AI Streamlines, Standardizes Cell Counting

Since virtually all biomedical research and development programs involve cells at some stage, cell counting has become a foundational, almost enabling operation.

According to Imarc Group, a market research firm, demand for cell analysis equipment and services, including cell counting, reached $117 billion in 2023 and will grow by more than 8% annually through 2032, to $237 billion.

As slow, tedious, manual counting gives way to more accurate, dedicated automated counting systems a further evolution is also occurring through the introduction of artificial intelligence- (AI) assisted counting. Since AI In cell count applications is deployed through software, it becomes possible to upgrade an automated cell counting system through relatively low-cost software upgrades instead of having to replace expensive hardware.

Although the terms AI and “machine learning” are often used interchangeably they are not the same. Machine learning is a subset of AI through which a computer teaches itself without prior programming, whereas AI is the application of machine learning to human-like decision-making. This may seem like a distinction without a difference but in the context of cell counting machine learning is most often associated with throughput (testing more samples in less time) while AI is associated with qualitative issues (Is this a cell? Is this the cell I’m looking for?).

Cell counting applications fall into two broad categories. In basic research, medical diagnostics, and cell-based therapies where more than one cell type is likely to be present counting occurs after sorting the desired cell type from everything else.

Counting is simplified somewhat for activities involving monocultures that do not require sorting or separation before counting, such as transferring seed cultures to larger and larger growth vessels during cell expansion. But issues of identity (Am I counting cells or something else?) and cell status (Is the cell I’m counting healthy and production-worthy?) persist for those measurements as well.

Slow and tedious

Conventional, unautomated cell counting involves sampling the culture, detaching and re-suspending anchorage-dependent cells, and counting under a microscope using a hemocytometer or cell counting chamber. “Cell counting chamber” and “hemocytometer” sound like expensive instruments but they are just specially constructed microscope slides.

“Manual counting not only interrupts the growth process but may also lead to atypical cell behavior during the experiment,” said Dr. Benjamin-Maximilian Schwarz, Market Sector Manager at ZEISS Microscopy. “Cell counting can easily take fifteen minutes or more, including preparation and follow-up time.”

Like all things biotech, especially in processes involving fatigue or boredom, automation has been a welcome improvement to cell counting.

Automated cell counters are typically small, standalone instruments that count cells using impedance (a type of resistance to electrical flow) or light scattering. Both techniques require the application of fudge factors and assumptions on how cells under study behave under test conditions so their outputs should be considered estimates, not exact numbers.

Impedance-based Coulter counters enumerate cells and particles of all sizes and shapes, but that is also their main drawback. They have difficulty differentiating between viable and non-viable cells, often miss very small particles, and tend to count aggregates as single cells. The choice of suspension buffer is also somewhat limited since the buffers must sustain cells while providing a medium for making electrical measurements.

Optical methods are sensitive to shapes, sizes, and the cells’ environment. While microscopy provides counts but cannot sort cells like a flow cytometer, the technology is much cheaper to acquire and operate than flow instruments.

Optical or image-based methods include both flow-based standalone devices (including flow cytometers) and microscope-based methods. Because of their high cost (relative to microscopes) and the expertise level required to operate them, flow cytometers are considered overkill for simple production-cell counting.

Reducing error, improving quality

However, when combined with AI, microscopy-based counting assumes new capabilities while retaining its relative simplicity and low cost. In addition to counting all objects within prescribed dimensional domains, AI-enabled microscope-based cell counting can distinguish target cells from contaminating species or debris, and provide further details on dead cells or those undergoing mitosis.

AI-based cell counting takes much less time than manual counts, so it preserves the cells’ ideal growth conditions for the duration of the experiment. For example, Zeiss’s AI-based counting and confluency software counts cells in seconds directly in the culture vessel, eliminating sampling steps while saving time and preventing atypical cell growth behavior, according to the company.

“This reduces errors significantly compared to conventional methods,” Schwarz said.

AI’s relevance to cell counting extends beyond throughput and accuracy, however.

“AI’s greatest benefit lies in its potential to gradually replace or facilitate all automatable work steps in laboratory routines while improving quality and leaving more time for qualitative research,” Schwarz said. It thereby enables cell counting that may have previously been impossible due to what he terms “human factors.”

For Brian Napora, VP for AI Innovation at Gestalt Diagnostics, the main shortcomings of current cell counting technology are limited “generalizability” and what he terms “look-alike cells.” Generalizability refers to the ability of a counting process to identify more than one type of cell.

A detection or counting system “stuck” on one type of cell may fail to detect the appearance of unexpected cells, for example during the initial stages of bacterial contamination. Yeast and mycoplasma, which are much smaller than CHO cells, could easily be missed.

This may seem like a trivial point since bioprocessors already use culture methods to detect microbial contamination but those methods take days, and the culture may already be beyond salvage by the time results come in. This is doubly unfortunate since the manual sampling required for most manual and automated cell counting operations is itself a source of bacterial contamination.

“Variation must be anticipated and trained into a counting model, but this is a slow and costly process,” Napora told Biocompare.

Look-alike cells are a related problem whereby a detection process, whether human or computer-based, fails to distinguish between cells with similar size and morphology, or cells and particulate contamination.

Counting in the cloud?

With cloud-based applications now part of many big-data projects it was only a matter of time before cell counting caught up.

Corning’s automated, cloud-based Corning^® Cell Counter and ThermoFisher’s Thermo Fisher™ Connect cloud-based platform are two examples. Both are based on the companies’ respective automated cell counting systems and operate similarly.

Cell samples are counted normally but the microscope image is processed and massaged within the cloud service to eliminate look-alikes, contaminants, and artifacts. Users enjoy the usual benefits of cloud-based applications: automatic feature and capability upgrades, no internal computing or information issues to deal with, lab-to-lab and site-to-site consistency, and sharing or collaborative features.