Chip With AI Spots Lab-Grown Breast Cancer Cells Among Blood Cells, No Dye Labels Needed

Somewhere between one and ten cells in every billion blood cells may be a cancer cell that broke away from a tumor, which is a big reason spotting them has been so difficult. Researchers in China now report a two-part lab system built for the job: a spiral-shaped chip that sorts cells by size, and an AI program that identified breast cancer cells in plain, unstained microscope images with 96% accuracy.

Finding those rare cells usually means tagging them with fluorescent dyes or antibodies, a step that can damage the cells and limit what scientists can do with them afterward. Skipping the tagging was the goal. Their chip separates cells by size using nothing but fluid movement, and a separate AI program learns to tell cancer cells from white blood cells by looking at ordinary microscope images.

All of it was tested on lab-prepared mixtures of human breast cancer cells and white blood cells from healthy volunteers, not on blood from people with cancer. Even so, the early numbers give a sense of what a dye-free approach to catching circulating cancer cells could one day look like.

Sorting cancer cells from blood by size with a spiral chip

Blood is a crowded place. Red blood cells, white blood cells, and any escaped cancer cells all flow together, and the cancer cells make up a tiny fraction of the mix. Circulating tumor cells typically measure 12 to 25 millionths of a meter across, noticeably larger than white blood cells (roughly 7 to 12 millionths of a meter) and red blood cells (roughly 6 to 8).

That size gap is what the chip exploits. Its channel winds in a spiral, and as fluid flows around the curves, cells of different sizes drift toward different lanes. Narrow and wide sections repeat along the channel, making the fluid speed up and slow down and creating small swirling currents that nudge cells into place faster. According to the authors, this lets the chip sort cells over a shorter distance, and in a smaller device, than a plain spiral would. Its design was first reported in 2020, per the paper, so the new work centers on pairing it with AI and testing it on cancer cells.

Before the chip was fabricated, computer simulations used particles of three sizes standing in for red cells, white cells, and cancer cells. In the results, the largest particles settled near the center of the channel while the two smaller sizes stayed toward the inner side, a pattern that pointed to size-based sorting being feasible.

AI cancer cell detection without stains

Sorting leaves a mix of cancer cells and leftover white blood cells, and telling them apart has usually meant staining cells with fluorescent dyes. Dyes add steps and can harm the cells, which limits later culturing or molecular analysis.

Instead, the team trained an image-recognition program called YOLOv8 to pick out cancer cells in ordinary brightfield images, the kind a standard microscope produces. Dyes still played a role in building the training set: breast cancer cells (a lab-grown line called MCF-7) were stained green and white blood cells blue, and researchers used those colors to label which cells were which. During training, the AI itself only saw the plain, unstained versions. In all, researchers gathered 200 sets of matched images, splitting 160 for training and 40 for validation. Once trained, the program needed no dyes to make its calls.

Putting the chip and the AI to the test

To check the chip, researchers mixed stained breast cancer cells with white blood cells (the red cells had been removed beforehand) and ran the mixture through the device. Fluid was collected at the chip’s three exits, and cells at each one were counted under a fluorescence microscope. Across three repeat experiments, the chip recovered 89.2% of the cancer cells at the middle exit and removed 86.9% of the white blood cells. Cancer cells made up about 9% of the cells entering the chip and 38% of those at the middle exit, a jump, though still well short of a pure sample.

Evaluation of the AI used brightfield images of prepared mixtures of cancer and white blood cells, with matching fluorescence images serving as the answer key. On those images it reached a precision of 94.6%, so a cell it labeled as cancer was usually right, and a recall of 94.6%, so it found nearly all of the real cancer cells. Overall accuracy was 96.0%, and specificity was 95.4%, the share of normal cells it correctly ruled out.

Stain-free cancer cell detection and what comes next

Cancer cells found in blood can give doctors a look at how a tumor is behaving without surgery, and may help track whether a treatment is working. Authors of the paper argue that skipping stains keeps cells healthier, which could leave them available for later steps such as single-cell sequencing and drug susceptibility testing. That remains a hope rather than something this study tested.

Real patient blood is far messier than the lab mixtures used here, and cancer cells differ from one patient to the next. As the study authors put it, “this integrated platform has significant potential for precision cancer diagnosis and therapeutic monitoring.” The next tests, using patient blood and several cancer cell types, will show whether the chip and the AI can hold up outside the lab.

---

Disclaimer: This article is for general information only and does not provide medical advice. This early laboratory research has not been tested as a diagnostic tool in patients.

---