Research map / Image generation / Image restoration
Single Exposure Quantitative Phase Imaging with a Conventional Microscope Using Diffusion Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes single-exposure quantitative phase imaging by using chromatic aberration in a white-light brightfield microscope with an RGB camera, and a Zero-Mean Diffusion model trained on synthetic data to recover phase.
Problem
Transport-of-Intensity Equation (TIE) phase retrieval normally needs multiple through-focus acquisitions, which is impractical in clinical settings; polychromatic illumination introduces a fixed small defocus and blur that conventional TIE solvers handle poorly.
Approach
Use chromatic aberration of white light and an RGB camera to obtain a single-exposure through-focus-like stack, then solve phase with a conditional diffusion model (Zero-Mean Diffusion) that diffuses residuals around a learned mean predictor, rather than normalizing data to [-1,1], with a theoretical upper-bound argument favoring the centered process.
Key results
- On synthetic HCOCO, ZMD (p) achieved MS-SSIM 0.97 and MAE 0.06, versus 0.81 and 0.14 for US TIE (p).
- On through-focus urine stacks, ZMD (p) gave the best MS-SSIM in samples 2-4 (e.g., 0.86 to 0.89) and best MAE in several samples, outperforming US TIE (p).
- On clinical urine images, reconstructed RBC phase values were reported as consistent with literature values of 1.8 to 2.3 rad.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ZMD (p) instead of MP (p) (ablation) | phase reconstruction · Synthetic HCOCO | MS-SSIM: 0.95 $\pm$ 0.02 -> 0.97 $\pm$ 0.02 | Helped |
| Add the ZMD residual diffusion prediction to Mean Prediction (ablation) | phase reconstruction · Synthetic HCOCO | MAE: 0.06 $\pm$ 0.02 -> 0.06 $\pm$ 0.02 | No clear effect |
| Add the ZMD residual diffusion prediction to Mean Prediction (ablation) | phase reconstruction · Through-Focus Brightfield, samples 1–4 | MS-SSIM: 0.78 -> 0.82; 0.88 -> 0.89; 0.83 -> 0.88; 0.84 -> 0.86 | Helped |
| Add the ZMD residual diffusion prediction to Mean Prediction (ablation) | phase reconstruction · Through-Focus Brightfield, samples 1–4 | MAE: 0.09 -> 0.08; 0.06 -> 0.05; 0.10 -> 0.10; 0.09 -> 0.09 | Mixed |
| Use ZMD (p) instead of the strongest non-ZMD method | phase reconstruction · Through-Focus Brightfield red blood cells, sample 1 | MS-SSIM: 0.90 -> 0.80 | Hurt |
| Use ZMD (p) instead of MP (p) (ablation) | phase reconstruction · Through-Focus Brightfield red blood cells, sample 2 | MS-SSIM: 0.78 -> 0.83 | Helped |
| Use ZMD (p) instead of MP (p) (ablation) | phase reconstruction · Through-Focus Brightfield red blood cells, sample 3 | MS-SSIM: 0.88 -> 0.90 | Helped |
Surprising findings
- On the through-focus red blood cell stacks, US TIE with 2-shot acquisition beat ZMD (p) in sample 1 MS-SSIM (0.90 vs 0.80).
Limitations
- Paraxial approximation and TIE assumptions limit accuracy; future work suggests Helmholtz-based PINNs.
- Blurring from polychromatic source incoherence degrades detail in the recovered phase.
- Ground truth for clinical data unavailable; validation relies on literature values.
Experimental setup
Datasets: ImageNet HCOCO UTI clinical urine microscopy dataset
Benchmarks: HCOCO synthetic images through-focus brightfield urine and red blood cell stacks UTI clinical microscopy dataset
Metrics: MS-SSIM MAE
Baselines: US TIE (2-shot) US TIE (polychromatic) Mean Prediction regression model
Builds on: Conditional Variational Diffusion Models Transport-of-Intensity Equation US TIE
Closest papers by meaning
- MicroDiffusion: Implicit Representation-Guided Diffusion for 3D Reconstruction from Limited 2D Microscopy Projections (CVPR 2024 · similarity 0.84)
- Volume Tells: Dual Cycle-Consistent Diffusion for 3D Fluorescence Microscopy De-noising and Super-Resolution (CVPR 2025 · similarity 0.82)
- OpticalNet: An Optical Imaging Dataset and Benchmark Beyond the Diffraction Limit (CVPR 2025 · similarity 0.82)
- CD-RCM: Generalizable Continuous-Depth Novel View Synthesis for Reflectance Confocal Microscopy (NeurIPS 2026 · similarity 0.81)
- Step-Calibrated Diffusion for Biomedical Optical Image Restoration (AAAI 2025 · similarity 0.81)
- Understanding Multi-layered Transmission Matrices (CVPR 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.