Multiclass classification for chest x-ray images based on lesion location in lung zones

• Main Authors: Saad, Mohd Nizam, Muda, Zurina, Sahari, Noraidah, Abd Hamid, Hamzaini
• Format: Conference or Workshop Item
• Language: English
• Published: 2015
• Subjects: QA75 Electronic computers. Computer science
• Online Access: http://repo.uum.edu.my/15542/1/PID068.pdf; http://repo.uum.edu.my/15542/; http://www.icoci.cms.net.my/proceedings/2015/TOC.html
• Institution: Universiti Utara Malaysia

Innovation in radiology technology has generated numerous kinds of medical images like the chest X-ray (CXR).This image is used to find common problem in lung like the lesion through scanning process in lung area which is divided into six zones.By classifying the CXR images with common feature like the lesion location, we can ensure efficient image retrieval.Recently, Support Vector Machine (SVM) has turn out to be a well-known method for image classification.While many previous studies have reported the achievement of SVM in classifying images, yet there is still problem with this technique for multiclass classification.Since SVM is a binary classification technique, its ability is limited to classifying features between two classes at one time. Therefore, it is difficult to classify CXR images which contain many image features.Realizing the problem, we proposed an application method for multiclass classification with SVM to the CXR images based on the lesion position in the lung zones.The multiclass classification application is executed on the CXR images taken from Japan Society of Radiology Technology dataset.Lesion coordinates were selected as the classification input while the lung zones becomes the labels. The multiclass classification is performed with RBF kernel and the classification accuracy is tested to attain the classifiers performance.Overall, it can be concluded that the percentage of the classification accuracy is high with the highest accuracy percentage recorded at 98.7% while the lowest was 94.8%.Meanwhile, the average classification accuracy was recorded at 96.9%. The result obtained revealed that the SVM classifiers generated have successfully classified the lesion location correctly according to the lung zones.