Prospective Assessment of Embryoid Body by Deep Learning on Label-Free Time-Lapse Images from the Microwell Array.

Area of Science:

• Organoid Engineering
• Biotechnology
• Artificial Intelligence in Medicine

Background:

• Embryoid bodies (EBs) are crucial for organoid engineering, but their formation and size impact differentiation outcomes.
• Current methods rely on retrospective quality assessment, hindering reproducibility in high-throughput systems.

Purpose of the Study:

• Develop a prospective, non-invasive framework using bright-field time-lapse imaging and 3D convolutional neural networks (3D-CNNs).
• Predict EB formation success and final diameter within microwell platforms for improved quality control.

Main Methods:

• Trained 3D-CNN models on early-phase time-lapse image sequences for classification and regression tasks.
• Utilized under-sampling for dataset balancing and five-fold cross-validation with data augmentation for performance evaluation.

Main Results:

• The classification model achieved 96.5% accuracy in predicting EB formation using short image sequences.
• The regression model accurately predicted final EB diameter with a mean absolute error of ±7.1 µm, capturing size variations.

Conclusions:

• Early aggregation dynamics from bright-field imaging provide sufficient data for accurate, prospective EB quality prediction.
• The label-free, automation-compatible framework boosts reproducibility in large-scale EB manufacturing.
• Supports development of adaptive, closed-loop organoid culture systems for clinical use.