Can the CellDrop Automated Cell Counter Reliably Count Hepatocytes?
As demonstrated by the study described in this technical note, the CellDrop’s machine learning-based, hepatocyte counting algorithm achieves a level of accuracy and statistical precision equivalent to the manual gold standard across human, canine, rat, and mouse species. Automating the process also provides significant operational and technical advantages over traditional methods.
Introduction
Primary hepatocytes are a critical cell type in toxicity and drug metabolism workflows. Given the sensitive nature of these experiments, accurate and precise hepatocyte counts are essential for achieving reproducible results. For automated cell counters, hepatocytes are a complex image analysis problem due to their irregular structures, tendency to clump, background debris, and multi-nucleated characteristics.
To address this issue, DeNovix developed a machine learning-based model for accurately counting hepatocytes and determining their viability using the CellDrop Automated Cell Counter. Automating this process aims to 1) help standardize results across laboratories, 2) reduce user to user variability, and 3) reduce the time requirement for counting.
This study aimed to test if an automated machine learning model could provide the same level of counting accuracy for hepatocyte quantification compared to the established gold standard method of manual hemocytometer counts by trained scientists. Data sets were generated independently by DeNovix scientists and collaborating scientists at BioIVT LLC, a leading supplier of commercial hepatocytes. This study evaluated the viability and concentration of cryopreserved cells from human, rat, mouse, and canine and freshly isolated hepatocytes from mouse.
Methods
Machine Learning Model Development:
The machine learning hepatocyte counting model was developed using both cryopreserved and freshly isolated hepatocytes from four species common to toxicology research: human, canine, rat, and mouse. The training dataset consisted of several hundred images across these four species to establish the proprietary cell counting model. Model performance was validated by DeNovix scientists and independently confirmed by BioIVT staff scientists.
Hepatocyte Preparation:
Cryopreserved human, canine, and rat hepatocytes were sourced from BioIVT (Baltimore, MD). Each sample was handled in accordance with the manufacturer’s protocol for thawing and subsequently resuspended in InVitroGRO KHB media prior to staining.
Manual Hemocytometer Hepatocyte Counting and Viability:
Hepatocytes were stained with 0.4% Trypan Blue (Sigma Aldrich, St. Louis, MO) by combining either an equal volume of cells and Trypan Blue, (50 µL of each), or 10 µL of Trypan Blue and 90 µL of cells, followed by mixing via gentle inversion. Each sample was counted in triplicate using a Neubauer grid hemocytometer on a 4X brightfield microscope. A fresh 10 µL sample volume was used for each replicate.
Automated CellDrop – Hepatocyte Counting and Viability:
Hepatocyte samples were stained with a common dual fluorescent stain consisting of a mixture of 12 µM Acridine Orange and 140 µM Propidium Iodide (AO/PI). Equal volumes of cells and AO/PI were combined (50 µL of each) mixed with gentle inversion, and incubated for less than 2 minutes. Stained samples were mixed by gentle rocking versus pipette mixing to mitigate damaging of the hepatocytes. Each sample was measured in triplicate using a fresh 10 µL volume per each sample.
Common Co-Cultured Liver-on-a-chip Non-parenchymal Cells – Preparation, Counting, and Viability Determination:
Primary human Stellate and Kupffer cells were thawed in 37°C water bath, centrifuged at 250 x g for 5 minutes at room temperature and resuspended in NPC medium (modified recipe without ITS+premix #354352, Corning). Primary human Liver Endothelial cells were similarly thawed but maintained in T-75 flasks until reaching approximately 85% confluency.
All three non-parenchymal cell types were stained by combining equal volumes of AO/PI (DeNovix, Wilmington DE) and cell suspension. After briefly mixing, 10 µL of stained cells were loaded onto the CellDrop FLi Automated Cell Counter to determine percentage viability and live cell concentrations.
Optimization was managed via specific software application profiles.
- Stellate Cells: Assessed using the dedicated Hepatocytes App
- Kupffer Cells: Assessed using the Primary Cell AOPI App with default protocol settings
- Lever Endothelial Cells: Assessed using the Primary Cell AOPI App with a modified protocol setting restricting the maximum cell diameter to 40 µm
Figure 1: Machine learning-based software interface for automated hepatocyte quantification. Full screen capture displaying a freshly isolated mouse hepatocyte sample analyzed on the CellDrop Automated Cell Counter. The callout box provides a high-magnification view of the counting classification, demonstrating color-coded identification of live hepatocytes (green), dead hepatocytes (red), live non-parenchymal cells (blue), free nuclei or dead non-parenchymal cells (yellow), and non-cellular debris (white).
Results
Figure 2: Internal cross-species accuracy validation of the machine learning model. Comparative evaluation of live cell density (10^6 cells/mL) and percentage viability (%) between the CellDrop automated algorithm and manual hemocytometer counts across three mammalian species: human, canine, and rat. Data represent the mean of triplicate measurements (n = 3) Brackets labeled with ns denote no statistically significant difference between the automated and manual quantification methods according to a paired, two-tailed Welch’s t-test (p > 0.05). Error bars represent the standard error of the mean (SEM).
Figure 3: Independent collaborative validation of model performance by BioIVT. Multi-site accuracy assessment tracking live cell density (10^6 cells/mL) and viability percentages (%) for cryopreserved human and freshly isolated mouse hepatocyte samples. Mean values (n = 3) demonstrate close alignment between automated machine learning counts and the manual hemocytometer gold standard. Statistical equivalence is indicated by ns (p > 0.05), paired two-tailed Welch’s t-test). Error bars denote the standard error of the mean (SEM).
Figure 4: Automated quantification and profiling of co-cultured liver non-parenchymal cells (NPCs). Performance metrics evaluating cell densities (10^6 cells/mL) and viability percentages (%) across three cell types common to liver-on-a-chip (LOC) platforms: (A) Stellate cells analyzed via the Hepatocytes App, (B) Kupffer cells evaluated with default Primary Cell AOPI App protocols, and (C) Liver Endothelial Cells (LECs) measured using a modified size gating protocol (<40) µm maximum diameter). Replicate evaluations (n = 3) demonstrate high precision, yielding a coefficient of variation (% CV) of (<8) for density measurements and <5% for viability. Error bars reflect the standard error of the mean (SEM).
Results
Statistical Analysis Methodology
To evaluate the performance of the automated machine learning model against the manual gold standard, all hepatocyte counts and sample viability determinations were statistically compared using a paired, two-tailed Welch’s t-test (analyzed via GraphPad Prism, version 10.0.1). For all evaluated species and sample types, the calculated p-values were strictly greater than the significance threshold (ɑ = 0.05). These findings were further validated by the calculated 95% confidence intervals for each independent t-test. Overall, the comparative data demonstrated no statistically significant differences between the automated CellDrop method and manual hemocytometer counts regarding mean live cell concentration or percentage viability.
Internal Cross-Species Validation (DeNovix Data)
Internal testing performed by DeNovix scientists evaluated the accuracy of the machine learning model using frozen-thawed (cryopreserved) hepatocyte samples across multiple reference species, including human, canine, and rat.
- Human Hepatocytes: Automated counting successfully matched manual tracking for both live cell density (10^6 cells/mL) and percentage viability, with overlapping error bars indicating statistical equivalence.
- Canine & Rat Hepatocytes: The machine learning model demonstrated robust cross-species flexibility. Both live cell density and percentage viability measurements closely mirrored manual hemocytometer results, demonstrating that the algorithm effectively accounts for species-specific morphology and structural variations without manual parameter adjustments.
Independent Collaborative Validation (BioIVT Data)
To confirm reproducibility and eliminate site-specific bias, staff scientists at BioIVT independently verified the machine learning model’s performance using both human and mouse hepatocyte lines.
- Accuracy Verification: The BioIVT data sets confirmed that the automated model captured the full sample distribution.
- Fresh vs. Cryopreserved Performance: The model maintained its high level of accuracy across different preparation methods, delivering precise concentration and viability metrics for both cryopreserved human lines and freshly isolated mouse hepatocyte samples.
Non-Parenchymal Cell (NPC) Quantification
Beyond primary hepatocytes, the automated system’s specialized application profiles were evaluated using common liver-on-a-chip (LOC) co-cultured non-parenchymal cells (NPCs). Cell densities and viability percentages were determined from the sums of duplicate counts across 3 to 4 independent batches for each isolated cell type:
- Stellate Cells: Successfully quantified using the dedicated Hepatocytes App, delivering precise cell density and viability data despite irregular cell shapes.
- Kupffer Cells: Effectively evaluated using the default settings of the Primary Cell AOPI App.
- Liver Endothelial Cells (LECs): Accurately assessed using the Primary Cell AOPI App with a customized protocol setting (restricting the maximum analysis diameter to 40 µm) to accommodate their specific size distribution.
Conclusion
Automating hepatocyte quantification using the CellDrop Automated Cell Counter provides significant operational and technical advantages over traditional manual hemocytometer counting. This study demonstrates that the proprietary, machine learning-based model achieves a level of accuracy and statistical precision equivalent to the established manual gold standard across human, canine, rat, and mouse species.
By replacing manual counting, this automated approach addresses critical industry challenges by:
- Eliminating Human Bias: Mitigating user-to-user variability and variations in subjective technique, thereby promoting strict standardization within and between collaborating laboratories.
- Enhancing Efficiency: Dramatically reducing the time required for routine cell counting workflows, allowing scientists to maximize throughput.
- Delivering Advanced Sample Insights: Providing valuable, automated readouts on sample quality metrics that are typically omitted or unquantifiable during manual counting, such as the level of background debris, the presence of free-floating nuclei, and the accurate differentiation of non-parenchymal cell populations in complex hepatic cultures.
For more information about the Hepatocytes App or to request a free trial of the CellDrop Automated Cell Counter, visit denovix.com/hepatocytes.
16-JUL-2026
