Deep learning-based image classification and quantification models for tablet sticking.

Area of Science:

• Pharmaceutical Manufacturing
• Drug Product Quality Control
• Computational Imaging

Background:

• Tablet sticking is a critical issue impacting drug quality, manufacturing efficiency, and therapeutic efficacy.
• Traditional methods like visual inspection and quality attribute testing may fail to detect subtle sticking problems.
• Developing automated, sensitive detection methods for tablet sticking is essential for robust pharmaceutical production.

Purpose of the Study:

• To develop and validate a novel integrated model for classifying and quantifying tablet sticking in pharmaceutical manufacturing.
• To evaluate the performance of different convolutional neural network (CNN) architectures for tablet sticking classification.
• To assess the capability of the integrated model in detecting mild sticking that might be missed by conventional quality control measures.

Main Methods:

• An integrated model combining convolutional neural network (CNN) architectures (AlexNet, VGG 16, ResNet 50, GoogLeNet) with gray-level co-occurrence matrix (GLCM) features and a support vector machine was developed.
• GoogLeNet demonstrated superior performance in classifying tablet sticking.
• GLCM features were analyzed to quantify sticking severity, and an integrated classification-quantification model was validated on a rotary tablet press.

Main Results:

• GoogLeNet achieved the highest classification accuracy (99.39%), precision (100.00%), recall (98.78%), and F1-score (99.38%).
• The quantification model successfully identified and measured sticking regions using a sticking index, showing significant differences between sticking and non-sticking areas.
• Validation demonstrated the model's ability to detect minimal and classified sticking levels with high repeatability, even when tablet quality attributes remained within acceptable limits.

Conclusions:

• The proposed integrated CNN-GLCM-SVM model effectively classifies and quantifies tablet sticking, surpassing the limitations of visual inspection and standard quality attribute testing.
• This advanced detection system can significantly improve pharmaceutical manufacturing efficiency and ensure consistent drug product quality by identifying mild sticking issues.
• The model offers a robust solution for enhancing quality control in tablet manufacturing, ensuring patient safety and drug efficacy.