FDA-Unet: A Feature fusional U-Net with Deep Supervision and Attention Mechanism for COVID-19 Lung Infection Segmentation from CT Images

Abstract

The COVID-19 infections segmentation is a challenging task due to the high variation in shape, size and position of infections or lesions in medical images. To solve it, we propose a deep learning-based segmentation method for COVID-19 chest CT images that can automatically segment COVID-19 lung lesions. Based on the U-Net model, we introduce a feature fusion and an attention block for increasing the multi-scale feature learning capacity. Moreover, the network is also equipped with a residual block and a deep supervision mechanism to improve model segmentation accuracy and completeness rate. Experimental results show that the method has a good test effect after training, and the Dice index can reach 63.26%, which is beneficial for the diagnosis of the coronary pneumonia.