Correct identification of mitosis phase of individual cells in a large population imaged via time-lapse fluorescence microscopy is important for drug discovery and cell cycle study. The large amount of image data makes manually analysis unrealistic, which calls for automatic systems for mitosis cell identification. The automatic system has to be able to handle two challenges: small size of training samples and datasets obtained under different conditions. Existing methods are rather limited in dealing with these two challenges. The paper introduces a Conditional Random Fields (CRFs) model, which can well handle the two requirements.