Introduction

Cell and gene therapy manufacturing combines the precision demands of pharma QC with the complexity of biological variability and the low batch counts of autologous workflows. Vision analytics — CV-based inspection for cell viability, container integrity, fill verification, and process monitoring — has emerged as a critical tool, but with caveats specific to the CGT (cell and gene therapy) context. This article maps where automated visual inspection replaces manual review in CGT, which defect classes it handles, what validation looks like for autologous-batch-size workflows under GMP, and where human judgement still wins. See the life sciences landing and the computer vision landing for the broader programme.

What this means in practice

• CGT inspection inherits pharma QC requirements with extra biological variability.
• Autologous workflows mean batch size of one — validation strategy adapts.
• CV reduces fatigue-driven false negatives; humans still own edge cases.
• GMP validation for AI inspection requires golden datasets and PQ.

How does computer vision replace manual visual inspection in CGT QC without losing defect sensitivity?

The replacement is not wholesale but per-defect-class. The pattern (2026):

Container inspection (bag integrity, vial seals, ports). CV-based inspection ships and outperforms manual on consistency. Defect sensitivity equal to or better than human for well-defined defects (cracks, leaks, missing seals); throughput much higher.

Fill-volume verification. CV-based volumetric measurement using calibrated optics; ships at scale. Sensitivity to under-fill / over-fill superior to visual estimation.

Particulate detection in cell suspensions. CV-based detection of foreign particulates ships; sensitivity depends on particulate size, suspension turbidity, container geometry. Better than human for sub-millimetre particles in well-controlled imaging conditions.

Label verification. OCR + CV for label correctness, orientation, lot/serial matching. Ships at scale; defect-free at production speeds humans cannot match.

Cell-state inspection (viability, morphology, contamination). Hybrid. CV-based screening triages; human reviews flagged exceptions. Pure CV is not yet at the level of pathologist-trained eye for complex cell-state judgements.

Process-step verification (correct manipulations, no missed steps). CV monitoring of operator actions; ships in some manufacturing lines.

The sensitivity preservation strategy. The CV system is calibrated against a golden dataset of known-good and known-defective samples. False-negative rate is held at or below the manual baseline by design. False-positive rate is tuned to be slightly higher than manual (CV errs on the side of escalating for human review).

The “no loss of sensitivity” requirement. The CV system’s PQ (performance qualification) explicitly compares to manual inspection on the same defect distribution. Approval is contingent on equivalent or better detection.

The CGT-specific tuning. Small batch sizes mean defect statistics are thin; the golden dataset must be carefully curated; ongoing-monitoring data accumulates slowly. Performance validation is more conservative than for high-volume pharma.

Which defect classes (particulates, cracks, fill level, labelling) can automated visual inspection reliably detect today in cell-and-gene-therapy products?

The reliably-detected defects:

Container defects:

Cracks in glass/plastic containers. Reliable with appropriate lighting and CV algorithms. Production-validated.

Seal integrity (caps, plugs, bag seals). Reliable for well-controlled defect classes.

Surface scratches and contamination. Reliable for visible defects above a size threshold.

Container shape and dimension. Reliable.

Fill defects:

Under-fill / over-fill. Reliable with calibrated volumetric measurement.

Fill-level variation across batch. Reliable.

Foam / bubble detection. Variable — depends on bubble size, persistence, container geometry.

Particulate defects:

Sub-millimetre foreign particles. Reliable in clear suspensions with good imaging.

Particulate in cloudy / turbid suspensions. Variable — sensitivity drops with turbidity.

Particle size distribution. Reliable for size estimation; classification (intrinsic vs extrinsic) limited.

Labelling defects:

Missing label. Reliable.

Wrong label / wrong lot. Reliable with OCR validation against expected.

Label orientation, position. Reliable.

Print quality (smudging, missing barcode). Reliable.

The CGT-specific defect classes:

Cell-suspension uniformity. CV measurement of suspension homogeneity; production-deployed.

Aggregate / clump detection in cell suspensions. Detection reliable; classification (acceptable vs critical) requires human judgement.

Colour change (oxidation, contamination indicators). Reliable for measurable colour shifts; subtle changes require correlation with chemistry.

The defects requiring human or hybrid:

Cell morphology abnormalities at fine resolution. Requires hybrid CV + pathologist review.

Subtle contamination signs (early bacterial growth before visible turbidity). Detection requires correlated tests; CV alone insufficient.

Functional defects (viable cell count, potency). Not visual; requires functional assays.

The 2026 reality. The CV-detectable defect set covers most visual inspection items. Edge cases and CGT-specific morphological assessments retain human review. The labour saving is substantial; the human role shifts from primary inspection to exception handling and complex judgement.

What does an automated visual inspection deployment cost compared with manual inspection at the same throughput in CGT?

The cost components:

CV system capital. Inspection station hardware (cameras, optics, lighting, motion); software and validation; integration with line. Capital cost: €100k–€1M+ per station depending on complexity and validation requirement.

Software licence. Annual licence for inspection software; vendor-dependent.

Validation and qualification. Initial PQ studies, ongoing IQ/OQ/PQ documentation; significant cost for first deployment.

Maintenance and support. Vendor support contracts; in-house support resources.

Operator training. Reduced for primary inspection (less manual labour), increased for exception-handling and validation.

Re-validation. Each major change (camera, model, line modification) requires re-validation; ongoing cost.

The manual inspection cost components:

Operator labour. Trained inspector wages and benefits; cost-per-hour.

Operator training and certification. Ongoing requirement.

Fatigue management. Rotation, break schedules, error mitigation.

Re-inspection (failed batches). Manual second-review cost when defect found.

Documentation overhead. Manual recording, batch records.

The cost comparison pattern:

At high throughput. CV system pays back in months — operator labour at 24/7 inspection of high-throughput line is substantial.

At low throughput (CGT autologous, low batch counts). Payback period extends; CV cost-justified more on quality consistency and audit-trail benefit than on labour reduction.

At medium throughput. Mixed; depends on labour cost and CV cost.

The CGT-specific cost factors:

Batch size of one. CV per-batch amortisation is unfavourable for small batches; system utilisation is the lever.

Multi-product lines. CV system that handles multiple product types amortises better; flexibility matters.

Validation cost dominance. For CGT, validation cost can exceed equipment cost; the validation discipline is the gating factor.

The total-cost-of-ownership pattern. For high-volume conventional pharma, CV is dramatically cheaper. For low-volume CGT, the financial justification rests more on quality outcomes (audit trail, consistency, reduced miss rate) than on labour cost. The decision is risk-management as much as cost.

The 2026 commercial reality. Most CGT manufacturing facilities deploy CV-based inspection for at least some inspection stages; full coverage is not yet universal because the cost-benefit varies by stage and the validation overhead is significant. The trajectory is toward greater CV coverage as validation costs decrease through reusable patterns.

How is a CV-based inspection system validated under GMP — golden datasets, performance qualification, ongoing monitoring?

The validation lifecycle:

Design qualification (DQ). User requirements specification; design specification; risk assessment; vendor qualification.

Installation qualification (IQ). Hardware installed per specification; cabling, lighting, motion correct; environmental conditions met.

Operational qualification (OQ). Software runs correctly; calibration successful; interface to MES (Manufacturing Execution System) functional; alarms and alerts trigger appropriately.

Performance qualification (PQ). The critical AI/CV-specific stage. Inspection performance evaluated against golden dataset; per-defect-class sensitivity and specificity measured; compared to acceptance criteria.

Ongoing monitoring. Daily / per-batch validation samples; drift detection; trend analysis.

Change control. Any change (model retraining, camera replacement, lighting modification) triggers re-validation per impact assessment.

The golden dataset:

Composition. Mix of known-good and known-defective samples, covering all defect classes the system is expected to detect; representative of production distribution.

Labelling. Each sample labelled with ground truth — defect class, severity, location. Labels reviewed and agreed by quality team.

Maintenance. Updated as new defect classes identified; aged samples replaced; provenance documented.

Use. PQ baseline; periodic re-test; new-model validation; change-control validation.

The performance qualification:

Acceptance criteria. Defined before PQ — sensitivity per defect class, specificity, throughput, false-positive and false-negative limits.

Test execution. Run inspection system on golden dataset; compare to ground truth; calculate performance metrics.

Documentation. Test protocol, results, deviation report (if any), conclusion. Signed off by quality and validation.

The ongoing-monitoring elements:

Per-batch validation samples. Known-good and known-defective samples included in routine batches; system performance tracked over time.

Drift detection. Monitor for changes in detection rate, false-positive rate; trigger re-validation if drift exceeds threshold.

Trend analysis. Long-term performance trends; identify degradation, environmental impact, model staleness.

The AI/ML-specific governance:

Predetermined change control plan (PCCP). For AI-based systems, define in advance what changes are pre-approved (e.g., retraining on additional data with same architecture) and what changes trigger full re-validation. Aligned with FDA’s emerging AI/ML guidance.

Model versioning. Each model version has a unique identifier; validation results tied to version; rollback capability.

Training data provenance. Documentation of training data sources, composition, labelling provenance.

The CGT-specific validation:

Small-sample challenges. Limited production data; golden dataset must be deliberately curated; synthetic data augmentation may supplement.

Cross-site reproducibility. CGT manufacturing distributed across multiple sites; validation must establish per-site performance.

Product-specific variation. Each CGT product may have unique inspection requirements; per-product validation.

When does AI-based inspection outperform deterministic machine vision, and when is the simpler approach correct?

The criteria:

AI wins when:

Defect classes are visually variable. Particulates, contamination patterns, cell morphology — no fixed template; learned representations outperform rules.

Lighting / contrast variation is significant. AI models robust to variation outperform rule-based methods that depend on tight imaging control.

Edge cases are common. AI models trained on broad data handle edge cases better than rule sets that must enumerate them.

Multi-product line with rapid changeover. AI models handle multiple product types with less per-product tuning than rule-based.

Defects evolve. AI models can be retrained on new defect classes; rule-based systems require expert rule re-engineering.

Deterministic machine vision wins when:

Defects are well-defined and stable. Specific dimensions, specific colour ranges, specific shape requirements — rules outperform learned models in interpretability, validation simplicity, and cost.

Audit requirement is paramount. Rule-based logic is fully auditable; AI models require additional explainability infrastructure for the same audit-level confidence.

Throughput is extremely high. Rule-based systems with simple checks can run at higher throughputs with deterministic latency.

Validation simplicity matters. Rule-based systems have simpler validation profiles; for short-product-lifecycle items, the simpler validation amortises faster.

Compute or cost constraints. Edge-deployed inspection on low-cost hardware favours rule-based; AI models require more capable compute.

The hybrid pattern (most common in practice):

Stage 1: Rule-based pre-screening. Fast, deterministic, handles 70-80% of inspection items with clear pass/fail.

Stage 2: AI for ambiguous cases. AI model handles cases where rule-based produces uncertain or borderline result.

Stage 3: Human for AI-flagged exceptions. Quality team reviews AI-uncertain or AI-flagged exceptions.

The CGT-specific pattern. Multi-stage hybrid is the production norm. Rule-based for high-throughput, well-defined defects; AI for biological variability and complex pattern recognition; human review for novel or critical exceptions.

The 2026 trend. The “AI vs machine vision” framing is increasingly obsolete; mature deployments combine both with workflow that uses each where it’s strongest.

How do CV systems handle difficult-to-inspect products (suspensions, opaque vials, lyophilised cake) where humans also struggle?

The hard cases:

Opaque or turbid suspensions. Cells in growth media, suspensions with significant scattering. CV approach: multi-angle illumination, polarised light, hyperspectral imaging; specialised systems exist but cost and complexity are higher.

Opaque vials. Amber glass, plastic with opacifiers, freeze-dried preparations. CV approach: X-ray or near-IR imaging; specialised equipment; limited to large defects.

Lyophilised cake (freeze-dried product). Visual inspection of cake quality (crack, melt-back, collapse, contamination). CV approach: shape analysis, surface texture analysis, colour analysis; production deployments exist for major defects, subtle cake-quality assessment remains hybrid.

Suspensions with cell clumps. Detecting abnormal aggregates vs normal cell distribution. CV approach: image analysis with size-distribution algorithms; works well for clear distribution but requires per-product calibration.

Frozen products. Visual inspection of frozen formulations (ice morphology, container integrity at low temperature). CV approach: imaging through cold containers, specialised optics; emerging deployments.

Multi-layer or stratified products. Suspensions with separation layers; visual inspection of layer interface, layer thickness. CV approach: vertical-imaging with layer-detection algorithms; production-deployed for specific products.

The strategic patterns:

Imaging modality matters. The right imaging (multi-angle, polarised, IR, X-ray) extends CV reach to products that simple visible-light cameras can’t handle. Investment in imaging hardware enables CV deployment.

Pre-processing matters. For turbid or opaque products, image pre-processing (deconvolution, denoising, contrast enhancement) can improve CV performance significantly.

Hybrid is the answer for the hardest cases. Even with advanced imaging, the most complex inspection (subtle morphological abnormalities, novel contamination patterns) retains human review. The CV system handles the routine; humans handle the unknown.

Limits of CV for biological variability. Cell-state assessment, viability indicators, functional defects — these are not pure visual problems; they require correlated tests (chemistry, functional assays). CV provides one input; the full inspection is multi-modal.

The CGT-specific challenge. CGT products often combine multiple difficult inspection categories — cell suspensions in container with frozen handling and visual cake-quality requirements. The inspection station design must accommodate this complexity.

The 2026 trajectory. CV deployment for difficult products expands as imaging technology becomes more capable and as validation patterns become reusable. The fundamental hard cases (subtle biology, novel contamination) remain hybrid, but the CV-handleable share grows over time.

Limitations that remained

Cell-state assessment retains human judgement. Subtle morphological abnormalities and functional defects require pathologist-trained review and correlated tests; CV alone insufficient.

Small-batch statistics limit validation. CGT autologous workflows produce thin defect data; validation requires conservative interpretation and ongoing monitoring with curated golden datasets.

Validation cost is high. Initial PQ for CV inspection in CGT is significant; re-validation on change is non-trivial; this gates rapid iteration.

Cross-site reproducibility is harder. Distributed CGT manufacturing requires per-site validation; harmonising performance across sites takes deliberate effort.

Difficult-imaging products still hybrid. Opaque vials, lyophilised cake, complex suspensions — CV handles routine; humans handle the hardest exceptions.

How TechnoLynx Can Help

TechnoLynx works with CGT and pharma manufacturing teams on production CV inspection — system design, defect-class scoping, GMP validation patterns, hybrid AI + rule-based pipelines, cross-site deployment. We focus on validation-discipline-first deployments that scale. If your team is scoping CV inspection in CGT or pharma, contact us.

Image credits: Freepik