Automatic Detection of COVID-19 Pneumonia Through Artificial Intelligence Applied to Chest X-rays

Artificial intelligence techniques, such as Deep Learning, aimed at the analysis of radiological images are having a continuous advance, which will allow an optimization in radiological diagnosis. The pandemic caused by COVID-19 has been a major diagnostic challenge, where chest radiography is a crucial technique due to its availability and accessibility. However, it is sometimes difficult to differentiate pneumonia caused by COVID-19 from that caused by other germs. To evaluate different architectures based on convolutional neural networks and Deep Learning techniques for the diagnosis of coronavirus pneumonia and its differentiation from pneumonia of other origins. We have retrospectively analyzed 1.341 normal chest X-rays, 1.200 X-rays of pneumonia caused by COVID-19, and 1.345 X-rays of pneumonia of bacterial or non-coronavirus viral origin. The Deep Learning architectures applied for image analysis were RestNet50, ResNet101, VGG, and inception. Explainability techniques were applied to choose the most suitable model according to the clinical interpretation of the image. The best Deep Learning-based model was built with the architecture ResNet50 with a diagnostic efficiency of 0.91. This model correctly diagnosed 83.1% of normal chest X-rays and 100% of pneumonias caused by COVID-19. Both accuracy and understandability were considered to choose the best-performing model. Because of that, ResNet101 based model was discarded even being the diagnostic efficiency of 0.94. Deep Learning using the architecture RestNet50 based on a convolutional neural network allows a diagnosis of COVID-19 pneumonia with high diagnostic efficiency and could be used in routine clinical practice.