Advanced 3D deep non-local embedded system for self-augmented X-ray-based COVID-19 assessment

COVID-19 diagnosis using chest x-ray (CXR) imaging has a greater sensitivity and faster acquisition procedures than the Real-Time Polimerase Chain Reaction (RT-PCR) test, also requiring radiology machinery that is cheap and widely available. To process the CXR images, methods based on Deep Learning (DL) are being increasingly used, often in combination with data augmentation techniques. However, no method in the literature performs data augmentation in which the augmented training samples are processed collectively as a multi-channel image. Furthermore, no approach has yet considered a combination of attention-based networks with Convolutional Neural Networks (CNN) for COVID-19 detection. In this paper, we propose the first method for COVID-19 detection from CXR images that uses an innovative self-augmentation scheme based on reinforcement learning, which combines all the augmented images in a 3D deep volume and processes them together using a novel non-local deep CNN, which integrates convolutional and attention layers based on non-local blocks. Results on publicly-available databases exhibit a greater accuracy than the state of the art, also showing that the regions of CXR images influencing the decision are consistent with radiologists’ observations.