Multi-scale Feature Fusion Residual Shrinkage Network for COVID-19 Diagnosis

With the normalization of prevention and control of COVID-19, the market of smart healthcare will be further opened. Smart healthcare and active assisted living have important applications in infectious disease prevention and control, hierarchical diagnosis and treatment, home monitoring, and other scenarios. CT scans have proven to be an effective way to appraise whether a patient is infected with COVID-19. However, due to the analysis process of CT detection being very complicated and requiring the attention of domain experts, computer-aided diagnosis systems based on artificial intelligence have received more attention from industry and academia. In this article, we propose an innovative deep learning model to detect COVID-19 from CT images, the Multi-scale Feature Fusion Residual Shrinkage Network (MSRSN). For the sake of eliminating the interference of noise information in CT images and enhancing the accuracy of CT scan image recognition, we insert soft thresholding as a nonlinear transform layer into the deep structure to effectively remove noise-related features. Furthermore, considering that reasonable setting of thresholds is often challenging, the MSRSN has been developed to integrate the MS-CAM module as a trainable module to adaptively determine the threshold, so it does not need professional knowledge in noise processing. The MSRSN achieved an accuracy of 0.9722, a specificity of 0.95 as well as an F1 score of 0.97 on the SARS-CoV-2 CT-Scan dataset. Compared with CNNs in existence as well as related state-of the-art works, the results achieved by the MSRSN were excellent. Our results demonstrate that the MSRSN can assist specialists in screening as well as can aid in diagnosing patients with suspected COVID-19.