Deep Implicit Supervision with Multiscale Feature Mapping for COVID-19 Detection in Chest Radiography Images

Abstract The COVID-19 pandemic is considered one of the major outbreaks all over the world, having a serious impact on human health and state economies. One of the important steps involved in fighting against COVID-19 is the early detection of positive patients and keeping them under observation in special care. Detecting COVID-19 from chest X-ray (CX) images is an efficient way to diagnose patients. Therefore, researchers investigated the specific abnormalities in the CX of COVID-19-positive patients. However, the detection accuracy of these methods is not enough for real-time implementation therefore, we develop an effective and efficient model for COVID-19 detection that obtains a better balance among accuracy, specificity, and sensitivity using three benchmark datasets. In the proposed work, a multiscale feature extraction mechanism is used to capture rich spatial information, which improves the discriminative ability of the model to detect COVID-19. Afterward, an implicit deep supervision mechanism is used to increase the interaction among information flows through dense connections. Lastly, a channel attention module selectively highlights the contribution between different feature maps. The experimental results of our model using three benchmark datasets including CXI, XDC, and CRD, demonstrate that our model surpassed the state-of-the-art approaches by achieving higher accuracy, specificity, and sensitivity.