(372f) Detecting and Evaluating Irregular Objects in OCT Images with Unsupervised Machine Learning

Recently, optical coherence tomography (OCT) has gained significant attention in the pharmaceutical industry due to its capability for real-time monitoring of coating processes, such as film coating of tablets and pellets. Accurately detecting the boundaries of objects in OCT is crucial to a meaningful evaluation of the image. Thus far, the focus has been on evaluating the mean coating layer thickness, using algorithms which presume a spherical, smooth shape of the product or - in the case of Convolutional Neural Networks (CNNs) - require large amounts of hand-annotated training data specific for each product. These methods reach their limits, if the shapes depicted on the OCT images are not spherical or cannot be trained easily, as is the case of irregularly shaped tables or pellets.

Significant work has been undertaken to address these challenges and a method was developed, which is able to evaluate a variety of product shapes by adjusting parameters accordingly, hence avoiding lengthy manual annotation of many OCT images. This new algorithm makes use of unsupervised machine learning and applies it to OCT image segmentation. Specifically, by working with an iterative spatial clustering approach, we are capable of capturing details, such as speckles and artifacts, while discarding noise. This technological approach enables us to accurately evaluate the surface roughness of non-smooth, non-spherical products. Furthermore, by distinguishing tablet features from image noise we facilitate an accurate representation of a coating layer or a product’s homogeneity. Ongoing research aims to utilise the results of these new algorithms by investigating a broader range of pharmaceutical products, beyond tablets and pellets.