Automatic Prognosis of COVID-19 from CT Scan using Super-convergence CNN Algorithm

Due to the high incident rate of the novel corona virus popularly known as COVID-19, the number of suspected patients needing diagnosis presents overwhelming pressure on hospital and health management systems. This has led to global pandemic and eventual lockdown in many countries. More so, the infected patients present a higher risk of infecting the healthcare workers. This is because once a patient is positive of the virus, the recovery progress or deterioration needs to be monitored by medical experts and other health workers, which eventually exposes them to the infection. In this paper, we present an automatic prognosis of COVID-19 from a computed tomography (CT) scan using deep convolution neural networks (CNN). The models were trained using a super-convergence discriminative fine-tuning algorithm, which uses a layer-specific learning rate to fine-tune a deep CNN model;this learning rate is increased or decreased per iteration to avoid the saddle-point problem and achieve the best performance within few training epochs. The best performance results of our model were obtained as 98.57% accuracy, 98.59% precision and 98.55% recall rate. This work is therefore, presented to aid radiologist to safely and conveniently monitor the recovery of infected patients. © 2021 IEEE.

Deep Learning-based Classification of COVID-19 Lung Ultrasound for Tele-operative Robot-assisted diagnosis

Despite the implementation of strict COVID-19 guideline, over 300,000 healthcare workers has been infected with COVID-19 globally with over 7,000 deaths. This risk of infection and loss of vital healthcare workers can be eliminated by deploying a deep learning enhanced teleoperated robot. The robot for this study was developed by Worchester Polytechnic Institute, US, to be deployed for COVID-19 at the Nigerian National Hospital Abuja. In this paper, we develop a deep learning-based automatic classification of lung ultrasound images for rapid, efficient and accurate diagnosis of patients for the developed teleoperated robot. Two lightweight models (SqueezeNet and MobileNetV2) were trained on COVID-US benchmark dataset with a computational-and memory-efficient mixed-precision training. The models achieve 99.74% (± 1) accuracy, 99.39% (± 1) recall and 99.58% (± 2) precision rate. We believe that a timely deployment of this model on the teleoperated robot will remove the risk of infection of healthcare workers. © 2021 IEEE.