IMPROVED LUNG PATTERN CLASSIFICATION FOR INTERSTITIAL LUNG DISEASE USING DEEP LEARNING

Abstract

In this research work, methods for classification and characterization of a computer aided diagnosis (CAD) system for interstitial lung disease have been executed. Although a lot of research has already been done in this regard, the growing popularity of the Deep learning techniques have evoked expectations that they might be applied in the field of image analysis as well. In the present work, a lot of methodologies have been applied on the chosen dataset and their results have been evaluated post which a network has been proposed. It consists of a layer of segmentation deploying the UNET architecture which is then connected and followed by a layer similar to that of CNN but with little modifications. It essentially consists of 5 convolutional layers and ReLU activations, followed by maxpooling which is then followed by 3 fully connected layers and a softmax layer. The last dense layer has got 5 outputs which form the classes to be considered. These are: healthy, ground glass opacity, emphysema, fibrosis and micronodules. In order to train and evaluate various network methodologies, a dataset of approximately 18400 image patches have been taken. A comparative analysis established the effectiveness of the proposed framework against all the other methods in a challenging dataset.