Deep learning-based image classification and quantification models for tablet sticking
Abstract
Sticking can significantly affect drug product quality, manufacturing efficiency, and therapeutic efficacy in pharmaceutical tablet manufacturing. This study presents a novel integrated model with a convolutional neural network (CNN) and gray-level co-occurrence matrix (GLCM) based features combined with a support vector machine to classify and quantify tablet sticking. The classification model was developed and evaluated using CNN architectures, including AlexNet, VGG 16, ResNet 50, and GoogLeNet. GoogLeNet showed the best performance in terms of accuracy (99.39%), precision (100.00%), recall (98.78%), F1-score (99.38%), and computational efficiency. GLCM features such as energy, homogeneity, contrast, and correlation were analyzed to develop an optimal quantification model, revealing a significant difference between the sticking and non-sticking regions. Based on these differences, the sticking regions were detected and quantified using a sticking index. To validate the final model, which integrated the classification and quantification models, 10 batches of tablets were produced using a rotary tablet press. The validation confirmed high measurement repeatability with minimal and classified sticking levels. Tablet quality attributes such as assay, content uniformity, and weight were evaluated. Despite the occurrence of sticking, the tablet quality attributes met their criteria. These results suggest that measuring tablet quality attributes and visual inspection may not be sufficient to detect mild sticking. However, the integrated model proposed in this study could detect mild sticking, even if the tablet quality attributes remained within the acceptable criteria. This study demonstrated that the proposed integrated model could improve pharmaceutical manufacturing efficiency, ensure consistent drug product quality, and overcome visual inspection limitations.
Introduction
Generally, sticking may occur when tablet components adhere to the punch surfaces or die walls during compression. The factors influencing sticking include the adhesion of the active pharmaceutical ingredient (API), its interaction with excipients, and the surface characteristics of the tooling material (i.e., punch and die) (Gunawardana et al., 2023, Patel et al., 2006). Moreover, inadequate lubrication, high moisture content, and inappropriate particle size distribution accelerate sticking, making the manufacturing process more complex. Sticking is a significant defect that affects drug product quality and manufacturing efficiency, leading to patient safety risks, production delays, and additional costs (Gunawardana et al., 2023, Paul et al., 2017). Therefore, it is necessary to prevent and manage sticking to obtain high-quality products and improve manufacturing efficiency. Although sticking is a significant phenomenon in the pharmaceutical industry, there are no standardized regulatory methods for precisely observing and quantitatively assessing it. One of the most popular methods to detect sticking is visual inspection; however, this increases the variability of results owing to its dependence on the inspectors’ subjective judgment. In addition, visual inspection has limitations in detecting mild sticking at an early stage or quantifying its severity. Advanced techniques such as atomic force microscopy analysis (Wang et al., 2003), fluid dynamics analysis (Christodoulou et al., 2020), and force–displacement measurements (Bunker et al., 2011) have been applied to overcome these limitations. However, these methods are primarily applied at the laboratory scale and are unsuitable for real-time monitoring in pharmaceutical manufacturing environments.
Recently, machine learning, particularly deep learning-based analysis, has been widely used as a powerful tool for visual defect detection across various industries (Hütten et al., 2024, Lin et al., 2024, Saberironaghi et al., 2023). Deep learning models can process large volumes of image data to automatically extract and analyze specific patterns, enabling more precise defect detection than conventional feature-based analyses (Bengio et al., 2013, Jha et al., 2018, Krizhevsky et al., 2012). In particular, convolutional neural network (CNN) can process high-dimensional image data effectively, automatically extract features, adapt to various types of defects, and have recently been widely applied for quality control in the pharmaceutical industry. For example, Ma et al. analyzed X-ray computed tomography images of tablets using a CNN model and confirmed that it could achieve 94 % accuracy in detecting internal cracks (Ma et al., 2020). In addition, this model significantly reduced processing time compared with conventional manual analysis methods. Ficzere et al. demonstrated that defects in film-coated tablets can be detected and classified in real-time using YOLOv5, a state-of-the-art CNN algorithm (Ficzere et al., 2022). The results showed that the accuracy of this method reached 98.2 %, and the processing speed was scalable to match the high productivity of continuous film coaters.
Previous studies have employed CNN to detect external defects in tablets (e.g., cracks and discoloration). However, research on the quantitative analysis of surface defects like sticking remains limited. Unlike cracks and discoloration, which exhibit clear boundaries in images, sticking appears as subtle texture changes, making it difficult to detect them effectively using CNN-based analysis. These limitations can be addressed by incorporating gray-level co-occurrence matrix (GLCM) and a support vector machine (SVM). GLCM quantitatively characterizes texture patterns in an image by analyzing the co-occurrence relationship of gray levels between pixels. This method is also useful for analyzing subtle textural changes on a surface where sticking occurs (Vijayakumar et al., 2024, Zhao et al., 2023). By utilizing GLCM features, such as contrast, homogeneity, entropy, and correlation, the features of the defect region can be evaluated with greater precision. However, because GLCM analyzes only textural features, it is difficult to classify the detection defects. This limitation can be overcome by effectively classifying the features extracted from the GLCM using SVM. SVM is a powerful machine-learning technique that classifies data into high-dimensional features and can effectively classify nonlinear data by utilizing the kernel trick. The combination of GLCM and SVM provides high accuracy, interpretability, and computational efficiency. In addition, it can be practically applied in pharmaceutical manufacturing environments (Dong et al., 2022, Maurya et al., 2014, Zhao et al., 2023).
This study aims to develop a novel integrated model that combines a CNN-based classification model with a GLCM-SVM-based quantification model for detecting tablet sticking. The model optimized for sticking detection was selected and classified according to the presence or absence of sticking with high accuracy by comparing various CNN architectures. In addition, a quantification model was developed to evaluate the degree of sticking using GLCM and SVM-based texture analyses. The integrated model was verified in an actual manufacturing environment, and sticking was detected and quantified in several batches. In addtion, tablet quality attributes were evaluated to establish the relationship between sticking and tablet quality attributes.
Read more here
Materials
Amlodipine besylate was supplied from Daewon Pharmaceutical Co., Ltd. (Seoul, Korea). Microcrystalline cellulose (MCC) was purchased from FMC BioPolymer (Avicel® PH-102, Philadelphia, PA, USA). Calcium hydrogen phosphate, sodium starch glycolate, colloidal silicon dioxide, and magnesium stearate (St-Mg) were purchased from Sigma-Aldrich Co. (St. Louis, MO, USA).
Ji Yeon Kim, Du Hyung Choi, Deep learning-based image classification and quantification models for tablet sticking, International Journal of Pharmaceutics, 2025, 125690, ISSN 0378-5173, https://doi.org/10.1016/j.ijpharm.2025.125690.
Read also our introduction article on Microcrystalline Cellulose here:
