COVID-WideNet-A capsule network for COVID-19 detection

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Infectious Disease Diagnostics

Background:

• The rapid spread of COVID-19 necessitates accurate and timely diagnostic methods.
• Chest X-ray (CXR) imaging is a key modality for identifying COVID-19, but performance of current deep learning models requires improvement.
• Conventional RT-PCR testing can be time-consuming and may not be suitable for all situations.

Purpose of the Study:

• To propose a novel deep learning model, COVID-WideNet, for enhanced COVID-19 diagnosis using CXR images.
• To evaluate the performance of COVID-WideNet against existing Convolutional Neural Network (CNN) based approaches.
• To develop a computationally efficient model for rapid COVID-19 detection.

Main Methods:

• Development of a multi-layer capsule network, termed COVID-WideNet.
• Training and validation of the model on the COVIDx dataset using CXR images.
• Comparative analysis with other CNN-based deep learning models for COVID-19 diagnosis.

Main Results:

• COVID-WideNet achieved state-of-the-art performance on the COVIDx dataset.
• The model demonstrated superior diagnostic accuracy compared to other CNN-based approaches.
• Achieved an Area Under Curve (AUC) of 0.95, with 91% accuracy, sensitivity, and specificity.
• The model has 20 times fewer trainable parameters than comparable CNN models, ensuring efficiency.

Conclusions:

• COVID-WideNet offers a highly accurate and efficient method for diagnosing COVID-19 from CXR images.
• The model's reduced parameter count leads to faster and more efficient diagnostic capabilities.
• This approach can serve as a valuable tool for radiologists in detecting COVID-19 and its variants.