Lung disease classification using GLCM and deep features from different deep learning architectures with principal component analysis

Abstract

Lung disease classification is an important stage in implementing a Computer Aided Diagnosis (CADx) system. CADx systems can aid doctors as a second rater to increase diagnostic accuracy for medical applications. It has also potential to reduce waiting time and increasing patient throughput when hospitals high workload. Conventional lung classification systems utilize textural features. However textural features may not be enough to describe properties of an image. Deep features are an emerging source of features that can combat the weaknesses of textural features. The goal of this study is to propose a lung disease classification framework using deep features from five different deep networks and comparing its results with the conventional Gray-level Co-occurrence Matrix (GLCM). This study used a dataset of 81 diseased and 15 normal patients with five levels of High Resolution Computed Tomography (HRCT) slices. A comparison of five different deep learning networks namely, Alexnet, VGG16, VGG19, Res50 and Res101, with textural features from Gray-level Co-occurrence Matrix (GLCM) was performed. This study used a K-fold validation protocol with K = 2, 3, 5 and 10. This study also compared using five classifiers; Decision Tree, Support Vector Machine, Linear Discriminant Analysis, Regression and k-nearest neighbor (k-NN) classifiers. The usage of PCA increased the classification accuracy from 92.01% to 97.40% when using k-NN classifier. This was achieved with only using 14 features instead of the initial 1000 features. Using SVM classifier, a maximum accuracy of 100% was achieved when using all five of the deep learning features. Thus deep features show a promising application for classifying diseased and normal lungs.