ReCoNet: Multi-level Preprocessing of Chest X-rays for COVID-19 Detection Using Convolutional Neural Networks

Abstract

AbstractLife-threatening COVID-19 detection from radiomic features has become a dire need of the present time for infection control and socio-economic crisis management around the world. In this paper, a novel convolutional neural network (CNN) architecture, ReCoNet (residual image-based COVID-19 detection network), is proposed for COVID-19 detection. This is achieved from chest X-ray (CXR) images shedding light on the preprocessing task considered to be very useful for enhancing the COVID-19 fingerprints. The proposed modular architecture consists of a CNN-based multi-level preprocessing filter block in cascade with a multi-layer CNN-based feature extractor and a classification block. A multi-task learning loss function is adopted for optimization of the preprocessing block trained end-to-end with the rest of the proposed network. Additionally, a data augmentation technique is applied for boosting the network performance. The whole network when pre-trained end-to-end on the CheXpert open source dataset, and trained and tested with the COVIDx dataset of 15,134 original CXR images yielded an overall benchmark accuracy, sensitivity, and specificity of 97.48%, 96.39%, and 97.53%, respectively. The immense potential of ReCoNet may be exploited in clinics for rapid and safe detection of COVID-19 globally, in particular in the low and middle income countries where RT-PCR labs and/or kits are in a serious crisis.