of 11

IRJET- Computer Assisted Lung Nodule Detection in Digital Chest Radiographs - A Survey

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.
International Research Journal of Engineering and Technology (IRJET)e-ISSN: 2395-0056Volume: 06 Issue: 06 | June 2019p-ISSN: 2395-0072www.irjet.netCOMPUTER ASSISTED…
International Research Journal of Engineering and Technology (IRJET)e-ISSN: 2395-0056Volume: 06 Issue: 06 | June 2019p-ISSN: 2395-0072www.irjet.netCOMPUTER ASSISTED LUNG NODULE DETECTION IN DIGITAL CHEST RADIOGRAPHS - A SURVEY Prashant A. Athavale1, Arpitha S. A2, B. L. Ashritha3, Latha N4, Niharika S5 1PrashantA. Athavale : Professor, Dept. of Electrical and Electronics Engineering, BMSIT&M college, Karnataka, India 2,3,4,5Student, Dept. of Electrical and Electronics Engineering, BMSIT&M college, Karnataka, India ---------------------------------------------------------------------***---------------------------------------------------------------------Abstract - This survey paper explains about the various methods for feature extraction and detection of pulmonary nodules in digital chest radiographs (CXR)s, as an early and accurate detection of the nodules could be saving lives. This survey focuses on various techniques which are used to spot and classify the lung nodules which in turn will assist the domain experts for better diagnosis. In most of the surveyed papers, four stages of works are carried out such as: (i) Image Acquisition (ii) Image Pre-processing (iii) Image Segmentation and (iv) Feature Extraction. Since the volume of the X-rays is very large, Computer Aided Detection/Diagnosis (CAD/x) has more advantages in addition to manual interpretation with respect to speed and accuracy. This paper aims at summarizing various methods that have been introduced by several authors over the years of research in this field. Key Words: Lung, Solitary pulmonary nodule, Thresholding, Segmentation, Feature extraction. 1. INTRODUCTION Cancer is one of the deleterious diseases that a person can contract. It is extremely difficult to treat, and can take a toll on the psyche of the sufferer. The lung cancer is the major cause of the cancer related death in India [1]. The report says that India, with 18% of the worlds’ population also has 32.0% of the global burden of COPDs. In India approximately 63,000 lung cancer cases are reported every year. Early diagnosis of lung cancer is the key to provide the best possible clinical treatment for patients. It can be diagnosed at an early stage by regular health screening. As an initial diagnostic tool for a variety of clinical conditions, Chest X- Ray (CXR) is the most commonly used radiological examination in health screening by far making up at least one third of all examinations in a typical radiology department [2], [3]. Radiologists define a lung nodule in chest x-ray radiographs as “solitary white nodule-like blob” [4], [5]. It is usually round or oval in shape. This definition contains two different descriptions, i.e. “white nodule-like blob” and “solitary”. The “white nodule-like blob” has been represented by many texture features. Early researches about lung nodule detection used the difference of candidates’ shape under various thresholds as the features to identify nodules from other candidates [6]–[8]. Present day research about lung nodule detection [9]–[13] added gradient features (including intensity and direction) and texture features to identify lung nodules from pre-detected candidates. Wei et al [9] extracted 210 texture features for lung nodule representation. Performing chest radiography is the primary step in the diagnosis of nodule, if a patient shows symptoms such as: long-term tobacco smoking, exposure to radon gas, asbestos, second-hand smoke, or other forms of air pollution and often caused by a combination of genetic factors that may suggest lung cancer. Because of its simplicity, it gives clear images of bones, is non-ionizing so aren't dangerous on developing fetuses and are cost effective. Despite its advantages, interpreting abnormalities in a CXR is difficult. Sometimes even radiologists can fail to detect nodules on CXR, because of its complexities like: 1) Nodule appearance in radio graphs varied in size which ranges from few millimeters to several centimeters. 2) Some nodules are slightly denser than the surrounding lung tissue (less visible). 3) Nodules can appear anywhere in the lung field, and can be obscured by ribs and structures below the diaphragm and heart (CXRs are projection images contains superimposed structures). Amid various non-invasive medical imaging modalities such as CT scan, MRI and Ultrasound, X-ray is commonly used as initial diagnostic tool for detection of nodule. Computed tomography (CT) is another widely used radiological examination. CT scan is not feasible due to its high cost and high dose. However, MRI is more expensive than x-ray and is less detailed than x-ray. The CXR has low noise when compared to CT and MRI images. Nodules are found in 1 out of every 4 CXRs. Most nodules (more than 90%) are benign and not cancerous which are caused by previous infections or old surgery scars. Characteristics of nodule: Most nodules that exceed 3 cm in diameter are malignant. Nodules below 1 cm are less likely to be cancerous although in some cases, there is a risk of malignancy if the person is a smoker. A nodule that is rough (spiculated irregular margin with lines) is more likely to be cancerous. Malignant lesions are usually partly solid whereas benign tumors are generally solid. Calcified nodules are less likely to be cancerous.© 2019, IRJET|Impact Factor value: 7.211|ISO 9001:2008 Certified Journal|Page 467International Research Journal of Engineering and Technology (IRJET)e-ISSN: 2395-0056Volume: 06 Issue: 06 | June 2019p-ISSN: 2395-0072www.irjet.net1.1 DIFFERENT STAGES IN DETECTION OF LUNG NODULES ON CXRs 1. Image Acquisition: The CXRs are acquired from the digital X-ray machines, with the main advantage being better clarity, low noise and distortion. 2. Image Preprocessing: Image pre-processing is a way to improve the quality of image and remove the irrelevant information, so that the consequential image is better than the original one. Contrast, brightness and intensity problem are removed by using contrast stretching, histogram equalization, negativity and power law transformation etc. Using modified thresholding, labeling algorithm and edge detection, segmentation of the lung nodule X-ray image is carried out. Features such as geometrical, textural and mathematical properties are used as input to the Computer Aided Diagnostic (CAD) system. Present CAD system will not reduce the role of radiologist but it will provide a second opinion. The result of this system along with the analysis of the diagnostician will increase the accuracy of the diagnosis. Various types of selective enhancement filters are used to enhance blob like structures and to suppress vessel like structures by [14] [15] [16] [17] and [18] recommended a selective enhancement filter to enhance dot like objects and to repress lung vessels. Cylindrical and spherical filters were combined for a better visualization of nodules by [19]. The Laplacian of Guassian (LoG) filter is preferred in enhancing blob like structures whose intensity is differs from that of background. [20] used LoG filters to enhance the input image. [21] and [22] recommended LoG filter for enhancement. Median filter: Median filter was used by [23] and [24] to remove the noise from the image. The advantage of this filter is that it doesn’t distort the edges. 3. Image segmentation: Image segmentation (division of an image into its constituent region or object) is an important process that is required to perform eventual tasks for image analysis. The goal of segmentation is to make simpler or change the representation of an image into more meaningful and easier to analyze. The segmentation algorithms either partition an image based on abrupt change in intensity, such as edges, or partitioning an image into regions that are similar according to a set of pre-defined criteria. Hence segmentation is based on the two basic properties of intensity values: discontinuity and similarity. 4. Feature Extraction: In the present study, image feature extraction is very important stage of computer based system. This extraction provides certain parameters, on the basis of which computer system takes decision. The entire feature which are calculated from the image, convey information regarding lung nodule. In this literature, the features extracted from the X-ray image can be used as diagnostics indicators: Area, Perimeter, Irregularity index, Equivalent diameter, Gray-level Co-occurrence Matrix properties, Solidity, Contrast. Frequently used methods and some of the earlier studies for detection of lung nodules are: A. Pixel Based White Nodule-Likeness Map Extraction: 1) Stationary Wavelet Transform, 2) Convergence-Index Filter, 3) AdaBoost Based White Nodule-Likeness Extraction. B. Solitary Degree Based Lung Nodule Blob Ranking: 1) Lung Nodule Blob Detection, 2) Solitary Degree Based Blob Ranking, 3) SVM Based Blob Classification.Š 2019, IRJET|Impact Factor value: 7.211|ISO 9001:2008 Certified Journal|Page 468International Research Journal of Engineering and Technology (IRJET)e-ISSN: 2395-0056Volume: 06 Issue: 06 | June 2019p-ISSN: 2395-0072www.irjet.netC. A feature-based classification approach to nodule detection has shown promising results in preliminary studies. A segmentation algorithm is applied to generate candidate nodules, and then features are calculated and used to classify each candidate as nodule or non-nodule (usually a broncho-vascular structure). Most features relate to size and shape on the premise that nodules are of greater diameter and tend to be spherical, while vessels are tubular. Giger et al have used multiple gray-level thresholds for extracting nodules, and then found 2-D geometric features such as perimeter, compactness and circularity for every nodule at each threshold. A rule-based approach was used to assign a confidence rating, in the range 1–5, to each 2-D candidate, with 1 being definite vessel and 5 being definitely nodule. Confidence ratings were altered based on ratings of the nodule in adjacent slices. Armato et al. included some 3-D features such as sphericity, and gray-level features such as mean and standard deviation, to classify nodules using a linear discriminant analysis. Kanazawa et al. segmented candidates by using fuzzy clustering to partition the histogram of pixels within the lung fields into two classes: “air part” and “blood vessels and tumors.” They then used similar features in a heuristic, rulebased approach to classify nodules and vessels. 2. LITERATURE SURVEY The following section elaborates on some of the papers presented in the literature. Xuechen Li et al [27] proposed an automated method for detecting lung nodules in chest x-ray radiographs using JSRT [67] database. Method shows three steps which employ stationary wavelet transform and convergence index filter to extract the texture feature using Multi resolution method and further used Adaboost to produce white nodule-likeness map using Laplace of Gaussian (LoG) blob detection method. From reported studies of JSRT gives 78% for false positive of 2 and 90% for false positive of 4 above results BJH method showed 80% for false positive of 2 and 93% for false positive 0f 5. Comparing with the best previously reported study, the proposed method obtained equal sensitivity when the FPs is two and much higher sensitivity when FPS is 5 along with a good lung nodule sensitivity. M. Mercy Theresa et al [29] proposed to detect lung nodules in chest radiography for finding and mapping the presence of small nodules using JSRT database. The proposed approach involves 4 stages: Stage 1-Image registration using geometrical transformation method, is performed to correct the GT for the input images. This provides errorless diagnosis. Stage 2- Lung segmentation using thresholding method. Lung region is segmented from x-ray image by conducting thresholding and other reconstruction processes. Stage 3-Feature extraction using CWT and Shearlet transform methods used to extract the feature and then classified as normal & abnormal random forest classification method. Stage 4- Image classification using random forest and classification. RFC has many trees where each tree grown using some form of randomization. All the trees considered are binary tree and in top down structure. Images are tested and classified based on proposed algorithm. Cesar Supanta et al [36] extracted the feature and detected pulmonary nodules in CXRs of JSRT [67] database. This approach involves 4 steps :A) Pre-processing- By comparing original image with correlated image, the changes in the image will make impact on processing and are performed by the gamma correction and re-quantization. B) Lung segmentation performed using OTSU method and projection analysis. C) Identification of pulmonary nodules, are made in the region of interest using convergence and ring filter. D) Extraction of feature-Image obtained from the convergence filter is analyzed for differentiating the nodule from noise and minimize false positive and extracting characteristics. The reported performance indices are: sensitivity (91%), precision (94%) and specificity (96%). Haoyan Guo et al [40] proposed an optimal feature subset used for detecting lung nodules. The two main methods for feature involves: filter and wrapper. Filter is performed before the classification task. The selection of attributes is independent of learning algorithm in filter approach while it is dependent in wrapper approach. Multilevel optimal feature selection (MOFS) eliminates the redundant and irrelevant features from the candidate features. It reduces the computation time. Evaluation is done using bagging construction by reduced features. This model consists of three major steps: 1. Selection of initial nodule candidates which includes 198 RoI’s (6464 pixels) taken from a private collection. 2. Extraction of morphological features based on wavelet snake model. 3. Reduction of false positives based on combination of features by artificial neural network and then detecting nodules. This paper establishes the accuracy of the ensemble using a MOFS method over a single CAD system. Vinod Kumar et al [46] proposed a method for diagnosing the X-Ray image using statistical features. This method uses filters, segmentation, threshold and edge detection approaches. This is classified into 5 stages. First stage is Image segmentation to group the similar characteristic image parts using thresholding. Second stage is median filter used for noise reduction and also replaces the worth of a pixel by the median of the intensity level within the surrounding of that pixel. Third stage is edge detection. This technique reduces the irrelevant information which will be used for further© 2019, IRJET|Impact Factor value: 7.211|ISO 9001:2008 Certified Journal|Page 469International Research Journal of Engineering and Technology (IRJET)e-ISSN: 2395-0056Volume: 06 Issue: 06 | June 2019p-ISSN: 2395-0072www.irjet.netprocessing. Fourth stage is Region of Interest. This outlines the boundaries and also limits the cancerous cells during detection process. Finally the detection is done with the statistical features which include: energy, entropy, mean, variance, standard deviation, skewness and kurtosis. The result shows that the skewness is high which means feel of respiratory organ becomes symmetry before the energy increases and in turn entropy decreases. The quality of the gray level pixels within the X-Ray image increase gray level. The tissue of respiratory organ becomes lumpy which makes detection easier. Zaw Zawhtike et al. [43] performed a three layer work to detect nodule in chest X-ray. Images were collected from JSRT [67] database. The first layer executes pre-processing of CXRs by Laplacian filter followed by second layer extract texture feature based on gray-level co-occurrence matrix. Third layer checks whether image contains nodules or not using rotation forest technique. Detection of nodule is done with the help of 44 dimensional GLCM texture feature and rotation forest. The author reports an accuracy of 75.6% in detecting the nodule. B. Nemade et al [48] proposes an advanced computerized scheme for detection of lung nodules, incorporating the CAD scheme on VDE images. The proposed method involves the following steps: A) the images collected were in the form of DICOM images. B) The images were enhanced using histogram equalization in order to obtain the images with better contrast. C) The feature was extracted on considering the Region of Interest. D) With the help of Support Vector Machine classifiers the nodules were classified as benign, or malignant. Due to the effective classification, the false positives in detecting the nodules were found to be reduced. Since this method is cost effective, it can be used as primary tool in detecting the lung nodules. Arnold M.R Schilham et al [51] proposed the method of Multiscale nodule detection in chest radiographs. The objective of this paper in the chest radiographs used computer algorithm to detect nodule as taken into account of wide size range of lung nodule with the help of multiscale image processing technique on JSRT CXRs. This method consists of following stages are: 1.The CXRs are subsampled to 1024X1024 pixels. 2. Segmentation of lung nodule, extracted with an active shape model. 3. The preprocessing stage of blob detector, in which contrast image is locally enhanced with local normalization filter (LN). 4. To find blob in lung field in the LN image by using Lindeberg’s multiscale blob detector method. 5. Based on feature with simple classification that result blob detector in which number of nodule are reduced.6. After this, finally a selected set of feature to find the probability that a candidate shows nodule with the help of K nearest neighbor (KNN) classifier from this complete JSRT database shows that as average of 2 false positive per image, but gives 50.6% of all nodules are detected, with some more false positive increased to 10 gives 69.5%. Changmiao Wang et al [53] propose using deep feature fusion method from the non-medical training and hand crafted feature in order to reduce false positive value. Based on performing public dataset, this method can obtain a result in terms of sensitivity and specificity (69.3% & 96.2%) then false positive per image at 1.19. But in case of hand crafted features (62% & 95.4%), this reduce false positive per image at 1.45 on made use of machine learning. Deep fusion feature method improves current CAD scheme and more effectively detect the presence of lung nodule. Haoyan Guo et al [60], proposes a practical feature selection approach that is based on an optimal feature subset of a single CAD system, which is referred to as a multilevel optimal feature selection method (MOFS). CAD systems consist of three major phases: 1) initial nodule candidate selection from lung nodule-enhanced images,2) extraction of features from these nodule candidates and nodule feature subset selection, and 3) discrimination of nodules from false positives based on these feature. Thus, the different optimal feature subsets are selected in order to eliminate features that are irrelevant and obtain optimal features. And also it is seen that the accuracy and sensitivity obtained is high. S. Savithri et al [28] proposed three methods to detect the nodule region from PA chest radiograph images which employs JSRT [67] database. Method 1 extracts the nodule region from PA chest radiographic image and detects geometrical features for nodule region. This Method involves: preprocessing, lung field segmentation and nodule extraction. Lung region is segmented from the original image. Based on the region growing, thresholding and morphological operations are applied to extract nodule region from the image. Method 2 involves enhancement and segmentation, thresholding and nodule detection (m
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