Single Exposure Quantitative Phase Imaging with a Conventional Microscope using Diffusion Models

1. Problem Statement

Quantitative Phase Imaging (QPI) is essential for biomedical applications (e.g., analyzing cell morphology) as it provides quantitative data on transparent objects without the need for staining. The Transport-of-Intensity Equation (TIE) is a standard method for phase retrieval, but it faces significant limitations in clinical settings:

• Multi-shot Requirement: Traditional TIE requires acquiring multiple images at different defocus distances (a through-focus stack), which is time-consuming and mechanically difficult to implement in point-of-care or high-throughput clinical environments.
• Low-Frequency Artifacts: Single-shot TIE approaches often suffer from low-frequency artifacts and noise amplification.
• Chromatic Aberration Challenges: While using white light (polychromatic sources) with a standard RGB camera can theoretically generate a "through-focus stack" via chromatic aberrations (different wavelengths focus at different planes), this introduces blurring and incoherence artifacts. Conventional TIE solvers fail to handle these specific artifacts effectively.

2. Methodology

The authors propose a novel framework combining chromatic aberration-induced defocus with a Deep Learning-based diffusion model to achieve single-exposure phase retrieval.

A. Optical Setup & Forward Model

• Hardware: A conventional brightfield microscope equipped with a standard white light source and a commercially available RGB (polychromatic) camera.
• Mechanism: Chromatic aberration causes different wavelengths to focus at different axial distances. The RGB camera captures a single image where the Red, Green, and Blue channels effectively represent different defocus distances ($z$).
• Mathematical Formulation: The intensity image $I(x; z)$ is modeled as a superposition of coherent images across the spectrum. By treating wavelength $\lambda$ and defocus $z$ as coupled variables ($\xi = \lambda z$), the TIE can be reformulated to allow phase recovery from a single polychromatic exposure.

B. Data Generation (Synthetic Training)

Since acquiring ground-truth phase data for real clinical samples is difficult, the authors generated a large synthetic dataset:

• Source: Images from ImageNet converted to grayscale and treated as pure phase objects.
• Simulation: They simulated polychromatic acquisitions using the Fresnel propagator for wavelengths between 400nm and 700nm.
• Sensor Modeling: The simulation accounts for the specific quantum efficiency of RGB sensors (modeled as Gaussian distributions centered at 630nm, 550nm, and 450nm) and adds realistic noise.

C. Zero-Mean Diffusion (ZMD)

The core algorithmic contribution is Zero-Mean Diffusion, a modification of Conditional Variational Diffusion Models (CVDM) designed specifically for quantitative data.

• The Problem with Standard Diffusion: Standard diffusion models typically normalize data to the $[-1, 1]$ interval. This destroys the physical meaning of the pixel values, making them unsuitable for quantitative phase imaging where absolute values matter.
• The ZMD Solution:
  1. Decomposition: The model learns to predict the phase in two parts:
     • A Mean Predictor ($\mu_\theta$): A regression model (U-Net) that estimates the expected value (mean) of the phase.
     • A Diffusion Model ($\tilde{\epsilon}_\nu$): A noise-prediction model that learns the residual (the difference between the actual phase and the predicted mean).
  2. Training: The diffusion process operates on the centered data ($\tilde{Y} = Y - \mu_\theta$). The total loss function combines the diffusion loss with a mean prediction loss ($L_{ZMD} = L_{CVDM} + \omega L_{mean}$).
  3. Inference: The final phase is reconstructed by adding the predicted mean back to the diffusion output: $\hat{Y} = \tilde{Z}_1 + \mu_\theta(X)$.
• Theoretical Basis: The authors prove that using a centered forward process (ZMD) provides a tighter upper bound on the Kullback-Leibler divergence between the true distribution and the generated distribution compared to standard non-centered diffusion, theoretically justifying its superior performance for quantitative tasks.

3. Key Contributions

1. Single-Exposure QPI: A practical method to perform quantitative phase imaging using a single white-light exposure on a standard microscope, eliminating the need for mechanical defocusing.
2. Zero-Mean Diffusion (ZMD): A new class of diffusion models that avoids data normalization, preserving physical interpretability for quantitative imaging. The paper provides theoretical proofs showing ZMD offers better convergence bounds than standard approaches.
3. Physics-Informed Synthetic Data: A robust pipeline for generating training data that simulates the complex physics of polychromatic diffraction and sensor characteristics, reducing reliance on scarce real-world ground-truth data.
4. Clinical Validation: Successful application to real-world clinical urine samples (red blood cells and epithelial cells) using off-the-shelf hardware.

4. Results

The method was evaluated on synthetic data (HCOCO dataset), through-focus clinical stacks, and uncurated clinical urine samples.

• Performance Metrics:
  • Synthetic Data: ZMD (polychromatic) achieved an MS-SSIM of 0.97 and MAE of 0.06, outperforming both standard US TIE (polychromatic) and 2-shot TIE methods.
  • Clinical Stacks: ZMD consistently outperformed US TIE in MS-SSIM and MAE across four different urine samples. For example, in Sample 1, ZMD (polychromatic) achieved an MS-SSIM of 0.82 vs. 0.65 for US TIE (polychromatic).
  • Real Clinical Images: The model successfully reconstructed the morphology of red blood cells (including biconcave shapes) and epithelial cells. The quantitative phase values (1.8–2.3 radians) matched literature values.
• Qualitative Improvements:
  • ZMD recovered fine structural details (e.g., the ring morphology of red blood cells) that were lost or blurred in US TIE reconstructions.
  • It significantly reduced "cloud artifacts" and noise common in traditional TIE solvers when applied to polychromatic data.

5. Significance

• Clinical Feasibility: By leveraging widely available hardware (standard microscopes and RGB cameras) and a single exposure, this method removes the primary barrier to adopting QPI in routine clinical diagnostics (e.g., Urinary Tract Infection screening).
• Robustness: The approach is robust against the blurring and incoherence inherent in white-light microscopy, a problem that has historically limited single-shot TIE.
• Generalizability: The introduction of Zero-Mean Diffusion offers a new paradigm for applying generative AI to any scientific imaging task requiring quantitative, non-normalized output, bridging the gap between deep learning and physical measurement.

In conclusion, the paper demonstrates that combining chromatic aberration with a specialized diffusion model allows for high-accuracy, single-shot quantitative phase imaging, making advanced microscopy accessible for real-world clinical applications.