Supervised contrastive learning for cell stage classification of animal embryos

Area of Science:

• Embryology
• Computer Vision
• Machine Learning

Background:

• Video microscopy combined with machine learning shows promise for studying early embryo development.
• Manual annotation of embryonic developmental events, particularly cell divisions, is time-consuming and limits scalability for practical applications.

Purpose of the Study:

• To develop a deep learning approach for automatic cell stage classification of bovine embryos from 2D time-lapse microscopy videos.
• To address challenges including low-quality images, ambiguous stage boundaries, and imbalanced data distribution in bovine embryonic development analysis.

Main Methods:

• A novel method, CLEmbryo, was introduced, utilizing supervised contrastive learning and focal loss for training.
• A lightweight 3D neural network, CSN-50, was employed as an encoder.
• A new dataset, Bovine Embryos Cell Stages (ECS), was created for bovine embryonic development analysis.

Main Results:

• CLEmbryo demonstrated strong performance in classifying embryo cell stages.
• The method showed good generalization capabilities across different datasets.
• CLEmbryo outperformed existing state-of-the-art methods on both the Bovine ECS dataset and the NYU Mouse Embryos dataset.

Conclusions:

• The developed deep learning method, CLEmbryo, offers an effective solution for automated cell stage classification in bovine embryos.
• This approach has significant potential for applications in cattle breeding and developmental biology research.
• The method's superior performance highlights the advancement in analyzing complex biological imaging data.