Self-supervised generation of realistic training data enables nanoscale localization in challenging conditions

Self-supervised, physics-informed generative model produces highly realistic training data, boosting nanoscale localization accuracy in challenging microscopy conditions.

• Why it matters: Accurate training data is crucial for deep learning-based localization microscopy, but creating simulations that match complex experimental environments remains difficult, limiting performance.
• What they did: The authors developed a self-supervised, physics-informed generative model based on the Deep Latent Particles framework, incorporating a physical Point Spread Function model trained on unlabeled experimental images.
• The result: This approach generates realistic, fully labeled training datasets that significantly enhance localization precision and emitter detection, especially in low signal-to-noise and complex background scenarios.