Chip Enriches and Identifies Circulating Tumor Cells Using AI

Finding a few tumor cells hidden among billions of blood cells is one of the biggest challenges in liquid biopsy. Circulating tumor cells (CTCs), which break away from primary tumors and enter the bloodstream, can provide important clues about cancer progression and treatment response, but their extremely low abundance—often only one to ten cells among a billion blood cells—makes them difficult to isolate and accurately identify.

A study published in Biomedical Analysis, led by first author Junyi Ouyang, describes a platform built to address that problem from two angles at once: enrichment and identification. The system integrates inertial microfluidic enrichment with a YOLOv8-based deep learning model for bright-field image recognition, allowing rapid enrichment and identification of tumor cells within the complex blood background. The separation step does not rely on specific cell-surface markers, and the identification step uses only bright-field images.

Most current CTC isolation methods depend on molecular markers on the surface of tumor cells, but tumor cells vary widely and different CTC populations may not share the same markers. To sidestep that issue, the team built a spiral microfluidic chip with contraction-expansion structures that sort cells by physical size. Because CTCs are generally larger (roughly 12–25 μm) than white blood cells (7–12 μm) and red blood cells (6–8 μm), the different cell types experience different forces as they move through the channel, guiding tumor cells and blood cells into separate outlets. In artificial blood samples containing MCF-7 breast cancer cells and white blood cells, the chip achieved an MCF-7 recovery rate of 89.2 ± 3.1% and a white blood cell removal rate of 86.9 ± 1.4%, raising the proportion of tumor cells from 9% before separation to 38% after enrichment.

Identifying which collected cells are true tumor cells is the second half of the challenge, and many existing workflows require fluorescent staining or antibody labeling to do it. The researchers instead trained a YOLOv8 model to recognize tumor cells directly from bright-field microscope images, using fluorescence only to establish ground truth during training. The model located, classified, and counted cells within full microscope images, without requiring individual cells to be separated first, reaching 96.0% accuracy, 94.6% precision and recall, and 95.4% specificity in distinguishing tumor cells from non-tumor cells.

The authors describe the work as a proof-of-concept study, validated with MCF-7 cell lines and artificial blood samples rather than patient-derived samples, with the enrichment and identification modules not yet combined into a single automated system. Future work will test additional tumor types and patient samples while improving integration.

“The challenge of circulating tumor cell analysis is not only finding these rare cells, but also identifying them accurately after separation,” said corresponding author Xiaochun Li. “By combining microfluidics with artificial intelligence, we hope to provide a simpler and more flexible approach for label-free CTC analysis and future downstream applications.”