Automatic segmentation of images, which is now feasible through an increase in available computing power, has become an important challenge in many fields. A key technology for obtaining such images is optical coherence tomography (OCT), which is already widely applied in ophthalmology and more recently in the pharmaceutical industry, as a method for real-time monitoring of solid oral dosage form coating processes. Accurately detecting the boundaries of objects in OCT images is required for a meaningful automatic evaluation. During in-line monitoring, the evaluation time for each image is a crucial factor to enable the real-time analysis of large amounts of data. The segmentation of images has previously been achieved via machine learning methods, which generally require a large number of training examples. This work aims to overcome this limitation by employing unsupervised machine learning for the segmentation of OCT images of coated pharmaceutical tablets. An adapted clustering method was specifically developed to achieve the fast real-time detection of the coating layer’s boundaries in OCT-generated images. A newly developed parallel implementation of DBSCAN, that is well suited for image evaluation, makes it possible to use this novel method for real-time process analytical technology (PAT) applications. This approach has been shown to be significantly faster than so far established methods for segmenting similar OCT images. Furthermore, the image-specific parallelized DBSCAN algorithm has been shown to be around three times faster than other parallel implementations.

图像的自动分割，现在可以通过增加可用的计算能力来实现，已经成为许多领域的重要挑战。获得此类图像的一项关键技术是光学相干断层扫描 (OCT)，它已广泛应用于眼科以及最近在制药行业中，作为一种实时监测固体口服剂型包衣过程的方法。有意义的自动评估需要准确检测 OCT 图像中对象的边界。在在线监测过程中，每个图像的评估时间是实现大量数据实时分析的关键因素。图像的分割以前是通过机器学习方法实现的，这通常需要大量的训练样本。这项工作旨在通过使用无监督机器学习来分割包衣药片的 OCT 图像来克服这一限制。专门开发了一种适用的聚类方法，以实现对 OCT 生成图像中涂层边界的快速实时检测。新开发的 DBSCAN 并行实现非常适合图像评估，使得将这种新方法用于实时过程分析技术 (PAT) 应用成为可能。这种方法已被证明比迄今为止建立的分割相似 OCT 图像的方法要快得多。此外，图像特定的并行化 DBSCAN 算法已被证明比其他并行实现快三倍左右。