DNN based approach to classify Covid'19 using convolutional neural network and transfer learning

• Published in: International journal of computers & applications Vol. 44; no. 10; pp. 907 - 919
• Main Authors: Joshi, Bhavya, Sharma, Akhilesh Kumar, Yadav, Narendra Singh, Tiwari, Shamik
• Format: Journal Article
• Language: English
• Published: Calgary Taylor & Francis 03.10.2022; Taylor & Francis Ltd
• Subjects: Abnormalities; Artificial neural networks; convolutional neural network; Coronaviruses; COVID-19; Datasets; Deep learning; machine learning; Medical imaging; Neural networks; Root-mean-square errors; transfer learning
• ISSN: 1206-212X; 1925-7074
• DOI: 10.1080/1206212X.2021.1983289

The Coronavirus pandemic has taken the world by surprise in its ability to change the way everything functions today. It has severely impacted the health, both mental as well as physical, of citizens around the world. Every country's health care systems have faced a mounting challenge to provide sufficient testing and treatment options to patients as the disease has progressed around the world. Covid-19 is said to affect the lungs of patients and early studies have shown abnormalities in the chest radiograms of infected patients. Inspired by earlier works, this research aims at applying deep learning methods in order to diagnose patients with Covid-19 by making use of X-Ray images. A comparison is made between three different proposed methods, a convolutional neural network with a novel custom architecture and two transfer learning based approaches. The first transfer learning approach uses a MobileNetV2 architecture and the second uses a DenseNet121 architecture with an unsharp mask applied on the images. The achieved performance is highly encouraging and sets a base for wider networks that can be trained on much larger datasets in the future. The experiments were conducted on images from publicly available datasets comprising multiple X-ray images. All three models achieved state-of-the-art results and outperformed other models in various metrics. The performance of the proposed models was compared to previous work and showed better accuracy, precision, recall, and also showed excellent results in certain metrics that are not explored in previous work, such as Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE).