An Enhanced Feature Extraction Network for Medical Image Segmentation

Abstract

The major challenges for medical image segmentation tasks are complex backgrounds and fuzzy boundaries. In order to reduce their negative impacts on medical image segmentation tasks, we propose an enhanced feature extraction network (EFEN), which is based on U-Net. Our network is designed with the structure of feature re-extraction to strengthen the feature extraction ability. In the process of decoding, we use improved skip-connection, which includes positional encoding and a cross-attention mechanism. By embedding positional information, absolute information and relative information between organs can be captured. Meanwhile, useful information will be strengthened and useless information will be weakened by using the cross-attention mechanism. Our network can finely identify the features of each skip-connection and cause the features in the process of decoding to have less noise in order to reduce the effect of fuzzy object boundaries in medical images. Experiments on the CVC-ClinicDB, the task1 from ISIC-2018, and the 2018 Data Science Bowl challenge dataset demonstrate that EFEN outperforms U-Net and some recent networks. For example, our method obtains 5.23% and 2.46% DSC improvements compared to U-Net on CVC-ClinicDB and ISIC-2018, respectively. Compared with recent works, such as DoubleU-Net, we obtain 0.65% and 0.3% DSC improvements on CVC-ClinicDB and ISIC-2018, respectively.