Label-Free Holographic Imaging Flow Cytometry With Deep-Learning-Based Detection and Classification of Thousands of Cells Per Second.

We present a new end-to-end neural network approach for real-time biological cell detection and classification via label-free quantitative imaging flow cytometry based on digital holography, offering a comprehensive representation of cellular structures without the need for chemical cell staining. In contrast to previous studies, our method is the first to obtain classification and detection of cells, imaged during flow using large-magnification microscopy, in 0.44 msec, allowing real-time label-free imaging flow cytometry, with more than 10× speedup compared to YOLOv8n. The custom-made two-stage neural network consists of fixed convolution layers using image processing filters to detect a single location per object, followed by two convolutional layers that classify each detected cell. This approach enables reducing computational complexity and offers high-throughput, label-free imaging-based analysis suitable for real-time imaging flow cytometry. We validate the method on two cell datasets, T-cells at different activation stages and cancer cells of different metastatic potentials, demonstrating the method's adaptability. Our results show the ability to image, detect, and classify thousands of cells per second during flow, highlighting the potential of label-free imaging flow cytometry for real-time cell monitoring, early disease detection, and high-speed diagnostics.