Small U-Net for Fast and Reliable Segmentation in Imaging Flow Cytometry.

Abstract

Imaging flow cytometry requires rapid and accurate segmentation methods to ensure high-quality cellular morphology analysis and cell counting. In deformability cytometry (DC), a specific type of imaging flow cytometry, accurately detecting cell contours is critical for evaluating mechanical properties that serve as disease markers. Traditional thresholding methods, commonly used for their speed in high-throughput applications, often struggle with low-contrast images, leading to inaccuracies in detecting the object contour. Conversely, standard neural network approaches like U-Net, though effective in medical imaging, are less suitable for high-speed imaging applications due to long inference times. To address these issues, we present a solution that enables both fast and accurate segmentation, designed for imaging flow cytometry. Our method employs a small U-Net model trained on high-quality, curated, and annotated data. This optimized model outperforms traditional thresholding methods and other neural networks, delivering a 35× speed improvement on CPU over the standard U-Net. The enhanced performance is demonstrated by a significant reduction in systematic measurement errors in blood samples analyzed using DC. The tools developed in this study are adaptable for various imaging flow cytometry applications. This approach improves segmentation quality while maintaining the rapid processing necessary for high-throughput environments.