Imaging flow cytometry has the ability to capture detailed images of thousands of cells per second. One of its most powerful modes is DC, developed by the group of Prof. Jochen Guck†, Director at the Max Planck Institute for the Science of Light (MPL) and Head of the Division “Cell Physics” at MPZPM and MPL. Cells are pushed through narrow channels on a microfluidic chip with tightly defined forces squeezing them. This deforms cells slightly, allowing researchers to evaluate their mechanical properties, which can serve as markers for various diseases.

Before this can be done, cells in the DC images must be recognized and outlined, a process called segmentation. After segmentation, blood cells are classified by type, for example erythrocytes, lymphocytes or neutrophils. The team around Guck has refined both of these processes with two different machine learning approaches.

For segmentation, they developed a compact U-Net neural network that identifies cells in bright-field images with high precision. Unlike traditional thresholding methods, which work very quickly but often miss faint cells such as lymphocytes, the new approach works reliably whilst processing large datasets quickly — even on standard clinical computers. This would not be possible with a standard U-Net network, which can take almost an entire day to process a single DC measurement on a CPU.

The reduction in the size of the network requires the careful selection and annotation of training images, allowing the network to learn efficiently. As a result, systematic errors in mechanical characterization of the cells were significantly reduced.

“This study shows that small convolutional neural networks can achieve the same accuracy as the larger models while being faster and cheaper. This makes them suitable for real-time clinical use,” says Dr. Sara Kaliman, Postdoctoral Fellow at MPL/MPZPM in the Guck division.

“Equally important, the tools we developed are not limited to this study. Any researcher using imaging flow cytometry can adopt our approach for their own applications. By making these resources openly available, we aim to help standardize and advance the field, bringing more accuracy and reproducibility to cell-based diagnostics,” adds Dr. Shada Hofemeier Abuhattum, Postdoctoral Fellow at MPL/MPZPM in the Guck division.

The team also automated the classification of white blood cells, a process that previously required manual work by trained experts. The algorithm, that utilizes unsupervised clustering, can distinguish major white blood cell types based on cell shape and texture. Together, these advances mean that not only millions of cells from a single experiment can be analyzed precisely, but hundreds of measurements per day can now be processed reliably and reproducibly — a major leap toward practical clinical use.