U-net gland cell image segmentation method combined with spatial attention

Glandular cell image segmentation is an important auxiliary analysis method for judging whether glandular cells are diseased. The segmentation of gland cell images helps doctors make reliable disease diagnosis and improve diagnosis efficiency. U-net is a convolutional neural network commonly used in the field of medical image segmentation. It surpasses traditional image segmentation methods in performance of a variety of medical image segmentation tasks. However, U-net still has certain limitations. Because U-net is a symmetrical convolutional neural network model, while increasing the input image resolution, the number of convolutional layers in the network will double, which will lead to the network The deepening of the level makes the training of the network more difficult. Although U-net uses layer jump connections to combine low-level features and high-level features to improve network performance, since low-level features contain a large number of redundant features and background noise, direct splicing of low-level features and high-level features will bring a lot of redundancy. Excess information, which easily leads to a decrease in the accuracy and robustness of the network model. In order to solve these problems, this paper proposes a U-net model based on spatial attention. This model uses a new lightweight spatial attention module in the layer jump connection, which can effectively eliminate low-level features. Redundant information and highlighting the key features in low-level features will ultimately enable the improved spatial attention U-net to have higher segmentation accuracy and robustness. The method proposed in this paper has been experimentally verified on the Warwick-QU dataset. The experimental results show that compared with other improved U-net and traditional segmentation methods, the U-net based on spatial attention proposed in this paper has higher segmentation accuracy with only a small increase in training parameters.