Q&As on Automated Visual Inspection (AVI)

The visual inspection of parenteral products is described in detail in pharmacopoeias. Nevertheless, there remains room for interpretation in practical application, especially in automated inspection, which regularly gives rise to questions. The questions and answers compiled below address typical uncertainties encountered in practice and provide guidance based on current regulatory requirements and best practices.

Q: Annex 1 requires challenge testing prior to start-up and at regular intervals. How often should challenge testing be performed during batch production?
A: EU GMP Annex 1 (8.32) states that the performance of automated inspection equipment should be challenged using representative defects prior to start-up and at regular intervals throughout the batch. However, the regulation does not prescribe a fixed testing frequency or require routine in-batch testing at predefined intervals.

In practice, a functional or system suitability test is typically performed before the start of batch inspection and repeated after completion of the batch. If the system demonstrates acceptable performance at both time points, it can reasonably be concluded that system performance was maintained throughout the batch.

These tests should primarily be understood as functional verifications confirming that the inspection system is operating as intended (e.g. camera functionality, lighting performance, piping of cables and connectors, selection of the correct recipe, stable trigger and signal transmission, and proper communication between system components), rather than as repeated small-scale performance qualifications. The purpose of these checks is to exclude mechanical, electrical, or hardware-related disturbances that may occur during routine operation, format part changes, cleaning, maintenance activities, or other technical interventions, as such influences may affect individual system components.

In contrast, the inspection concept itself - including the design of the optical system, such as camera resolution, selection of lighting, lens type, as well as the implemented parameterization of the image processing algorithms and associated recipes - does not change on a day-to-day basis as long as the system remains under a controlled change management process. Verification that this inspection concept is capable of achieving the required detection performance is performed as part of Performance Qualification (PQ) using appropriate representative defect samples.

Routine interruption of the inspection process may introduce operational risks, including the potential mix-up of test units with product units, while typically providing limited additional assurance when modern inspection systems are properly qualified and equipped with continuous internal monitoring and self-diagnostic functions (e.g. camera checks, lighting intensity monitoring, and reject rate alarms).
However, functional in-batch challenge testing may be appropriate in specific situations, including:

• After maintenance or technical intervention during a batch
• After a machine restart following a system crash or software reset
• In the event of a malfunction, alarm, or suspected performance drift

In such cases, event-driven verification is consistent with Annex 1 expectations and Quality Risk Management principles.

Q: What aspects should be considered when assessing artificially created test kits? What constitutes adequate justification?
A: Artificially created test kits are used when the real product cannot be used due to technical or safety-related reasons, or when the number of samples required for evaluation, training or qualification purposes cannot be obtained in sufficient quantity from routine production.
The assessment of the suitability of artificially created test kits or individual samples should be based on the inspection technology used and the resulting required degree of representativeness. The key requirement is that defective and acceptable samples used for qualification exhibit behavior under routine conditions that is comparable to the real product with regard to their visual and physical characteristics.

For AI-based inspection systems, higher requirements regarding representativeness may apply, as such systems often utilize very subtle features and characteristics to distinguish between acceptable and defective units. For example, a genuine lyophilized cake with crystalline reflective properties can generally be distinguished from a matte plaster-based imitation, provided that these differences are visible within the image data.

If artificially created defects differ in relevant characteristics - such as texture, reflectivity, contrast, size, movement behavior, or other optical or physical dimensions - there is a risk that the neural network learns to use the specific characteristics resulting from the artificial manufacturing process for differentiation, rather than evaluating the representation of the actual defect itself.

In conventional rule-based image processing systems, the risk of overfitting to specific characteristics is generally lower, as the vision engineer typically designs the algorithm to assess potential defects based on relatively independent and simple image features, such as size, contrast, grayscale values, morphology, or similar parameters. These approaches are generally less selective and may be less suitable for highly challenging inspection tasks, but they often demonstrate greater robustness with regard to variations and realism drifts within artificially created test kits.

Nevertheless, the assessment of the suitability of a test kit for the intended application remains a quality-relevant requirement. For example, it is evident that the detection of particles may be directly influenced by factors such as viscosity, turbidity, and fill height of the solution, as well as by the contrast and size of the particles themselves.

The suitability of artificially created kits should therefore always be evaluated and appropriately documented with consideration of the described inspection risks and the inspection technology used, particularly with regard to optical, physical, and mechanical behavior in comparison to real production material.

In summary, the decisive factor is not whether the test set contains artificially created or real samples taken from production, but whether it reliably represents the real-world behavior of the final drug product in relation to the selected inspection strategy.

Q: Should automated systems be qualified against fixed acceptance criteria, or should they only be compared to human inspection performance?
A: According to EU GMP Annex 1, automated inspection systems must be equal to or better than manual inspection.
If human operators are qualified against predefined acceptance criteria, it is reasonable and consistent to apply similar qualification principles to machines. However, it must be understood that qualification represents only a snapshot of performance at a specific point in time. Long-term process reliability depends much more on continuous monitoring, trending analysis, and robust process control than on a one-time qualification exercise. Qualification is necessary, but ongoing process oversight should be rated even more important.