Abstract
Human islets of Langerhans are composed mostly of glucagon-secreting α cells and insulin-secreting β cells closely intermingled one another. Current methods for identifying α and β cells involve either fixing islets and using immunostaining or disaggregating islets and employing flow cytometry for classifying α and β cells based on their size and autofluorescence. Neither approach, however, allows investigating the dynamic behavior of α and β cells in a living and intact islet. To tackle this issue, we present a machine-learning-based strategy for identification α and β cells in label-free infrared micrographs of living human islets without immunostaining. Intrinsic autofluorescence is stimulated by infrared light and collected both in intensity and lifetime in the visible range, dominated by NAD(P)H and lipofuscin signals. Descriptive parameters are derived from micrographs for ~ 10(3) cells. These parameters are used as input for a boosted decision-tree model (XGBoost) pre-trained with immunofluorescence-derived cell-type information. The model displays an optimized-metrics performance of 0.86 (i.e. area under a ROC curve), with an associated precision of 0.94 for the recognition of β cells and 0.75 for α cells. This tool promises to enable longitudinal studies on the dynamic behavior of individual cell types at single-cell resolution within the intact tissue.