AI-based outlier detection in optical coherence tomography (OCT) measurements for pharmaceutical applications

In the world of making medicine, ensuring that every pill is perfect is a matter of both safety and economics. One of the most powerful tools scientists use to check the quality of these pills is a technique called optical coherence tomography. Imagine shining a special kind of light onto a tablet as it tumbles inside a machine, allowing the light to bounce back and create a detailed cross-sectional picture of the pill's surface and the thin coating wrapped around it. This process happens incredibly fast, capturing thousands of these cross-sectional images every hour as the pills move through the manufacturing line. The goal is to measure the thickness of the coating to ensure it is even and consistent, which is vital for how the medicine releases into the body. However, because the pills move randomly and quickly, the machine sometimes takes pictures that are confusing or misleading. These bad pictures can look like real problems with the medicine, or they can hide real problems, creating a massive pile of data that human experts must sift through by hand to find the truth.

A team of researchers at the Research Center Pharmaceutical Engineering in Austria set out to solve this bottleneck. They focused on the early stages of developing new medicines, a time when the machines and the pill formulas are still being tuned and are prone to generating these confusing images. In these early phases, the equipment might misinterpret a speck of dust, a reflection from the machine wall, or a clump of ingredients as a flaw in the coating. A human expert can spot these errors easily if they look at one image, but when a single experiment produces thousands of images, checking them one by one becomes a tedious task that can take days. The researchers wanted to know if they could teach a computer to do this sorting work automatically, separating the genuine problems from the false alarms so that scientists could focus only on the images that truly mattered.

To do this, the team built a digital brain using a type of artificial intelligence known as a convolutional neural network. They started with a model that had already learned to recognize thousands of everyday objects, like cats and cars, and then they taught it specifically to look at the cross-sectional pictures of the pills. They showed the computer thousands of examples of good pictures and bad pictures, letting it learn the subtle visual differences between a real coating and a mistake. The bad pictures they encountered fell into a few specific categories. Sometimes, a pill would pass too close to the sensor, creating a mirror image that looked like a ghost reflection. Other times, certain ingredients in the pill that scatter light strongly would trick the machine into thinking it saw a coating layer where there was none. In other instances, coating material that had stuck to the inside of the mixing drum would be mistaken for a pill coating. The computer learned to recognize these patterns just as a human expert would, but with the speed of a machine.

When the researchers tested their new tool, the results were striking. They applied the system to massive datasets containing thousands of images from real coating trials. In one experiment involving a drum coater, the system analyzed over six thousand data points. Before the computer helped, nearly seven hundred of these points looked like they were outside the normal range, suggesting the coating was uneven. After the AI sorted through them, it identified that more than half of those suspicious points were actually just measurement errors. By removing these false alarms, the data became much cleaner, and the remaining outliers represented real issues with the process rather than glitches in the camera. In another trial with a fluidized-bed coater, the system reduced the number of suspicious points by more than half as well. The average variation in the coating thickness measurements dropped significantly, giving the scientists a much clearer picture of how the process was actually performing.

The accuracy of the system was also impressive. When tested on a set of images it had never seen before, the computer correctly identified whether an image was a good measurement or a mistake in nearly ninety-six percent of cases. To ensure the computer was not just guessing or looking at the wrong parts of the image, the researchers used a technique that highlights exactly which areas of the picture the computer focused on to make its decision. These visual maps showed that the computer was indeed looking at the pill and the coating, ignoring the background noise and artifacts. This transparency is crucial, as it allows human experts to trust the system and verify its decisions when necessary.

The true value of this work lies in how it changes the workflow for pharmaceutical developers. Instead of spending hours or days manually reviewing thousands of images, a scientist can now let the computer do the heavy lifting in a matter of minutes. The system flags only the questionable images for human review, while discarding the obvious errors and keeping the clear, useful data. This does not mean the computer replaces the expert; rather, it acts as a highly efficient filter that removes the noise, allowing the human to focus on the genuine problems that need solving. As a manufacturing process becomes more stable and moves from development to routine production, these measurement errors naturally become rarer, and the need for such a filter diminishes. However, during the critical early stages of creating new medicines, this tool helps scientists make faster, more confident decisions, speeding up the journey from the laboratory to the pharmacy shelf.