Deep Phase Decoder: Self-calibrating phase microscopy with an untrained\n deep neural network

Emrah Bostan, Reinhard Heckel, Michael Chen, Michael E. Kellman, Laura WallerarXiv 2020 · DOI · arXiv

Abstract

Deep neural networks have emerged as effective tools for computational imaging including quantitative phase microscopy of transparent samples. To reconstruct phase from intensity, current approaches rely on supervised learning with training examples; consequently, their performance is sensitive to a match of training and imaging settings. Here we propose a new approach to phase microscopy by using an untrained deep neural network for measurement formation, encapsulating the image prior and imaging physics. Our approach does not require any training data and simultaneously reconstructs the sought phase and pupil-plane aberrations by fitting the weights of the network to the captured images. To demonstrate experimentally, we reconstruct quantitative phase from through-focus images blindly (i.e. no explicit knowledge of the aberrations).

Introduction: The Challenge of Phase Retrieval

Quantitative Phase Microscopy (QPM) is essential for imaging transparent "phase objects," such as living cells, which do not significantly absorb light but do alter its phase. Because standard image sensors only record intensity (brightness), retrieving this phase information requires computational reconstruction.

Currently, researchers face a trade-off between two paradigms:

  1. Traditional Iterative Methods: These require precise knowledge of the microscope's physics (e.g., lens aberrations) and "hand-crafted" mathematical priors. If the physical model is slightly inaccurate, the reconstruction fails.
  2. Supervised Deep Learning: While powerful, these models are "data-hungry," requiring thousands of labeled examples. They are also notoriously fragile; if the experimental setup changes slightly from the training data, the network can produce artifacts or "hallucinations."

The Deep Phase Decoder (DPD) Framework

The authors introduce the Deep Phase Decoder (DPD), a training-free approach that leverages the structural priors of neural networks without requiring a pre-existing dataset. Instead of learning from a database, the DPD optimizes its weights to fit a single set of experimental measurements.

The DPD framework combines a deep decoder for phase generation ($G_p$) and a Zernike-based aberration generator ($G_a$) with a physical model of image formation to minimize the loss between generated and raw measurements.

The architecture is split into two sub-generators:

  • Phase Generator ($G_p$): An untrained "Deep Decoder" network that synthesizes the sample's phase image. Its under-parameterized nature acts as a regularizer, favoring natural image structures over random noise.
  • Aberration Generator ($G_a$): A fully-connected layer that models system-specific aberrations using Zernike polynomials.

By integrating these generators with a mathematical model of the microscope's physics, the system can "self-calibrate," simultaneously recovering the phase image and the unknown pupil-plane aberrations by minimizing the difference between generated and raw intensity images.

Experimental Validation and Results

The DPD was tested using an experimental through-focus intensity stack of a phase target captured with a Nikon TE300 microscope. Unlike traditional methods that require explicit knowledge of defocus distances, DPD performs "blind" reconstruction, identifying the microscope's state on the fly.

Comparison of DPD reconstructions using 8 versus 4 intensity measurements. The method maintains high fidelity even with reduced data, demonstrating significant data efficiency.

Key Findings

  • High-Fidelity Reconstruction: DPD successfully recovered the phase target with high fidelity, showing strong agreement with established iterative methods like accelerated Wirtinger flow.
  • Blind Self-Calibration: The network accurately identified "defocus-like" pupil functions without being told the defocus distances, effectively calibrating the system during the reconstruction process.
  • Data Efficiency: The method remains robust even with reduced data, producing high-quality results with as few as four intensity measurements.

Discussion and Future Directions

While DPD offers a robust, training-free alternative to supervised learning, it is currently limited by computational speed. The optimization process requires approximately $5 \times 10^4$ iterations, making it significantly slower than the near-instantaneous inference of a pre-trained CNN.

Future work will focus on testing the approach with more complex, highly scattering biological tissues and 3D volumes. Additionally, finding a "universal" architecture that generalizes across different microscope types remains an open challenge for untrained neural network approaches.

Summary updated

Abstract

Deep neural networks have emerged as effective tools for computational\nimaging including quantitative phase microscopy of transparent samples. To\nreconstruct phase from intensity, current approaches rely on supervised\nlearning with training examples; consequently, their performance is sensitive\nto a match of training and imaging settings. Here we propose a new approach to\nphase microscopy by using an untrained deep neural network for measurement\nformation, encapsulating the image prior and imaging physics. Our approach does\nnot require any training data and simultaneously reconstructs the sought phase\nand pupil-plane aberrations by fitting the weights of the network to the\ncaptured images. To demonstrate experimentally, we reconstruct quantitative\nphase from through-focus images blindly (i.e. no explicit knowledge of the\naberrations).\n
BibTeX
@misc{bostan2020deep,
  author = {Emrah Bostan and Reinhard Heckel and Michael Chen and Michael E. Kellman and Laura Waller},
  title = {{Deep Phase Decoder: Self-calibrating phase microscopy with an untrained\textbackslash\{\}n deep neural network}},
  year = {2020},
  eprint = {2001.09803},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2001.09803},
}
Figure