Poisson-Gaussian Holographic Phase Retrieval with Score-based Image Prior

Zongyu Li, Jason Hu, Xiaojian Xu, Liyue Shen, Jeffrey A. Fessler

TL;DR

This paper introduces AWFS, a new algorithm for holographic phase retrieval that explicitly models the realistic, mixed Poisson-Gaussian noise found in optical sensors. By combining an accelerated optimization framework with a score-based generative image prior, the authors achieve state-of-the-art reconstruction quality and provide theoretical convergence guarantees for this notoriously difficult non-convex problem.

What Problem Does This Paper Address?

Phase retrieval (PR) is the challenge of recovering a signal or image from magnitude-only measurements—essentially trying to find the "lost" phase information. In holographic PR, a known reference beam is used to help resolve this ambiguity.

However, real-world optical systems suffer from two distinct types of noise simultaneously:

  1. Poisson Noise: Arising from discrete photon counting and dark current.
  2. Gaussian Noise: Arising from electronic readout circuitry (like ADCs).

Most existing algorithms simplify the math by assuming only one type of noise exists. This mismatch between the model and reality leads to blurry or artifact-heavy reconstructions. Furthermore, while Deep Learning (DL) has improved image priors, many DL methods require massive datasets that aren't always available in specialized scientific imaging fields.

Methods

The authors propose the Accelerated Wirtinger Flow with Score-based prior (AWFS). Their approach rests on three pillars:

  • Realistic Likelihood Modeling: Instead of approximating noise, they use the full Poisson-Gaussian (PG) log-likelihood. They derive the gradient for this complex function and, crucially, determine its Lipschitz constant, which is necessary for stable optimization.
  • Score-based Generative Prior: Rather than using a simple denoiser (like in Plug-and-Play methods) or a full diffusion model (which can be data-hungry), they use a Noise Conditional Score Network (NCSN). This network learns the "score function"—the gradient of the log-probability density of the image data. This acts as a powerful regularizer that keeps the reconstruction on the manifold of "natural" images.
  • Optimization Framework: They integrate these components into an Accelerated Wirtinger Flow (AWF) framework, which uses Nesterov-style momentum to speed up convergence.
Fig. 1: Illustration of Poisson and Gaussian noise statistics in holographic phase retrieval.

Datasets

The researchers evaluated AWFS on three distinct datasets to test its versatility:

  • Histopathology: 162 breast cancer tissue images.
  • CelebA: 920 celebrity face images.
  • CT-Density: 720 images derived from SPECT/CT medical scans.

These datasets represent relatively small-data regimes, highlighting the algorithm's ability to learn effective priors without millions of training examples.

Key Insights and Ideas

  • The "PG" Advantage: Explicitly modeling the mixed Poisson-Gaussian statistics consistently outperforms models that assume only one noise type. Accurate physics leads to better math.
  • Score Matching > DDPM for Small Data: The authors found that score-based priors are more efficient and effective than Denoising Diffusion Probabilistic Models (DDPM) when training data is limited.
  • Theoretical Reliability: Unlike many "black box" deep learning imaging methods, the authors provide a proof that their algorithm converges to a critical point of the posterior distribution, offering much-needed reliability for scientific applications.
  • Parameter Robustness: AWFS proved to be more robust to varying noise levels without requiring the tedious manual parameter tuning that plagues methods like Total Variation (TV) or PnP-ADMM.

Key Results

AWFS demonstrated superior performance across both qualitative and quantitative metrics (SSIM and NRMSE):

  • Visual Quality: As seen in the histopathology results (Figure 2), AWFS recovers fine details and textures that other methods like TV or RED-SD blur out or lose in the noise.
  • Quantitative Gains: In the CT-Density dataset, AWFS achieved an SSIM of 0.88, significantly higher than the Unregularized PG model (0.24) or even the diffusion-based DOLPH (0.55).
  • Noise Robustness: The algorithm maintained high performance even as Gaussian readout noise ($\sigma$) increased, showing much smaller performance drops compared to competing methods.
Fig. 2: Reconstructed images on dataset [97]. The bottom left/right subfigures correspond to the zoomed in area and the error map for each image. α and σ were set to 0.02 and 1, respectively.

Limitations and Future Work

While powerful, the authors note a few hurdles:

  • Computational Complexity: Calculating the infinite sum required for the exact Poisson-Gaussian likelihood is computationally expensive. The authors currently use a precise approximation, but further acceleration is needed.
  • Dataset Specificity: While AWFS dominated in structured scientific images (Histopathology and CT), standard denoising priors (PnP-Prox) performed slightly better on the highly varied CelebA dataset. This suggests that the choice of prior should be tailored to the "randomness" of the target data.
  • Future Directions: The team aims to explore Diffusion Posterior Sampling (DPS) to further refine the generative process and find ways to speed up the likelihood gradient calculations.

Summary updated

Abstract

Phase retrieval (PR) is a crucial problem in many imaging applications. This study focuses on resolving the holographic phase retrieval problem in situations where the measurements are affected by a combination of Poisson and Gaussian noise, which commonly occurs in optical imaging systems. To address this problem, we propose a new algorithm called "AWFS" that uses the accelerated Wirtinger flow (AWF) with a score function as generative prior. Specifically, we formulate the PR problem as an optimization problem that incorporates both data fidelity and regularization terms. We calculate the gradient of the log-likelihood function for PR and determine its corresponding Lipschitz constant. Additionally, we introduce a generative prior in our regularization framework by using score matching to capture information about the gradient of image prior distributions. We provide theoretical analysis that establishes a critical-point convergence guarantee for the proposed algorithm. The results of our simulation experiments on three different datasets show the following: 1) By using the PG likelihood model, the proposed algorithm improves reconstruction compared to algorithms based solely on Gaussian or Poisson likelihood. 2) The proposed score-based image prior method, performs better than the method based on denoising diffusion probabilistic model (DDPM), as well as plug-and-play alternating direction method of multipliers (PnP-ADMM) and regularization by denoising (RED).
BibTeX
@inproceedings{li2023poisson,
  author = {Zongyu Li and Jason Hu and Xiaojian Xu and Liyue Shen and Jeffrey A. Fessler},
  title = {{Poisson-Gaussian Holographic Phase Retrieval with Score-based Image Prior}},
  booktitle = {Deep Inverse @ NeurIPS},
  year = {2023},
  url = {https://openreview.net/forum?id=j6YOCS0cyM},
}
Figure