Cell Counting Gets Smarter, With Help From AI

Almost everything downstream of an experiment depends on getting the cell count right. In CAR-T processing, for example, viable cell density at critical steps such as transfection can significantly affect downstream product quality. Inaccurate counts can throw off both the engineering step and the final dose a patient receives. Stakes like these are now driving a wave of AI-powered counting technology that goes well beyond faster hemocytometers. Here’s what’s new, and how to think about applying it.

Why manual counting falls short

Despite known weaknesses, including inter-operator variability, subjective calls on what counts as a cell, and a steep time investment when working with lots of samples, manual hemocytometer counting persists in cell therapy manufacturing. These counting chambers are still in routine use for T cells and other cell therapy products, alongside newer automated and flow-based methods, where cells stream past a sensor for measurement instead of sitting still under a microscope.

Automated counters introduced in the 1950s, based on the Coulter principle of measuring electrical resistance as cells pass through a tiny aperture, solved some problems but introduced new ones, particularly around distinguishing cell clumps and providing viability data. Modern optical counters use a camera and software instead, or edge detection that looks for sharp jumps in brightness between pixels, to decide what counts as a cell. That means evaluating pixels one at a time rather than recognizing a cell’s overall shape the way a trained eye would. Irregular morphology, debris-heavy samples, and clustered cells, all common in cell therapy and organoid work, are where the approach falls short.

How deep learning finds cells

Deep learning models, by contrast, are built to recognize a cell’s overall shape and pattern, rather than checking pixel brightness against a fixed cutoff. Convolutional neural network (CNN) and fully convolutional network (FCN) models learn what cells of a given type look like from large sets of labeled training images, then apply that learned recognition to new samples.

Consider a recent study using digital holographic microscopy, a specialized imaging technique used to observe live cells without staining them. Rather than capturing a normal photo, it records how much light slows down as it passes through cells, producing a pattern rather than a picture. That pattern usually has to be computationally reconstructed into a viewable image before cells can be counted, a slow step in itself. Researchers trained a deep FCN to skip that step entirely, reading the raw pattern directly to detect and count cells, achieving counting accuracy of 99% at a throughput rate of greater than 288 cells per second.¹

Vendors are reporting similar throughput improvements commercially. Leica’s Mateo FL, for instance, completes an AI-based count in about 5 seconds versus roughly 15 minutes for manual counting. Much of that gain comes from a segmentation capability. Rather than applying a single brightness or size cutoff to the whole image, a CNN trained for segmentation traces the actual boundary of each individual cell based on its shape and texture. That lets it separate clumped cells that a threshold-based system would count as one, and recognize debris that doesn’t match a cell’s shape.

Label-free imaging catches up

A related trend is the push toward label-free counting, obtaining viability and morphology data without Trypan blue or fluorescent dyes. Staining introduces its own error sources, including dye toxicity with prolonged exposure and batch-to-batch reagent variability.

The same digital holographic microscopy technique described above is emerging as a leading label-free platform. Researchers at Tel Aviv University recently built a custom CNN for holographic imaging flow cytometry. It detected and classified cells in 0.44 milliseconds, more than 10 times faster than YOLOv8n, part of the YOLO (“You Only Look Once”) family widely used as a speed benchmark for object detection.² It’s worth noting that the generalization capabilities of learning-based models can be limited by biases from data collected under different imaging settings, a reminder that a model trained on one cell type or instrument configuration may not transfer cleanly to another.

In-line monitoring for CAR-T manufacturing

Few workflows depend as much on accurate cell counting as CAR-T manufacturing, where count and viability checks are required quality-control checkpoints under GMP conditions. Counting cells typically means pulling a sample from the sealed production batch and testing it separately, which costs time and adds contamination risk each time the vessel is opened. A new generation of systems is eliminating that manual step, counting cells without ever removing a sample.

The iLine F, a holography-based microscope from Ovizio Imaging Systems, connects directly to a bioreactor. It continuously images cells as the culture grows, generating a machine-learning-ready holographic fingerprint for each one without ever opening the vessel. The iLine F's counts have held up as accurate across cell populations at different viability and activation states. They also closely match a conventional reference count run in parallel, and the same fingerprinting flags contaminants such as beads before they compromise the final product.

This kind of continuous monitoring also solves an accuracy problem. A validation study of automated counting for current GMP manufacturing found that inter-operator variation in manual counting can reach nearly 20%, even among highly experienced counters.³

Single-cell sequencing’s counting problem

On droplet-based platforms such as 10x Genomics’ Chromium system, cells are encapsulated one at a time into tiny droplets for sequencing, and the number of cells loaded has to be calibrated carefully. Load too few, and many droplets come out empty. Load too many, and more droplets end up capturing two or more cells instead of one. When that happens, two different cells’ sequencing reads get lumped together under a single barcode, making them look like the gene expression profile of one cell that doesn’t actually exist.

Even with a well-calibrated load, some droplets end up ambiguous. A low-quality cell can look a lot like an empty droplet contaminated with background RNA, and a fixed cutoff struggles to tell the two apart. AI is now being used to sort that out computationally, after sequencing rather than at the loading stage. Researchers at the Broad Institute of MIT and Harvard built CellBender, an open-source tool that uses a deep generative model to learn each dataset’s specific background noise pattern and subtract it, improving on fixed-cutoff methods particularly in ambiguous droplets and single-nucleus datasets.⁴

3D samples test AI’s limits

Organoids, spheroids, and other 3D culture systems create counting challenges that neither manual hemocytometry nor early automated counters were designed for. Irregular shapes, clustering, and variable optical density all confound simple thresholding. Segmenters based on deep learning are a potential solution. Researchers applied Segment Anything, a general-purpose image segmentation model built by Meta AI and released in 2023, to trace the outline of each individual organoid. That single step yields both a count and measurements like size and shape.⁵

Even so, purpose-built, non-AI tools remain competitive. A recent study found that OrganoSeg2, a non-AI tool that automatically adjusts its detection threshold across different parts of an image rather than using one fixed setting, surpassed several deep-learning-based segmenters in generalized accuracy across organoid types. AI, in other words, isn’t always the better choice when it comes to 3D counting.⁶

Accessible AI for tissue imaging

Cell counting isn’t limited to suspensions. Quantifying cells in fixed tissue sections is a long-standing task across biological and biomedical research, and it’s an area where AI-assisted counting looks different from the deep-learning systems built for flow-based or holographic platforms.

ACCT (Automatic Cell Counting with Trainable Weka Segmentation) illustrates the point well. Rather than a large CNN, it uses classical machine learning to assign each pixel in a microscopy image a probability of belonging to a cell, then groups those pixels into countable objects. Tested on real tissue images, its counts correlated closely with researchers’ own manual counts.

A comparable deep-learning tool, c-ResUnet, required four V100 GPUs,⁷ specialized chips well beyond typical lab hardware, while ACCT was developed and validated entirely on ordinary consumer laptops. It also has real limitations. Across its test datasets, it consistently identified real cells accurately when it flagged them, but it missed more true cells than expert human counters did, tending to undercount rather than overcount.⁸

What to look for in a new counter

• Is the model pretrained, retrainable, or both? A model trained on lymphocytes may not generalize well to organoids or debris-heavy primary tissue. Retrainable systems tend to stay more accurate across diverse workflows.
• Is staining required? Label-free options reduce reagent cost and eliminate stain-timing errors, but confirm viability results have been validated against a stained reference method for your cell type.
• What’s the actual automation level? Some systems still require pulling a sample out for testing, even when marketed as automated. Ask directly.
• How is accuracy validated? Ask for published or internal comparisons against manual counts or an independent method, such as flow cytometry.

Cell counting is moving away from a single, fixed method toward a wider set of specialized tools, AI-based and otherwise, each suited to a particular kind of sample. This is a stark departure from just the hemocytometer and a trained eye. For labs sorting through that broader field, speed shouldn’t be the deciding factor. The more useful question is whether a given tool’s accuracy has actually been demonstrated on your specific cell type and imaging setup.

References

1. Yi F, Park S, Moon I. High-throughput label-free cell detection and counting from diffraction patterns with deep fully convolutional neural networks. J Biomed Opt. 2021;26(3):036001. doi:10.1117/1.JBO.26.3.036001

2. Yagoda-Aharoni D, Dotan E, Dudaie M, Shaked NT. Label-Free Holographic Imaging Flow Cytometry With Deep-Learning-Based Detection and Classification of Thousands of Cells Per Second. Cytometry A. 2026;109(2):89-97. doi:10.1002/cytoa.70008

3. Manzini P, Peli V, Rivera-Ordaz A, Budelli S, Barilani M, Lazzari L. Validation of an automated cell counting method for cGMP manufacturing of human induced pluripotent stem cells. Biotechnol Rep (Amst). 2022;33:e00708. Published 2022 Feb 7. doi:10.1016/j.btre.2022.e00708

4. Fleming SJ, Chaffin MD, Arduini A, et al. Unsupervised removal of systematic background noise from droplet-based single-cell experiments using CellBender. Nat Methods. 2023;20(9):1323-1335. doi:10.1038/s41592-023-01943-7

5. Xing X, Tang C, Guo Y, Kurniawan N, Yang G. SegmentAnything helps microscopy images based automatic and quantitative organoid detection and analysis. Proc SPIE. 2024;12930:129301N. doi:10.1117/12.3006022

6. Wells CJ, Labban N, Showalter SL, Przanowska RK, Janes KA. Fast learning-free organoid quantification and tracking with OrganoSeg2. Sci Rep. 2026;16(1):7928. Published 2026 Feb 9. doi:10.1038/s41598-026-37526-7

7. Morelli R, Clissa L, Amici R, et al. Automating cell counting in fluorescent microscopy through deep learning with c-ResUnet. Sci Rep. 2021;11(1):22920. Published 2021 Nov 25. doi:10.1038/s41598-021-01929-5

8. Kataras TJ, Jang TJ, Koury J, Singh H, Fok D, Kaul M. ACCT is a fast and accessible automatic cell counting tool using machine learning for 2D image segmentation. Sci Rep. 2023;13(1):8213. Published 2023 May 22. doi:10.1038/s41598-023-34943-w