Return to top
Unlocking the Potential of Artificial Intelligence in the Next Generation of Pharma Visual Inspection
Unlocking the Potential of Artificial Intelligence in the Next Generation of Pharma Visual Inspection

Share on


29 June 2021

Unlocking the Potential of Artificial Intelligence in the Next Generation of Pharma Visual Inspection

Inspection is a particularly challenging and primordial stage in the good pharma manufacturing process to ensure product integrity thus patient safety

The COVID-19 crisis has stressed-out badly healthcare system thus the pharma industry to accelerate their readiness in delivering continuously safe and genuine products worldwide. One of the optimistic technologies to leverage is the adoption of Artificial Intelligence (AI) for total quality management.


Discover the Antares Vision Radio, click here to listen to this article:

English Version - Italian Version

Despite its huge potential, fewer than 5% of healthcare organizations were using AI, based on a study by HIMSS Analytics in 2017. At that moment, one of the main barriers is the difficulty of explaining how decisions are made by AI because companies and regulators, with quality validation specialists often want documentation and explanations of such decisions as patient lives could be at risk.

AI-driven fully automated Visual Inspection machines


In a strategic reflection for EMA (European Medicines Agency) regulatory science to 2025[1], it was encouraged to unlock the potential of Artificial Intelligence in healthcare for many reasons and complete its validation process.  More precisely, they define in “Goal 2: Driving collaborative evidence generation — improving the scientific quality of evaluations.” how Artificial Intelligence can be used in decision-making so it should be exploited in automating product packaging and content inspection with algorithms to perfect quality, minimize false rejects, improve throughput, and minimize the risk of recall. Now, is the pharma capable of running continuous production in a critical time like this where human intervention cannot be scalable and efficient for each product inspection? Most pharma companies are implementing fully automated vision inspection systems with smart cameras to take pictures and collect images. Furthermore, within this automation, they are passing from traditional rule-based algorithms (machine learning) to the application of deep learning models to cover all problematic situations and avoid those false rejects deemed to be the only trade-off for good but misclassified products. 


Harnessing the power of AI in drug design and development, diagnostics and clinical trials, or personalized medicines to facilitate treatment are some use cases. Many other areas had to be explored better such as highly accurate inspection of tablets, capsules, lyophilized products, and in the form of suspensions or prefilled syringes, let us say products with difficult characteristics that require more intelligent visual systems, especially for high-cost pharmaceuticals, in which every single false reject is costing a lot (i.e., orphan drugs).  AI-equipped visual inspection machines will substantially increase defect detection accuracy by clearing out misinterpretations of supposed defects. Traditional systems can, for example, misclassify cosmetic defects or air bubbles as foreign particles at the rubber stopper of a PFS. Artificial Intelligence with DL mitigates such misclassification risk and reduces costly re-inspection for verification purposes. As consequence, the ROI is increased when the risk of recalls is decreased.

Now, let us consider a use case related to the blister and the usefulness of TCI (tablets and capsule automatic inspection machine) equipped with AI modeling algorithms. With traditional automatic systems, computer vision algorithms trained on a set of images to parametrize the threshold for good and defective capsules can misinterpret some cases leading to a high ratio of false rejects after a training of 15-30 minutes to set each parameter. Some common problems to inspect are the presence of foreign particles, crushed/broken product, only body/cap in capsule, changes in shape/ size and form and Spots or discolorations. In this case, an intelligent camera-based inspection system for blister packaging machines can ensure defect–less product packaging, with minimal human intervention and no requirement for rework. So, the TCI machine had to leverage deep learning modeling on a very qualitative and highly representative training dataset to cover all possible defects for accurate detection.

Moreover, and to speed up its adoption, many machine vision software companies are offering deep learning algorithms to reuse so few efforts are required to do some software modifications on a higher computer processing power with GPUs (Graphical Processing Units, already used in gaming industry). Besides, PDA[2] 2021 had revealed how US FDA and EMA are encouraging the use of artificial intelligence for faster and accurate decision-making with all the regulatory requirements to 2025.
Newer revisions of USP <1790> Chapter[3] for Injections and EU Annex 1 for sterile products, are allowing to benefit from new technologies and expected to be published in 2021. There is still one crucial point that must be considered to enable validation: contrary to other industries, the deep learning model for pharmaceutical use cases must be “frozen” once the development phase is finalized. It must be static and can no longer change to make it version-controlled for validation. A recent paper published by the FDA about the regulatory framework[4] for Software as a Medical Device (SaMD) based on AI/ML, is a good reference where the versioning had to be frozen when the system is validated. That is why most pharma companies had two machines, one physical for live production and one for testing.


Now how to implement it? As “one size fits all” approach cannot work with deep learning projects for visual inspection, the first step consists in collaborating with the customer to make accessible but secure, many images about his product as reference samples to pre-assess and keep only the very qualitative and representative ones to train the machine DL-based. For example, this could be images of good units with bubbles, different stopper positions, and filling volumes for body inspection, as well as different types of particles, and for any kind of product: liquid and lyophilized, capsules or tablets and softgels, in dual-chamber syringes, vials or blisters. Based on the available image data, offline verification studies and testing will support the integration of deep learning model into the existing software. The recipe parameters are still validated according to GMP requirements. The only changes are the tool used to develop the process and the required hardware with GPUs to be able to process very complex and large amounts of data. In a deterministic deep learning model, small packages are trained up to a certain “level of intelligence considered of highest accuracy “and then freezing the model at this version so it can be validated and gets the regulatory approval in qualification strategy and implementation.

Antares Vision Group,
 a global leader in Track & Trace and Inspection machines, complemented with smart data management platforms, is offering a wide range of solutions that implement AI in automatic visual inspection for pharma manufactures but not limited to. Because packaging and product inspections are critical to ensuring product quality and safety, we had developed our TCI - Tablet and Capsule automatic Inspection machine with deep learning for pharma drugs and nicotine gums and the UPI - Universal Packaging Inspection for blisters, bottles, cartons, etc. and labelling inspection, powered by NOCR (Neural OCR) to address problems which cannot be tackled using traditional OCRs or outperform the existing solutions for variable printing quality, distortions on the image, uneven surfaces, variable lighting conditions and non-uniform backgrounds, just to cite few. Besides, other AI-equipped machines are developed for PFS (prefilled syringes) and lyophilized/biopharmaceutical products to precisely catch any micro-cosmetics defect and/or product alteration. Concerning the governance of AI-enabled systems, Antares Vision Group - through Orobix - can provide fully managed AI governance and lifecycle management for its clients.® allows to get a complete picture of the AI processes at any point in time, ensuring compliance, and establishing proper governance when implementing AI mission-critical processes (such as pharmaceuticals). This is perfectly compliant with the proposal for a REGULATION OF THE EUROPEAN PARLIAMENT AND OF THE COUNCIL LAYING DOWN HARMONISED RULES ON ARTIFICIAL INTELLIGENCE (ARTIFICIAL INTELLIGENCE ACT)[5] and the European regulatory strategy to 2025 for the use of AI technology.

Stay tuned for the next upcoming scenarios on how to optimize AI in our industries.







No comments have been posted yet. Please feel free to comment first!

Leave or Reply a comment for this article