Real-time component-based particle size measurement and dissolution prediction during continuous powder feeding using machine vision and artificial intelligence-based object detection

This work presents a system, in which machine vision combined with artificial intelligence-based image analysis was used to determine the component-based particle size distribution of pharmaceutical powder blends. The blends consisted of acetylsalicylic acid (ASA) and calcium hydrogen phosphate (CHP). Images of powders were recorded with a digital camera in-line during feeding from a continuous feeder. The component-based particle size distributions determined with the system correlated well with those measured using a microscope as a reference method. This novel method proved to be effective in the real-time determination of particle size distribution of different components in the same blend. It was also possible to predict the in vitro dissolution profile of capsules filled with this blend by using the measured particle size distribution of ASA as input in a population balance model. The method could provide valuable information on the blends used in the pharmaceutical industry and could play a key role in the development of pharmaceutical quality control.

Introduction

Powder blends are integral to many areas of the pharmaceutical industry. Tablets represent up to 70% of all drugs on the market. Powder blends are the main intermediate products of tablets. Therefore, knowing and controlling their properties is of utmost importance. It’s particularly important considering that the malfunctions of the equipment used in the manufacturing of tablets are mainly tied to powder behaviour (Brubaker and Moghtadernejad, 2024). The pharmaceutical excipients can vary in size from the nanoscale colloidal range to granules of millimetre size in various pharmaceutical formulations. The two most important parameters of pharmaceutical powder particles are their size and shape, since these can have a huge effect on the quality attributes of the powders and a considerable impact on the end product. Particle size can affect content uniformity since the particles of different size and shape can segregate and cause inhomogeneity. Its effect on flow properties should not be overlooked (Goh et al., 2018, Fu et al., 2012). Particle size can also strongly influence dissolution and absorption rates (Shekunov et al., 2007), especially in the case of BCS II compounds, while for BCS I/III compounds particle size distribution is not the limiting factor, making the development of the dissolution method simpler. Given this great impact, particle size is in many pharmaceutical processes considered a Critical Quality Attribute (CQA) (Madarász et al., 2023) typically measured using the distribution of parameters such as equivalent diameter or Feret diameter.

Determining and controlling the component-based particle size distributions is crucial in understanding the segregation of powder blends, where the sticking behaviour is dependent on material properties, such as surface area and crystal morphology (Abdel-Hamid et al., 2011). Although there are as many as 13 mechanisms related to the segregation of ingredients identified in engineering areas (de Silva et al., 1999) only a few are relevant in handling pharmaceutical solids. Amongst those the most relevant ones are sifting, fluidization and rolling. In these three phenomena the varying behaviour of the particles is attributed to the difference in their sizes, rather than the differences in the materials themselves (Jakubowska and Ciepluch, 2021). As a result, particles of the same component but of varying sizes and shapes may exhibit different behaviours, which can lead to further segregation and inhomogeneity. In their research, Péterfi et al. have investigated the surface powder sticking of blends containing amlodipine as API, highlighting that the concentration of amlodipine increased in the material adhered to the inner wall, showing that in certain cases different materials may show diverse behaviour when contacting various surfaces (Péterfi et al., 2024).

Traditionally, particle size has been determined using off-line methods such as sieve analysis, laser diffraction (LD), and microscopy. These methods often require skilled operators and can be time consuming, therefore they are typically not suitable for real-time process control (Madarász et al., 2023). Recently, as Process Analytical Technology (PAT) and continuous manufacturing (CM) are becoming more widespread in the industry, there has been a significant shift towards in-line methods in particle size analysis. These methods include in-line LD (Ma et al., 2001), spatial filtering velocimetry (SFV) (Wiegel et al., 2016) and Focused Beam Reflectance Measurement (FBRM) (Muhaimin et al., 2021). LD is still mostly used as an off-line method. While LD can measure the size of relatively small, even nanoscale particles, the primary limitations of the technology are its destructivity and the fact that the particles are always assumed to be spherical. For non-spherical particles, this leads to a distribution of equivalent light scatter diameters, which is dependent on particle orientation (Merkus and Merkus, 2009). While SFV can easily be installed into manufacturing equipment, it requires special attention and maintenance and in many cases the use of pressurised air in order to ensure representative sampling. In FBRM a major challenge in interpreting its results is that the chord length distributions measured by the method may not correlate directly with particle size (Madarász et al., 2023). Neither of the aforementioned methods are capable of differentiating between particles of varying materials, thus they are not able to determine the component-based particle size distribution of a powder blend. This leads to limitations in the application of such methods in CM.

Machine vision is a simple and effective tool for the in-line analysis of pharmaceutical processes. A machine vision system recovers useful information about a scene from its two-dimensional projections (Jain et al., 1995). In past studies machine vision systems have been used to classify the quality of coated tablets (Hirschberg et al., 2020), to predict bulk powder flowability of pharmaceutical materials based on their particle shape and size distributions (Yu et al., 2011) and to characterize the particle size distribution of granular products (Soprana et al., 2018). A machine vision system comprises of several components, including the camera and light source, each serving the purpose of collecting reproducible data (Galata et al., 2021). For the light source, the most important factor is the consistency in the quality of the emitted light, which is primarily required due to the strict pharmaceutical and GMP standards. In their publication, Martin et al. provided an in-depth analysis of selecting an appropriate light source and on the importance of optimal setup in machine vision systems (Martin, 2007). Selecting the appropriate camera is crucial for capturing images of the required quality, which can be utilized not only for object detection, but for determining shapes and leveraging properties (colour, transparency etc.) to distinguish between different components (Hannan et al., 2009). A limitation of image analysis and machine vision for monitoring particle parameters during production is the low efficiency of data evaluation with classical algorithms, such as thresholding (Seelaboyina and Vishwakarma, 2023).

The combination of artificial intelligence and neural networks applied to the processing of information obtained through machine vision has the potential to yield significant advancements. For the automation and efficiency improvement of image analysis YOLOv5, an object detecting algorithm combined with a convolutional neural network (CNN), may prove useful. It uses a convolutional neural network to detect different objects, thereby enhancing its performance and efficiency in real-time detection. Its ability to learn many appearances of an object makes it better at managing real-world situations where the appearance of objects changes dynamically (Vijayakumar and Vairavasundaram, 2024, Lu et al., 2024).

Compared to the object detection algorithms used in the past, it also has the advantage of using instance segmentation instead of bounding boxes. Instance segmentation is capable of accurately following the contours of detected objects, whereas the use of bounding boxes is most effective when the primary parameter of interest is the objects’ location or quality. With the use of instance segmentation, a significant improvement could be achieved in determining object size and shape, and it has already been applied in various fields, including agriculture to enhance fruit production (Lawal, 2023). As for its use in the pharmaceutical industry a recent study by Fazekas et al. presented a quality assurance system, where CQAs were monitored with fast at-line techniques. The developed technique measured the diameter of electrospun fibrous samples using camera images and instance segmentation performed by a trained AI model (Fazekas et al., 2024).

Another valuable reference is the study by Ficzere et al., who introduced a system, where object detection was used on images captured with a digital camera to simultaneously determine the API concentration and the particle size distribution of two components of a powder blend (Ficzere et al., 2023). They utilized a setup where the powder blend was transferred onto a conveyor belt providing a surface to capture images with a digital camera and then the images were analysed using the method of bounding boxes.

The bioavailability of orally administered drugs depends on their absorption from the gastrointestinal tract (GI). The Biopharmaceutical Classification System (BCS) is a system used to classify drugs based on their solubility and permeability. It helps predict the rate and extent of drug absorption in the body. For drugs belonging to BCS II (characterized by high permeability and low solubility), the dissolution rate in GI fluids serves as the rate limiting step in the absorption of these drugs rather than their diffusion through the GI membrane (Dizaj et al., 2015). It is therefore an integral part of preclinical and clinical development.

Dissolution testing constitutes one of the most important analytical tools in the pharmaceutical industrial quality control laboratories. This analytical method typically involves numerous steps (Shekunov and Montgomery, 2016). For the dissolution testing, the tablets are placed in the dissolution medium and samples are collected at predetermined time intervals for analysis of the concentration of dissolved API. This method is relatively slow, destructive, and unsuitable for examining a large number of tablets or for in-line application. Significant improvements could be achieved by integrating an in-line sensor with a mathematical model. It is therefore advantageous to develop reliable dissolution models that are capable of predicting the dissolution rate of the API based on its properties such as particle size distribution. Population balance equation could be a useful tool to predict the dissolution rate based on particle size distribution for products, if the particle size distribution has been established as a critical quality attribute for the product.

The population balance equation describes the nucleation, growth, aggregation/ agglomeration and breakage. Population balance equations can be formulated using an Eulerian or a Lagrangian approach, where the Lagrangian viewpoint tracks a finite number of particles in a flow field, while the Eulerian viewpoint tracks the particles as a bulk continuous phase (Omar and Rohani, 2017). These equations can be applied in the determination of dissolution rate and could be combined with an in-line particle size distribution measurement technique to enable real-time prediction of dissolution rates.

Given the significant results achieved with the technology introduced by Ficzere et al. (Ficzere et al., 2023) where the powder blends were transferred onto a conveyor belt and images were analysed using bounding boxes, a setup in which the blend is examined directly at the feeder output, making it possible for an ideal in-line application, along with the use of instance segmentation, could lead to notable advancements in determining particle shape and size distribution. The determined distribution could be combined with population balance equations to achieve real-time dissolution rate prediction.

Download the research paper as PDF(pre-proof): Real-time component-based particle size measurement and dissolution prediction during continuous powder feeding

or continue reading on Real-time component-based particle size measurement

Materials

The model API used in the study, acetylsalicylic acid (ASA) was acquired from Sigma-Aldrich (St. Louis, Missouri, USA). Anhydrous calcium hydrogen phosphate (CHP) was obtained from JRS Pharma (Rosenberg, Germany). Microscopic images of the used materials are presented in Fig. 1. The presented images illustrate the differences between the two components in terms of shape (CHP being more circular), size (CHP having a smaller diameter) and transparency (ASA being more transparent).

Áron Kálnai, Máté Ficzere, Brigitta Nagy, Orsolya Péterfi, Máté Benczúr, Zsombor Kristóf Nagy, Dorián László Galata,
Real-time component-based particle size measurement and dissolution prediction during continuous powder feeding using machine vision and artificial intelligence-based object detection,
European Journal of Pharmaceutical Sciences, 2025, 107080, ISSN 0928-0987,
https://doi.org/10.1016/j.ejps.2025.107080.