Phase-Contrast Magnitude and Phase Priors

• Phase-contrast specific magnitude and phase priors are tailored constraints on the amplitude and phase, integrating physical bounds with statistical models for precise phase retrieval.
• They combine explicit physics-based limitations with deep and generative learning approaches to stabilize highly ill-posed imaging problems.
• Applications in X-ray, optical, and MRI modalities demonstrate significant improvements in RMSE, SSIM, and noise robustness for artifact suppression.

Phase-contrast specific magnitude and phase priors are application- and modality-tailored assumptions, constraints, or learned distributions imposed on the amplitude (magnitude) and phase of complex fields in phase retrieval or phase-contrast imaging. These priors act to regularize highly ill-posed inverse problems, leveraging the distinct statistical or physical behaviors of the magnitude and phase channels, and have been rigorously formulated in contemporary nonlinear optimization, deep image prior, and @@@@1@@@@ frameworks for X-ray, optical, and MRI systems.

1. Physical Basis and Mathematical Formulation

Modern phase-contrast imaging treats the object as a spatially-varying complex refractive index, $$n(\mathbf{r}) = 1 - \delta(\mathbf{r}) + i \beta(\mathbf{r})$$, yielding an exit-surface transmission function $T(u,v) = e^{-A(u,v) - i \phi(u,v)}$ where $A(u,v)$ is the integrated absorption and $\phi(u,v)$ is the integrated phase shift. The imaging system propagates $T(u,v)$ through a physically-motivated forward model—Fresnel or Fraunhofer propagation, Zernike or DPC contrast, or Fourier encoding as in MRI—where only intensity or magnitude is measured directly.

Priors are imposed on $A(u,v) = -\log|T(u,v)|$ and $\phi(u,v) = \arg T(u,v)$, based on physical or empirical knowledge:

• Feasible physical bounds (box constraints), e.g., $A_{\min} \le A(u,v) \le A_{\max}$, $\phi_{\min} \le \phi(u,v) \le \phi_{\max}$, derived from maximum thickness, refractive-index range, or specific imaging conditions (Mohan et al., 2023).
• Proportionality constraints for (quasi-)homogeneous objects, $x = z^{1 + i(\delta/\beta)}$, reducing the parameter space (Mohan et al., 2023, Gureyev et al., 2015).
• Sparse or smoothness assumptions on the phase (for modalities canceling DC/background), e.g., "dark-field sparse priors" in DPC (Zhang et al., 2022).
• Statistical generative models of complex amplitude/phase, either learned (PixelCNN, diffusion) or implicit (deep image priors) (Luo et al., 2023, Zhuang et al., 2022, Zhou et al., 24 Jan 2025).

2. Classes of Priors in Phase-Contrast Imaging

The construction and role of magnitude and phase priors depend on imaging modality and physical characteristics.

Explicit Physics-based Box and Ratio Priors

Physical knowledge yields explicit bounds on absorption and phase, parametrized by material properties and imaging setup:

• For X-ray or neutron phase-contrast tomography, maximal absorption and phase are set by the product of maximal $\beta$/$\delta$ and the thickest path through the object, $$A_{\max} \approx (2\pi/\lambda)\beta_{\max}w_{\max}$$, $$\phi_{\max} \approx (2\pi/\lambda)\delta_{\max}w_{\max}$$ (Mohan et al., 2023).
• In the monomorphous decomposition, the ratio $\delta/\beta$ is bounded globally by the set of plausible materials, and this serves as both a phase and a magnitude prior, providing minimal but stabilizing envelope constraints for TIE-based reconstructions (Gureyev et al., 2015).

Structural/Penalty-based Priors

Structurally, phase-contrast images under certain modalities exhibit sparsity or piecewise regularity:

• Dark-field sparse prior (DSP): DPC and related modalities yield "dark-field" images where the odd kernel eliminates background, resulting in intrinsic sparsity in the measurement domain; the prior is encoded as an $\ell_0$ penalty on the forward-transformed phase $K_n \phi$ (Zhang et al., 2022).
• Hessian/TV smoothness: Additional magnitude or phase priors may involve second-order smoothness (Hessian-$\ell_1$) penalties, regularizing phase in space (Zhang et al., 2022).

Deep and Generative Priors

Modern approaches leverage the expressive power of (overparameterized) neural networks or generative models:

• Double deep image priors (Double-DIP): Parameterize amplitude and phase as outputs of separate convolutional neural networks, with each acting as an implicit prior (favoring spatially structured, naturalistic solutions) and allowing joint optimization with physically-constrained forward models (Zhuang et al., 2022).
• Generative priors (PixelCNN, diffusion): Statistical models learned on large corpora of magnitude-phase pairs, trained via phase augmentation if needed, enforce high-dimensional, realistic constraints on complex image structure; can be plugged into linear or nonlinear MRI reconstruction as $-\log p(x)$ penalties (Luo et al., 2023).
• Untrained neural network decoders: For Zernike PCM and similar modalities, an untrained deep decoder acts as a self-regularizer for the phase, shaping the solution to the statistics favored by the NN architecture without requiring explicit hand-tuned regularization (Zhou et al., 24 Jan 2025).

3. Optimization and Algorithmic Incorporation

Priors are integrated with data-fidelity losses—often maximum-likelihood or least-squares forms (e.g., $\|\mathbf{y} - |\mathcal{F}\mathbf{x}|\|_2^2$, $\ell(\mathbf{y}; \mathbf{x})$)—yielding composite optimization objectives:

• Projection/Indicator Penalties: Box constraints are enforced using indicator functions and projected-gradient (or L-BFGS with projection) iterations; e.g., $$x^{k+1} = \mathrm{Proj}_C(x^k - \tau_k \nabla \ell(y;x^k))$$ where $C$ encodes the constraints in $(A,\phi)$ (Mohan et al., 2023).
• Hard/Soft Penalties for Sparsity: Nonconvex $\ell_0$ sparsity in DSP prior is handled via half-quadratic splitting (hard-thresholding in auxiliary variables) or via smooth surrogates (e.g., $1-e^{-c|x|}$) with gradient descent (N-Adam) (Zhang et al., 2022).
• Deep Prior-based Optimization: Parameters of deep networks ($\theta_1, \theta_2$ in Double-DIP, $W$ in decoder priors) are trained from scratch on a single measurement under the physical loss function, with auto-differentiation through the full forward model; no explicit regularization terms or hand-tuned hyperparameters are needed (Zhuang et al., 2022, Zhou et al., 24 Jan 2025).
• Generative Priors in MAP Reconstruction: For generative priors, reconstruction minimizes $\|A x - y\|^2 + \lambda R(x)$ using proximal-gradient or Gauss-Newton schemes; both linear ADMM/FISTA and nonlinear joint estimation are supported (Luo et al., 2023).

4. Empirical and Theoretical Impact

Across diverse modalities, explicit phase-contrast specific priors have been shown to yield substantial improvements in reconstruction accuracy and robustness:

• Quantitative Gains: Up to 2-fold normalized RMSE reduction, 5–10% SSIM increase, and significant high-frequency (MTF) gains (10–20%) in phase-contrast tomography with explicit $(A, \phi)$ bounds (Mohan et al., 2023).
• Artifact Suppression: Artifact and streak suppression, especially in the absence of explicit spatial regularizers—pure constraints suffice where regularizers would otherwise need delicate tuning (Mohan et al., 2023, Zhang et al., 2022).
• Noise Robustness: DSP prior yields 5–15 dB higher LSNR than $\ell_2$ or TV at typical DPC SNRs; Richardson–Lucy with N-Adam is faster and more robust to non-convexities than strict HQS (Zhang et al., 2022).
• Convergence and Stability: Projection/proximal approaches ensure physical feasibility at every step; monomorphous decomposition lends operator invertibility and stabilizes TIE reconstructions under noise and misalignment (Gureyev et al., 2015).
• Deep Priors' Generalization: Double-DIP outperforms both single-instance and classical support-constrained methods in Bragg CDI, breaking all major symmetries and yielding lower MSE (Zhuang et al., 2022).
• MRI Generative Priors: Phase-augmented diffusion priors deliver robust PSNR/SSIM improvements over $\ell_1$-wavelet and magnitude-only priors, with phase-information retention essential for diagnostic fidelity (Luo et al., 2023).

5. Implementation Guidelines and Practical Considerations

Successful implementation of magnitude and phase priors is grounded in system physics, empirical tuning (where necessary), and computational pragmatism.

• Physical Bound Selection: Magnitude and phase bounds derive from physical object dimensions, maximal absorption/phase coefficients, and estimated noise floor; e.g., $A_{\min} \approx 0$, $A_{\max}$ from maximal path length (Mohan et al., 2023).
• Sparsity Priors Tuning: The DSP regularization weights adapt according to a Laplacian-based image noise metric, tying prior strength to SNR directly (Zhang et al., 2022).
• Computation: Main bottlenecks are multiple forward/backward passes (e.g., repeated Fresnel or Fourier transforms); efficient implementation (FFT batching, GPU parallelism) is critical (Mohan et al., 2023, Zhuang et al., 2022).
• Deep Priors: Architecture hyperparameters (channel count, depth, upsampling schedule) influence implicit bias toward smooth/structured images; fixed architectures enable tuning-free application across sample classes (Zhou et al., 24 Jan 2025).

6. Modalities and Transferability

Phase-contrast specific magnitude and phase priors have distinct realizations in each imaging modality:

Modality	Priors Used	Representative References
X-ray/Ptychography	Box constraints, monomorphous ratios	(Mohan et al., 2023, Gureyev et al., 2015)
DPC Optics	Dark-field sparsity ($\ell_0$), Hessian	(Zhang et al., 2022)
MRI (PC, general)	Generative models, phase augmentation	(Luo et al., 2023)
Bragg CDI	Double-DIP CNN priors (mag., phase)	(Zhuang et al., 2022)
PCM (Zernike)	Untrained decoder (phase prior)	(Zhou et al., 24 Jan 2025)

A plausible implication is that while the physics dictates certain universal priors (e.g., sparsity in dark-field modalities), the technical realization and optimization algorithm must be tailored to the data model and noise characteristics—deep priors afford modality-independent flexibility as long as the forward operator is differentiable (Zhou et al., 24 Jan 2025).

7. Conceptual and Practical Significance

Imposing phase-contrast specific magnitude and phase priors transforms ill-posed inverse imaging into tractable, stable, and noise-robust reconstructions, aligning physical feasibility with statistical optimality. The incorporation of explicit, physically-motivated bounds, algorithmic adaptivity via deep or generative models, and cross-modal transfer through statistical augmentation are developments that advance both theoretical understanding and practical imaging quality across applications. Contemporary research confirms that the integration of such priors is essential for high-fidelity, artifact-resistant phase retrieval in demanding experimental settings (Mohan et al., 2023, Zhang et al., 2022, Zhuang et al., 2022, Luo et al., 2023, Zhou et al., 24 Jan 2025, Gureyev et al., 2015).