Enhanced COVID-19 pneumonia diagnosis through optimized transformers and chest X-rays: post-pandemic insights

Abstract

The COVID-19 pandemic led to an intense global health and economic impact, prompting researchers to seek rapid and effective diagnostic tools, such as Chest X-rays (CXR), especially in regions with limited resources. To alleviate the burden of manual CXR analysis, our study focuses on developing an optimized deep learning model to discern COVID-19 infections from other respiratory illnesses. We introduced an advanced screening tool comprising a Convolutional Neural Network (CNN) equipped with an optimized attention module. The training and validation of the model used a comprehensive private dataset of 5572 CXR images, illustrating various COVID-19 pneumonia severities, other pneumonia types, and healthy cases. Performance metrics showed training accuracy varying between 87% and 99.7%, with validation accuracy ranging from 87% to 93%. These results highlight the influence of dataset composition, especially the distribution of COVID-19 severity sub-classes, on the model's effectiveness. In the testing phase, the model performed optimally with a data set configuration rich in COVID-19 images, encompassing all sub-classes (sensitivity: 83%, false-negative rate: 17%). This study underscores the vital role of interdisciplinary cooperation in delivering expedient screening solutions during health crises. The proposed model achieved over 90% accuracy in identifying COVID-19 cases in CXRs, providing novel insights into the potential of imaging configurations for enhanced detection and differentiation of COVID-19 severities from other pneumonia instances.

Ähnliche Arbeiten