010 * Radiologic aspects of orthopedic diseases DFTamisiea Practical Orthopedics Philadelphia, Pa Mosby Elsevier 2008 16 6th ed. * Computing neck-shaft angle of femur for x-ray fracture detection TPTian YChen WKLeow WHsu TSHowe MAPng Proc. Int. Conf on Computer Analysis of Images and Patterns Int. Conf on Computer Analysis of Images and Patterns 2003 LNCS 2756