Cellular structure image classification with small targeted training samples.

Area of Science:

• Developmental Biology
• Computational Biology
• Machine Learning in Biology

Background:

• Cell shape and collective shape changes are critical indicators of morphogenetic events in complex tissues.
• Automated detection of these changes in 3D time-lapse images is essential but hindered by the difficulty of generating sufficient training data for deep learning.
• Limited annotated image datasets pose a significant challenge for machine learning model development in biological imaging.

Purpose of the Study:

• To develop a deep learning approach for accurate identification of multicellular rosettes in Caenorhabditis elegans embryos using minimal annotated training samples.
• To leverage generative adversarial networks (GANs) and feature transfer to overcome data limitations in biological image classification.
• To establish a public benchmark dataset for rosette detection to facilitate future research.

Main Methods:

• Utilized a generative adversarial network (GAN) framework for unsupervised training on 11,250 unlabeled 3D live images of C. elegans embryos with fluorescently labeled cell membranes.
• Employed feature transfer by incorporating the GAN discriminator structure into an Alex-style neural network for subsequent learning with a small set of labeled samples (dozens of images).
• Combined GAN-based unsupervised learning with supervised transfer learning to enhance image classification accuracy with minimal annotations.

Main Results:

• Achieved over 80% accuracy in identifying multicellular rosettes using only 10-15 labeled rosette images and 30-40 non-rosette images.
• The developed approach captured over 90% of the model accuracy compared to using a training dataset five times larger.
• Successfully created and released a public benchmark dataset for rosette detection, aiding future studies.

Conclusions:

• The proposed GAN-based transfer learning strategy effectively identifies cellular structures like multicellular rosettes with minimal training data.
• This method significantly reduces the time and effort required for analyzing large-scale 3D live imaging datasets in developmental biology.
• The approach is broadly applicable to the study of other cellular structures and morphogenetic events where annotated data is scarce.