Issuu

International Research Journal of Engineering and Technology (IRJET)

e-ISSN: 2395-0056

Volume: 07 Issue: 01 | Jan 2020

p-ISSN: 2395-0072

www.irjet.net

Classifying Chest Pathology Images using Deep Learning Techniques Vrushali Dhanokar1, Prof. A.S. Gaikwad2 1M.tech Student, Deogiri College of Engineering and Management studies, Aurangabad. 2Professor, Deogiri College of Engineering and Management studies, Aurangabad.

----------------------------------------------------------------------***--------------------------------------------------------------------Abstract - Chest radiographs are the most common examination in radiology in today’s era. They are essential and very helpful for the management of various diseases associated with high mortality and display a wide range of potential information about various diseases, many of which is subtle. Most of the research in computer-aided detection and diagnosis in CNN in chest radiography has focused on lung nodule detection. Although the target of most research attention, lung nodules and chest are a relatively rare finding in the lungs. The most common findings in chest X-rays include lung infiltrates and Cardiomegaly, Mediastinum of the size of the large heart. Distinguishing the various chest pathologies is a difficult task even to the human observer and also for radiologist. Therefore, there is an interest in developing computer system diagnosis to assist radiologists in reading chest images through machine. The healthy versus pathology detection i.e. Tuberculosis and Cardiomegaly in chest radiography was explored using Laplacian of Gaussian (LoG), Local Binary Patterns (LBP), Speed up Robust Features (SURF) and also used the Bag-of-Visual-Words (BoVW) model using Artificial Neural Network (ANN) technique that discriminates between healthy pathological cases. Key Words: Artificial Neural Network, Chest Radiographs, Deep Learning, Feature Selection, Image Enhancement. 1. INTRODUCTION Chest radiography is a preferred examination of radiology. They are critical for the management of diseases related to high mortality and show a variety of capability records, maximum of which might be sensitive data. Most research on pc-assisted detection and analysis in chest radiography has focused on lung tumor detection. Although the intention of maximum researches, the hobby of acne lumps is rare in the lungs. The maximum not unusual locating in chest X-rays, together with the infiltration of the lungs, catheters and abnormalities in the size or shape of the heart guy. Therefore, there's an hobby in developing system diagnostics to help radiologists in reading chest pictures. Deep neural networks have acquired extremely good interest because of the improvement of new species CNN and the development of green parallel software program that is appropriate for modern GPUs. Therefore, there is a passion in growing system diagnostics to help radiologists in studying chest photos. Deep neural networks have obtained awesome interest because of the improvement of latest species CNN and the improvement of efficient parallel software that is appropriate for present day GPUs. (CNNs) that show intermediate and advanced abstracts obtained from raw data (such as images) [2]. The latest results indicate that the general explanations drawn from CNN are very effective in object recognition and are now leading technology. [3, 4] Deep learning methods are most effective when used with size training sets. Big in the medical field, large data sets are often not available. Preliminary studies can be found in the medical field that uses deep architectural methods [5,6]. However, we are not aware of any work that uses general training sets that are not physicians to identify medical imaging tasks. Moreover, we are not aware of deep architectural methods for specific tasks of pathological examination in chest radiography. In this work, we examine the strength of the deep learning method for pathological examination in chest radiography. We also explore the classification of health and pathology, which is an important screening mission. In our experiments, we explored the possibility of using neural networks (CNNs) learned from Image Net, a large medical non- medical database for medical image analysis. Neural networks have progressed at a remarkable rate, and they have found practical applications in various industries. [1] Deep neural networks define input for output through complex elements of layers that present building blocks, which include To conversion and nonlinear functions [2]. Now, deep learning can solve problems that are difficult to solve with traditional artificial intelligence [3]. A. You can use unlabeled data during training. Therefore, it is highly suitable to deal with different information and information to learn and receive knowledge. [4]. In this work we utilize the strength of deep learning approaches in a wide range of chest-related diseases and in addition we explore of a pre-trained CNN [8] that is learned from a large scale real life and non-medical image database (ImageNet [6]). Deep learning methods are most effective when applied on large training sets. However, since such large data are generally not available in the medical domain, we explore the feasibility of using a deep learning approach based on non-medical learning. We also apply a feature selection technique to the deep learning representations and show that it can improve the performance in our task. Preliminary studies can be found in the medical field that uses deep architecture methods (e.g., [12], [3]). In our earlier work [2] we showed initial results on 3 pathologies in a small image dataset. We used a leave-one-out (LOO) cross validation method and explored the combination of low-level features (GIST) with features extracted from deep architecture network. In the current work, we have an augmented dataset that includes 6 pathologies. We explore feature selection on a large set of features are the more informative for the task. Finally, we provide more clinically relevant results using a separate training and test set scenario. © 2020, IRJET

Impact Factor value: 7.34

ISO 9001:2008 Certified Journal

Page 1500

International Research Journal of Engineering and Technology (IRJET)

e-ISSN: 2395-0056

Volume: 07 Issue: 01 | Jan 2020

p-ISSN: 2395-0072

www.irjet.net

1.1 Literature Survey In order to examine the use of deep learning in medical diagnosis, 263 articles published in the domain have been analyzed. The data collection process consists of extensive research of articles that discuss the application of depth learning in the medical field. These articles were downloaded and analyzed to obtain sufficient theoretical information on the subject. The results in this article are of natural quality and the main focus is to review the use of profound learning and to answer the research questions outlined in the Introduction section of this article. In summary, the data collection process was conducted in four main steps: Phase 1: Finding reliable journal articles including using keywords displayed under section 2.4 of this article. At this point, the article is thoroughly analyzed. Phase 2: Literature analysis and article separation that are not suitable for eligibility criteria because there is no special screening during the search process. At this point, the article has been analyzed and selected for further analysis. Phase 3: Detailed analysis of eligible articles and qualitative data classified by purpose of review At this stage, it is possible to have a bias towards writing and make a clear research article. Phase 4: Qualitative data received and recorded to present information in the results of this article concisely Information is collected in the notes, forms and records of the type of information and methods used and which applications. The strength of the deep network is to learn how to display multiple levels of concepts that are consistent with the abstract level. For a set of data that sees low-level abstraction, it may explain the edge in the image, while the high layer in the network refers to the object's part and even the type of object that looks at CNN. The deep layer, at the middle layer, is received when entering properties created by the old layer and passing the output to the next layer. The two popular options are CNN. Recommended by [7] and [8] for Take the challenge of ImageNet's large image recognition [9]. ImageNet is a comprehensive, lifelike, large- scale image database consisting of approximately 15 million images with more than 20,000 categories (such as musical instruments, instruments, fruits). A few layers that learn persuasion interleaved with nonlinear operations and profit combinations, followed by layers in space or fully connected. Properties had drawn from the middle layer of these networks results in image classification work. [3] In this work, we test the capabilities of in-depth learning networks in the detection of chest pathology. We separate the different details and compare among them Our main explanations were pulled out using CNN's Decaf Implement [10] which follows CNN in [7] CNN in[10] receiving more than one million subsets of images from ImageNet. Is 1,000 categories. Use the symbol of [10] to display the activation of the hidden layer n of the network received as Decaf n, Class 5 (Decaf5), Tier 6 (Decaf6) and Class 7 (Decaf7) Isolated. DeCAF5 consists of 9216 activation of the final layer and is the first set of activation that has been completely published. In this work we can explore the outcomes of various parameters on system Performance, and show satisfactory consequences using LBP (Local Binary Pattern) classifier. We will make use of the strength of text, interior, exterior layers, and Features Deep learning approaches in a wide range of chest-related Sicknesses. 3. We will even explore categorization of healthful versus pathology that is important screening mission.

Impact Factor value: 7.34

ISO 9001:2008 Certified Journal

Page 1501

Sr. No 1

International Research Journal of Engineering and Technology (IRJET)

e-ISSN: 2395-0056

Volume: 07 Issue: 01 | Jan 2020

p-ISSN: 2395-0072

www.irjet.net

Year

Author-Title

Approach

Metrics

Dataset

April2010

B. van Ginneken, L. Hogeweg, and M. Prokop, "Computeraided diagnosis in chest radiography: Beyond nodules," J.M. Carrillo-de-Gea, G. García -Mateos, "Detection of Normality / Pathology on Chest Radiographs using LBP," U. Avni, H. Greenspan, E. Konen, M. Sharon, and J. Goldberger, "X-ray categorization and retrieval on the organ and pathology level, using patch-based visual words," U. Avni, H. Greenspan, and J. Goldberger, "X-ray categorization and spatial localization of chest pathologies,"

Using CAD development technique worked upon interstitial Infiltrates, catheter tip detection, size measurement, detection and quantification of emphysema.

Precision, Recall

DICOM, Imagenet

The main novelty of our contribution is the application of transformation of Local Binary Pattern to these areas and LBP Histograms are used as input features for a classification system which is ultimately responsible for the decision of normality pathology.

Specificity, Sensitivity

DICOM, Imagenet

The system discrimination between healthy and pathological cases, and is also shown to successfully identify specific pathologies in a set of chest radiograph taken from routine hospital examination.

Precision, Recall

DICOM, Imagenet

The system find between healthy and pathological cases indicates the sub region in the image that is automatically found to the most relevant for the decisions.

Specificity, Sensitivity

DICOM, Imagenet

April2012

March 2014

April 2016

The strength of the deep network is to learn how to display multiple levels of concepts that are consistent with the abstract level. For a set of data that sees low-level abstraction, it may explain the edge in the image, while the high layer in the network refers to the object's part and even the type of object that looks at CNN. The deep layer, at the middle layer, is received when entering properties created by the old layer and passing the output to the next layer. The two popular options are CNN. Recommended by [7] and [8] for Take the challenge of ImageNet's large image recognition [9]. ImageNet is a comprehensive, lifelike, large- scale image database consisting of Approximately 15 million images with more than 20,000 categories (such as musical instruments, instruments, fruits) A few layers that learn persuasion interleaved with nonlinear operations and profit combinations, followed by layers in space or fully connected. Properties drawn from the middle layer of these networks. create highly selective features that achieve effective results in image classification work. [3] In this work, we test the capabilities of in-depth learning networks In the detection of chest pathology. We separate the different details and compare among them. Our main explanations were pulled out using CNN's Decaf Implement [10] which follows CNN in [7] CNN in[10] receiving more than one million subsets of images from ImageNet. Is 1,000 categories. Use the symbol of [10] to display the activation of the hidden layer n of the network received as Decaf n, Class 5 (Decaf5), Tier 6 (Decaf6) and Class 7 (Decaf7). Isolated.DeCAF5 consists of 9216 activation of the final layer and is the first set of activation that has been completely published through the layer. The Convolutional of the DeCAF6 network consists of the activation of the first fully connected layer - such as before propagating through the fully connected layer to create a class prediction. Figure 1 shows the CNN Decaf usage map view.[10] The second basic indicator covered in this work is the "Image Code" indicator (PiCoDes).[9] PiCoDes is a high-level, compact display of low-level features that are (SIFTs, GIST, PHOG, and SSIM) which are optimized for a subset of ImageNet datasets with approximately 70,000 images. For PiCoDes, an offline step is executed, which creates a classification criteria that consists of seeds that learn about the low-level image properties received from the ImageNet subset. The PiCodes then use this classification basis to determine the image. Recognition model for classification of object categories by converting image data using the classification criteria in which the item in the image description is a projection. Situation of low-level image features pulled from the image. This encoding schema yields a binary image descriptor with high performance rates on object category recognition. As a benchmark for our approach, we tested many general explanations. These © 2020, IRJET

Impact Factor value: 7.34

ISO 9001:2008 Certified Journal

Page 1502

International Research Journal of Engineering and Technology (IRJET)

e-ISSN: 2395-0056

Volume: 07 Issue: 01 | Jan 2020

p-ISSN: 2395-0072

www.irjet.net

include local binary format (LBP) [12] and GIST [13]. GIST descriptor, which initially proposed for scene recognition [13], is derived from the orientation data, color and intensity of the histogram. Above different levels and cell division. Classification is performed using SVM with a linear kernel using single-use validation. There are three measurement accuracy checks: sensitivity, specificity and area under the ROC curve (AUC). Sensitivity and specificity are derived from the most appropriate intersection of the ROC. That is the point on the curve that is nearby. Most (0,1) for all properties except binary numbers. Values will be standardized: each column has an average value removed and divided by standard deviation. 1.2 Application of our system • Deep learning practical applications. • Deep learning and medical diagnosis. • Deep learning and MRI. • Deep learning CT. • Deep learning segmentation in medicine. • Deep learning classification in medicine. • Deep learning diagnosis medicine. • Deep learning application medicine. 2. Proposed System We use the strength of deep learning methods in various chest related diseases. We also explore the classification of health and pathology, which is an important screening mission. We explore empirically how to use CNN for these tasks, focusing on CNN that has been trained before, learning from real and non-medical images. CNN consists of a feed-forward family of deep networks, where the intermediate layer is received when entering properties created by the old layer and passing the output to the next layer. The depth of the network is deep in learning. Know the hierarchy layers of concepts that correspond to different levels of abstraction. For low-level image data, abstract things may describe different edges in the image. The middle level may describe various parts of the object, while the high layer refers to the larger part of the object and even the object itself. In this event, we test the ability of the deep learning network to detect chest pathology. We focus on the CNN model that has been practiced before Decaf [12], the CNN adaptation, which follows CNN, which was created by Krizhevsky and the faculty closely.[13] Except for small differences in input data and Cancellation of network breaks into two CNN routes in [12, 13] has been learned through a subset of images from ImageNet [14], which is a comprehensive real-life large image database (> 20M) arranged according to Concepts / Categories (> 10K ) Especially [12] learning CNN in over one million images that are divided into 1,000 categories. To represent an image using the BoVW model, it must be treated as a document, which means that the image is considered to be an image composition. We therefore need to search for these image elements

Figure 1 Proposed System

Impact Factor value: 7.34

ISO 9001:2008 Certified Journal

Page 1503

International Research Journal of Engineering and Technology (IRJET)

e-ISSN: 2395-0056

Volume: 07 Issue: 01 | Jan 2020

p-ISSN: 2395-0072

www.irjet.net

And separate their areas to create a word dictionary. When we create a word dictionary that is visible, the image can be displayed as a histogram of the visual term based on the collection of the machine, providing a local description. Steps in our proposed method 1. Separate patches from each training image 2. Application of core component analysis (PCA) to reduce data dimensions 3. To reduce noise levels and computational complexity 4. Adding patch center coordinates to property vectors 5. Using spatial data to display images 6. The combination of all modified data sets using K-mean is a representation of the image that created the K-Mean dictionary as a general grouping method that groups the input attribute vectors into the K group at their centers.

• Training set Module - ANN access and prepares data for creating neural network. Configure the network input and outputs and train the network. Validate network result.

• Testing set Module – Test the image with train system and Integrate the network into a production system. • Feature Extraction Module- Object Recognition using Speeded-Up Robust Features (SURF) is composed of three steps -

feature extraction, feature description, and feature matching. The BoVW image representation is adapted from the bag-ofwords (BoW) representation of text documents

• Classification Module- Local Binary Pattern (LBP) is classify the efficient and pixel wise LBP with medical and non medical image.

3. Experimental Setup Our data set contains 443 chest X-ray images (DICOM format). The images are obtained from the diagnostic department of Sheba Medical Center (Tel-Hashomer Israel). Two radiologists interpret X-rays and this is a reference gold standard. Radiologists inspect all images freely. Then they talked and achieved consensus on the label of every image. For each type of image and pathology, there are labels, positive or negative values. The picture shows 3 conditions: chest pathology: Effural right lung (44 images), Cardiomegaly (99 images) and abnormal Mediastinum (110 images). Overall, the data set contains 219 images with at least one pathological condition. Digital images are cropped and centered. The images are of variable size. They are cropped, centered and contain several artifacts such as reading directives (e.g. arrows, left/right indicators) and medical equipment but otherwise were not preprocessed (e.g. equalization). We have replicated the Intensity channel to support the CNN 3-channel RGB input data expectations. The pertained CNN resizes the images into specific accepted resolution automatically. The images were collected from the Diagnostic Imaging Department of Sheba Medical Center, Tel Hashomer, Israel. Gold standard was achieved using image interpretation done by two expert radiologists. The radiologists examined all of the images independently and then reached a decision regarding the label of every image. For each image and pathology type, a positive or negative label was assigned. The images depict 6 chest pathology conditions: Right Pleural Effusion (73 images), Left Pleural Effusion (74 images), Right Consolidation (58 images), Left Consolidation (45 images), Cardiomegaly (154 images) and Abnormal Mediastinum (145 images). Overall, the dataset contains 325 images with at least one pathology condition. We split the data randomly into a training set of 443 images and an independent testing set of 194 images, reflecting a 70-30 split. The Laplacian of an image highlights regions of rapid intensity change and therefore often used for edge detection and image enhancement. Following are the steps of LoG algorithm:

1. Gaussian and Laplacian Pyramid used. 2. Noise filtering and smoothing the image. 3. Finding the edge direction. 4. Tracing the edge as per direction. 5. Use thresholding to eliminate streaking © 2020, IRJET

Impact Factor value: 7.34

ISO 9001:2008 Certified Journal

Page 1504

International Research Journal of Engineering and Technology (IRJET)

e-ISSN: 2395-0056

Volume: 07 Issue: 01 | Jan 2020

p-ISSN: 2395-0072

www.irjet.net

Object Recognition using Speeded-Up Robust Features (SURF) is composed of three steps - feature extraction, features description, and feature matching. Following are the steps of SURF algorithm:

1. Interest point detection. 2. Local neighborhood description. 3. Matching pairs. A local binary pattern (LBP) is a sort of visible descriptor used for class in pc vision. It has considering that been observed to be a powerful characteristic for texture type. It has further been determined that once LBP is mixed with the Histogram of oriented gradients (HOG) descriptor, it improves the detection performance considerably on some datasets. This classifies whether image is medical and non-medical. We next conduct a performance comparison across the feature sets used, on a set of binary categorization task, per pathology. For each task, cases diagnosed with the examined pathology were labeled as positive cases, while cases that were not diagnosed with this pathology were labeled as negative cases. Classification was performed using a Support Vector Machine (SVM) with a non-linear intersection kernel. We build a model from the training set and evaluate it on the testing set. Accuracy measures that were examined include: sensitivity, specificity and the area under the ROC curve (AUC). Sensitivity and Specificity are derived based on the optimal cut point on the ROC the point on the curve closest to (0,1).

Figure 2 LBP Histogram

Figure 3 Classification of Pathology Image

Impact Factor value: 7.34

ISO 9001:2008 Certified Journal

Page 1505

International Research Journal of Engineering and Technology (IRJET)

e-ISSN: 2395-0056

Volume: 07 Issue: 01 | Jan 2020

p-ISSN: 2395-0072

www.irjet.net Table 1 Confusion Matrix Table

Actual Value

Prediction Outcome Pathology NonPathology Pathology TP (105) FN (04) NonFP (03) TN (20) Pathology Table 2 Key Index Parameters

Accuracy 95.62 %

Precision 96.57 %

Specificity 92.91 %

Figure 4 Result Values Table 3 System Analysis

Pathology Detection With Deep Learning

Accuracy

Precision

Specificity

Existing System

90%

83.75%

84%

Improved Method

95.62%

96.57%

92.91%

4. CONCLUSION In conclusion, in this work will present a system for the medical application of chest pathology detection in radiography images which makes use of ANN that is discovered from a non-clinical dataset (Image Net) and scientific dataset (DICOM). Chest radiography are the maximum not unusual exam in radiology. They will critical for the management of various sicknesses related to excessive mortality and show a wide variety of capacity facts unlike previous work on the usage of pertrained CNN and GIST as a feature extraction method. In our case ANN and SURF will use the main illustration. This will represent alone is an effective off-the-shelf descriptor for chest x-ray retrieval tasks. In our case will SURF and LBP assuming that the combination captures information that classifies functions or features. These will a standard technique this is additionally applicable to other medical type responsibilities. By obtain features of medical snaps will generate layers for neurons and classify image as Healthy or Pathology image according the diseases. Artificial neural networks (ANN) deep architecture class techniques have won recognition due to their potential to study mid and high degree image representations with provide more accurate result.

Impact Factor value: 7.34

ISO 9001:2008 Certified Journal

Page 1506

International Research Journal of Engineering and Technology (IRJET)

e-ISSN: 2395-0056

Volume: 07 Issue: 01 | Jan 2020

p-ISSN: 2395-0072

www.irjet.net

REFERENCES [1]Vrushali Rajesh Dhanokar, Prof. A. S Gaikwad "Classifying Chest Pathology Images Using Deep Learning Techniques", International Journal of Science & Engineering Development Research , ISSN:2455-2631, Vol.4, Issue 12, page no.80 - 84, December-2019. [2] B. van Ginneken, L. Hogeweg, and M. Prokop, "Computer-aided diagnosis in chest radiography: Beyond nodules," European Journal of Radiology, vol. 72, no. 2, pp. 226- 230, 2010. [3] J.M. Carrillo-de-Gea, G. García -Mateos, "Detection of Normality / Pathology on Chest Radiographs using LBP," In Bioinformatics, pp. 167-172, 2012 [4] U. Avni, H. Greenspan, E. Konen, M. Sharon, and J. Goldberger, "X-ray categorization and retrieval on the organ and pathology level, using patch-based visual words," IEEE Transactions on Medical Imaging, vol. 30, no. 3, pp. 733-746, 2014. [5]Yaniv Bar, Idit Diamant, Lior Wolf, Sivan Lieberman “Chest Pathology Detection using Deep Learning with Non- Medical Training”, IEEE 2015. [6] ] U. Avni, H. Greenspan, and J. Goldberger, "X-ray categorization and spatial localization of chest pathologies," Medical Image Computing and Computer-Assisted Intervention–MICCAI, Springer Berlin Heidelberg, pp. 199-206, 2016. [7] G. Csurka, C. Dance, L. Fan, J. Willamowski, and C. Bray, "Visual categorization with bags of keypoints," Workshop on statistical learning in computer vision, ECCV, vol. 1, no. 1-22, pp. 1-2, 2004. [8] Y. LeCun, K. Kavukcuoglu, and C. Farabet, "Convolutional networks and applications in vision," Proceedings of IEEE International Symposium on Circuits and Systems (ISCAS), pp. 253-256, 2010. [9] A.S. Razavian, H. Azizpour, J. Sullivan, and S. Carlsson, "CNN Features off-the- shelf: an Astounding Baseline for Recognition," IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pp. 512-519, 2014. [10] M. Oquab, L. Bottou, I. Laptev, and J. Sivic, "Learning and transferring mid-level image representations using convolutional neural networks," IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1717-1724, 2014. [11] A. Prasoon, K. Petersen, C. Igel, F. Lauze, E. Dam, and M. Nielsen, "Deep feature learning for knee cartilage segmentation using a triplanar convolutional neural network," Medical Image Computing and Computer-Assisted Intervention–MICCAI, Springer Berlin Heidelberg , pp. 246-253, 2013. [12] D. C. Cireşan, A. Giusti, L. M. Gambardella, and J. Schmidhuber, "Mitosis detection in breast cancer histology images with deep neural networks," Medical Image Computing and Computer-Assisted Intervention–MICCAI, Springer Berlin Heidelberg, pp. 411-418, 2013. [13] J. Donahue, Y. Jia, O. Vinyals, J. Hoffman, N. Zhang, E. Tzeng, and T. Darrell. "Decaf: A deep convolutional activation feature for generic visual recognition," arXiv preprint arXiv:1310.1531, 2013. [14] A. Krizhevsky, I. Sutskever, and G. E. Hinton, "Imagenet classification with deep convolutional neural networks," Advances in neural information processing systems, pp. 1097-1105, 2012.

Impact Factor value: 7.34

ISO 9001:2008 Certified Journal

Page 1507