A U-Net-Based Model with Attention Module and Contourlet Transform for Covid-19 Segmentation

This paper introduces a semantic segmentation neural network to detect lung abnormality due to the damage of COVID-19 via lung CT images. The model is based on U-Net architecture which is improved by four techniques. First, we apply a soft-gate skip connection in the residual dilated block. Second, the attention module is utilized to retain the useful features and produce better feature representations. Third, we apply deep supervision, which combines all the results at all levels in the decoder of the U-Net architecture. Fourth, the Contourlet transform is applied to capture the smooth contours at the encoder of the U-Net architecture to produce better feature representations. According to experimental results, the proposed model achieves a Dice Score Efficient of 78.5%, Sensitivity of 88.6%, and Specificity of 99.6% on the lung CT images affected with COVID-19 on the test set. This model improves 0.9% in Dice, 0.3% in Sensitivity, and 0.2% in Specificity, compared to the U-Net model.