Four years ago we faced a serious problem. Our team of computer vision scientists was developing a medication adherence monitoring system that could identify tablets and capsules using only an image taken with an iPhone. We quickly learned that due to manufacturing variation there are often several visual “variants” of each medication in circulation – substantially increasing the complexity of this task. However, the recognition that this visual variation was systematic and could be statistically captured through a straightforward sampling process would lead to the development of MedSnap’s highly scalable system for precision product authentication.
Systematic variation can be modeled and used for authentication
Early on we recognized that if you’ve seen one pill, you’ve seen one pill. At the lot or batch level, individual instances can have slight variations in texture or imprint. However, using images of 100 pills from a lot we create composite statistical models which generalize across the lot, and in most cases across that entire production line. Variation between lots from the same line, albeit rare, are accommodated through more frequent lot sampling.
Using the iPhone camera our effective resolution at the surface of the tablet or capsule is 415 pixels per square millimeter. With our Snap Surface background we can measure size and shape to within 0.1mm and differentiate over 250,000 shades of color. Therefore our ability to measure size, color, imprint and texture are well within the tolerances of the production equipment making the product.
The net result is we capture the small unique visual characteristics of a product that make it distinctively authentic. Often we can differentiate authentic products made at different facilities – and attribute them to their site of origin when necessary.
Sampling visual output as a part of lot QA
Dissolution and other tests are performed at the end of the manufacturing process. Visual assessment using our VR testing app can be easily added to this workflow – it takes less than a minute. Any failures represent a statistically significant visual change in the product. QA staff can then decide whether to add this new variant to the set of statistical models which define authentic for that medication, or reject the batch. In this way the definition of authentic is described by a set of models which can be continuously updated and expanded across an entire network of production facilities, keeping pace with changes in tooling or excipients.
Counterfeiters are surprisingly precise but not accurate
Like any manufacturer, in order to make a margin counterfeiters must scale their operations to produce product in sufficient quantities to meet demand and with visual quality high enough to fool a non-expert on inspection. We also sample counterfeit products and compare their visual relatedness. Surprisingly, some counterfeit products show visual consistency across and within samples rivaling that of authentic products. Visual clustering therefore can provide a “fingerprint” of these operations, quantify the product they have produced, and provide important intelligence for investigators.
It can be intimidating to recognize the industrial production capacity that many counterfeiters possess. However, given that two authentic production facilities cannot easily make pills that look exactly the same due to slight differences in equipment, tooling, and excipients, it would be extremely difficult for someone outside your company using different materials and equipment to produce a perfect visual counterfeit. To date we have not found an example.
Setbacks can help you take a big leap forward
The systematic variation in the production process for solid oral medications was at first a big problem for our team, but tackling it led to insights which have allowed for the creation of tools that utilize this variation to authenticate the physical product. With this unique, novel approach there is no need to modify the product to capture its inherent authenticity.