Abstract: Lung pathological images are characterized by fuzzy boundary and overlapping and interweining cells. In order to solve the problem of cell segmentation, a lung cell image segmentation method combining Attention U-Net network and bottleneck detection was proposed. Firstly, bilateral filtering and Laplacian sharpening were performed on the collected images to highlight the details of cell edges and increase the contrast between the target and the background while removing the noise. Then the Attention U-Netnetwork was trained, and the pathological images were segmented using the trained model to obtain the cell regions. Based on the segmentation results of the model, the discriminant model was established with area, circumference and roundness as screening conditions to distinguish single cells from overlapping cells. The bottleneck detection method was used to determine the separation point in the overlapping region of cells, and the ellipse fitting method was used to modify the boundary, and the final segmentation result was obtained. Experimental results show that this method can segment complex lung cell pathological images (including single cells and overlapping cells) and achieve good segmentation results.

Key words: lung pathology image, cell segmentation, Attention U-Net, bottleneck detection, ellipse fitting