PaperlayerView exact revision on arXiv

Source-linked AI summary

Label-free cell counting and viability prediction with brightfield imaging and deep learning

Amir Reza Vazifeh, Christian Zeigler, Sornanathan Meyyappan, Richard Jeske, Jason W. Fleischer

arXiv:2610.10473v1cs.CVq-bio.CB

TL;DR

Cell viability is important in cell culture, but whether models trained on stained images can generalize to unstained ones has rarely been examined. ViabiLens combines cell detection and CNN classification to predict viability from brightfield images, achieving a 2.68% mean absolute error on unstained CHO samples.

  • Problem

    Whether models trained on stained images can generalize to unstained brightfield images has rarely been examined, despite the importance of cell health measurement in cell culture.

    Source: §1 · pp. 1–2

  • Method

    ViabiLens trains a cell detector and CNN viability classifier on stained samples and applies them to unstained brightfield images.

    Source: §1 · p. 3

  • Results

    2.68% mean absolute error on unstained CHO cells across a wide range of viability conditions.

    Source: §5 · p. 13

  • Takeaways & Limitations

    The results support using unstained transmitted-light images for label-free viability prediction and open possibilities for continuous, non-invasive cell-health monitoring.

    Source: §5 · p. 13

    Predictions can vary with focal plane and bounding-box placement, which changes the image crop presented to the classifier.

    Source: §4 · p. 12

Abstract

from arXiv · show

Cell viability assessment is a core requirement in cell culture systems, with critical applications in biopharmaceutical manufacturing and drug development. Conventionally, it is measured by adding membrane-impermeable dyes to a sample (a process called staining), which allows compromised cell membranes to be distinguished from intact ones. However, staining has several limitations: (a) chemical agents can perturb normal cellular processes of the cells being measured, (b) it is often ambiguous to assign viability to individual cells whose membrane integrity is only partially compromised. (c) photobleaching can undermine measurement accuracy over time when using fluorescent stains, and (d) staining cannot be performed in situ or in real time. Here, we show that (1) stained cells captured under brightfield imaging contain sufficient information to distinguish live and dead cells, and (2) cells captured under unstained brightfield imaging exhibit similar image features to their stained counterparts, enabling models trained on stained cells to generalize to unstained ones. We then report the development and validation of ViabiLens, an AI-assisted software for label-free cell viability analysis. The ViabiLens combines a cell detection model for localizing individual cells with a convolutional neural network (CNN) classifier for live/dead prediction, paired with an interactive UMAP-based viewer for visualizing and exploring individual cells across the sample. Evaluated on Chinese Hamster Ovary (CHO) cells spanning a wide range of viability conditions, ViabiLens achieves a mean absolute error of 2.68\% on unstained samples against fluorescence-based reference measurements. We also release a benchmark dataset for label-free cell viability analysis to facilitate future research, available at https://amirrezavazifeh.github.io/ViabiLens-Project-Page/.

1. Introduction

Cell viability measurement matters in biopharmaceutical production and cell therapies, but conventional staining can perturb samples and complicate interpretation. ViabiLens addresses the limited evidence for transferring models trained on stained images to unstained brightfield images.

1. Introduction

  • Accurate cell counts and viability measurements support predictable cell culture operation and assessment of therapeutic-cell health.The introduction highlights biopharmaceutical production and cell therapies, including risks from dying cells.Source: §1 · p. 1
  • Staining can disrupt or kill cells, cause dead-cell under-counting, leave viability ambiguous, and prevent in-situ or real-time measurement.Photobleaching can also reduce measurement reliability over time.Source: §1 · p. 2
  • Virtual staining requires accurately co-registered image pairs and still requires cell counting, while existing detection-and-classification studies often use narrow viability ranges or stained samples.Some studies also rely on complicated optical setups.Source: §1 · p. 2
  • ViabiLens tests stained-cell brightfield discrimination and stained-to-unstained feature similarity, then evaluates detection and CNN classification on held-out CHO image pairs.It reports 2.68% mean absolute error against fluorescence-based measurements across a wide range of viability conditions.Source: §1 · p. 3

2. Methods

The methods describe a two-stage pipeline that detects individual cells and classifies each cell’s viability. Stained CHO images provide training annotations, while paired unstained images are reserved for end-to-end evaluation.

2. Methods

  • The pipeline localizes individual cells with an object detector, classifies them with a CNN, and evaluates both stages plus end-to-end viability analysis.The methods cover image acquisition, training and test splits, model details, and evaluation metrics.Source: §2 · p. 3

2.1. Sample Preparation

  • CHO cultures were sampled across 14 days at multiple time points to span viability conditions encountered in bioprocessing.The cultures were maintained in shake flasks under fed-batch conditions.Source: §2.1 · p. 3

2.2. Image Acquisition

  • The dataset includes 75 stained images, with 15 paired pre-staining brightfield images held out for testing; PLX annotations train the models.The stained images include brightfield and two fluorescence channels, cell bounding boxes, and live/dead labels.Source: §2.2 · p. 3

2.3. Cell Detection

  • Faster R-CNN was trained on 4,500 overlapping crops from 45 stained images because full-resolution training performed poorly for small cells.At inference, overlapping crops, safe-zone filtering, and two NMS passes remove incomplete and duplicate detections; detected cells are resized to 32 × 32 pixels.Source: §2.3 · pp. 4–5

2.4. Cell Classification

  • A compact CNN trained on an 80% subset of 137,434 annotated cells from 60 stained images predicts viability; its two-element live/dead softmax output is converted to labels using a 0.5 threshold.Across the full dataset, there were 120,983 live and 16,451 dead cells; the split preserved label distribution across training and validation.Source: §2.4 · p. 5

2.5. Evaluation Metrics

  • Evaluation covers detector metrics on stained images, classifier metrics with the dead minority class as the positive class, and MAE and R² for paired-sample viability.Unstained-image evaluation is limited to total cell counts and qualitative visual comparison because reference bounding boxes are unavailable.Source: §2.5 · p. 5

3. Results

Brightfield morphology distinguishes live and dead cells, while stained and unstained cells show similar image features. The detector and CNN support cell-level analysis and end-to-end viability estimates on unstained samples.

3.1. Live and Dead Cell Discrimination in Brightfield Imaging

  • UMAP separated dead cells into featureless white and blurred-outline subclusters, possibly reflecting distinct degradation stages; irregular crops also emerged as outliers.Source: §3.1 · pp. 6–7
  • ∼95% accuracy and ∼0.80 F1-score were reached by t-SNE/UMAP KNN; these embeddings separated populations better than PCA, and live cells showed higher-frequency spectra.Source: §3.1 · p. 6, Figure 3 · p. 7

3.2. Similarity of Stained and Unstained Cells in Brightfield Imaging

  • Across PCA, t-SNE, UMAP, and original cells, stained/unstained classification stayed ∼56–68%, and Fourier profiles were nearly identical; paired samples were not cell-matched.Source: §3.2 · pp. 7–8, Figure 5 · p. 9

3.3. Cell Detection Performance

  • The detector counted approximately 17% more cells than PLX on training samples and achieved mAP≥0.9 on training and test sets, with best performance across most evaluation metrics.Source: §3.3 · pp. 8–9

3.4. Cell Classification Performance

  • CNN accuracy was 98.64% (F1-score = 0.94) on training and 98.10% (F1-score = 0.92) on validation, with close learning-curve agreement.Source: §3.4 · p. 9

3.5. End-to-End Viability Prediction

  • Stained samples had MAE 3.57% (90% CI: [2.75, 4.47]) and R2 0.82 (90% CI: [0.60, 0.92]).Source: §3.5 · p. 10
  • Unstained-sample viability prediction achieved MAE 2.68% (90% CI: [1.72, 3.89]) and R2 0.85 (90% CI: [0.73, 0.93]).Source: §3.5 · p. 10
  • UMAP clusters linked sharp cells with predominantly live labels, fuzzy or white cells with dead labels, and adjoining-cell patches with likely division.Source: §3.5 · p. 10

4. Discussion

ViabiLens predicts sample-level viability from unstained brightfield images, closely matching stained-sample reference measurements. The discussion highlights per-cell probability outputs, sensitivity to imaging conditions, and complementary visualization and modality-transfer tools.

4. Discussion

  • The pipeline predicts sample-level viability from unstained brightfield images, closely matching reference measurements from stained samples.The classifier also produces continuous live/dead probabilities rather than only hard labels.Source: §4 · pp. 10–11
  • The figure compares detection outputs across stained and unstained images; PLX cannot provide detections for unstained samples because it relies on fluorescence.The figure also presents classifier training and validation confusion matrices, ROC and precision-recall curves, and loss and accuracy curves over 50 epochs.Source: Figure 6 · p. 11, Figure 8 · p. 13
  • Per-cell detection and classification provide direct viability predictions, avoiding the ambiguity of cells appearing in both fluorescent channels in virtual staining.Classifier probabilities offer a more quantitative characterization than hard binary labels.Source: §4 · pp. 11–12
  • Cell detection remains robust across focal planes, but viability classification can be more sensitive to planes outside its training distribution.Slight cell movement also prevents perfect registration across the focal stack.Source: Figure 7 · p. 12, §4 · p. 12
  • Experiments across imaging modalities and focal planes remained consistent, while UMAP enables unsupervised phenotyping of live and dead populations.Manifold alignment can map representations from different imaging conditions to a shared low-dimensional space, supporting discriminability and knowledge transfer.Source: §4 · pp. 12–13

5. Conclusion

The conclusion presents a pipeline for assessing cell viability from unstained brightfield images, justified by shared image features across stained and unstained cells. On unstained CHO cells across a wide range of viability conditions, it reports a mean absolute error of 2.68%.

5. Conclusion

  • 2.68% mean absolute error was achieved on unstained CHO cells across a wide range of viability conditions.The pipeline combines a cell detector with a compact CNN classifier for live/dead prediction.Source: §5 · p. 13
  • Live and dead cells showed distinct morphological signatures under stained brightfield imaging, and stained and unstained cells had closely matched image features.The conclusion states these findings jointly justify training on stained samples and deploying on unstained ones.Source: §5 · p. 13
  • The authors conclude that the pipeline opens possibilities for continuous, non-invasive monitoring of cell health.Source: §5 · p. 13

Funding

The authors acknowledge financial support from Waters Corporation.

Funding

  • The authors acknowledge financial support from Waters Corporation.Source: Funding · p. 13

Data Availability

The study dataset is publicly available, and supporting content is provided in Supplement 1.

Data Availability

Loading 2610.10473v1…