Second Harmonic Imaging Enhanced by Deep Learning Decipher

Weiru Fan, Tianrun Chen, Eddie M. Gil, Shi‐Yao Zhu, Vladislav V. Yakovlev, Da‐Wei Wang, Delong ZhangACS 2021 · DOI · arXiv

Abstract

Wavefront sensing and reconstruction are widely used for adaptive optics, aberration correction, and high-resolution optical phase imaging. Traditionally, interference and/or microlens arrays are used to convert the optical phase into intensity variation. Direct imaging of distorted wavefront usually results in complicated phase retrieval with low contrast and low sensitivity. Here, a novel nonlinear optical encoding approach has been developed and experimentally demonstrated using optical second harmonic generation to sharpen the phase information carried by the probe beam. By designing and implementing a deep neural network, we demonstrate the second harmonic imaging enhanced by a deep learning decipher (SHIELD) for efficient and resilient phase retrieval. Inheriting the advantages of two-photon microscopy, SHIELD demonstrates single-shot, reference-free, and video-rate phase imaging with sensitivity better than $\lambda/100$ and high robustness against noise, facilitating numerous applications from biological imaging to wavefront sensing.

Introduction

Capturing the phase of light—the timing of its oscillations—is fundamental to biological imaging and adaptive optics. However, standard image sensors only detect intensity, necessitating indirect methods to recover phase information. Traditional techniques like interferometry or Shack-Hartmann sensors suffer from mechanical sensitivity, complex alignment requirements, or limited spatial resolution. While computational methods like ptychography offer high resolution, they typically require multiple exposures, hindering real-time observation of dynamic processes.

SHIELD (Second Harmonic Imaging Enhanced by Deep Learning Decipher) addresses these challenges by combining nonlinear optical encoding with a specialized deep neural network. This framework enables high-resolution, single-shot, reference-free phase reconstruction at video-rate speeds.

Methodology: Nonlinear Encoding and Deep Learning Decipher

The SHIELD framework operates through a two-stage process: nonlinear encoding followed by a deep learning-based reconstruction.

Nonlinear Encoding

The system utilizes a fundamental laser beam (1064 nm) that passes through a sample, acquiring spatial phase modulations. This beam then enters a periodically poled potassium titanyl phosphate (PPKTP) crystal. Within the crystal, Second Harmonic Generation (SHG) occurs, producing light at 532 nm. The intensity of the resulting SHG signal is intrinsically sensitive to the spatial phase modulations of the fundamental beam. Effectively, the SHG process acts as a natural high-pass filter, encoding high-frequency details—specifically the second derivative of the phase—into a 2D intensity pattern.

Densely Connected Reconstruction Network (DCRN)

Because the relationship between input phase and SHG intensity is governed by complex, nonlinear Maxwell equations, analytical solutions are often impractical. To solve this, the researchers developed the DCRN, a neural network architecture designed to map SHG intensity patterns back to the original phase distributions. Unlike standard U-Net architectures, the DCRN employs dense blocks to extract hidden nonlinear features, ensuring high-fidelity reconstruction.

Schematic of the SHIELD framework, showing (a) the nonlinear encoding process via SHG and (b) the Densely Connected Reconstruction Network (DCRN) architecture.

The model was trained using the EMNIST (handwritten characters) and Quick Draw (sketches) datasets. A 90/10 split was used for training and testing, demonstrating that the network could generalize across different types of spatial patterns.

Results and Discussion

SHIELD demonstrates significant improvements over traditional phase imaging techniques in terms of sensitivity, speed, and robustness.

Performance Metrics

  • Sensitivity: The system can detect phase changes smaller than 10.6 nm, achieving a sensitivity better than $\lambda/100$.
  • Speed: Reconstruction occurs at approximately 22.8 frames per second (44 ms per image), enabling near video-rate imaging.
  • Robustness: The DCRN maintains high reconstruction accuracy (measured by the Jaccard Index) even under adverse conditions, such as 30% Gaussian noise or partial overexposure.
Performance of SHIELD: (a) Phase reconstruction examples, (b) Jaccard Index over training steps, and (c) robustness across different modulation depths.

Advantages of the SHIELD Approach

By operating in a reference-free, lens-free configuration, SHIELD is immune to the "fringe hopping" and alignment errors typical of interferometers. Furthermore, it inherits the benefits of two-photon microscopy, including reduced photodamage and enhanced penetration through scattering media. The system's ability to capture sharp edges and high-spatial-frequency details stems from the SHG process's inherent sensitivity to phase gradients.

Conclusion and Future Outlook

SHIELD represents a robust alternative to traditional wavefront sensors, offering a compact and efficient setup for high-resolution phase imaging. While current implementations are limited by the relatively low efficiency of SHG, future improvements could involve more efficient nonlinear crystals or single-photon sensitive detectors.

Moreover, because the SHG process is nearly instantaneous, the integration of femtosecond laser pulses could enable chemically selective phase imaging with femtosecond temporal resolution. This would represent a six-order-of-magnitude improvement over current transient phase imaging techniques, opening new avenues for studying ultrafast biological and chemical dynamics.

Summary updated

Abstract

Wavefront sensing and reconstruction are widely used for adaptive optics, aberration correction, and high-resolution optical phase imaging. Traditionally, interference and/or microlens arrays are used to convert the optical phase into intensity variation. Direct imaging of distorted wavefront usually results in complicated phase retrieval with low contrast and low sensitivity. Here, a novel nonlinear optical encoding approach has been developed and experimentally demonstrated using optical second harmonic generation to sharpen the phase information carried by the probe beam. By designing and implementing a deep neural network, we demonstrate the second harmonic imaging enhanced by a deep learning decipher (SHIELD) for efficient and resilient phase retrieval. Inheriting the advantages of two-photon microscopy, SHIELD demonstrates single-shot, reference-free, and video-rate phase imaging with sensitivity better than λ100 and high robustness against noise, facilitating numerous applications from biological imaging to wavefront sensing.
BibTeX
@inproceedings{fan2021second,
  author = {Weiru Fan and Tianrun Chen and Eddie M. Gil and Shi‐Yao Zhu and Vladislav V. Yakovlev and Da‐Wei Wang and Delong Zhang},
  title = {{Second Harmonic Imaging Enhanced by Deep Learning Decipher}},
  booktitle = {ACS},
  year = {2021},
  doi = {10.1021/acsphotonics.1c00395},
}
Figure