“…Some literatures have reported that DNNs have achieved competitive performance in lung cancer [ 27 , 28 ], corneal endothelial cells [ 29 ], pancreas [ 30 , 31 ], polyp segmentation [ 32 , 33 ], etc. But, they are always prone to overfitting [ 34 , 35 ], poor robustness [ 36 ] and lack of generalization [ [37] , [38] , [39] ]. Furthermore, the lesion area of COVID-19 is usually small and blurred in CXR images, which needs us to pay more attention to these critical regions.…”