A deep learning approach for staging embryonic tissue isolates with small data.

Area of Science:

• Developmental Biology
• Computational Biology
• Genomics

Background:

• Machine learning (ML) models typically require large datasets for training.
• Data scarcity in fields like developmental biology limits ML applications.
• Recent advances suggest ML can be effective with smaller, information-rich datasets.

Purpose of the Study:

• To train a convolutional neural network (CNN) classifier for staging zebrafish tail buds at four developmental stages.
• To demonstrate the efficacy of ML in data-scarce research areas.
• To validate ML staging using small, curated datasets.

Main Methods:

• Developed a CNN-based classifier for staging zebrafish tail buds.
• Utilized both morphological and gene expression confocal microscopy images.
• Trained and tested the classifier on small, information-rich datasets (under 100 images).

Main Results:

• Achieved up to 100% test accuracy in staging zebrafish tail buds.
• Demonstrated high accuracy is possible with significantly smaller training datasets than typically required for CNNs.
• Showed the classifier can stage isolated embryonic structures without whole-embryo landmarks.

Conclusions:

• CNNs can effectively stage zebrafish development using limited data, decoupling ML from Big Data requirements.
• This approach is valuable for data-scarce fields and for staging 3D in vitro systems like organoids.
• Encourages the application of ML in research areas previously limited by data availability.