The common approach to solve CGH problems relies on iterative algorithms. The simplest and most popular approach is an iterative search with the Gerchberg-Saxton algorithm [1] (Check out our introduction to holography lecture). When better computational capabilities became available, new algorithms such as NOVO-CGH [2] have been developed that can improve hologram quality with direct optimization methods. However, in all these approaches, the accuracy of holograms improves with the number of steps but requires longer computation time.
The algorithm we developed, DeepCGH [3], pushes the envelope with both greater speed and accuracy. To accelerate computation, we rely on a trained convolutional neural network (CNN) that yields holograms without iterations. The computation time is fixed and only depends on hologram dimensions and CNN model size. Gains in speed and accuracy are also made possible with a task-driven CNN model design. First, our model includes an interleaving step that reorganizes the input data to reduce the CNN size. This maximizes speed and allows for the computation of large holograms (~10M Voxels) in milliseconds.
To maximize accuracy, the CNN estimates the complex field in real space, and the output undergoes a Fourier transform to yield phase at the SLM plane. This ensures that the input and output of the CNN share spatial features, and best leverages the CNN's capabilities for nonlinear mapping.
Since no algorithm can calculate optimal CGH solutions, Matched I/O training datasets for supervised learning are not an option. Instead, DeepCGH is trained unsupervised, by comparing the input image to a simulated propagation of the estimated solution. An explicit loss function (e.g. accuracy) compares input and simulation and updates the CNN parameters to minimize errors.
Unsupervised learning allows DeepCGH to explore potential solutions without restrictions on hologram feasibility. As a result, DeepCGH can identify solutions that are more accurate than state-of-the-art CGH methods, including some of our prior work. Improving accuracy is a major advantage for multiphoton holography, because with the same amount of laser power it is possible to yield larger amounts of two-photon absorption.
For holographic optogenetics, this result indicates that DeepCGH can activate more neurons with the same amount of IR light below brain heating and photo-damage thresholds.
This work has been made possible by Hossein Eybposh (Ph.D. student), Nicholas Caira (Postdoc), Mathew Atisa (UNC Computer Sciences Major), and Praneeth Chakravarthula (Ph.D. Student) The team received generous support from the Burroughs Wellcome Fund (Career Award at the Scientific Interface, PI: N.Pegard), and the Nvidia GPU research grant program.
Hossein Eybposh received an OSA Student Paper Award for this work at the 2020 OSA Biophotonics congress).
Mathew Atisa received a first poster prize for his contributions at the 2020 Duke Research Computing Symnposium.