Automated Viability Estimation from Digital Holographic Microscopy: Validation on Heterogeneous Industrial Bioproduction Cultures

This paper presents and validates a robust, label-free Digital Holographic Microscopy pipeline capable of accurately estimating cell viability across diverse industrial bioproduction conditions without requiring calibration, while also offering early detection of viability decline and correlation with protein titer.

The Gist

The Big Picture: The "Crystal Ball" for Medicine Making

Imagine a massive factory that makes life-saving medicines (like antibodies) using tiny living factories inside: cells. These cells are like hardworking chefs in a kitchen. If the chefs are healthy, they make delicious food (medicine). If they get sick or die, the kitchen shuts down, and the batch is ruined.

For decades, checking if these "cell chefs" are healthy has been a slow, messy, and risky job. Scientists had to take a spoonful of the soup, dye the cells with toxic chemicals (like trying to see if a fish is alive by poking it with a sharp stick), and count them under a microscope. This is like checking the health of a whole army by pulling out one soldier every day, giving him a shot, and asking, "Are you okay?" It's slow, expensive, and you lose the soldier in the process.

This paper introduces a new, magic trick: A camera system that can look at the whole army of cells without touching them, without using any dyes, and tell you instantly who is alive and who is dying.

---

The Magic Tool: Digital Holographic Microscopy (DHM)

The authors built a special camera based on Digital Holographic Microscopy (DHM).

• The Old Way (Bright Field): Imagine looking at a clear glass marble in a glass of water. It's almost invisible because it's so transparent. You can barely see it. This is what traditional microscopes see with cells.
• The New Way (DHM): Now, imagine shining a laser through that marble. The light bends slightly as it passes through the marble. DHM doesn't just take a picture; it measures exactly how much the light bends.
  • The Analogy: Think of it like a sonar system for a submarine. A submarine is invisible in the dark ocean, but sonar bounces sound off it to tell you exactly where it is, how big it is, and what shape it has. DHM uses light instead of sound to create a "3D map" of every single cell, measuring its thickness and density.

The Challenge: The "Chameleon" Problem

The researchers faced a huge problem. In a real factory, the cells aren't all the same.

• Some cells are big, some are small.
• Some are swimming in salty water, some in sugary water.
• Some are growing fast, some are slowing down.

If you build a robot to count "healthy" cells based on a simple rule (e.g., "If the cell is bigger than X, it's healthy"), the robot will get confused. A healthy cell in one batch might look like a dying cell in another batch. It's like trying to identify a friend in a crowd where everyone is wearing different costumes and masks.

The Solution: Instead of looking at one cell at a time, the team built an AI "Sherlock Holmes."

1. It looks at the entire crowd of cells at once.
2. It creates a "fingerprint" of the whole group (a 2D map of cell sizes and densities).
3. It uses a smart computer brain (a Neural Network) to say, "Ah, I recognize this pattern! This group is mostly healthy," or "Uh oh, this group is shifting toward sickness."

The Results: A Super-Reliable Detective

The team tested this system on 40 different batches of cells from different factories and labs. They threw everything at it:

• Different types of cells.
• Different liquids (media).
• Different ways of growing (some were fed constantly, some were left alone).
• The Ultimate Test: They tested it on super-dense crowds (100 million cells in a single drop!). Usually, when cells are packed this tight, cameras get confused because the cells overlap like a traffic jam. But this system handled it like a pro.

The Verdict: The new system was almost as accurate as the "Gold Standard" (the old, toxic dye method) but did it faster, cheaper, and without killing the cells. It could spot when the cells started to get sick before the old method even noticed.

The Bonus Features: Seeing the Future

The coolest part? The system does more than just count. Because it sees so much detail about the cells, it can predict other things:

1. The Crystal Ball for Medicine Yield: By looking at the cells' "mood" (their shape and density), the system could guess how much medicine the factory would produce. It's like looking at a baker's hands and knowing exactly how many loaves of bread they will bake before they even go into the oven.
2. The Early Warning System: The system noticed tiny changes in the cells hours before they actually started dying. It's like a smoke detector that smells the smoke before the fire even starts, giving the factory managers time to save the batch.

Why This Matters

This isn't just a cool science experiment. It's a game-changer for making medicine.

• Faster: No more waiting 24 hours for lab results.
• Cheaper: No more expensive dyes or wasted samples.
• Safer: No more opening the tank to take samples (which risks contamination).
• Smarter: It helps factories run at maximum speed without crashing.

In short, the authors have built a non-invasive, all-seeing eye for the biotech industry, turning the chaotic world of cell cultures into a clear, manageable dashboard. They are moving us from "guessing and checking" to "knowing and optimizing."

Technical Summary

1. Problem Statement

Cell viability is a critical parameter in biopharmaceutical production (specifically for CHO cells producing monoclonal antibodies), yet most facilities rely on manual, offline assays (e.g., trypan blue exclusion or flow cytometry). These traditional methods suffer from:

• Low Temporal Resolution: Measurements are typically spaced 12–24 hours apart, missing rapid events like apoptosis or contamination.
• Operational Risks: Manual sampling introduces contamination risks and operator variability.
• Cost and Labor: They are labor-intensive and require toxic reagents.
• Limitations of Existing Automated Methods: Current inline/online alternatives (optical density, biocapacitance, spectroscopy) often derive viability indirectly, making them sensitive to cell morphology changes, debris, and environmental fluctuations. They often require complex calibration and struggle at high cell densities (>100 million cells/mL).
• Lack of Generalizability: Previous Digital Holographic Microscopy (DHM) studies were validated only under controlled, homogeneous lab conditions, failing to address the variability of industrial bioprocesses (different cell lines, media, and process modes).

2. Methodology

A. Imaging System and Acquisition

• Technique: The authors utilized Digital Holographic Microscopy (DHM), a label-free technique that records interference patterns to reconstruct both intensity (absorption) and optical phase shift (Optical Path Difference - OPD).
• Optical Setup: A simplified, cost-effective defocused transmission microscope configuration was used. The beam interferes with itself (no dedicated reference arm), utilizing a low-numerical-aperture (0.25 NA) air objective to balance resolution and field of view.
• Hardware: Integrated into a custom optical bench with a motorized scanning stage for automated acquisition.
• Reconstruction: An advanced algorithm combining inverse-problem optimization and Convolutional Neural Networks (CNNs) was used to reconstruct phase and absorption images. The CNN specifically reduces phase-wrapping errors and improves resolution in high-density scenarios.
• Segmentation: Individual cells were identified and segmented using a fine-tuned Cellpose deep-learning framework.

B. The Viability Estimation Algorithm

The core innovation is a population-based, multiparametric analysis designed to work without per-experiment calibration.

1. Feature Extraction: Single-cell metrics are extracted, including cell diameter, mean OPD, mean absorption, optical volume difference (OVD), and granularity.
2. 2D Histogram Construction: At each time point, a 2D histogram is generated mapping Mean OPD vs. Mean Absorption.
3. Population Detection: A local maxima detection algorithm analyzes the histogram to identify distinct cell populations (viable vs. non-viable).
   • Two Peaks: If two populations are detected, a fixed-slope decision boundary separates them based on their centers.
   • One Peak: If only one peak is detected (indicating high overlap or a single dominant state), a Prediction CNN classifies the dominant state (viable vs. non-viable). This CNN takes the primary OPD/Absorption histogram plus nine additional 2D histograms derived from other features as input.
4. Decision Boundary: Once the state is determined, a parameterized boundary is applied to count cells and calculate the final viability percentage.

C. Dataset and Validation

• Scale: The pipeline was validated on a massive, heterogeneous dataset comprising 40 cell cultures from three industrial sites and one academic lab.
• Variability: The dataset included 6 different CHO cell lines, 5 media formulations, and 3 process modes (batch, fed-batch, perfusion).
• Density: Cultures ranged from low density to extremely high densities (up to 100–110 million cells/mL).
• Ground Truth: Validated against Vi-CELL (Beckman Coulter) trypan blue exclusion measurements.

3. Key Contributions

1. Generalizable Pipeline: Development of a DHM-based viability estimation algorithm that operates across diverse industrial conditions (cell lines, media, processes) without requiring recalibration or parameter tuning for new experiments.
2. High-Density Capability: Successful operation at cell densities up to 100 million cells/mL, a regime where traditional optical sensors fail due to signal saturation and overlapping cells. This was achieved by optimizing the defocus distance (50 µm) and reconstruction algorithms.
3. Advanced Optical Design: Implementation of a simplified, self-interfering DHM setup suitable for future inline/on-line probe integration, reducing cost and complexity compared to traditional interferometric systems.
4. Beyond Viability: Demonstrated the potential of DHM features to predict recombinant protein titer (IgG) and detect early signs of viability decline before the drop is visible in standard metrics.

4. Key Results

• Viability Accuracy:
  • High agreement with Vi-CELL measurements, particularly in the critical high-viability range (90–100%).
  • Mean Error: 1.8% ± 1.5% for viability between 90–100%.
  • Overall Error: 3.7% ± 5.7% across the full 0–100% range.
  • The method successfully detected the onset of viability decline simultaneously with standard methods.
• High-Density Performance:
  • Tested on manually mixed populations and a perfusion culture (110 M cells/mL).
  • Deviation from reference values was consistent across low and high densities (0–6% error at high viability; 2–12% at low viability).
• Exploratory Insights:
  • Titer Prediction: A gradient boosting model using DHM features successfully predicted IgG titer trends, capturing the dynamics of production increases and plateaus.
  • Early Warning: Using UMAP (Uniform Manifold Approximation and Projection), the authors observed that cell populations shift in a 2D feature space before viability drops. This shift could serve as an early warning signal for culture stress.

5. Significance and Future Outlook

• Industrial Impact: This work bridges the gap between academic DHM research and industrial bioprocessing. It offers a non-invasive, label-free, and automated solution for real-time monitoring, potentially replacing manual sampling.
• Process Control: By providing high-temporal-resolution data, the system enables better control over feeding, induction, and harvest timing, directly impacting yield and product quality.
• Scalability: The simplified optical design is a crucial step toward miniaturizing the system for inline or online probes (e.g., using flow cells with controlled thickness to prevent multiple scattering).
• Future Work: The authors plan to integrate the system into inline probes and further validate the exploratory results (titer prediction and early warning systems) across a broader range of industrial conditions to move from "exploratory" to "routine" application.

In conclusion, this paper presents a robust, generalizable, and high-density-capable DHM pipeline that significantly advances the state-of-the-art in bioprocess monitoring, offering a pathway toward fully automated, non-invasive control of industrial biomanufacturing.