Abstract: Auto-segmentation of cell is one of the most interesting segmentation problems due to the complex nature of the cell tissues and to the inherent problems of video microscopic image. Objects, which are variant, narrow range of gray levels, non-random noise, are ubiquitous problems presented in this kind of images. Considering above characteristics, an adaptive min-distance algorithm is proposed in this paper, which is available to segment suspected cell and nucleus from the complex background in the microscopic image of cells fallen into peritoneal effusion. 15 features of cancer cell and calculating formulas are presented respectively. These features are employed to construct a BP neural network classifier, which classifies and recognizes the cancer cells fallen into peritoneal effusion. Tests are performed using clinic cases recommended by the pathologists, results show that the proposed algorithm can efficiently segment cell image and receive higher accuracy of cancer cell diagnosis.