A new AI-based approach for computational imaging and microscopy that does not require any experimental objects or real data has been developed by researchers from the UCLA Samueli School of Engineering.
The approach, called GedankenNet, is a self-supervised AI model that learns from the principles of physics and mental experiments. The model uses only the universal laws of physics that describe how electromagnetic waves travel in space to reconstruct microscopic images from random artificial holograms — generated entirely from ‘imagination’ without depending on any real-world experiments, actual sample similarities or real data.
GedankenNet, coming from the German word for thought, is new self- supervised deep learning model for hologram reconstruction. Hologram reconstruction in machine learning refers to the use of machine learning techniques to reconstruct images from raw holograms. What makes GedankenNet special is that it does not need labeled or experimental training data.
In a Nutshell
– GedankenNet eliminates the need for large labeled training datasets and generalizes well to reconstructing experimental holograms of various tissue samples, even though it never saw real samples during training.
– The key innovation is training GedankenNet to match the input holograms by predicting holograms from its outputs using physics-based forward models, not relying on ground truth sample images.
– This physics-consistency loss encodes the wave propagation physics into the network, making its outputs compatible with Maxwell’s equations.
– GedankenNet showed superior generalization compared to supervised learning models trained on the same synthetic datasets. It also outperformed iterative phase retrieval algorithms.
– GedankenNet successfully reconstructed both the amplitude and phase images of various tissue types like lung, prostate, kidney from experimental holograms, despite training only on random synthetic data.
– The network can digitally autofocus defocused holograms and showed resilience to unknown shifts in parameters like pixel size and wavelength.
– This self-supervised deep learning approach eliminates needs for large labeled training datasets and has broad applicability for computational imaging and microscopy.
