[论文解读] Towards single-shot coherent imaging via overlap-free ptychography

Oliver Hoidn, Aashwin Mishra+3 more|arXiv (Cornell University)|Feb 24, 2026

Advanced X-ray Imaging Techniques被引用 0

一句话总结

作者将 PtychoPINN 扩展为在加速传统多次扫描的同时实现无重叠、单 shot 的 Fresnel 相干衍射成像，利用物理约束、自监督框架与可微前向模型和泊松似然。

ABSTRACT

Ptychographic imaging at synchrotron and XFEL sources requires dense overlapping scans, limiting throughput and increasing dose. Extending coherent diffractive imaging to overlap-free operation on extended samples remains an open problem. Here, we extend PtychoPINN (O. Hoidn \emph{et al.}, \emph{Scientific Reports} extbf{13}, 22789, 2023) to deliver \emph{overlap-free, single-shot} reconstructions in a Fresnel coherent diffraction imaging (CDI) geometry while also accelerating conventional multi-shot ptychography. The framework couples a differentiable forward model of coherent scattering with a Poisson photon-counting likelihood; real-space overlap enters as a tunable parameter via coordinate-based grouping rather than a hard requirement. On synthetic benchmarks, reconstructions remain accurate at low counts ($\sim\!10^4$ photons/frame), and overlap-free single-shot reconstruction with an experimental probe reaches amplitude structural similarity (SSIM) 0.904, compared with 0.968 for overlap-constrained reconstruction. Against a data-saturated supervised model with the same backbone (16,384 training images), PtychoPINN achieves higher SSIM with only 1,024 images and generalizes to unseen illumination profiles. Per-graphics processing unit (GPU) throughput is approximately $40 imes$ that of least-squares maximum-likelihood (LSQ-ML) reconstruction at matched $128 imes128$ resolution. These results, validated on experimental data from the Advanced Photon Source and the Linac Coherent Light Source, unify single-exposure Fresnel CDI and overlapped ptychography within one framework, supporting dose-efficient, high-throughput imaging at modern light sources.

研究动机与目标

• Motivate reducing or removing overlap requirements in ptychographic imaging to increase throughput and reduce dose.
• Develop a self-supervised, physics-constrained framework that unifies single-shot Fresnel CDI and overlapped ptychography.
• Enable arbitrary scan geometries and handle realistic probe shapes while maintaining reconstruction fidelity.

提出的方法

• Use an inverse map G: X -> Y coupled with a differentiable forward model F: Y -> X to form an autoencoder F ∘ G trained with diffraction-domain losses.
• Represent overlap as a tunable, coordinate-based grouping rather than a hard constraint, enabling overlap-free reconstruction (Cg = 1) in Fresnel CDI geometry.
• Employ a translation-aware fusion (constraint map Fc) to merge per-patch reconstructions in a common frame across arbitrary scan geometries.
• Model diffraction with a coherent scattering forward model, including a Poisson photon-counting likelihood to connect predicted and measured diffraction amplitudes.
• Normalize diffraction data and learn a global log-intensity scale αlog to connect network outputs to absolute photon counts.
• Adopt an encoder–decoder backbone that reconstructs central high-resolution regions and a low-resolution periphery to accommodate extended probes.
• Train with Poisson NLL loss (or MAE when counts are unknown) in the diffraction domain, without real-space supervision.

(a) Idealized — CDI

实验结果

研究问题

• RQ1 Can overlap-free single-shot reconstruction be achieved in Fresnel CDI by replacing hard overlap constraints with a tunable, physics-informed grouping mechanism?
• RQ2 How does overlap-free PtychoPINN perform under low photon counts, position jitter, and with experimental probes across real data (APS, LCLS) versus synthetic data?
• RQ3 What are the comparative advantages in data efficiency and generalization versus a supervised baseline using the same backbone?
• RQ4 How does the method scale in throughput relative to conventional LSQ-ML ptychography solvers at common resolutions?
• RQ5 Can the approach generalize to out-of-distribution illumination and scan conditions (e.g., cross-facility transfer)?

主要发现

• Overlap-free reconstruction with an experimental probe achieves amplitude SSIM 0.904 versus 0.968 for overlap-constrained reconstruction in synthetic line-pattern data.
• Poisson NLL training at ~10^4 photons/frame yields comparable resolution to MAE at ~10^5 photons/frame (FRC50), indicating ~10x dose efficiency.
• PtychoPINN attains higher SSIM with only 1,024 training images versus 16,384 in the supervised baseline, showing improved data efficiency.
• Across APS and LCLS experimental data, the method reaches ~6.1×10^3 diffraction patterns/s at 64×64 and ~2.6×10^3 at 128×128 on a single GPU.
• Out-of-distribution transfer (APS-trained to LCLS data) degrades the supervised baseline but preserves edge structure in PtychoPINN, indicating better generalization.
• Single-GPU inference throughput of PtychoPINN is ~40× higher than LSQ-ML at matched 128×128 resolution.

(b) Idealized — Ptycho