Cellular Image Segmentation Model Based on Multiscale and Attention Mechanisms

Aiming at the difficulties in cell segmentation, such as brightness imbalance and boundary-blurring in cell images, an advanced UNet model combining the Squeeze-and-Excitation (SE) block and atrous spatial pyramid pooling (ASPP) modules (SAP-UNet) is proposed. SE module is used to increase the weight of cellular information so that the model focuses on the key infor-mation and reduces the interference of brightness imbalance on the cell segmentation. Atrous Spatial Pyramid Pooling (ASPP) is introduced to capture a wider range of contextual information through hybrid atrous convolution in multiscale to extend the sensor field of the model and capture multiscale cellular features. Compared with some existing models of U-Net++, Attention U-Net, U2-Net, and CE-Net, the experimental results on the dataset of Data Science Bowl-2018 (DSB-2018) show the superiorities of the proposed SAP- UNet. It can achieve excellent segmentation results close to the ground truth with high quantitative indicators, proving that a more accurate segmentation effect can be achieved by our method on the data with brightness imbalance and boundary blurring.