Search for a command to run...
The segmentation of cervical cell in liquid-based smear image plays an important role in cervical cancer detection. Despite of research for many years, it is still a challenge for the complexity of cell images such as poor contrast, cell irregularity, and overlapping. To solve this problem, a novel method is proposed based on Cell-GAN - a generative adversarial network. Firstly, the Cell-GAN is trained to learn a probability distribution of cell morphology by comparing the difference between the generated single-cell images and annotated single-cell images. Thus, the Cell-GAN has the ability to judge the integrity of a cell and treat other cellular information of a cell image, except for overlapping parts, as the background. Then, a complete single-cell image is generated by the trained Cell-GAN for each cell, which is located by a guide factor. The guide factor is constructed by a part of the cell to be segmented, such as the nucleus, to help Cell-GAN locate the cell and avoid generating a multi-cell image in the presence of overlapping, which means the contours of cells still cannot be distinguished. Finally, the segmentation line is defined by the contour of the generated cell, and the input image is cropped using the cell size information. The cropped image is reused for image generation until the area of generated cell varies within a small range.The proposed method is evaluated on the segmentation of single-cell images and overlapping cell images and obtained significant values of 94.3% DC, 7.9% FNR <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><i>o</i></sub> for single-cell images and 89.9% DC, 6.4% FNR <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><i>o</i></sub> for overlapping cell images respectively. The experimental results indicate that the proposed method can adaptively approach the boundary lines of cells to handle with different cases of overlapping in cervical cell images through what learned by the Cell-GAN. The proposed method outperforms most current methods in both segmentation accuracy and robustness.