A residual network-based framework for COVID-19 detection from CXR images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Infectious Diseases

Background:

• The COVID-19 pandemic, caused by a novel coronavirus, emerged in late 2019.
• Current diagnostic methods like Polymerase Chain Reaction (PCR) have limitations in sensitivity.
• Automated detection from Chest X-rays (CXR) is crucial, but traditional machine learning requires extensive manual processing.

Purpose of the Study:

• To explore and validate Convolutional Neural Network (CNN)-based methods for classifying COVID-19 from CXR images.
• To develop an efficient and automated framework for reliable COVID-19 detection.
• To compare the performance of different CNN architectures and machine learning classifiers.

Main Methods:

• Dataset preprocessing followed by deep feature extraction using Residual Network (ResNet) architectures (ResNet18 and ResNet50).
• Classification of extracted features using Support Vector Machines (SVM) with various kernels (Quadratic, Linear, Gaussian, Cubic).
• Experimental validation of the proposed framework for COVID-19 detection and classification from CXR.

Main Results:

• The proposed CNN-based framework demonstrated efficient detection of COVID-19 from CXR images.
• The highest accuracy achieved was 97.3% when utilizing the ResNet50 architecture for feature extraction.
• The study validates the viability of the developed framework for automated COVID-19 diagnosis.

Conclusions:

• CNN-based approaches, particularly using ResNet50, are highly effective for COVID-19 classification from CXR.
• The developed framework offers a promising automated solution for timely and accurate COVID-19 diagnosis.
• This research contributes to improving diagnostic capabilities for infectious respiratory diseases using deep learning.