Robust single-cell classification in hyperspectral stimulated raman scattering imaging by machine learning

Label-free phenotypic classification at a single-cell level is a challenging yet important task in cell biology. Stimulated Raman scattering (SRS) microscopy provides high chemical selectivity and sensitivity for label-free imaging of biological samples. With the capability to record hyperspectral SRS images with high-speed, mapping of biomolecules inside living cells enables label-free phenotyping. However, like all high-dimensional data, it remains challenging to fully exploit the excessive amount of information contained in hyperspectral data for single-cell analysis. Here, we developed and compared two machine-learning-based methods - the convolutional neural network and support vector machine - to automatically extract important features from high dimensional data and achieve a high-accuracy label-free single-cell classification. These methods serve as a robust approach to classify cells based on their molecular features, allowing unbiased, high-throughput data analysis.