Label-Free Imaging Flow Cytometry for Cell Classification Based on Multiple Interferometric Projections Using Deep Learning

A new label-free imaging flow cytometry method for noninvasive and automated biological cell classification is presented. Each cell is rolled during flow, and its off-axis holograms from multiple viewpoints are acquired. Using the reconstructed quantitative phase profiles of the cell projections, highly discriminating features, enabling cell detection, classification, and differentiation, are extracted via a modified ResNet-18 deep convolutional neural network architecture. The model is first validated by classifying metastatic breast carcinoma cells (MCF-7) and normal human mammary epithelial cells (MCF-10A). An increase in classification accuracy by 1% is achieved when processing five interferometric projections versus processing a single interferometric projection. This model is further tested on four types of white blood cells and exhibits an accuracy increase of 5% when processing 12 interferometric projections versus processing a single interferometric projection. This approach is shown to be superior to that of using conventional 2D-rotation augmentation, and can be used to decrease substantially the number of cell examples needed for training the classification model without impairing the results. This novel concept has great potential to be incorporated into label-free imaging flow cytometry and improve cell classification, and be used to detect various types of medical conditions and diseases.