Pulmonary Disease Identification and Classification Using Deep Learning Approach

Abstract

Deep Learning (DL) based model has shown great performance in the medical field for the detection of diseases. We examine the difficulty of classifying pulmonary disease (PD) classification in X-ray images in order to address medical-related issues. PD is a disease that prevents the lungs from functioning properly. Various researches have been done to automate the detection of pulmonary diseases. However, most studies concentrate only on identifying the presence or absence of the disease. As well, almost all studies ignore the automatic classification of tuberculosis in the lungs with other diseases. This research work focuses on the major occurrence of respiratory diseases, which are pneumonia, pulmonary tuberculosis, and pleural effusion. We proposed a novel framework for the detection and classification of PD from Chest X-Ray (CXR) images. Noise reduction, image quality enhancement, data augmentation, segmentation, feature extraction, and classification are all the part of the proposed framework. During image preparation, we used a Gaussian filter to eliminate noise from X-ray images and an advanced histogram equalization technique to improve image quality. The Region of Interest (ROI) has been extracted through the segmentation technique. We proposed Otsu's threshold segmentation approach to extract the ROI of the lung. In the feature extraction technique, we utilized the Gabor filter to apply the raw image to extract texture features. For classification, we employed a Deep Convolutional Neural Network (DCNN). To classify into a given class (normal, pneumonia, pulmonary tuberculosis, and pleural effusion), a four-way SoftMax classification is utilized. We developed four DCNN models (VGG16, VGG19, ResNet50V2, and DenseNet201) and compared their performance. In our models, DenseNet201 model had a training accuracy of 97.80% and a testing accuracy of 95.73% in PD detection and classification. When compared to the state-of-the-art models, the DenseNet201 model has great accuracy and is better at detecting and classifying diseases.

Keywords: Feature Learning, Gabor filter, Pulmonary Disease, Segmentation, X-ray.