A Machine Learning Approach to Manufacturing Process Control
Open Access DepositedDownloadable Content
Reliable, objective monitoring of adherent cell cultures is essential for consistent, Good Manufacturing Practice (GMP) compliant production of stem‐cell therapies, yet today’s confluence, assessment for the area of the flask occupied by cells, and cell‐count assessments still rely on subjective visual inspection. In this research, the evaluation of whether modern image‐segmentation models can automate and improve the accuracy and repeatability of these critical measurements is performed. A diverse test dataset of bright‐field and fluorescence micrographs spanning multiple cell types and growth conditions is utilized to reflect real‐world manufacturing variability. Five variants of the Cellpose model, including trained model, and the latest released Segment Anything Model(SAM) were benchmarked on pixel‐level overlap (Dice coefficient), object‐level detection (Average Precision at 50 % overlap, AP50), and end‐to‐end process metrics such as mean absolute error (MAE) and root‐mean‐square error (RMSE) in confluence and Total Number of Cells(TVC) and Viable Cell Density(VCD) estimates. Without any additional training, the SAM achieved a Dice coefficient of 0.95 ± 0.03, confluence MAE/RMSE of 3.1 %/4.4 %, and average precision of 0.86, matching or exceeding our targets for both accuracy and precision. Fine-tuning produced marginal improvements in detection metrics but introduced occasional large counting errors, highlighting the importance of balancing overlap gains against enumeration stability. At the flask level, predicted versus ground-truth confluence, VCD showed excellent agreement (slopes near unity, R² ≥ 0.90), demonstrating that segmentation‐based pipelines can faithfully reproduce operator‐derived key performance indicators. Finally, an end-to-end Python toolkit for image aggregation, sparsity detection, and KPI reporting, establishing a foundation for more objective, high‐throughput monitoring of stem‐cell cultures is provided.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.