Out of Distribution Generalization via Interventional Style Transfer in Single-Cell Microscopy.

Area of Science:

• Biomedical research
• Computer vision
• Machine learning

Background:

• Real-world deployment of computer vision necessitates causal representations invariant to context.
• Generalization to new data is crucial for biomedical discovery processes.

Purpose of the Study:

• To propose tests for assessing causal representation learning in models.
• To evaluate model generalization across increasing out-of-distribution (OOD) generalization levels.
• To introduce a novel method for improving OOD generalization.

Main Methods:

• Leveraging internal replicate structure of two single-cell fluorescent microscopy datasets.
• Developing generally applicable tests for causal representation assessment.
• Introducing Interventional Style Transfer (IST) to generate interventional training distributions.

Main Results:

• Naive and contemporary baseline models, despite strong performance on other metrics, failed the proposed causal representation tests.
• Interventional Style Transfer (IST) substantially improved OOD generalization.
• IST mitigated spurious correlations between biological causes and contextual nuisances.

Conclusions:

• Current methods for causal representation learning in computer vision are insufficient for robust biomedical applications.
• IST offers a promising approach to enhance the generalization capabilities of models in complex biological data.
• The study provides valuable insights and resources (code, datasets) for advancing causal representation learning.