painting of cell in flow

An artist’s interpretation of a cell passing through the ghost cytometry image-free visual system. [Image: Sacco Fujishima / Creative Commons license CC BY-ND]

A group of researchers from Japan, led by University of Tokyo’s Sadao Ota, reports creating and demonstrating an image-free system called “ghost cytometry” (GC) that can quickly identify and sort biological cells in motion and in real time (Science, doi: 10.1126/science.aan0096).

Using machine learning, the researchers “teach” the GC system to recognize different kinds of cells by the way they flow over a patterned optical structure, rather than identifying cells by their reconstructed visual images. Ota and his team further shortened the time needed to classify cells via GC by combining image-free detection with compressive sensing, to reduce the data size of the spatial information collected by the instrument’s camera and single-pixel detector.

The researchers say they were able to demonstrate high-throughput cell classification as well as cell sorting based on cell morphology using only one type of fluorophore marker. They believe that image-free GC could find use in a variety of applications where studying individual cells is important, including immunology, oncology, neuroscience, hematology and development.

Machine learning and compressive sensing

To train the GC system to identify different types of cells, the researchers pass several cells of the same type, dyed with a single fluorophore, one-by-one across a patterned optical structure at a speed of 3,000 cells-per-second. A camera records the fluorescence and motion of each cell relative to the patterned optical structure. These data are compressed and shuttled to a single-pixel detector. A neighboring electrical circuit with machine-learning algorithms combines these compressed signals with the intensity distribution of the patterned optical structure to create a training dataset for this specific type of cell. Using the training set, the GC system learns to identify the cell type during experiments.

(The “ghost” in the technique’s name refers to the minimal light needed for GC analysis, since a cell can be classified directly from the compressed waveforms, compared with other cytology techniques that require transforming light data into a visual image of the cell.)

Testing the GC method

After completing a proof-of-concept test of the GC system using fluorescent beads mounted on a coverslip, the researchers conducted a series of validation experiments to demonstrate the system’s ability to classify and sort cells. Results from their demonstrations showed that the image-free GC system was able to identify cell type within 10 µs—a rate of 10,000 cells per second—with high accuracy. The team was also able to effectively classify and sort two different but similarly sized types of cells in a single test solution using ultrafast fluorescence imaging-activated cell sorting (FiCS). For the sorting demonstration, the circuit relayed an electrical signal that would disrupt the flow of the cells in solution and guide individual cells to their specific pathway.

The researchers are currently working on more advanced machine-learning programs and imaging techniques that do not require fluorescent staining. A few members of the team created a company called ThinkCyte to commercialize their image-free GC system. The company plans to launch their first research-use beta product next year.