Intensity and Phase Imaging Through Scattering Media Via Deep Despeckle Complex Neural Networks

Shuai Liu, Peng Li, Hao Sha, Jiuyang Dong, Yué Huang, Yanjing Zhao, Xu‐Ri Yao, Peng Qin, Xiu Li, Xing Lin, Yongbing ZhangSSRN 2022 · DOI

Introduction

Imaging through scattering media, such as biological tissue or plastic containers, remains a significant challenge in optical microscopy. Light passing through these media becomes scrambled, resulting in a "speckle pattern" that obscures the underlying object. While deep learning has successfully recovered object intensity from these patterns, phase information—critical for high-contrast, label-free imaging of transparent biological cells—is often lost. Traditional phase recovery requires multiple exposures or complex holographic setups, which are difficult to implement in scattering environments.

Methods

To address these challenges, researchers developed DespeckleNet, a deep complex-valued neural network designed to simultaneously reconstruct intensity and phase from a single-shot speckle image.

Network Architecture

DespeckleNet is a Generative Adversarial Network (GAN) that incorporates complex-valued convolution blocks. Unlike standard real-valued networks, this architecture models light as an electromagnetic wave with both amplitude and phase, more accurately reflecting the physics of light propagation. The network utilizes a dual-branch decoder: one branch reconstructs the intensity image, while the other recovers the phase map. A PatchGAN discriminator is employed to maintain high-frequency details and local textures.

Overview of the training dataset preparation and the Despeckle complex network architecture.

Training and Ground Truth

Ground truth phase maps were generated using the Transport of Intensity Equation (TIE), which calculates phase from a stack of clean intensity images captured at different focal planes. The model was trained on a diverse dataset including nasopharyngeal carcinoma cells, breast cancer cells, and living HeLa cells, imaged through various scattering media like thin tape and plastic Petri dishes.

Results

DespeckleNet demonstrated superior performance in reconstructing both intensity and phase compared to standard architectures like UNet and IDiffNet.

Fidelity and Generalization

The model achieved higher Structural Similarity Index (SSIM) scores, particularly in phase reconstruction. Notably, the network exhibited strong generalization; once trained on a specific class of diffuser (e.g., tape), it could accurately reconstruct images through different, unseen samples of that same material.

Demonstration of the system's versatility across different diffusers and samples, including long-term real-time phase observation of cell division.

Real-Time Biological Observation

A key achievement was the continuous monitoring of living HeLa cells through a Petri dish for 800 minutes. DespeckleNet successfully captured the entire cell division process in a scattering environment without the need for toxic stains or labels. The reconstructed phase images provided significantly higher contrast than intensity images, clearly showing morphological changes during mitosis.

Conclusion

The integration of complex-valued layers into neural network architectures represents a major advancement for optical tasks. By mimicking the mathematical form of coherent light propagation, DespeckleNet enables high-speed, single-shot phase recovery through scattering media. This approach opens new possibilities for non-invasive, real-time imaging in clinical and biological research, even in challenging environments that typically preclude high-resolution microscopy.

Summary updated

BibTeX
@misc{liu2022intensity,
  author = {Shuai Liu and Peng Li and Hao Sha and Jiuyang Dong and Yué Huang and Yanjing Zhao and Xu‐Ri Yao and Peng Qin and Xiu Li and Xing Lin and Yongbing Zhang},
  title = {{Intensity and Phase Imaging Through Scattering Media Via Deep Despeckle Complex Neural Networks}},
  year = {2022},
  doi = {10.2139/ssrn.4089089},
  howpublished = {SSRN},
}
Figure