Published Paper Details:

Abstract: One of the most dangerous viruses of the 20th century is COVID-19, automated diagnosis has become one of the most fashionable research topics to achieve faster mass screening. Deep learning-based approaches have been found to be the most promising methods for finding this type of disease. However, this paper proposes a two-step deep CNN-based method to detect COVID-19 from chest X-ray images to achieve optimal performance with limited training images. In the first step, an encoder-decoder-based autoencoder network trained on unsupervised lung X-ray images is proposed, and the network learns to reconstruct the X-ray images. In the second step, an encoder network is proposed, which consists of different layers of the encoder model and then the encoder. Here, the encoder model is initialized with the weights learned in the first step, and the outputs of the different layers of the encoder model are effectively used by combining them into the proposed component network. An intelligent functional redundancy system is implemented in the proposed redundancy network. Finally, the link network of the encoder is trained to extract the detected features from the X-ray images, and the resulting features are used in the classification layers of the proposed architecture. Considering the final classification task, the EfficientNet-B4 network is used in both stages. Full training is performed on datasets covering the following categories: COVID-19, Normal, Bacterial Pneumonia, Viral Pneumonia. The proposed method gives a very satisfactory performance compared to the state-of-the-art methods and achieves 91.23 accuracy in 4-class, 95.72% in 3-class and 98.87% in 2-class classification.