Placental Super Micro-vessels Segmentation Based on ResNeXt with Convolutional Block Attention and U-Net.

Abstract

Accurate placenta super micro-vessels segmentation is the key to diagnose placental diseases. However, the current automatic segmentation algorithm has issues of information redundancy and low information utilization, which reduces the segmentation accuracy. To solve this problem, we propose a model based on ResNeXt with convolutional block attention module (CBAM) and UNet (RC-UNet) for placental super micro-vessels segmentation. In the RC-UNet model, we choose the UNet as the backbone network for initial feature extraction. At the same time, we select ResNeXt-CBAM as the attention module for feature refinement and weighting. Specifically, we stack the blocks of the same topology following the split-transform-merge strategy to reduce the redundancy of hyperparameter. Moreover, we conduct CBAM processing on each group of the detailed features to get informative features and suppress unnecessary features, which improve the information utilization. The experiments on the self-collected data show that the proposed algorithm has better segmentation results for anatomical structures (umbilical cord blood (UC), stem villus (ST), maternal blood (MA)) than other selected algorithms.