Classification of Lung Tissue Patterns on HRCT Images: Nature of Region of Interest and Classifier Performance

Abstract

Diagnosis of Interstitial Lung Diseases (ILDs) patterns in High Resolution Computed Tomography (HRCT) image is a challenging and time consuming task because of wide variety of patterns and large volume of data. Computer based detection of ILD patterns may increase radiologist’s efficiency in terms of accuracy and time. In this article, we present texture based classification of ILD patterns using discrete wavelet (DWT). The effect of classifier performance due to variation in the shape of region of interest is studied extensively. Three different databases are created with varying shape of region of interest using a popular publicly available data set from University Hospital of Geneva, Switzerland.  First type of ROI is irregular or arbitrary in nature covering the entire spread of the disease. These leveling are done by expert radiologists and other two are inscribed and bounding rectangular masks designed with the help of irregular mask. Five different classifiers (viz. Artificial Neural Network, K-Nearest Neighbour, Radial Basis Function Network, Support Vector Machine and Nave Bayes) are trained using the features computed from ROI using Discrete Wavelet Transform (DWT). The optimal parameter values are set for each classifier using extensive experimentations. Finally, a comparative study is performed to find out the best classifier and nature of ROI that provides highest performance.