Classification and segmentation of chest radiography images using deep convolutional neural network

Abstract

Artificial intelligence has played a significant role in medical image analysis. Traditional computer vision techniques use handcrafted features that are domain-specific, whereas deep learning architectures do the automatic feature extraction without any human intervention. Another salient reason for the shift towards deep learning is the availability of medical datasets for automatic analysis. In this thesis, chest x-ray images are used for disease detection and lung segmentation. Chest x-ray is the most commonly used medical imaging technique due to its low cost, easy to capture, and non-invasive nature. However, the automatic diagnosis is a challenging task in chest x-rays as: (1) rib-cage and clavicle bones presence makes abnormalities difficult to detect because they may be located beneath them, (2) fuzzy intensity transitions near the lung and heart, dense abnormalities, rib-cages, and clavicle bones make the identification of lung contours subtle. The Convolutional Neural Network (CNN) is the popular deep learning architecture used in x-ray image analysis. Deep CNN architectures have a large number of parameters that need high computational power to train these models. Furthermore, the size of chest x-ray datasets is small, and there is a chance of overfitting during the training of the model. In this thesis, five deep CNNs are proposed for disease detection and lung segmentation in chest x-rays to address these challenges. newline newlineIn the first study, the parameter-efficient lightweight convolutional neural network (ALCNN) is developed for pneumothorax detection. The model uses the convolutional layer and attention mechanism to re-calibrate the features channel-wise. The proposed model achieved a comparable result in comparison with state-of-the-art deep models trained using three transfer learning approaches. The key highlight is that the proposed model has ten times fewer parameters than the state-of-the-art deep models. In the second study, the FocusCovid model is proposed for COVID-19 detection. This model us