Lightweight neural architectures to improve COVID-19 identification

Abstract

The COVID-19 pandemic has had a global impact, transforming how we manage infectious diseases and interact socially. Researchers from various fields have worked tirelessly to develop vaccines on an unprecedented scale, while different countries have developed various sanitary protocols to deal with more contagious variants. Machine learning-assisted diagnosis has emerged as a powerful tool that can help health professionals deliver faster and more accurate outcomes. However, medical systems that rely on deep learning often require extensive data, which may be impractical for real-world applications. This paper compares lightweight neural architectures for COVID-19 identification using chest X-rays, highlighting the strengths and weaknesses of each approach. Additionally, a web tool has been developed that accepts chest computer tomography images and outputs the probability of COVID-19 infection along with a heatmap of the regions used by the intelligent system to make this determination. The experiments indicate that most lightweight architectures considered in the study can identify COVID-19 correctly, but further investigation is necessary. Lightweight neural architectures show promise in computer-aided COVID-19 diagnosis using chest X-rays, but they did not reach accuracy rates above 88%, which is necessary for medical applications. These findings suggest that additional research is necessary to improve the accuracy of lightweight models and make them practical for real-world use.