The importance of quality in pharma manufacturing can't be overstated. Fortunately, the industry has stepped up drug quality. While much of this can be attributed to better science and clearer guidance, technology improvements — like the application of AI to visual inspection equipment — have potential. Chief content director Karen Langhauser and Giacomo Girotto, vision inspection project manager for Stevanato Group discuss.
Karen Langhauser: In 1950, British mathematician, Alan Turing, pondered a simple question, "Can machines think?" Now recognized as a founding father of artificial intelligence AI), Turing's published paper and its subsequent Turing Test helped establish the fundamental goal and vision of AI. Mostly the world has been familiarized with the concept of AI through science-fiction writing, where AI robots simultaneously amaze and terrify us with their abilities, but the AI being used in the pharma industry may play a very critical role in bringing safer and more reliable drugs to patients. The importance of quality in pharma manufacturing cannot be overstated as we're constantly reminded of the cost of poor quality with FDA warning letters, recalls, drug shortages, and in worst-case scenarios, public health crises. Fortunately, over the past decade, the industry has really stepped up drug quality. While much of this can be attributed to better science and clearer guidance on quality control and criteria, technology improvements — such as the application of AI to visual inspection equipment — have the potential to be a key part of the quality puzzle. And that specifically is what we're here to talk about today.
I'm Karen Langhauser, chief content director of Pharma Manufacturing magazine, and you're listening to a special solutions spotlight edition of Off Script: A Pharma Manufacturing podcast about what's happening behind the scenes in our magazine and what's trending in the drug industry. Joining me today to talk about this is Giacomo Girotto, the vision inspection project manager for Stevanato Group. Giacomo has extensive experience working on special projects, including application of AI in the manufacturing process. Welcome, Giacomo.
Giacomo Girotto: Hi, Karen. Hi, everybody. Thank you very much for inviting Stevanato Group and me to participate at this podcast.
Karen: Thanks for being here with us. Why don't we start by talking about quality challenges? Why do some drugs, such as drugs in the form of suspensions or lyophilized cakes, cause problems for traditional inspection systems?
Giacomo: Yes. So let's talk about some kind of difficult to inspect product as we recognize them in the market. So among these, we have, for example, heavy suspensions and the lyophilized drugs, which are considered difficult to inspect products in our business mainly because their features poses some challenges. A suspension is not clear like a transparent water-like liquid, for example. It is a heterogeneous mixture. This generates two issues. The first one is that you need to homogenize the product to make the solute particles get suspended in the solvent and, secondly, the contaminants can hide inside the product at the center of the container, and this makes difficult from the cameras to see if they are not spinning the correct way and make them visible. On the other hand, when we look into the lyo, so the lyophilized products present challenges. For example, when looking into the cake area to the viable surface, and topology, and the appearances that this can take. So for example, you can have...there are cracks that might create shadows or patterns that can be misclassified as contaminants and also even hide the particles, thus, causing high false rejection rates. Other products in this category typically causing some headaches are drugs prone to bubbles. In this case, standard vision software can be challenged as it can be complex to differentiate... They found blobs in the image between the bubbles and a particle. So, these are some of the challenges connected to difficult to inspect products that we experience in the market.
Karen: So it seems vision inspection technology has really evolved over the past decade, you know, to the point where you're able to apply different levels of AI, such as machine learning or deep learning models to vision inspection. Can you talk a little bit about these advancements?
Giacomo: Yes, you're right. Let me say that these technologies are not totally new as concepts. In fact, the first time the words artificial and intelligence had been put together was, like, back in 1950s when a professor, John McCarthy, together with some colleagues from the University of Dartmouth asked the Rockefeller Foundation to fund a summer research on "artificial intelligence," indeed. After that, several studies and investment have been performed also from big companies like IBM and so on. However, there was a limited spread of this technology due to some limitations that there were back then, and then it wasn't until the early 2000s when the terminology "machine learning" started to get used again. And just to give you an idea, machine learning is, in fact, dated back to the 1959, and this is now coming back again. Thanks to the advancement in high resolution cameras and improvement in the hardware, which increased the image quality, together with the capacity to collect and process data, and, in particular, to be able to store them.
So, to use a machine learning, deep learning, this type of technology, the main factor that allow to use them are the capacity of store the data and the availability of this data, something that today is much more available and easier to get. So, nowadays, we see that the first application for this machine learning are going to be in full operation in different type of sector, including the pharmaceutical and the healthcare ones. So, the pharma industry has always looked at the best and most cutting-edge technology available as it enables them to accomplish their mission, which is to bring safer drugs to the market, so it's a natural process. We can say that these are natural process in which we are in. So the best technology are becoming more available, and the pharma start to benefit from them.
Karen: So a long and interesting history. In present day, how can deep learning and visual inspection help improve performance?
Giacomo: Yes. So from different studies performed, we saw that the deep learning models can improve the overall inspection quality, which is often connected to the performance in terms of detection rate and false rejection rate. Basically, it optimizes the correct balance of detecting the rejects and incorrectly rejecting the good items, which are usually called "false rejects." This is achieved thanks to its feature and the capability to learn the different type of defects. So being able to determine if it's a black particle or if it's metal, if it's rubber, if it's fiber. So, we are able to identify these particular defects, and, in particular, this last point brings advantages, making this system more robust to the variability in the product, and/or variability coming through the process in which the problem goes through during this production. So basically, we can say that deep learning can help reducing the false rejects, but also the costly intervention to parameterize the machine in production. Summarizing, it helps pharma companies and contract manufacturers to the decrease their total cost of ownership. Just to give you an idea, some studies we recently ran gave promising results, demonstrating that we can reduce the false rejects tenfold.
Karen: Okay. So just so I'm understanding this, there's typically a trade-off then between the detection rate and the false rejection rate when it comes to vision inspection in pharma. And, of course, that could be costly, especially when it comes to high-value biopharmaceutical drugs. So, how has Stevanato overcome that trade-off?
Giacomo: Yeah. I would say that this is achieved thanks to a mix of innovations that we, as the Stevanato Group, achieved over the last 30 years in inspection that we have been working on. We have used several different strategies to find the right balance, starting from the machine and the mechanic design. For example, we engineer sophisticated product handling systems and recently even use the robotics extensively, for instance, in our newly developed Vision Robot Unit. We're also obviously working on the inspection setup and the vision software to ensure valuable performances. And this is the mix of innovation and technologies you must follow in order to work on this balance. And lately, also now integrating the deep learning is the latest development in this direction and it's a great tool to support further reduction of the trade-off between the detection rate and the false rejection rate. And likely, traditional rule-based systems, deep learning models, thanks also to the data augmentation technique, can generalize their prediction and be more flexible regarding all the case that can occur during the production. It will allow for cost savings, but maintaining and improving the overall quality.
Karen: So, Stevanato Group recently launched this AI platform and it's based on deep learning models. Can you talk specifically about the platform and what it brings to the pharma industry?
Giacomo: Yes, correct. At Stevanato Group, we have been developing a cloud-based platform in collaboration and working together with Microsoft and using Microsoft Azure, which features deep learning models. This product is called SG Vision AI. It can be used to store and share images, label them, and deploy deep learning models. A very important point is that the platform ensures full data integrity and security, and in compliant with the GAMP [SP] and the CFR 21 Part 11, basically in alignment with the pharma regulations. This platform can be applied, this technology can be applied to all our range of vision inspection product we have in our portfolio, and also working towards retrofitting existing equipment.
Karen: Yes. Speaking about retrofitting, sometimes incorporating new technologies can be challenging for pharma companies as it usually requires them to adapt internal processes, invest time, money, resources. How does your new platform help ease those challenges?
Giacomo: Yes, this was a great point of focus from our team. As a result, we based our entire business model and brought it for allowing the pharma companies to choose among different assistance and customization levels based on their needs. So we have a team of data scientists and vision engineers that work to support the customer throughout their journey into deep learning technology. Further, the platform is a cloud-based empowered by, as I said, Microsoft Azure. Therefore, it can be easily integrated with a wide range of existing environments, and delivering phase, which is another point of focus that can be extremely time-consuming. Thanks to our platform, we designed it to optimize this process and dramatically reduce the time through an assistant labeling tool. The operator can label the images. We are quick and user-friendly interface, or let the software make a pre-evaluation and then approve the results.
Karen: That sounds great. How about in terms of future developments? How do you foresee AI applied to inspection?
Giacomo: So, the pharmaceutical industry is moving towards a more predictive method, exploiting the advantages in technology. AI can increase the productivity and reduce production cost, leading to a structure root cause analysis. Rather than just saying whether an item is good or bad, we can now be more precise and categorize defects in detail, creating database that pharma company can use to improve their processes. And we're working on this side. Another feature, the development area is creating a pre-trained narrow networks based on different layers which could then adapt to specific defects and customers' drugs. For instance, in the pandemic situation, using the database of existing vaccine, AI could support the rapid realization of an inspectional recipe for new treatments and vaccines, engaging inspection procedure while reducing the time to market. To generate these next generation models, we should train the system with thousand of images from different customers, where it's actually this allowed. At Stevanato Group, anyway, we are currently exploring this approach and we have to say we have a huge advantage. In fact, as part of the group, we produce in-house billions of pharmaceutical glass containers every year, and this leads us to a privileged access to a huge image database that we can exploit for cosmetic defects. And this allows us to creating basically high performance and robust models that can be applied again in this master model and then can be specifically trained, having high performance and reducing the time to market.
Karen: Well, it sounds like you're definitely very busy over there at Stevanato Group and you have some exciting things going on, and we look forward to hearing about them. So thanks so much for joining me today and sharing your insights on AI and vision inspection with our audience.
Giacomo: Thank you very much to you.
Karen: Okay. This is Karen Langhauser, and you've been listening to a special edition of Off Script: A Pharma Manufacturing podcast. Stay healthy and stay informed.