Developmental Stage Classification of Embryos Using Two-Stream Neural Network with Linear-Chain Conditional Random...

Area of Science:

• Developmental Biology
• Computational Biology
• Machine Learning

Background:

• Embryo development stages (cleavage, morula, blastocyst) follow a monotonic order.
• Existing embryo stage classification methods often rely solely on individual image frames, facing challenges like cell overlap and stage imbalance.
• Temporal information from time-lapse videos is underutilized but valuable for capturing developmental dynamics.

Purpose of the Study:

• To develop an improved method for classifying embryo developmental stages using time-lapse videos.
• To leverage both temporal and visual information for more accurate embryo stage classification.
• To incorporate the inherent monotonic order of embryonic development into the classification model.

Main Methods:

• Proposed a novel two-stream model that processes both temporal and image data from embryo time-lapse videos.
• Developed a linear-chain conditional random field (CRF) integrated with neural network features from both streams.
• Explicitly incorporated monotonic development order constraints within the CRF framework for tractable sequential modeling.

Main Results:

• Achieved high classification accuracy: 98.1% for mouse embryos and 80.6% for human embryos.
• Demonstrated the effectiveness of the two-stream model on both mouse and human embryo datasets.
• Showcased the advantage of combining temporal and image data over image-only approaches.

Conclusions:

• The proposed two-stream model effectively classifies embryo developmental stages by integrating temporal and image information.
• The inclusion of monotonic developmental constraints enhances the model's performance and interpretability.
• This approach offers a promising new direction for developmental stage classification, enabling deeper clinical and biological insights.