MA-UNet++: A Multi-Attention Guided U-Net++ for COVID-19 CT Segmentation

At the beginning of 2020, Coronavirus Disease 2019 (COVID-19) spread widely all over the world, leading to a public health crisis in the world. Automatic COVID-19 CT segmentation can not only assist radiologists in understanding images, but also help physicians to calibrate diagnoses and provide image-guided clinical diagnosis. However, due to the inhomogeneous intensity distribution of COVID-19 in CT scans, the ambiguous and missing boundaries, and highly variable shapes of lesions, it is quite challenging to develop an automatic solution. Therefore, this paper proposes a novel Multi-Attention Guided U-Net++ (MA-UNet++) for COVID-19 segmentation. In this network, we design a novel long-skip channel-wise attention module and introduce a spatial-wise attention module to re-weight the feature representation and capture rich contextual relationships at different scales. The experiment evaluated on the COVID-19 CT Segmentation dataset, demonstrate the MA-UNet++ achieves higher segmentation accuracy than the state-of-art methods.