Automated microstructural characterization of hydrogels using deep instance segmentation and graph-based agglomerate...

Area of Science:

• Materials Science
• Computer Vision
• Pharmaceutical Sciences

Background:

• Pharmaceutical hydrogel performance relies on microstructural features like particle size and agglomeration.
• Manual microscopic assessment is subjective and challenging for indistinct boundaries or agglomerates.
• Automated analysis is needed to improve objectivity and efficiency in hydrogel characterization.

Purpose of the Study:

• To automate the analysis of hydrogel micrographs using computer vision.
• To apply and compare deep instance segmentation networks (Mask R-CNN, Mask2Former) for particle segmentation and agglomerate identification.
• To introduce novel methods for agglomerate identification and recommend appropriate performance metrics.

Main Methods:

• Applied Mask R-CNN and Mask2Former deep instance segmentation networks to hydrogel micrographs.
• Developed a particle dilation and graph-centric approach for agglomerate identification.
• Trained networks on a hydrogel database and evaluated performance using Average Precision (AP) and Average Recall (AR).

Main Results:

• Both networks achieved high Average Precision (Mask R-CNN: 92.47%, Mask2Former: 91.86%).
• Mask2Former demonstrated superior Average Recall (76.6%) compared to Mask R-CNN (72.32%).
• Extracted morphological features showed strong agreement with manual measurements.

Conclusions:

• Computer vision, particularly Mask2Former, offers an effective automated solution for hydrogel microstructural analysis.
• Average Recall is a recommended metric for model selection due to annotation subjectivity.
• This automated workflow can enhance pharmaceutical formulation development and quality control.