Classification of Osteoporosis using Fractal Texture Features

Transcription

1 Classification of Osteoporosis using Fractal Texture Features V.Srikant, C.Dines Kumar and A.Tobin Department of Electronics and Communication Engineering Panimalar Engineering College Cennai, Tamil Nadu, India Article History ABSTRACT Received on: Accepted on: In our proposed metod an automatic Osteoporosis classification Publised on: system is developed. Te input of te system is Lumbar spine digital radiograp, wic is subjected to pre-processing wic includes conversion of grayscale image to binary image and enancement using Keyword Osteoporosis Contrast Limited Adaptive Histogram Equalization tecnique(clahe). Osteopenia Furter Fractal Texture features(sfta) are extracted, ten te image is Lumbar Spine classified as Osteoporosis, Osteopenia and Normal using a Probabilistic CLAHE Neural Network(PNN). A total of 158 images ave been used, out of SFTA PNN wic 86 images are used for training te network and 3 images for testing and 40 images for validation. Te network is evaluated using a confusion matrix and evaluation parameters like Sensitivity, Specificity, precision and Accuracy are computed fractal feature extraction tecniques. Copyrigt 015 International Journal for Modern Trends in Science and Tecnology All rigts reserved. Classification of Osteoporosis using Fractal Texture Features Page 11

2 I. INTRODUCTION Osteoporosis is a progressive bone disease tat is caracterized by a decrease in bone mass and density wic can lead to an increased risk of fracture. In osteoporosis, te Bone Mineral Density (BMD) is reduced, bone micro arcitecture deteriorates, and te amount and variety of proteins in bone are altered. Te form of osteoporosis most common in women after menopause is referred to as primary type 1 or postmenopausal osteoporosis, wic is attributable to te decrease in oestrogen production after menopause. Primary type osteoporosis or senile osteoporosis occurs after age 75 and is seen in bot females and males at a ratio of :1. Secondary osteoporosis may arise at any age and affect men and women equally, tis form results from cronic predisposing medical problems or disease, or prolonged use of medications suc as glucocorticoid, wen te disease is called steroid or glucocorticoid-induced osteoporosis. So in order to detect te presence of osteoporosis we ave to use image processing tecniques on Digital radiograpers contribute to te test values used for training te neural network. Tus Image processing is used to monitor te traits of te X- rays for possible infirmities tat migt occur and provide te necessary data for furter treatment. II. LITERATURE SURVEY According to AlceuFerraz Costa[1], a new and efficient texture feature extraction metod called te Segmentation-based Fractal Texture Analysis, or SFTA is preferred tan Haralick and Gabor filter. Te extraction algoritm consists in decomposing te input image into a set of binary images from wic te fractal dimensions of te resulting regions are computed in order to describe segmented texture patterns. Te decomposition of te input image is acieved by te Two- Tresold Binary Decomposition (TTBD) algoritm, wic we also propose in tis work. Te SFTA is evaluated for te tasks of content-based image retrieval (CBIR) and image classification, comparing its performance to tat of oter widely employed feature extraction metods suc as Haralick and Gabor filter banks. SFTA acieved iger precision and accuracy for CBIR and image classification. Additionally, SFTA was at least 3.7 times faster tan Gabor and 1.6 times faster tan Haralick wit respect to feature extraction time. Histogram equalization, wic stretces te dynamic range of intensity, is te most common metod for enancing te contrast of an image. An adaptive metod to avoid tis drawback is blockbased processing of istogram equalization. In block-based processing, image is divided into subimages or blocks, and istogram equalization is performed to eac sub-images or blocks. Contrast Limited Adaptive Histogram Equalization (CLAHE), proposed by K. Zuierveld, as two key parameters: block size and clip limit. Tese parameters are mainly used to control image quality. In tis paper, a new novel metod was proposed by ByongSeok Min [] to determine two parameters of te CLAHE using entropy of an image. An Automatic Brain Tumor Classification using PNN and Clustering developed by P.Sangeeta [3] describes te Probabilistic Neural Network (PNN) will be employed to classify te various stages of Tumor cut levels suc as Benign, Malignant or Normal. Probabilistic Neural Network wit Radial Basis Function will be applied to implement tumor cells segmentation and classification. Decision sould be made to classify te input image as normal or abnormal cells. Prediction of malignant cells or non-tumor cells can be executed using two variants: i) Feature extraction and ii) classification using Probabilistic Neural Network (PNN). Te ability of teir proposed Brain Tumor Classification metod is demonstrated on te basis of obtained results on Brain Tumor image database. In teir proposed metod, only 5 classes of Brain tumors are considered, wit respect to an example of 0 test images for instance but tis metod can be extended to more classes of Brain tumors. Since medical X-Ray images are grayscale images wit almost te same texture caracteristics, conventional color or texture features cannot be used for appropriate categorization in medical X- Ray image arcives. Terefore, a novel feature is proposed by Seyyed Moammad Moammadi [4] wic is te combination of sape and texture features. Te feature extraction process is started by edge and sape information extraction from original medical X-Ray image. Finally, Gabor filter is used to extract spectral texture features from sape images. In order to study te effect of feature Classification of Osteoporosis using Fractal Texture Features Page 1

3 fusion on te classification performance, different effective features like local binary pattern. It provides low computation complexity and straigtforward implementation. Due to following advantages we can strongly claim tat tese features are te most powerful and reliable features for medical image X-Ray classification. III. PROPOSED METHOD Te proposed approac starts first from preprocessing. It is ten followed by Grayscale Conversion and te image is enanced using a CLAHE filter. Ten te features namely fractal features are extracted from te enanced image and te fractal features are used to determine te category in wic tey fall determined by te PNN classifier for one sample. Tis is repeated for te stock of samples in wic /3 of te samples are taken as Database samples. Metodologies A. Pre-processing B. Grayscale conversion C. Image Enancement. D. Feature Extraction E. PNN Classifier. Test Image Training enancement and resizing. Tese are done in order to make te image more suitable tan an original image for specific applications. B. Grayscale Conversion If te image selected is in tree dimensions, it is converted to a grayscale image using rgbgray conversion command for feature extraction. Ten te intensity variation of gray level image is sown in te grap. Its value varies from 0 to 55.Resizing of image is done for accurate processing of image. Image can be resized to any size of our interest. C, Image Enancement Image enancement is te process of adjusting digital images so tat te results are more suitable for display or furter image analysis. After resizing te image, we go for image enancement. We use CLAHE (Contrast Limited Adaptive Histogram Equalization) tecnique. Tis metod separates te image into a number of tiles, and ten adjusts te contrast suc tat te tile istogram as te desired sape. Te tiles are ten stitced togeter using bilinear interpolation. Te transformation function modifies te pixels based on te gray level content of an image. Tese tecniques are used to enance details over small areas in an image. Input X-Ray image Input X-Ray image Processing(Image Enancement) Pre- Pre- Processing(Image Enancement ) Figure 1. Image after Enancement Feature Extraction( SFTA Features) Classification (PNN Classifier) Normal Osteopenia Feature Extraction(SFTA Features) Features Database Osteoporosis Figure 1.1 Block Diagram of proposed work IV. RESEARCH METHODOLOGY A. Preprocessing Steps wic are done prior to processing of an image are called preprocessing. It includes image D. Feature Extraction Feature is a parameter of interest to describe an image. Transforming te input data into te set of features is called feature extraction. If te features extracted are carefully cosen it is expected tat te features set will extract te relevant information from te input data in order to perform te desired task. Te Segmentation-based Fractal Texture Analysis or SFTA metod is a feature extraction algoritm tat decomposes a given image into a set of binary images troug te application of wat te autors call te Two Tresold Binary Decomposition (TTBD). For eac resulting binary image, fractal dimensions of its Classification of Osteoporosis using Fractal Texture Features Page 13

4 region boundaries are calculated tat describe te texture patterns.ttbd takes an input greyscale image and returns a set of binary images by first computing a set of T tresold values from te gray level distribution information in an input image. Tis is accomplised by recursively applying to eac image region te multilevel Otsu algoritm, an algoritm tat quickly finds te tresold tat minimizes te input image intra-class variance until te desired number of tresolds is obtained. Te input image is decomposed into a set of binary images by selecting pairs of tresolds from T and applying two-tresold segmentation.fractal measurements are used to describe te boundary complexity of objects, wit eac region boundaries of a binary image represented as a border image. Te fractal dimension is computed from eac boarder image using a box counting algoritm. A. Haussdorf fractal dimension Haussdorf dimension serves as a measure of te local size of a set of numbers, taking into account te distance between eac of its members. Te Haussdorf dimension of an n-dimensional inner product space equals n. Tis underlies te earlier statement tat te Haussdorf dimension of a point is zero, of a line is one, etc., and tat irregular sets can ave non integer Haussdorf dimensions. Haussdorf fractal dimension of an object represented by te binary image. Non-zero pixel belongs to an object and zero pixel constitute te background. dim H X = inf {d 0: C H d X 0} (1) Were dim _H(X) is te infimum of te set of d [0, ) suc tat te d-dimensional Haussdorf measure of X is zero. B. OTSU S Segmentation Otsu's metod is used to automatically perform clustering-based image tresolding i.e., te reduction of a graylevel image to a binary image. Te algoritm assumes tat te image contains two classes of pixels following bi-modal istogram, it ten calculates te optimum tresold separating te two classes so tat teir combined spread is minimal.in Otsu's metod, we exaustively searc for te tresold tat minimizes te intra-class variance (te variance witin te class), defined as a weigted sum of variances of te two classes. σ w t = ω 1 t σ 1 t + ω t σ t () Weigts ω_i are te probabilities of te two classes separated by a tresold t and σ_i^variances of tese classes. Otsu tresolding returns a set of tresolds for te input image employing te multilevel Otsu algoritm. Te multilevel Otsu algoritm consist in finding te tresold tat minimizes te input image intraclass variants. Ten, recursively, te Otsu algoritm is applied to eac image region until total tresolds are found. Figure 1.3 Image after OTSU Segmentation C. Edge detection Edge detection is an image processing tecnique for finding te boundaries of objects witin images. It works by detecting discontinuities in brigtness. Edge detection is used for image segmentation and data extraction in areas suc as image processing, computer vision, and macine vision.edge detection returns a binary image wit te regions boundaries of te input image. Te input image must be a binary image. Te returned image takes te value 1 if te corresponding pixel in te image as te value 1 and at least one neigbouring pixel wit value 0. Oterwise takes value 0. E. Artificial Neural Network Artificial neural networks (ANNs) are a family of statistical learning algoritms inspired by biological neural networks. Tey are used to estimate or approximate functions tat can depend on a large number of inputs and are generally unknown. Artificial neural networks are generally presented as systems of interconnected "neurons" wic can compute values from inputs, and are capable of macine learning, as well as pattern recognition tanks to teir adaptive nature. After being weigted and transformed by a function (determined by te network's designer), te activations of tese neurons are ten passed on to oter neurons. Tis process is repeated until finally, an output neuron is activated. A.PNN classifier PNN is a useful neural network arcitecture wit Classification of Osteoporosis using Fractal Texture Features Page 14

5 sligtly different in fundamentals from te back propagation. Te arcitecture is feed forward in nature wic is similar to back propagation, but differs in te way tat learning occurs. PNN is supervised learning algoritm but includes no weigts in its idden layer. x1 x x Input nodes Hidden nodes Class nodes Decision node Figure 1.4 Block Diagram of PNN Classifier. Basically, PNN consists of an input layer, wic represents te input pattern or feature vector. Te input layer is fully interconnected wit te idden layer, wic consists of te example vectors (te training set for te PNN). Te actual example vector serves as te weigts as applied to te input layer. Finally, an output layer represents eac of te possible classes for wic te input data can be classified. Te output class node wit te largest activation represents te winning class. In PNN algoritm, calculating te class-node activations is a simple process. For eac class node, te example vector activations are summed, wic are te sum of te products of te example vector and te input vector. Te idden node activation, sown in te following equation, is simply te product of te two vectors (E is te example vector, and F is te input feature vector). i = E i F (3) Te class output activations are ten defined as: c1 c z cosen troug experimentation. If te smooting factor is too large, details can be lost, but if te smooting factor is too small, te classifier may not generalize well. It's also very easy to add new examples to te network by simply add te new idden node, and its output is used by te particular class node. Tis can be done dynamically as new classified examples are found. Te PNN also generalizes very well, even in te context of noisy data. V. RESULTS AND DISCUSSION Here, totally 158 images ave been used out of wic 86 are taken for training and remaining ave been used for testing and validation. Calculated feature values of various input X-Ray images are tabulated. Results sow tat images are normal, are Osteopenia and osteoporosis. Comparison of various features for normal, Osteopenia and osteoporosis images can be seen. Class Table 1.1 Table Sowing Result Set Number Of Images Normal 55 Osteopenia 46 Osteoporosis 57 Figure 1.5 GUI Result sows Normal c j = ( i 1) N e y i=1 N (4) Were N is te total number of example vectors for tis class, i is te idden-node activation, and γ is a smooting factor. Te smooting factor is Figure 1.6 GUI Result sows Osteopenia Classification of Osteoporosis using Fractal Texture Features Page 15

6 Figure 1.7 GUI Result sows Osteoporosis A. CONFUSION MATRIX Confusion matrix is a specific table layout tat allows visualization of te performance of an algoritm, typically a supervised learning one. Eac column of te matrix represents te instances in a predicted class, wile eac row represents te instances in an actual class. Evaluation parameters suc as Sensitivity, Specificity, Precision and Accuracy are calculated for te confusion matrix. Table 1. Tree Class Confusion Matrix of SFTA Table 1.3 Table Sowing Validation result of SFTA Parameter Formula SFTA VI. CONCLUSION AND FUTURE WORK In tis work, te suitability of texture features in classification of Lumbar spine digital radiograps is analyzed. Te Experimental results during testing and teoretical analysis prove in perspective of time, accuracy of system being a main concern te fractal features are more precise and more accurate. In medical imaging or in diagnosis, te important factor is accuracy rater tan speed, ence present fractal features as a more suitable tecnique for feature extraction. Te results on classification ave a combined accuracy of 93%. Tis value is obtained by validating our practical results wit DEXA results. Hence our system can assist in diagnosis of osteoporosis. In future te work can be implemented for distal and toracic digital radiograp, te system can also be embedded in a Digital radiograpy macine to diagnose osteoporosis on te fly. REFERENCES [1] AlceuFerraz Costa, Gabriel Humpire-Mamani andagmajuci Macado Traina, (01) An Efficient Algoritm for Fractal Analysis of Textures, Conference on Grapics, Patterns and Images (SIBGRAPI), pp [] ByongSeok Min, Dong Kyun Lim, Seung Jong Kim and Joo Heung Lee, (013) A Novel Metod of Determining Parameters of CLAHE Based on Image Entropy, International Journal of Software Engineering and Its Applications, Vol.7, No.5, pp [3] Sangeeta. P, (014) Brain Tumor Classification using PNN and Clustering, International Journal of Innovative Researc in Science, Engineering and Tecnology, Vol. 3, No.3, Mar 014. [4] Seyyed Moammad Moammadi, Moammad SadegHelfrous and Kamran kazemi, (01) Novel sape-texture feature extraction for medical X-Ray image classification, International Journal of Innovation Computing, Information and Control (ICIC), Vol.8, No.1, ISSN Sensitivity TP/ (TP+FN) 88.% Specificity TN/ (FP+TN) 94.% Precision TP/ (TP+FP) 97% Accuracy (TP+TN)/ (TP+TN+FP+FN) 94.3% Classification of Osteoporosis using Fractal Texture Features Page 16