Evolutionary algorithms for automatic lung disease detection

Abstract

The World Health Organization estimated that 210 million people are suffering from Chronic Obstructive Pulmonary Disease (COPD), causing 300 thousand deaths in 2005 with an increase of 30% in 2015. Also, it is estimated that by 2030, COPD will rank third worldwide among the leading causes of death. These statistics about lung diseases get worse when one considers fibrosis, calcifications and other diseases. Medical images analysis is of great importance for early and accurate diagnosis of pulmonary disease and assist medical doctors for effective treatments and prevents further deaths. This work aims to identify and classify lung Computerized Tomography (CT) scan images as healthy lungs and diseases as COPD and Fibrosis. Three steps are required to achieve these goals: Extracting relevant features from the lung images, Feature Selection and Identification of lung diseases using a machine learning classifier. In the first step, this work follows an approach that extracts Haralick texture features using Gray Level Co-occurrence Matrix, Zernike’s moments, Gabor features and Tamura texture features from the segmented lung images to compose a pool of features for selection. As to the second step, we propose three evolutionary algorithms, Improvised Crow Search Algorithm (ICSA), Improvised Grey Wolf Algorithm (IGWA) and Improvised Cuttlefish Algorithm (ICFA), as a feature selection methods, which selects an optimal features subset from a large pool of features extracted from medical images to improve the classification accuracy and reduce the computational costs. In the final step, four machine learning classifiers: k-Nearest Neighbor, Support Vector Machine, Random Forest Classifier and Decision Tree Classifier were applied to each feature subset selected by the proposed feature selection methods. The experimental results shows that ICSA eliminated the maximum amount of insignificant features of about 71% whereas IGWA removed only 52.3% out of the total extracted features. ICFA filtered out the least amount of features upto 40.6%. However, IGWA gave the best accuracy of 99.4% for classifying lung diseases followed by ICSA with an accuracy of 99.0% respectively. A comparatively lesser accuracy of 97.3% was achieved by ICFA. Our results led to conclude that the proposed feature selection methods are suitable for classification of diseases in medical images, can also be used in real-time applications due to their reduced computational cost and very high accuracy.