This thesis presents a novel approach that combines deep learning-based image segmentation and classification architectures to enhance the detection of COVID-19, pneumonia, and normal chest X-ray images. The proposed method addresses the challenges of accurate segmentation and classification by leveraging the capabilities of advanced deep learning techniques. To achieve accurate segmentation, a conditional generative adversarial network (C-GAN) with the pix-to-pix algorithm is employed. The network is trained using a dataset of chest X-ray images with ground truth masks, enabling effective segmentation of lung regions. The segmented lung images are then subjected to feature extraction using a combination of deep neural networks, including VGG-16, VGG-19, DenseNet-169, DenseNet-201, and sCNN. Additionally, keypoint detection methods such as SIFT and BRISK are utilized to capture local intensity information. This comprehensive approach allows the extraction of discriminative features from the segmented lung images, capturing both global and local characteristics. The extracted features are subsequently classified using various machine learning techniques, such as softmax, random forest (RF), support vector machine (SVM), and XG Boost. The classification step enables the identification of COVID-19, pneumonia, and normal chest X-ray images. Experimental results demonstrate the effectiveness of the proposed method, with the combination of the VGG-19 model, BRISK keypoint extraction, and RF as the final layer achieving the highest average classification accuracy of 96.6%. The model also exhibits low false positive rate (FPR) and false negative rate (FNR), making it suitable for accurate classification. Comparison with existing methods reveals that the proposed framework outperforms other approaches for the detection of COVID-19 and pneumonia using chest X-ray images. Future work can focus on further validating the method on larger datasets and exploring additional transfer learning techniques to enhance performance evaluation. Overall, this research contributes to the development of a cost-effective diagnostic tool that can aid radiologists in rapid and accurate identification of COVID-19 patients based on chest radiographs.
