CNN based classification of 5 cell types by diffraction images

Rapid and label-free cell assay presents a challenging and significant problem that have wide applications in life science and clinics. We report here a method that combines polarization diffraction imaging flow cytometry (p-DIFC) with deep convolutional neural network (CNN) based image analysis for solving the above problem. Cross-polarized diffraction image (p-DI) pairs were acquired from 6185 cells in 5 types to investigate their uses for accurate classification. Different CNN architects have been studied to develop a compact architect named DINet which has relatively small set of network parameter for fast training and test. The averaged accuracy among the 5 groups of p-DI data ranges from 98.7% to 99.2%. With the DINet, the strong potentials of the p-DIFC method for morphology based and label-free cell assay have been demonstrated.