Self-supervised learning of hologram reconstruction using physics consistency

Analysis

Why This Paper Matters

This paper addresses a critical bottleneck in deep learning for computational imaging: the reliance on large, labeled experimental datasets. By demonstrating that a self-supervised model can be trained entirely on synthetic random images and still generalize to real-world holograms, GedankenNet challenges the conventional wisdom that domain-specific training data is essential. This is particularly impactful for microscopy and holography, where acquiring diverse labeled data is labor-intensive and often impractical for rare or novel samples.

The use of a physics-consistency loss—enforcing the wave equation—is a clever way to embed domain knowledge into the learning process without requiring explicit labels. This approach not only eliminates the need for experimental data but also provides a principled way to handle variations in the imaging setup, such as changes in wavelength or pixel size. The paper's results suggest that self-supervised learning with physical constraints could become a standard paradigm for inverse problems in optics and beyond.

Technical Contributions

• Self-supervised training on synthetic data: The model is trained using artificially generated random images that do not resemble any real samples, completely avoiding the need for experimental data.
• Physics-consistency loss: A novel loss function that ensures the reconstructed complex field satisfies the wave equation in free space, providing a strong inductive bias.
• Generalization to unseen samples: After training, GedankenNet reconstructs phase and amplitude from experimental holograms of biological samples (e.g., cells, tissue) without any fine-tuning.
• Robustness to forward model perturbations: The model maintains performance under unknown changes in hologram distance, pixel size, and illumination wavelength, indicating resilience to experimental variability.

Results

The paper reports that GedankenNet achieves high-quality reconstructions on experimental holograms of various biological samples, including human lung tissue and blood cells. Quantitative metrics (e.g., SSIM, PSNR) are not explicitly stated in the abstract, but the qualitative results show clear phase and amplitude images consistent with wave propagation. The model outperforms traditional supervised methods when tested on out-of-distribution samples, highlighting its superior generalization. Additionally, the model's robustness to perturbations is demonstrated by reconstructing holograms with altered parameters without retraining.

Significance

GedankenNet represents a paradigm shift in computational imaging, showing that self-supervised learning with physical constraints can replace supervised learning for inverse problems. This could reduce the data acquisition burden in microscopy and enable rapid deployment of AI-based reconstruction in new settings. The approach may also inspire similar methods in other fields where physical laws govern the forward model, such as medical imaging, remote sensing, and acoustics. By decoupling training from experimental data, this work paves the way for more generalizable and practical AI solutions in science and engineering.