iXray: A machine learning-based digital radiograph pattern recognition system for lung pathology detection

• Main Authors: De La Cruz, Ria Rodette G., Roque, Trizia Roby-Ann C., Rosas, John Daryl G., Vera Cruz, Charles Vincent M.
• Format: text
• Language: English
• Published: Animo Repository 2013
• Online Access: https://animorepository.dlsu.edu.ph/etd_bachelors/14835
• Institution: De La Salle University

Radiography is a branch of Health Science that uses x-ray beams to picture out the bones and organs. Chest plain radiographs are used by experts to identify lung abnormalities using pattern recognition. Digitized x-ray images already available however, diagnosis, through the uses of pattern recognition, is done manually. In this research, the group presents a system that automates pattern recognition on digital chest radiographs utilizing image processing, feature extraction and machine learning algorithms, making early detection of symptoms of lung abnormalities more efficient. This paper focuses on 6 common lung conditions namely, Normal, Pleural Effusion, Pneumothorax, Cardiomegaly, Hyperaeration and Lung Nodules. The lung conditions were divided into 2: histogram-based (Normal, Pleural Effusion and Pneumothorax) and statistics-based (Possible Lung Nodules, Cardiomegaly and Hyperaeration). The database is composed of 743 x-ray images in total, 560 acquired from De La Salle-Health Sciences Institute (DLS-HIS) and 183 downloaded from the internet, which are all in TIFF image format. Furthermore, it follows a labeling scheme that is dependent on the lung condition it pertains to. Sequential Minimal Optimization, known in pattern recognition and is able to handle multi-class classification, is used for the modeling and classification of the histogram-based lung conditions. The SVM classifier is trained with features from 40 images each from the histogram-based lung conditions, and is tested 18 images. The statistics-based ling conditions are classified using logic operation. Classification of the histogram-based lung conditions implemented in WEKA showed 92.59% classification accuracy with both Radial Basis Function kernel and Polynomial kernel. For the statistics-based lung conditions, Normal vs Cardiomegaly attained an accuracy 0f 70%, Normal vs Hyperaeration attained an accuracy of 73.33%, and Normal vs Possible Lung Nodules attained an accuracy rate of 58.33%. The test performance of the system in classifying Normal vs Abnormal case achieved an accuracy of 67.22%. The system has to be modified to improve the classification for Cardiomegaly and possible Lung Nodules and to increase the False Negative rate of Normal vs Cardiomegaly which is Nodules and to increase the False Negative rate of Normal vs Cardiomegaly which is 33.33%.