Paper Overview
- Field: Machine Learning
- Authors: Ana Sanchez-Fernandez, Thomas Pinetz, Werner Zellinger
- Published: 2026-04-22
- arXiv: 2604.20824
- Evaluated on Mechanism-of-Action (MoA) classification using the large-scale JUMP-CP dataset.
- A standard ResNet's accuracy drops from 0.939±0.005 on the training domain to 0.862±0.060 on new experimental batches.
- Foundation models, even with typical variation normalization, cannot close this gap.
- CS-ARM-BN is the first method shown to close the domain gap, reaching 0.935±0.018.
- When new batches exhibit strong domain shift (e.g., data generated in a different laboratory), the meta-learning method can be stabilized using control samples, which are always available in biomedical experiments.
Summary
The central problem in biomedical imaging is batch effects: systematic technical variations unrelated to the biological signal of interest. These effects critically undermine experimental reproducibility and are the primary cause of failure of deep learning systems on new experimental batches, preventing their practical use in the real world. Despite years of research, no method has succeeded in closing this performance gap for deep learning models.
The authors propose CS-ARM-BN (Control-Stabilized Adaptive Risk Minimization via Batch Normalization), a meta-learning adaptation method that exploits negative control samples. Such unperturbed reference images are present in every experimental batch by design and serve as stable context for adaptation.
Key Results
Conclusion
Batch effects in biomedical imaging data can be effectively neutralized through principled in-context adaptation, making the approach both practical and efficient for real-world deployment.
--- *Auto-collected on 2026-04-24.*