Aberration-Corrected Iterative Phase Retrieval

• Aberration-corrected iterative phase retrieval algorithms are techniques that integrate the estimation and compensation of spatially varying optical aberrations to restore diffraction-limited imaging.
• They extend classical methods like Gerchberg–Saxton and Fienup by incorporating models such as Zernike polynomials and modal decompositions to refine both amplitude and phase reconstructions.
• Practical applications in digital holography, microscopy, and ultrasound imaging demonstrate significant improvements in resolution, SNR, and overall imaging fidelity while balancing computational cost.

An aberration-corrected iterative phase-retrieval algorithm is any phase retrieval procedure in which optical aberrations or non-idealities—typically modeled as spatially varying phase distortions in the imaging system’s transfer function or PSF—are directly estimated, compensated, or incorporated into the forward propagation scheme. These algorithms extend classical iterative phase retrieval methods (e.g., Gerchberg–Saxton, Fienup error reduction) to correct for degradations caused by defocus, lens aberrations, misalignment, system drift, or wavefront inhomogeneities. The principal aim is to accurately reconstruct the amplitude and phase of an object from intensity-only measurements in the presence of aberrations, thereby restoring diffraction-limited or otherwise optimal imaging performance.

1. Mathematical Models with Aberrations

The standard forward model in phase retrieval posits an object with complex transmission $O(x,y)=a(x,y)\exp[i\phi(x,y)]$, which after propagation through an aberration-free optical system yields a field $U(x',y')$ and measured intensity $I_0(x',y')=|U(x',y')|^2$. Incorporating aberrations requires modifying the system transfer function:

• Pupil Plane Aberrations: In the frequency domain, aberrations are described by a phase function $A(u,v)=\exp[i\phi_\mathrm{sys}(u,v)]$, often parameterized by Zernike polynomials $Z_n$ as $\phi_\mathrm{sys}(u,v) = \sum_n a_n Z_n(u,v)$. The transfer function becomes $H(u,v)=A(u,v)H_0(u,v)$, with $H_0(u,v)$ the aberration-free kernel.
• Space-Invariant vs. Space-Variant PSF (SI-PSF, SV-PSF): For SI-PSF, the convolutional model holds: $U(x',y')=(O*h)(x',y')$. For SV-PSF, the local PSF is $h_{x,y}(x',y')$ and the forward operator is integral: $U(x',y') = \iint O(x,y)\, h_{x,y}(x',y')\, dx\,dy$ (Brault et al., 7 Feb 2025).
• Separable Modal Decomposition for SV-PSF: The SV-PSF is approximated by a modal sum $$h_{x,y}(x',y') \approx \sum_{p=1}^Q w_p(x,y) m_p(x',y')$$ with global “mode” kernels $m_p$ and spatially variable weights $w_p(x,y)$, typically derived by truncated SVD on PSF measurements (Brault et al., 7 Feb 2025).

2. Iterative Phase-Retrieval Schemes with Aberration Correction

Aberration correction is introduced at specific points in the iterative update loop. The principal frameworks are:

• Modified Gerchberg–Saxton–Fienup Algorithm: Every iteration alternates between object and measurement domains. In the measurement domain, the amplitude is replaced by the measured $\sqrt{I_0}$. Back-propagation and forward-propagation are performed using the aberration-corrected transfer function: $F\{O^{(n)}\}(u,v) = G^{(n)}(u,v)A^*(u,v)H_0^*(u,v)$ and $V^{(n)}(x',y')=F^{-1}\{Q^{(n)}A(u,v)H_0(u,v)\}$ (Brault et al., 7 Feb 2025).
• Space-Variant Propagation: For SV-PSF, each iteration replaces the convolution with a modal sum, requiring $Q$ FFT-based convolutions per step, both in forward and back directions (Brault et al., 7 Feb 2025).
• Constraint Enforcement: Object domain constraints (support, known phase, amplitude positivity) are imposed after each back-propagation; measurement domain constraints are imposed by replacing amplitudes (Brault et al., 7 Feb 2025).
• Two-Stage “Background Compensation”: In systems with persistent, nonparametric aberrations, a “background” disturbance is first estimated via calibration (using a known object), and subsequent phase retrieval divides by the background estimate at each iteration (Migukin et al., 2012).
• Blind Joint Recovery (BPR): In ptychographic and coherent diffraction regimes, the probe or pupil function $p$—containing all aberration phase—is recovered simultaneously with the sample $u$ by alternately updating both via operator splitting (e.g., ADMM, PALM), block-coordinate descent, or second-order (Gauss–Newton/LM) steps (Chang et al., 2022).
• Adaptive Noise Models and Phase Diversity: In optics and microscopy, phase-diverse datasets with known defocus are used to more robustly retrieve aberration coefficients, modeling noise as either Gaussian or Poisson and updating object and phase coefficients in alternating ML/EM or gradient steps (Reiser et al., 2023).

3. Specific Algorithmic Strategies and Implementations

A variety of algorithmic strategies can be concretely instantiated for aberration-corrected iterative phase retrieval:

Scheme	Aberration Representation	Typical Update Mechanism
Modified IPR (Brault et al., 7 Feb 2025)	Zernike expansion in $A(u,v)$	FFT-based forward/backprops
SV-PSF Modal (Brault et al., 7 Feb 2025)	Modal PSF sum: $w_p(x,y), m_p$	$Q$ FFT convs per iteration
Background Compensation	Generalized pupil function $u_B$	AL/split Bregman approach
Blind BPR (Chang et al., 2022)	Probe $p$, updated at each step	Alternating minimization
Phase Diversity (Reiser et al., 2023)	Zernike and defocus terms	ML/EM, Gauss–Newton, Poisson

• FFT-Based Convolution and Propagation: All spatial propagation, both in the SI- and SV-PSF cases, is efficiently implemented via FFT, with padding to prevent wrap-around (Brault et al., 7 Feb 2025).
• Edge Artifacts and Padding: For 1D arrays or data with finite support (in ultrasound, for example), iterative zero-padding prevents edge-induced artifacts due to the periodic FFT assumption (Monjazebi et al., 2021).
• Support and Prior Constraints: Constraints on object size or imposed priors (e.g., positivity, known phase) enhance convergence and stability (Brault et al., 7 Feb 2025, Migukin et al., 2012).
• Regularization in Blind/BPR: Additional penalties on the probe (e.g., smoothness or Zernike basis projection) are incorporated to stabilize aberration attribution and prevent nonphysical phase artifacts (Chang et al., 2022).

4. Convergence, Complexity, and Practical Considerations

• Convergence Rates: Convergence for aberration-corrected IPR methods is typically empirically observed in $O(100$–$300)$ iterations for optical phase retrieval (Brault et al., 7 Feb 2025), while some specialized settings (e.g., ultrasound) converge in $2–3$ iterations due to simpler aberration structure (Monjazebi et al., 2021).
• Computational Scaling: In the SV-PSF formalism, computational cost scales as $2Q$ FFTs per iteration ($Q$ for forward, $Q$ for back-propagation), making the approach practical for $Q \approx 10$–$20$ modes on standard hardware (Brault et al., 7 Feb 2025). For blind BPR with large $n$, subspace and domain decomposition methods enable scaling to $10^5 \times 10^5$ pixels via parallelization (Chang et al., 2022).
• Memory Requirements: Storing all PSF mode kernels $m_p$ and weight maps $w_p(x,y)$ is required for SV-PSF correction; in blind BPR, probe and sample arrays must be maintained (Brault et al., 7 Feb 2025, Chang et al., 2022).
• Accuracy/Runtime Tradeoffs: Increasing the number of modal PSF terms $Q$ or Zernike bases improves fidelity (e.g., RMSE, SSIM) at a linear increase in computation (Brault et al., 7 Feb 2025, Reiser et al., 2023).
• Stopping Criteria: Convergence is typically monitored via the norm of parameter updates, the plateauing of data-fit cost, or predefined maximum iterations (20–300, depending on modality and data scale) (Migukin et al., 2012, Reiser et al., 2023).

5. Applications and Quantitative Performance

• Digital Holography: SV-PSF-corrected IPR suppresses twin-image artifacts and achieves higher fidelity reconstructions, with demonstrated resolution and contrast gains over uncorrected retrieval (Brault et al., 7 Feb 2025).
• Optical and Electron Microscopy: Phase-diverse, aberration-corrected retrieval enables recovery of high-order aberrations, yielding near-diffraction-limited images and sharp point-spread functions. Poisson-based models are statistically superior in high photon count regimes (Reiser et al., 2023).
• Ultrasound Imaging: Adaptive 2D spatiotemporal filtering combined with iterative delay updates significantly improves SNR, CNR, and spatial resolution in low SNR scenarios (e.g., PSNR from 5.4 dB to 8.2 dB, CNR from 1.54 dB to 1.92 dB) (Monjazebi et al., 2021).
• Coherent Diffraction Imaging/Ptychography: Blind phase retrieval with probe/pupil updates robustly corrects for system aberrations and enables high-quality reconstructions in large fields of view. Convergence and accuracy are further improved with probe-calibration terms or second-order refinement (Chang et al., 2022).
• Calibration-Based Systems: The two-step approach (background estimation then object phase retrieval) decouples aberration estimation from object regularization; SNR improvements (e.g., from 18 dB to 33 dB) and resolution gains (e.g., from 35 lp/mm to 50 lp/mm) are common (Migukin et al., 2012).

6. Limitations and Extensions

• Aberration Model Accuracy: Algorithm stability and reconstruction quality hinge on the accuracy of the aberration model or modal decomposition; model mismatch can slow convergence or leave residual artifacts (Brault et al., 7 Feb 2025).
• Computation vs. Fidelity: Modal truncation ($Q$) or the number of Zernike modes introduces a trade-off between computation time and aberration correction fidelity. Excessive modal truncation may underfit spatially complex aberrations (Brault et al., 7 Feb 2025, Reiser et al., 2023).
• Edge Effects: Periodicity assumptions in FFTs necessitate appropriate padding or extrapolation schemes to prevent edge distortions (Monjazebi et al., 2021).
• Adaptivity and Noise Models: Noise and signal statistics (Gaussian/Poisson) strongly influence optimal cost functions. Poisson models yield increased robustness to spatially variant and random phase noise in photon-limited microscopy (Reiser et al., 2023).
• Scan Overlap and Data Layout: In ptychography and BPR, scan overlap of at least $50$–$60\%$ is needed for robust probe and aberration recovery; sub-Nyquist overlap may induce probe drift or convergence stagnation (Chang et al., 2022).
• Parallelism and Multigrid: For large-scale problems, domain decomposition and multigrid methods enable scalable, parallel correction of both low- and high-order aberrations (Chang et al., 2022).

7. Summary Table of Core Algorithms

Algorithm	Aberration Handling	Primary Setting	Typical Iteration Count
Aberration-corrected IPR (Brault et al., 7 Feb 2025)	Zernike/modal correction	Holography, lens optics	100–300
SV-PSF Modal-Sum (Brault et al., 7 Feb 2025)	Modal low-rank PSFs	Wide-field imaging	100–200 ($Q$-dependent)
Phase diversity (ML/EM, Poisson/Gauss) (Reiser et al., 2023)	Zernike, defocus	Widefield microscopy	20–50
Blind ptychographic BPR (Chang et al., 2022)	Joint probe/object update	X-ray/electron imaging	50–300
Background-comp./augmented Lagrangian (Migukin et al., 2012)	Generalized pupil	4f coherent optics	20–50
f-NCC for ultrasound (Monjazebi et al., 2021)	Time delay estimation	Medical imaging	2–3

Significantly, implementation details such as convolution padding, edge zero-padding, careful mode-truncation, probe regularization, and noise-adaptive cost functions are critical for high-performance aberration-corrected phase retrieval across diverse optical, acoustical, and electron imaging modalities.