of 14


All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Medical images provide diagnostic evidence/information about anatomical pathology. The growth in database is enormous as medical digital image equipment's like Magnetic Resonance Images (MRI), Computed Tomography (CT), and Positron Emission
  International Journal on Soft Computing (IJSC) Vol.8, No. 1, February 2017 DOI:10.5121/ijsc.2017.8102  13                                     D. Chitra and M. Karthikeyan   Assistant Professor, Dept. of Computer Science, Govt. Arts College (A), Salem-7, Tamil Nadu, India  A  BSTRACT     Medical images provide diagnostic evidence/information about anatomical pathology. The growth in database is enormous as medical digital image equipment’s like Magnetic Resonance Images (MRI), Computed Tomography (CT), and Positron Emission Tomography CT (PET-CT) are part of clinical work. CT images distinguish various tissues according to gray levels to help medical diagnosis. Ct is more reliable for early tumours and haemorrhages detection as it provides anatomical information to plan radio therapy. Medical information systems goals are to deliver information to right persons at the right time and  place to improve care process quality and efficiency. This paper proposes an Artificial Immune System (AIS) classifier and proposed feature selection based on hybrid Bacterial Foraging Optimization (BFO) with Local Search (LS) for medical image classification.  K   EYWORDS   Computed Tomography (CT), Feature Selection, Artificial immune classifier, Correlation based Feature Selection (CFS), Bacterial Foraging Optimization (BFO), Local Search (LS) 1.   INTRODUCTION Medical images classification is fundamental in different applications in medical image retrieval systems [1]. Due to medical image data’s high variability it is important to use correct models in classification. Various medical image classification methods are discussed in literature. Voluminous medical images generated daily have valuable information useful for medical diagnosis, treatment, research, and education. Automatic image annotation extracts symbolic knowledge from images to facilitate text-based retrieval of relevant images having specific disease abnormalities. Image based Computer Aided Diagnosis (CAD) uses images and serves as a second source of opinion for clinicians on abnormality detection and pathological classification [2].The task of automated knowledge extraction from medical image databases remains a big technological challenge. CAD systems assist radiologists and increase diagnosis accuracy. Medical images are from varied modalities like Magnetic Resonance Images (MRI), Computed Tomography (CT), and Positron Emission Tomography CT (PET-CT) and Ultrasound. CT is a reliable technique for detection of abnormalities like tumours and haemorrhages [3] providing anatomical information to plan radio  International Journal on Soft Computing (IJSC) Vol.8, No. 1, February 2017 14   therapy. For such reasons, this research presents an abnormality detection method and classification for CT brain images. Medical images classification is a manual process [4]. The voluminous amount of medical images produced and limitations of current healthcare technologies in using an automatic classification method highlights the need for consistency in classification. This helps process a huge amount of images easily as this is costly when done manually. Generally, images have color, shape, texture, edge, shadows, and temporal features. Most promising features are color, texture and edge for the following reasons [5]: Color: Egeria occurs in two colors – black and pink (rusty rose) Texture: Texture is a neighbourhood feature; a region or block. Each pixels variation regarding its neighbouring pixels defines texture. Edge: Edge is a large frequency change. Feature extraction in image processing is a dimensionality reduction form. When an algorithm’s input data is too large to be processed and is supposed to be notoriously redundant (much data, but not information) then input data is transformed to a reduced representation features set (named feature vector).Transforming input data to a feature set is feature extraction. If extracted features are carefully chosen it is expected that features set will extract relevant information from input data to perform the desired task using less representation instead of full-size input. Image texture is an image area with repeated pixel intensities patterns arranged in some structural way. Textures are prominent in natural images (grasslands, brick walls, fabrics) and properties for interpretation and image description are revealed through texture observation and analysis like granularity, periodicity, coarseness, geometric structure, smoothness and orientation [6, 7]. Feature selection has two approaches. One is forward selection where a process begins with no attributes/features which are added one by one. At every step, feature which decreases most error is added and process continues till features addition does not decrease the error [8]. The Second approach is backward selection where the idea is to start with all attributes/features and remove them one by one. The feature removed at every step is that which decreases error the most. The process continues till any further removal increases error greatly. Feature selection methods are classified into filter, wrapper, and hybrid approaches. A filter approach is applied to data before classification where features are evaluated by heuristics based on general data characteristics. In a wrapper approach, features are evaluated using classification algorithms. Features in a Hybrid approach are evaluated using filter and wrapper approaches [9]. Feature selection seeks an optimal set of d features out of m. Many methods were previously used for feature selection on training/testing data. Among various methods proposed for FS, population-based optimization algorithms like Genetic Algorithm (GA) and Ant Colony Optimization (ACO) attracted attention. In the new feature reduction system, an evolutionary hybrid feature selection algorithm based on swarm intelligence called Bacteria Foraging Optimization [10] is used.  International Journal on Soft Computing (IJSC) Vol.8, No. 1, February 2017 15   In the study, CT images classification based on feature extraction using Gaussian wavelet and Gray Level Co-occurrence Matrix (GLCM) is used. Feature selection is by wrapper technique using Correlation based Feature Selection (CFS) and hybrid Bacterial Foraging Algorithm. Section 2 reviews related work, section 3 describes methodology, section 4 discusses experimental results and section 5 concludes the work. 2.   RELATED   WORK Efficient implementations of curve let transform for denoising and medical image segmentation was presented by Al Zubiet al., [11]. A comparison study was carried out between different transforms revealing that curvelet transform showed optimal region of interest (ROI) representation with better accuracy and less noise. An adaptive fusion algorithm of CT and MRI medical images based on NSCT was presented by Dai et al., [12]. Source images were decomposed multi-directionally using non-sub sampled pyramids (NSP) and non-sub sampled directional filter banks (NSDFBs). Combining adjustable parameter and adaptive fusion rules objective evaluation index were used for low-frequency sub-band fusion. The experiment verified the method’s feasibility regarding visual quality and objective evaluation criteria, standard deviation, entropy, space-frequency, and mutual-information. A low-dose scan protocol and an associated reconstruction algorithm allowing triple phase-correlated in vivo imaging of perfusion and associated processes was proposed by Sawallet al., [13]. The new reconstruction method keeps administered radiation dose as low as 500 mGy reducing metabolic inference to an animal ensuring longitudinal studies. It provided the first approach to phase-correlated perfusion CT imaging in mice, boosting preclinical research with new possibilities. A method called Feature Selection based on the Compactness Measure from Scatter plots (FSCoMS) to select best features from medical images to improve CBIR effectiveness was proposed by Humpire-Mamaniet al., [14]. The algorithm has a scatter plots compactness analysis to locate most relevant features with high separability abilities. A scatterplot’s high relevance value means better predictability among classes based on two features. This information generates a feature ranking for usefulness. This method was compared to 2 known feature selection methods using 3 real medical datasets. All were compared regarding final feature vector and retrieval effectiveness dimensionality measured by precision and recall graphs. Experiments show that the new method not only got the highest retrieval performance but also achieved smallest number of demanded features (dimensionality) than other analyzed methods. A method, to choose an optimal subset of statistical texture descriptors inefficient representation and ultrasound medical images retrieval, was presented by Sohail et al., [15]. The new feature selection based approach of image annotation and retrieval was tested with a database of 679 ultrasound ovarian images. Retrieval performance achieved was satisfactory. Also, ultrasound medical image retrieval performance with/without applying feature selection based image annotation technique was compared. A colon biopsy image classification technique, where 2 novel structural feature types: white run-length features and percentage cluster area features were introduced and proposed by Rathore et  International Journal on Soft Computing (IJSC) Vol.8, No. 1, February 2017 16   al., [16] were calculated for each colon biopsy image. Features were reduced using PCA. The new technique was evaluated on 174 colon biopsy images, and promising performance was observed regarding various well-known performance measures like accuracy, sensitivity and specificity. The new technique also proved to be better compared to earlier published techniques in the experimental section. Further, classifiers performance was evaluated using ROC curves, and area under the curve (AUC). Rotation boost classifier with PCA based feature selection showed better classification results compared to other classifiers. The ability to decompose 5 materials using energy discriminating detectors and multiple discriminant analysis (MDA) was investigated by Leeet al., [17]. A small field-of-view multi-energy CT system was built. Linear attenuation coefficient was considered a feature of multi- energy CT. MDA decomposes 5 materials with 6 measurements of energy dependent linear attenuation coefficients. The results showed that a CdTe detectors based CT system with MDA can decompose 5 materials. A new method for lung-motion tracking from 4-dimensional X-ray computed tomographic (4D-CT) images proposed by Kubota et al., [18] uses an enhanced 3D-KLT tracker. The feature point extraction algorithm is image gradients based. The new method adopted pyramidal image structure based hierarchical tracking. The proposed method’s performance for artificial 4D-CT images described quantification results of real 4D-CT images. Results showed that lung movement is not modelled by simple translation but by oval pattern. A new algorithm for medical image segmentation with reference to abdominal aortic aneurysm and degraded human brain imaging was presented by Pham et al., [19]. The new algorithm’s development was based on theoretic distance matrix implementation with spatial semi-variances. An alternative, approach to use energy sensitive CT imaging techniques, was proposed by Ghadiriet al., [20]. To accurately validate the new method, a new bone model based on cortical and marrow mixtures was proposed. A two-step energy mapping algorithm was implemented. Phantom tomographic projections in 5 energy bins were acquired and reconstructed for validation. The new energy mapping technique estimated LAC of different bone tissues at 511 keV. The results had 1.1% error maximum compared to true values. To test precision, 10% variation’s effect on effective energy was investigated. Medical image fusion for merging complementary diagnostic content using PCA and Wavelets was carried out by Krishnet al., [21]. The new fusion approach involved sub-band decomposition using 2D-Discrete Wavelet Transform (DWT) to preserve spectral and spatial information. Also, PCA was applied on decomposed coefficients to maximize spatial resolution. An optimal variant of daubechies wavelet family was selected for better results. Simulation showed improved visual quality in the fused image compared to other state-of-art fusion approaches. An automated tibial eminence extraction in MDCT image using shape matching was proposed by Uozumi et al., [22]. The new method evaluated 6 patients (Age 27 ± 7, four males / two females). Hence, the new method automatically extracted an eminence shape for all patients . 3.   METHODOLOGY This study uses classification process for CT images based on feature extraction using Gaussian wavelet and GLCM. The features extracted are reduced using Information Gain and CFS [32].  International Journal on Soft Computing (IJSC) Vol.8, No. 1, February 2017 17   The Hybrid bacterial foraging optimization (HBFO) is proposed in this paper for feature selection and the selected features are classified using artificial immune classifier. The proposed HBFO for feature selection and artificial immune classifier are discussed in this section. 3.1 Proposed hybrid Bacterial Foraging Optimization (BFO)   Global optimization combines a fast local optimization method with an initial search phase, showing improved robustness compared to local search strategies [26]. It uses randomization and local search to solve an optimization problem. Different meta-heuristic approaches solve the same optimization problem differently [27]. BFO algorithm is a new evolutionary computation algorithm based on Escherichia coli (E. coli) bacteria’s foraging behaviour in the human intestine. The BFO algorithm is a biologically inspired computing technique mimicking E.coli bacteria’s foraging behaviour [28].Natural selection removes animals with poor foraging strategies favouring the circulation of genes of animals with successful foraging strategies, as they are more likely to enjoy reproductive success. After generations, poor foraging strategies are removed or shaped into good ones. Foraging is used in optimization which is explained below: Chemotaxis: This simulates movement of an E.coli cell by swimming and tumbling via flagella. Biologically, an E.coli bacterium moves in two ways. It can swim in the same direction or tumble, or alternate between these two modes for an entire lifetime. Swarming: Interesting group behaviour was observed in many mobile bacteria species including E.coli and S. typhimurium, where stable spatio-temporal patterns (swarms) are formed in a semisolid nutrient medium [29]. Reproduction: least healthy bacteria eventually die while healthier bacteria (those yielding higher fitness function values) asexually split into two bacteria in the same location keeping the swarm size constant. Elimination and dispersal: Gradual / sudden changes in environment where a bacterium population lives occurs for various reasons; e.g. significant rise in local temperature may kill a group of bacteria that are in a region with a high nutrient gradients concentration. Events take place so that all bacteria in a region are killed, or a group   disperses to new locations. To simulate this, some bacteria are liquidated at random with very small probability while new replacements randomly initialize over search space. Chemotaxis is the basis for local search, and reproduction speeds up convergence simulated by a classical BFO. To a large extent, Chemotaxis and reproduction are not enough for global optima searching. As bacteria may get stuck in initial positions or local optima, it is possible for BFO diversity to change gradually or suddenly to eliminate being trapped in local optima [30]. Dispersion in BFO happens after certain reproduction processes. Then some bacteria are chosen, according to a preset probability Ped, to be killed and moved to another position in the environment. T  he algorithm’s steps are as follows [31]:
Related Search
Related Docs
View more...
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks