AI-Powered Image Analysis for Cell Culture Quality Control

At Cytion, we understand that visual assessment of cell culture health is fundamental to producing high-quality Cells and Cell lines. Traditional microscopy-based quality control relies heavily on human expertise and subjective interpretation, which can vary between operators and over time. Artificial intelligence-powered image analysis transforms this subjective process into an objective, quantitative, and scalable quality control system that ensures consistent product quality across all our cell line offerings. By leveraging deep learning algorithms including U-Net architectures for segmentation, ResNet-50 and EfficientNet models for classification, and advanced computer vision techniques like transfer learning and ensemble methods, we can detect subtle changes in cell morphology, identify contamination earlier, and make data-driven decisions about culture health and readiness for downstream applications. Our AI systems process over 50,000 images monthly from our production of HeLa Cells, HEK293 Cells, and other critical cell lines, providing consistent quality assessment with accuracy exceeding 95% across multiple parameters.

Deep Learning Revolution in Cell Image Analysis

The application of deep learning to cell culture imaging represents a fundamental shift in how we approach quality control. Unlike traditional image analysis algorithms that require explicit programming of features to detect, deep learning models can automatically learn relevant features from thousands of training images. At Cytion, we have developed custom convolutional neural network (CNN) architectures based on proven models like U-Net for semantic segmentation (identifying cell boundaries with pixel-level accuracy), ResNet-50 for feature extraction (learning hierarchical representations from raw pixels), and EfficientNetB4 for classification tasks (distinguishing healthy from stressed cells). Our models are trained on extensive image databases—currently >150,000 annotated images spanning 200+ cell types, multiple passage numbers (P2-P30), diverse culture conditions (standard, stressed, contaminated), and various imaging modalities (phase-contrast, brightfield, fluorescence). These models achieve >95% accuracy in confluence estimation, >92% sensitivity in contamination detection, and >88% accuracy in morphology classification. The training process employs data augmentation techniques (rotation, flipping, brightness adjustment, elastic deformation) to improve model robustness and transfer learning from ImageNet-pretrained weights to accelerate convergence. Model training is performed on NVIDIA A100 GPU clusters with batch sizes of 32-64 images and training times of 12-48 hours depending on model complexity, using Adam optimizer with learning rate scheduling and early stopping based on validation set performance.

Automated Confluence Measurement and Growth Tracking

Confluence measurement—determining what percentage of culture surface is covered by cells—is one of the most critical yet subjective assessments in cell culture. At Cytion, we employ U-Net convolutional neural network architectures specifically designed for semantic segmentation tasks, achieving pixel-level classification of cell vs background regions with Intersection over Union (IoU) scores exceeding 0.90. Our U-Net implementation features a contracting path (encoder) with 4 downsampling stages using 3×3 convolutions and 2×2 max-pooling, and an expansive path (decoder) with upsampling and skip connections that preserve spatial information from earlier layers. The network is trained on manually annotated images where expert cell culture scientists have labeled cell boundaries, using a combination of binary cross-entropy and Dice loss functions to handle class imbalance. The trained model processes 2048×2048 pixel images in <300ms on GPU, generating pixel-wise probability maps that are thresholded to create binary masks, from which confluence percentage is calculated as (cell pixels / total pixels) × 100. This automated confluence measurement achieves accuracy within ±2% when validated against manual expert annotation, compared to ±15-20% variation between different human observers. Beyond single-timepoint measurement, our system tracks confluence over time to generate growth curves (plotting confluence vs time with exponential curve fitting), enabling calculation of doubling times, prediction of optimal passage timing (typically at 80-90% confluence), and identification of cultures growing anomalously slowly (>2 standard deviations below expected growth rate) which may indicate cell line senescence, media quality issues, or incubator problems. For our Cells and Cell lines catalog, this precise growth tracking ensures optimal harvest timing that maximizes cell quality and viability.

Morphological Analysis and Phenotype Stability

Cell morphology provides rich information about cell health, identity, and functional state. At Cytion, we extract comprehensive morphological features using computer vision algorithms and deep learning-based feature extraction. Following cell segmentation, we calculate classical morphology descriptors including cell area (µm²), perimeter (µm), circularity (4π×area/perimeter²), aspect ratio (major axis/minor axis), solidity (area/convex hull area), and texture features based on Gray Level Co-occurrence Matrices (GLCM) including contrast, correlation, energy, and homogeneity. Additionally, we employ ResNet-50 convolutional networks pre-trained on ImageNet and fine-tuned on our cell image dataset to extract 2,048-dimensional deep feature vectors that capture subtle morphological patterns not easily described by handcrafted features. These multi-scale features (combining traditional morphometrics with deep features) are input to Random Forest classifiers (100 trees, Gini impurity criterion) or Support Vector Machines (RBF kernel, C=1.0, gamma=auto) that distinguish normal morphology from aberrant phenotypes with >92% accuracy. For quality control, we maintain reference morphology profiles for each cell line in our catalog—for example, HeLa Cells exhibit characteristic epithelial morphology with mean area 450±80 µm², circularity 0.65±0.12, while HEK293 Cells show 380±70 µm² area with higher circularity 0.72±0.10. Morphological drift detection uses Hotelling's T² statistic to test whether current batch morphology significantly deviates from reference distribution (p<0.05 threshold), flagging cultures for review when phenotypic changes are detected that may indicate unwanted differentiation, genetic drift, or suboptimal culture conditions.

Early Contamination Detection

Contamination is one of the most serious threats to cell culture operations, potentially resulting in lost cultures, wasted resources, and compromised experimental results. At Cytion, we have developed specialized contamination detection models trained on curated datasets of contaminated cultures including bacterial contamination (characterized by rapid increase in small particulate debris, media turbidity, pH shifts visible as color changes in phenol red-containing media), fungal contamination (visible as mycelial structures, spores), and mycoplasma infection (subtle morphological changes, reduced growth rate, increased granularity). Our detection system employs EfficientNetB4 architectures (16.8M parameters, compound scaling of depth, width, and resolution) trained using a two-stage approach: first, classification into clean vs contaminated categories (binary cross-entropy loss, achieving AUC-ROC 0.96); second, multi-class classification identifying contamination type (categorical cross-entropy, 85% accuracy across bacterial/fungal/mycoplasma/yeast categories). The models analyze multiple image features including unusual particle distributions (detected via blob detection algorithms), media appearance changes (color shifts quantified in LAB color space), and abnormal cell morphology patterns. Time-series analysis comparing current images to 24-48 hour historical baseline enables detection of developing contamination before it becomes visually obvious to operators, typically providing 24-48 hour earlier warning compared to manual inspection. When contamination probability exceeds 0.7 threshold, automated alerts notify QC personnel via email and LIMS notifications, triggering immediate investigation including visual confirmation, Gram staining (for bacterial contamination), and mycoplasma PCR testing. This AI-enhanced contamination surveillance has reduced contamination-related batch losses by 60% at Cytion through earlier detection and intervention, particularly valuable for long-term cultures and high-value cell line development projects where contamination late in the process would represent significant resource loss.

Non-Invasive Viability Assessment

Traditional viability assessment using trypan blue or other membrane-impermeant dyes requires sampling cells from culture, which is destructive and limits temporal resolution. At Cytion, we have developed morphology-based viability prediction models that estimate cell viability from label-free brightfield or phase-contrast images using machine learning. The approach is based on the observation that dying and dead cells exhibit characteristic morphological changes: cell shrinkage, membrane blebbing, cytoplasmic granulation, loss of cell-substrate adhesion, and increased light refraction. We extracted 156 morphological and texture features from individual segmented cells, then used feature selection (Recursive Feature Elimination with cross-validation) to identify the 35 most predictive features including cell area, perimeter irregularity, mean pixel intensity, intensity variance, and GLCM texture descriptors. Gradient Boosting Regression models (XGBoost with 200 estimators, learning rate 0.1, max depth 6) trained on these features predict viability percentage with R²=0.87 when validated against gold-standard trypan blue exclusion measurements performed on parallel samples. The model was trained on 12,000 image-viability pairs covering viability ranges from 50% to 99% across multiple cell types and passage numbers. For production monitoring, the system processes images captured every 2-4 hours by IncuCyte live-cell analysis systems, generating continuous viability trend data without disturbing cultures. Sudden viability drops (>10% decrease in 12 hours) trigger alerts for investigation, while gradual declining trends inform passage timing decisions—we typically passage at >90% predicted viability to maintain cell health. This non-invasive viability monitoring is particularly valuable for suspension cultures and bioreactor systems where traditional sampling is more disruptive, and for screening experiments where preserving culture integrity while monitoring cell health is essential.

Multi-Parameter Quality Scoring

Rather than relying on single metrics, AI systems can integrate multiple image-derived parameters into comprehensive quality scores. At Cytion, we have developed holistic quality assessment models that combine confluence (target 80-90% for passage), morphology score (0-100 scale, >75 indicates normal morphology), viability estimate (>90% target), contamination risk (<0.1 probability threshold), and culture uniformity (coefficient of variation in cell size/shape, <20% target) into an overall QC score using weighted ensemble methods. The ensemble combines predictions from specialized models: U-Net confluence (weight 0.25), ResNet-50 morphology classifier (weight 0.30), EfficientNet contamination detector (weight 0.25), XGBoost viability regression (weight 0.15), with weights optimized through grid search on held-out validation sets to maximize correlation with expert QC decisions. The final QC score ranges 0-100, with automated decision rules: score ≥85 = pass (proceed to passage/harvest), 70-84 = borderline (flag for manual review), <70 = fail (investigate or discard). These multi-parameter assessments provide objective, quantitative criteria for release decisions in production—at Cytion, cultures must achieve QC score ≥85 before progressing to next passage or final harvest, ensuring consistent product quality. Analysis of our production data shows strong correlation (r=0.82) between AI QC scores and downstream culture performance metrics including post-passage viability and expansion success, validating the predictive value of the integrated scoring approach. The automated scoring system processes complete microplate images (96 wells) in 8-12 minutes, compared to 45-60 minutes for manual microscopic inspection, enabling real-time QC decisions that keep production workflows moving efficiently.

Transfer Learning and Model Adaptation

One of the challenges in implementing AI for cell culture analysis is the need for large training datasets, particularly for specialized or rare cell lines. Transfer learning addresses this by starting with models pre-trained on large general image datasets (ImageNet with 1.4M images, 1000 categories), then fine-tuning on cell culture-specific images. At Cytion, we leverage transfer learning extensively: we initialize our models with ImageNet-pretrained weights (e.g., ResNet-50, EfficientNetB4), then fine-tune the final layers or entire network using our cell image datasets with significantly reduced training data requirements. For example, developing a new morphology classifier de novo might require 10,000+ annotated images, while transfer learning achieves comparable performance with 1,000-2,000 images. Our fine-tuning protocol uses lower learning rates (1e-4 to 1e-5) compared to training from scratch (1e-2 to 1e-3), typically trains for 20-50 epochs with early stopping based on validation loss plateau, and employs discriminative learning rates where earlier layers (general features) update slowly while later layers (cell-specific features) update faster. For new cell lines added to our Cells and Cell lines catalog, we implement continuous learning where models are periodically retrained with accumulated images from production batches, typically quarterly updates that incorporate 500-1000 new validated images, maintaining model accuracy as our cell line portfolio expands. Domain adaptation techniques like Maximum Mean Discrepancy (MMD) and adversarial training help models generalize across imaging platforms—we train on data from multiple microscope systems (IncuCyte, ImageXpress, EVOS) to ensure robust performance regardless of acquisition hardware.

Explainable AI and Quality Assurance

While deep learning models can achieve impressive accuracy, their "black box" nature can be problematic for quality control applications where understanding the basis for decisions is important. At Cytion, we implement explainable AI (XAI) techniques to make model decisions interpretable and trustworthy. We employ Grad-CAM (Gradient-weighted Class Activation Mapping) to generate heatmaps highlighting which image regions most influenced classification decisions—for example, showing that contamination detection focuses on debris particles and morphology changes rather than irrelevant background features. SHAP (SHapley Additive exPlanations) values quantify each feature's contribution to individual predictions, revealing that confluence predictions primarily depend on cell density and coverage metrics while viability predictions weight membrane integrity and cytoplasmic texture features heavily. For morphology classification, we visualize learned filters in convolutional layers, showing that early layers detect edges and textures while deeper layers recognize cell-specific patterns like epithelial sheet formation in HeLa cells or neuronal-like processes in differentiated cell types. These XAI visualizations serve multiple purposes: building trust among QC personnel who can verify the AI is making decisions based on biologically relevant features, facilitating troubleshooting when unexpected predictions occur by identifying what features drove the decision, and providing training material showing new personnel what features are important for quality assessment. We maintain an XAI dashboard displaying explanation visualizations for flagged or borderline cultures, enabling rapid expert review with context about why the AI made its assessment. This transparency has been crucial for regulatory acceptance of AI-based QC—our validation packages for GMP production include representative XAI visualizations demonstrating models make decisions based on scientifically sound criteria aligned with traditional expert assessment principles.

High-Content Analysis Integration

AI-powered image analysis integrates seamlessly with high-content imaging platforms that capture multiple fluorescent channels, perform automated Z-stacking, and image entire multi-well plates with precision robotics. At Cytion, we deploy Molecular Devices ImageXpress Micro Confocal systems that acquire up to 6 fluorescence channels (DAPI, FITC, TRITC, Texas Red, Cy5, Cy5.5) plus transmitted light, with automated Z-stacking (1-50 planes, 0.5-10 µm steps) and precise XY stage positioning (±1 µm accuracy). For high-content applications like assessing stem cell differentiation efficiency, we use immunofluorescence staining for lineage markers followed by AI-powered analysis: cell segmentation based on nuclear staining (DAPI channel, watershed algorithm), classification into marker-positive vs negative based on fluorescence intensity thresholds (optimized by Otsu's method), and quantification of differentiation efficiency as percentage of marker-positive cells. Multi-channel analysis enables sophisticated phenotyping—simultaneously quantifying nuclear morphology (size, shape, DNA condensation from DAPI), protein localization (nuclear vs cytoplasmic via channel colocalization analysis), and cell cycle state (based on DNA content histograms from integrated DAPI intensity). For engineered cell lines with reporter constructs, high-content imaging combined with AI analysis screens clone libraries: acquiring GFP fluorescence to confirm transgene expression, measuring expression intensity distribution to assess clonal heterogeneity (CV <25% target), and correlating expression with morphology to identify stable high-expressing clones. Our high-content workflows generate 50-100 GB of image data daily, requiring efficient data management (automatic compression, cloud storage on AWS S3) and high-performance computing (GPU-accelerated analysis on NVIDIA A100 clusters processing 200 images/minute). The combination of high-content imaging hardware generating rich multi-dimensional datasets and AI analysis software extracting maximum information from each imaging session enables us to perform sophisticated cell line characterization and quality control that would be impossible with manual analysis.

Time-Lapse Analysis and Dynamic Monitoring

Time-lapse microscopy provides valuable information about cell behavior over time, including division rates, migration patterns, and responses to environmental changes. At Cytion, we employ Sartorius IncuCyte S3 systems that capture images at 15-minute to 2-hour intervals for up to 14 days continuously, generating time-series datasets of 100-1000 images per culture well. AI analysis of these time-lapse sequences includes: single-cell tracking using algorithms like TrackMate or DeepCell to follow individual cells across frames, measuring division times by detecting mitotic events (cell rounding, subsequent daughter cell separation), quantifying cell migration speeds and directionality (mean squared displacement, persistence length), and identifying cell death events (characteristic morphology changes, cell detachment). For division tracking, we achieve 87% accuracy in mitosis detection using 3D convolutional networks (C3D architecture) that analyze spatiotemporal features across 5-frame windows, enabling automated calculation of population doubling times that correlate strongly (r=0.91) with manual cell counting measurements. Migration analysis uses optical flow algorithms and deep learning-based cell segmentation to track cell centroids frame-to-frame, calculating velocities (µm/hour) and chemotactic indices for migration assays. Time-lapse data reveals dynamic behaviors invisible in single timepoint images: we have identified cell lines with circadian oscillations in proliferation rate, detected heterogeneous division rates within cultures indicating subpopulation structure, and characterized response kinetics to Cell culture media changes or drug treatments. For quality control, time-lapse monitoring provides early warning of problems—we detect growth arrest (absence of divisions for >24 hours) or elevated death rates (>5% cells showing apoptotic morphology per 24 hours) much faster than endpoint measurements. The rich temporal data also enables predictive modeling: using early-phase growth kinetics (first 24-48 hours) to forecast final cell yields, trained via recurrent neural networks (LSTM architecture with 128 hidden units) achieving 82% accuracy in predicting whether cultures will reach target density at expected timing.

Standardization Across Imaging Platforms

Different microscopes, cameras, and imaging conditions can produce images with varying characteristics, potentially confusing AI models trained on images from a specific platform. At Cytion, we address cross-platform variability through comprehensive image preprocessing and normalization pipelines implemented in Python using OpenCV and scikit-image libraries. Our standardization workflow includes: flat-field correction to compensate for uneven illumination (dividing each image by reference flat-field image, subtracting dark current), color normalization for brightfield images using histogram matching or Reinhard color transfer, intensity rescaling to standardized dynamic range ([0,1] float or [0,255] uint8), and resolution harmonization via bicubic interpolation when images from different systems have different pixel sizes. For phase-contrast images which are particularly sensitive to optical settings, we employ CycleGAN-based domain adaptation that translates images from one microscope's appearance to match another's, trained on unpaired image sets from both systems. This preprocessing ensures models trained on IncuCyte images work equally well on ImageXpress or EVOS images after standardization. We validate standardization effectiveness by measuring model performance degradation when applied to new platforms: before standardization, accuracy drops 12-25% when models trained on one system are applied to another; after standardization, degradation reduces to <5%. Our standardization pipeline is automated in our image analysis infrastructure, applying appropriate transformations based on metadata tags indicating source microscope, so that images from all platforms flow through unified analysis workflows. This cross-platform robustness is essential for multi-site operations and enables sharing of trained models across the cell culture research community, advancing the field beyond individual laboratory implementations.

Integration with Laboratory Automation

AI-powered image analysis becomes even more powerful when integrated with automated cell culture systems. At Cytion, we have implemented closed-loop automation where IncuCyte imaging systems inside automated incubators (Liconic STX series) capture images every 2 hours, Python-based analysis pipelines process images within 5 minutes of acquisition using containerized inference services (Docker on Kubernetes), and analysis results feed into our Hamilton VENUS automation controller via REST APIs to trigger automated actions. For example, when confluence analysis indicates cultures have reached 85% (optimal passage density), the system automatically generates a worklist in VENUS that schedules the liquid handling robot to perform passage operations (aspirate media, add trypsin, neutralize, count cells, seed new flasks) within the next 4-hour window. Contamination detection probability >0.7 immediately quarantines affected cultures by moving them to isolated incubator zones and generating urgent alerts, preventing contamination spread. Viability estimates <80% pause automated processing and flag cultures for manual expert review. This integration creates autonomous culture management systems that maintain optimal cell health with minimal human intervention—our integrated systems successfully culture 200+ concurrent cell lines with 92% of passage operations performed completely automatically, human involvement required only for 8% of cultures flagged for exceptional conditions. The closed-loop operation includes safety interlocks: AI predictions below confidence thresholds (typically 0.75) trigger manual review rather than automatic actions, and all automated decisions are logged with explanation data for traceability and continuous improvement. System performance monitoring tracks key metrics including false positive rates for contamination detection (target <2%), accuracy of confluence-based passage timing (>90% of passages occur at 80-95% confluence), and correlation between predicted and measured post-passage viability (r>0.8), with quarterly reviews ensuring performance remains within specifications.

Training Data Generation and Annotation

The performance of AI models depends critically on the quality and quantity of training data. At Cytion, we maintain extensive, carefully annotated image databases covering all our cell line catalog under various conditions and passage numbers, currently totaling >150,000 annotated images representing >2,000 hours of expert annotation effort. Our annotation strategy combines multiple approaches: manual annotation by expert cell culture scientists using tools like LabelImg and VGG Image Annotator (VIA) for segmentation masks and class labels, semi-automated annotation where initial AI predictions are reviewed and corrected by experts (reducing annotation time by 60% while maintaining accuracy), and active learning where models identify images with high prediction uncertainty for prioritized annotation effort focused on edge cases. We maintain rigorous annotation quality control with inter-rater reliability testing—three independent annotators label subsets of 100 images, achieving Cohen's kappa >0.85 agreement for classification tasks and IoU >0.90 for segmentation annotations, validating annotation consistency. For continuous improvement, we implement systematic data collection protocols: all production images are automatically archived with metadata (cell line, passage, date, imaging system, culture conditions), regular batches undergo expert annotation adding diversity to training sets, and images associated with QC failures or unusual events are prioritized for annotation to improve edge case handling. Data augmentation expands effective training set size: rotations (0-360°), horizontal/vertical flips, brightness/contrast adjustment (±20%), elastic deformations (simulating microscope field variations), and Gaussian noise addition (σ=0.1) generate augmented variants during training, effectively 10x multiplying training data while improving model robustness to natural image variations. We also curate specialized datasets for particular challenges: contamination detection dataset includes 5,000+ images of bacterial, fungal, and mycoplasma-contaminated cultures; rare morphology dataset captures unusual phenotypes, debris, artifacts; multi-passage dataset tracks individual cell lines across P5-P30 documenting senescence and phenotypic drift. This comprehensive, well-curated training data infrastructure is fundamental to the accuracy and reliability of our AI-powered quality control systems.

Model Validation and Performance Monitoring

Before deploying AI models for quality control decisions, rigorous validation is essential. At Cytion, we follow structured validation protocols aligned with FDA guidance on software validation and machine learning for medical devices (applicable principles for GMP cell production): we partition datasets into training (70%), validation (15%), and test (15%) sets with stratification ensuring all cell lines and conditions are represented proportionally; perform k-fold cross-validation (k=5) during development to assess model generalizability; evaluate performance on held-out test sets never seen during training using comprehensive metrics including accuracy, precision, recall, F1-score for classification tasks, R², MAE, RMSE for regression tasks, and AUC-ROC curves for probability predictions; compare AI predictions against gold-standard measurements (expert manual assessment, flow cytometry for viability, microscope grid counting for confluence) across diverse test conditions; and conduct prospective validation where models run in shadow mode parallel to standard QC for 3 months before deployment, comparing predictions to actual QC outcomes. Once deployed, we implement continuous performance monitoring: automated comparison of AI predictions against periodic expert reviews (20% of cultures undergo parallel expert assessment), tracking of prediction confidence scores over time (declining confidence may indicate data drift), correlation analysis between AI quality scores and downstream batch performance metrics (post-passage viability, expansion success), and quarterly validation reviews examining model performance across cell lines and operating conditions. We maintain detailed validation documentation including model architecture specifications, training data characteristics (size, diversity, annotation quality), performance benchmark results, and change control records for model updates. When model performance degrades below acceptance criteria (e.g., confluence accuracy drops below ±5%, contamination detection AUC <0.90), we trigger retraining or recalibration: collecting additional training data from recent production batches, retraining models with updated datasets, validating updated models on new test sets, and implementing controlled deployment where updated models initially run in shadow mode before full deployment. This rigorous validation and monitoring framework ensures our AI-powered QC maintains accuracy and reliability over time despite evolving cell line portfolios, imaging equipment changes, and natural data drift.

Future Developments in AI Image Analysis

The field of AI-powered cell image analysis continues to advance rapidly with emerging techniques promising even greater capabilities. Developments we are actively tracking and piloting at Cytion include: 3D image analysis using volumetric segmentation networks (3D U-Net) for organoid and spheroid cultures, enabling measurement of organoid size, morphology, and internal structure from Z-stack images; label-free fluorescence prediction where models trained on paired brightfield/fluorescence images learn to predict fluorescence patterns from brightfield images alone, potentially replacing some staining requirements; self-supervised learning techniques (SimCLR, BYOL) that learn useful representations from unlabeled images, reducing annotation requirements by learning general cell image features without manual labels; foundation models for cell biology (analogous to GPT for language) pre-trained on massive diverse cell image datasets that can be fine-tuned for specific tasks with minimal data; real-time analysis during live imaging with inference latency <1 second enabling immediate feedback for automated experiments; and predictive models forecasting culture outcomes hours or days in advance from early-phase images, trained on longitudinal datasets linking early imaging features to final batch quality. We are also exploring multi-modal integration combining microscopy images with molecular profiling data (RNA-seq, proteomics) to discover imaging biomarkers predicting molecular phenotypes, and physics-informed neural networks incorporating biological constraints (cell cycle dynamics, nutrient consumption kinetics) to improve prediction accuracy and reduce data requirements. As these technologies mature, we expect to achieve even earlier problem detection through subtle pre-symptomatic changes invisible to current methods, more precise quality assessments through integration of diverse data modalities, and deeper insights into factors influencing culture success. These advances will enable Cytion to continue delivering the highest quality Cells and Cell lines with even greater consistency and efficiency, maintaining our leadership in quality and innovation.