A technique for automatically and accurately determining life or death of cells in a sample prepared without any labeling, not relying on the subjectivity of a determiner, is intended. A fluorescent specimen in which respective fluorescence generation modes in response to excitation light are different from each other between a living cell and a dead cell is prepared, and fluorescence imaging of the fluorescent specimen is performed. Bright field imaging of the fluorescent specimen is performed, and respective positions of the living cell and the dead cell are specified from a fluorescent image. Image objects in a bright field image, which are located at positions corresponding to the specified positions of the living cell and the dead cell are extracted as a living cell object and a dead cell object, respectively, and a classifier is constructed through machine learning using supervisor data including respective feature vectors of the extracted objects.