DEEP LEARNING TECHNIQUES FOR COVID-19 DETECTION: A FOCUS ON DENSENET ARCHITECTURE

Abstract

According to World Health Organization statistics, COVID-19 has been the most challenging pandemic to date, spreading to nearly every nation on the earth and killing millions of people. Reverse transcription-polymerase chain reaction (RT-PCR), a prolonged and extortionate technique, is mostly used in the diagnosis of this disease. Presently, the infectious disease diagnosis often relies on reverse transcription-polymerase chain reaction (RT-PCR), a method that is both time-intensive and exorbitant. Detecting COVID-19 accurately and swiftly is crucial for effective disease management and treatment. Deep learning methods for COVID-19 identification with increased accuracy are examined in this study and DenseNet-169 model with fine-tuned hyper parameters is selected as the optimal architecture for the same. The chest X-Ray images dataset taken are preprocessed and augmented to avoid over fitting. Also transfer learning is exploited since the dataset is limited. Experimentation of the DenseNet-169 model using the public dataset showed that it can be chosen as the best architecture for detecting COVID-19. The performance is assessed against popular deep learning models including DenseNet-101, ResNet-50, EfficientNet-B0 and EfficientNet-B1. To assess the model's effectiveness, metrics such as accuracy, precision, recall, and F1-score are used. The DenseNet-169 model achieves a commendable accuracy of 89% in distinguishing COVID-19 cases. The findings support ongoing efforts to create robust and efficient diagnostic tools for COVID-19 aiming to support early detection and prompt medical intervention.
¬