COVID-19 Diagnosis from Chest Radiography Images using Deep Residual Network
| Autor: | Mita Paunwala, Himansh Mulchandani, Chirag N. Paunwala, Poojan Dalal, Ojas A. Ramwala |
|---|---|
| Rok vydání: | 2020 |
| Předmět: |
Coronavirus disease 2019 (COVID-19)
business.industry Computer science Radiography Deep learning Feature extraction Pattern recognition Residual 030218 nuclear medicine & medical imaging 03 medical and health sciences Class imbalance 0302 clinical medicine 030220 oncology & carcinogenesis Artificial intelligence Sensitivity (control systems) business Reliability (statistics) |
| Zdroj: | ICCCNT 2020 11th International Conference on Computing, Communication and Networking Technologies (ICCCNT) |
| DOI: | 10.1109/icccnt49239.2020.9225521 |
| Popis: | The outbreak of COVID-19 has received much international attention due to its life-threatening repercussions. This pandemic has taken an enormous toll on the social, psychological, and economic stability of the humans. With the Coronavirus being extremely contagious, it becomes essential to automate the process of detecting the presence of this virus in humans. The diagnosis using Reverse Transcript Polymerase Chain Reaction (RT-PCR) is arduous and time-consuming; thus, the utilization of the chest X-rays has been proposed. Deep Learning algorithms often suffer from vanishing gradients and accuracy reduction with the increase in depth of the network. To efficiently tackle this issue on a limited dataset with severe class imbalance, this paper proposes an optimized variant of ResNet50, a Residual Network with Weighted Cross-Entropy loss to predict the presence of Coronavirus accurately in susceptible patients by analyzing their chest X-rays. The model yields reliable and stable results, with 97.5% Accuracy, 99% Positive Predictive Value, 96% Negative Predictive Value, 98.96% Specificity, and 96.11% sensitivity. The clinical reliability of the results has been validated by the precise feature extraction that has been highlighted in the heat maps of the predicted results. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|