A cell image segmentation method based on multiple attention mechanism

Cell image segmentation is important for diagnosing diseases and studying disease processes. Accurate segmentation is key for doctors to diagnose patients correctly. Today, developments in deep learning have greatly advanced the field of medical image segmentation, making it easier to differentiate cells from the background. However, in practice, when the cell boundaries are not clear or when the cells are glued together, the methods commonly used today still have the problem of not being accurate in segmentation, which is indeed more troublesome. The proposed solution to this problem is a cell image segmentation method based on multiple attention mechanism. Based on Unet, the Monte Carlo Attention is added to the encoder to enhance the extraction of features from small targets. Adaptive Rectangular Convolution is employed in lieu of traditional convolution to improve the perception of targets that vary in size and form. Convolutional Block Attention Module integration in the skip connection is conducive to the effective suppression of background interference and the highlighting of key features, which in turn improves segmentation accuracy. In order to verify whether this method is good or not, we found two different cell image datasets to do experiments and found that our method outperforms other deep learning approaches, which shows that our method is well segmented and can maintain stable performance on different datasets.