Abstract
Time-lapse volume images of cells obtained by the latest high-resolution light sheet microscope, “lattice light-sheet microscope (LLSM)”, exhibit superior imaging speed and three-dimensional resolution, enabling precise measurement of the shape and movement of nano-sized cellular structures. In this study, an algorithm was developed to extract structural features from LLSM images using machine learning and applied it to the analysis of mitosis. In this approach, two different sets of cell datasets were first trained, and the generated filters were then applied to the test data to generate images where the extracted features were highlighted. After iterative learning while evaluating the generated images with a loss function, the generated images were visually interpreted by humans. Using this method, a comparison of the chromosome structures between control cells and cells overexpressing the cancer gene AURKA revealed differences in chromosome thickness that were not perceptible to the human eye. Further investigation of the histone modification states that could affect chromosome thickness biochemically confirmed differences in histone modifications in cells overexpressing AURKA. These results demonstrated that feature extraction from LLSM images using machine learning can detect subtle changes that may not be recognizable to the human eye.
