Continuous Local Histogram Descriptor For Diagnosis of Bronchiolitis Obliterans K. Muzzammil Saipullah, Mai Mariam M. Aminuddin, Izadora Mustaffa, Nuraishah Sarimin, Ammar Anuar Faculty of Electronic and Computer Engineering, Universiti Teknikal Malaysia Melaka, Melaka, Malaysia {muzzammil, maimariam, izadora, nuraishah.sarimin}@utem.edu.my, m021110015@student.utem.edu.my Abstract— Texture feature is an important feature analysis method in computer-aided diagnosis systems for disease diagnosis. However, texture feature itself could not provide an overall description of the diseases. In this paper, we propose Continuous Local Feature (CLH) to diagnose the Bronchiolitis Obliterans (BO) lung diseases in the chest computer tomography images. CLH is based on the continuous combination of histograms of local texture feature, local shape feature, and the brightness feature. Because CLH extracts more information, it has high discriminating power and is able to classify between the BO lung disease and normal lung region effectively. The experimental results in classifying between BO and normal lung region show that CLH achieves 98.15% of average sensitivity whereas Local Binary Patterns and Gray Level Run Length Matrix achieve 73% and 75.8% of average sensitivities, respectively. In the receiver operating curve analysis, CLH archives 0.9 of area under curve (AUC) whereas LBP and GLRLM achieve 0.78 and 0.86 of AUCs. Keywords-component; continuous local feature, CAD system, bronchiolitis obliterans, CT images I. INTRODUCTION Computed tomography (CT) scans are usually applied to examine the pathological change of the tissues inside the body. However, for examining the pathological change of the tissues, CT scans generate a large number of images. Therefore diagnosing pathological changes using CT image are exhausting for radiologists. Recently, a number of computer-aided detection (CAD) systems [1, 2, 3, 4] have been developed to help the radiologists to diagnose diseases. Thus, using CAD systems to detect lung diseases such as emphysema, honeycombing, and lung cancer has become a significant part in the medical image processing in nowadays [5, 6, 7]. Feature extraction in CAD system is one of the most important steps in recognizing the abnormal regions from the medical image. Texture feature is a fundamental feature for image segmentation [1, 8], classification [8], image retrieval systems [10, 11] etc. In the past decades, texture features such as the gray level difference method (GLDM), the gray level run-length method (GLRLM), and the special gray level dependent method (SGLDM) [12] have been widely used to represent the medical image characteristics that are inaccessible to human observers. Recently, local binary patterns (LBP) have been widely used for medical image analysis. The combination of LBP and gray level generates a powerful texture descriptor in classifying three types of Emphysema and lung regions [13, 14]. Classifying diseases in lung CT images is not a simple task. The diseases are similar and only can be classified by professional radiologist. However a lot of CAD systems have been developed to assist radiologist in identifying abnormal regions. Kim et al. [17] and Park et al. [16] proposed an implementation of shape feature in the detection of obstructive lung diseases and the results shows improvement in classification sensitivity compared to feature based on texture only. However, their proposed system is dependent on the region size, e.g., 16x16, 32x32 and 64x64 pixels. Gathering region images from CT image is not an easy task especially with fixed size of region. Their system shows a good performance in diagnosing large number of disease classes compared to those of other systems [4, 13] which diagnose only specific diseases. In order to discriminate large number of diseases, most of the CAD systems combine large number of features [16, 17, 5]. The disadvantage of this kind of system is low efficiency. This is because the system needs to execute a lot of feature extraction algorithms to diagnosis the disease. To increase the efficiency while preserving the high sensitivity, a new feature is desired. In this paper, a novel feature is called the continuous local histogram (CLH) is introduced. CLH integrates three basic types of features which are texture feature, shape feature and brightness to increase the discrimination power. Compared to the Kim et al. system, CLH only use one powerful feature from each type of feature explained before while Kim et al. utilized 13 texture features and 11 shape features in their system. CLH is a dense descriptor since it is constructed by analyzing the smallest region in an image which is a 3x3 pixels region. Because of that, CLH can be generated from the all size of images and does not possess the region size dependency setback. The proposed system has been tested in classifying hardly classified bronchiolitis obliterans (BO) lung diseases and normal lung region which results in nearly perfect classification sensitivity. 394 978-1-4577-2152-6/11/$26.00 c 2011 IEEE