Early Detection of COVID-19 from CT Scans Using Deep Learning Techniques

Limna Das, P.; Manoj, A. S.; Sharma, S.; Jayaraj, P. B.

The novel coronavirus 2019 (COVID-2019) which began from China, further spread to all over the planet and was announced as a pandemic by WHO. It has blocked our daily lives and world economy to a large extent. In the lack of any particular vaccine for present pandemic COVID-19, it is necessary to recognize the disease at an early stage and quarantine these infected patients to stop the further spread. The popular diagnosis of COVID-19 is being done using polymerase chain reaction (PCR), but there are some cases of false interpretation. Rapid antibody test also has faulty/wrong implications. Till now, we have witnessed the global deficit of testing labs and testing kits for COVID-19. There is an urgent requirement for developing quick and reliable devices that can help doctors in diagnosing COVID-19. Developing a computer-based COVID detection tool will be very useful as it can screen the positive cases from a mass collection. Radiological imaging like computed tomography (CT) scans can be used for the early diagnosis. With the invention of AI algorithms, we can apply learning algorithms for early detection of COVID-19. 2016 on wards, deep learning, a deep neural network-based learning technique is widely applied in biomedical problems. In this article, we suggest a fast and reliable diagnostic tool using deep learning algorithms for identifying this pandemic. We have built two models for this purpose;one with an EfficientNet architecture using focal loss and a GradCam heatmap for testing its reliability in practical use. We also built a model using ResNet by custom vision AI of Microsoft Azure. Data was collected from different sources and the highly scaled EfficientNet architecture outperformed the Resnet architecture of MS Azure for classifying the COVID CT scans by an increase in accuracy of 10%. We are planning to deploy this software in the form of a chatbot. Also, our model continuously learns from data regularly and would attain better accuracy in future. © 2021, The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd.