Computer Vision Development in Zurich, Switzerland

Pharmaceutical / Biotechnology

A Swiss biologics manufacturer producing injectable medications (monoclonal antibodies in pre-filled vials — high-value products at CHF 800-2,000 per vial) needed to replace its semi-automated visual inspection system. The existing system combined: automated particle detection (a machine that spun vials and detected particulate matter using light scattering), followed by manual visual inspection (trained inspectors examining each vial for: fill level accuracy, vial integrity (cracks, chips), stopper seating, cap crimping, label presence and alignment, and cosmetic defects (scratches, stains on vial exterior)). The problems: (1) Inspection throughput: manual visual inspection limited throughput to 2,400 vials per hour per inspector. The production line produced 4,000 vials per hour — requiring 2 inspectors running in parallel, creating a bottleneck that slowed production during peak demand. (2) Inspector fatigue: 8-hour inspection shifts with required breaks every 30 minutes (industry best practice for visual inspection). Even with breaks, inspector accuracy degraded measurably in hours 6-8 compared to hours 1-2 — studies at the facility showed 15% higher false negative rates (missed defects) in late-shift inspection. (3) Subjectivity: "acceptable" cosmetic defects (minor scratches, small stains) were judged subjectively. Different inspectors had different thresholds, creating inconsistent reject rates across shifts (ranging from 1.8% to 3.2% — a significant variance for a product worth CHF 800-2,000 per unit). (4) Regulatory documentation: manual inspection relied on inspector training records and periodic competency assessments. Regulatory auditors (Swissmedic, FDA) increasingly questioned whether manual visual inspection was sufficiently controlled and documented. The company wanted to demonstrate regulatory leadership by implementing validated automated inspection.

We developed a GMP-validated computer vision vial inspection system over 18 weeks: (1) Multi-camera inspection station: 8 synchronized cameras (4 Mpx Basler ace2 with telecentric lenses) arranged around a rotating vial holder: 2 side cameras (detecting cracks, chips, label placement, surface defects), 1 bottom camera (vial base integrity), 1 top camera (stopper seating, cap crimping quality), 2 meniscus cameras (fill level measurement through the transparent vial wall — detecting underfill and overfill with ±0.5% volume accuracy), and 2 label cameras (OCR reading lot number, expiry date, product name, barcode — verifying against batch record data). Controlled LED illumination (dark field for particle and crack detection, bright field for surface defects, backlighting for fill level) eliminated environmental lighting dependency. (2) Multi-task CV model: the system runs 6 specialized models sequentially on each vial (total inference time: 120ms per vial — supporting 6,000 vials/hour throughput): fill level model (regression model predicting fill volume from meniscus image — trained on 8,000 images with calibrated fill volumes), crack/chip detection (object detection model identifying structural defects — trained on 5,000 defect images including synthetically generated crack patterns overlaid on clean vial images, because real defect images are rare for high-value biologics), stopper/cap model (classification model verifying correct stopper seating depth and cap crimping — trained on 12,000 images across 4 stopper/cap combinations the facility uses), label verification (OCR + layout verification — reading printed text and comparing against expected batch data, checking label alignment within ±1mm tolerance), surface defect detection (segmentation model identifying scratches, stains, and foreign material on vial exterior — trained with augmented data to handle the transparent/reflective properties of borosilicate glass), and overall classification (an ensemble model combining outputs from all 5 models to produce Accept/Reject/Review decisions — with configurable sensitivity thresholds per defect category). (3) GMP validation package: complete GAMP 5 validation documentation: URS (23 user requirements covering inspection scope, accuracy targets, throughput requirements, and regulatory compliance), FS and DS (detailed system specification including model architectures, training data descriptions, and hardware specifications), IQ protocol and report (documenting hardware installation, camera calibration with NIST-traceable reference vials, and software installation verification), OQ protocol and report (testing system performance across the full operating range — fill level accuracy at minimum and maximum fill, defect detection at minimum and maximum line speed, performance with all product/vial/stopper combinations), PQ protocol and report (3 consecutive production runs demonstrating consistent performance — inspecting 12,000+ vials per run with comparison against parallel manual inspection), and validation summary report (consolidating all validation evidence for regulatory review). (4) Continuous monitoring: the system logs every inspection decision with annotated images — creating a complete audit trail. Statistical process control charts track: reject rate over time (detecting drift from expected reject rate), defect category distribution (detecting shifts in defect types that might indicate production issues), and model confidence distribution (detecting degradation in model certainty that might indicate calibration drift or model degradation).

Inspection throughput: 2,400 vials/hour (manual, per inspector) → 6,000 vials/hour (automated). The production line bottleneck was eliminated — the inspection station now runs faster than the filling line, providing buffer capacity. Inspection consistency: reject rate variance across shifts reduced from 1.8-3.2% (manual) to 2.1-2.3% (automated) — the system applies identical criteria 24/7 regardless of operator, shift, or fatigue. Defect detection: the CV system detected 99.4% of known defects in a challenge test set (100 vials with known planted defects) — compared to 94.1% detection by the best human inspector and 89.3% by the average inspector. Critically: the CV system detected 100% of structural defects (cracks, chips) — the highest-risk defect category for patient safety. False reject rate: 0.8% (CV system) vs. 2.1% average (manual inspection). At CHF 800-2,000 per vial, the reduced false reject rate recovers approximately CHF 520,000 annually in product that would have been incorrectly rejected by manual inspection. Regulatory response: Swissmedic inspectors reviewed the validated CV system during a routine GMP inspection and noted it as a "positive observation" — the validated system with complete documentation and continuous monitoring exceeded the regulatory expectation for visual inspection control. FDA pre-approval inspection for a new product similarly noted the CV inspection system favorably. Inspector redeployment: the 4 visual inspectors (2 per shift) were redeployed to upstream quality roles (in-process monitoring, environmental monitoring) — not laid off. The company had unfilled quality positions that the inspectors filled after retraining. Development cost: CHF 420,000 (including cameras, industrial PC, custom inspection station fabrication, software development, and GMP validation). Annual savings: CHF 520,000 in reduced false rejects + CHF 340,000 in labor redeployment value + unquantified regulatory risk reduction. Payback: under 12 months.