Spatial Reconstruction Fidelity (SRF)

• Spatial Reconstruction Fidelity (SRF) is a metric that quantifies how accurately ODT algorithms reconstruct known structured illumination patterns using optical phase conjugation.
• The approach employs a spatial light modulator to generate high-resolution phase masks, using mean-square error between the back-propagated field and the reference as the fidelity measure.
• Experimental benchmarks demonstrate SRF's efficacy, highlighting the superior performance of Rytov-based reconstructions over Born and Radon methods in replicating complex illumination patterns.

Spatial Reconstruction Fidelity (SRF) is a quantitative metric introduced by Ayoub et al. for evaluating the performance of optical diffraction tomography (ODT) reconstruction algorithms. It measures the degree to which a tomographic reconstruction algorithm, combined with optical phase conjugation, can reproduce an arbitrarily chosen two-dimensional spatial illumination pattern without requiring knowledge of the actual three-dimensional object under investigation. SRF assesses the fidelity of the reconstructed refractive-index distribution by using structured illumination as a form of “ground truth” in the input transverse plane. The method enables rigorous, experimental benchmarking of ODT algorithms in complex, real-world imaging scenarios, including live biological cells, entirely without access to physical phantoms or a priori knowledge of the true object (Ayoub et al., 2019).

1. Principle of Structured-Illumination-Based Assessment

SRF employs a spatial light modulator (SLM) to display high-resolution phase masks that generate known two-dimensional illumination patterns, such as the Einstein portrait and the 1951 USAF resolution target. These patterns serve as ground truth for the transverse input plane. When the system records the complex field transmitted through a sample, it later numerically propagates the conjugate field backward through a reconstructed refractive-index distribution. The core idea is that, for a perfect reconstruction and ideal noiseless system, this back-propagation will exactly reproduce the original SLM pattern. Any deviation quantifies errors strictly attributable to the ODT algorithm, rather than to unknown or unmeasured sample structure or system imperfections (Ayoub et al., 2019).

2. Experimental Apparatus and Optical Setup

The setup consists of a 0.3 W, 532 nm DPSS laser spatially cleaned by a pinhole, split into a signal and reference arm by a non-polarizing beamsplitter. In the signal arm, a reflective LCOS SLM (Holoeye PLUTO VIS, 1080×1920 pixels, 8 μm pitch) steers the illumination beam across 361 angles (±37° in 1° steps, including normal incidence) and simultaneously imprints high-resolution phase patterns. Two 4f relay systems eliminate higher-order diffracted light and provide ~240× magnification onto the object plane (100×, NA 1.4 objective). After passing through the sample, a third 4f system with a 100×, NA 1.45 objective collects the transmitted light. This is interfered with the reference beam and recorded by a scientific CMOS camera (Andor Neo 5.5, 2150×2650 pixels, 6.5 μm pitch). This configuration enables acquisition of high-fidelity holographic data stacks and precise ground-plane pattern reproduction (Ayoub et al., 2019).

3. Optical Phase Conjugation and Numerical Propagation

Central to SRF is optical phase conjugation (OPC) through complex media. The system records a hologram of the sample under structured illumination, retrieves the complex transmitted field, and then digitally conjugates it. By numerically propagating this conjugated field backward through the estimated refractive-index distribution (as reconstructed by an ODT algorithm), the method seeks to undo all distortions induced by the sample. In theory, perfect knowledge of the refractive index would yield perfect pattern recovery. The numerical propagation employs the scalar Lippmann–Schwinger integral equation:

$$E(\mathbf{r}) = E_{\rm inc}(\mathbf{r}) + \int G(\mathbf{r}-\mathbf{r}') n(\mathbf{r}') E(\mathbf{r}')\, \mathrm{d}\mathbf{r}'$$

where $E_{\rm inc}$ is the incident field, $E$ the total field, $n(\mathbf{r})$ the scattering function, $k=2\pi n_m/\lambda$ the wavenumber, $G$ the relevant Green’s function. For numerical computation, the discretized linear system

$\mathbf{E} = (I - G N)^{-1} \mathbf{E}_{\rm inc}$

is solved using BiConjugate Gradients Stabilized (BiCGStab) iteration. The resultant field is transformed to the detector plane for direct comparison with experimental holograms (Ayoub et al., 2019).

4. Definition and Calculation of the SRF Metric

SRF is quantified by the mean-square error (MSE) between the back-propagated field and the reference pattern recorded in the absence of the sample. Let $E_{\rm orig}(x, y)$ denote the complex SLM field recorded through a clear medium and $E_{\rm ret}(x, y)$ the back-propagated field after OPC through the ODT reconstruction. Then the SRF error is given by:

$$\mathrm{Error} = \mathrm{MSE}(E_{\rm ret}, E_{\rm orig}) = \frac{\displaystyle \sum_{x,y}|E_{\rm ret}(x,y) - E_{\rm orig}(x,y)|^2}{\displaystyle \sum_{x,y}|E_{\rm orig}(x,y)|^2} \times 100\%$$

A lower MSE percentage corresponds to higher spatial reconstruction fidelity. This metric is direct, interpretable, and does not require access to any a priori or ground-truth structural information about the actual three-dimensional object (Ayoub et al., 2019).

5. Evaluation Across ODT Reconstruction Algorithms

Ayoub et al. benchmarked three established ODT reconstruction schemes using SRF:

• Radon (Filtered back‐projection): Models projections as line integrals of phase, ignores diffraction, operates on unwrapped phase using the classical Radon–Wolf transform.
• Born Approximation: Applies the first-order Born series on the complex field, linearizes the scattering but is limited for thick samples or high index contrast, and does not unwrap phase.
• Rytov Approximation: Linearizes after taking the complex logarithm (enabling explicit phase–amplitude separation) and uses the Wolf inversion formula after phase unwrapping.

Each algorithm reconstructs a 3D refractive‐index map from a holographic stack (361 projection angles) of biological cells (HCT‐116 or Panc-1). The reconstructed maps are then used for numerically back-propagating structured-illumination holograms to compute SRF.

6. Comparative Results and Conclusions

Quantitative assessment using SRF yielded the following mean-square error percentages:

Test Pattern & Cell Type	Radon (%)	Born (%)	Rytov (%)
Einstein portrait, HCT-116 cell	8.83	34.73	6.39
USAF chart, Panc-1 cell	16.19	24.58	7.97

The Born method consistently produced the highest SRF errors, typically more than three times that of Rytov. Rytov approximation delivered the lowest error percentages, with Radon intermediate. These findings are corroborated by qualitative assessment of pattern restoration fidelity, where Rytov-based ODT reconstructions yielded sharper, less distorted structured images. This demonstrates SRF’s capability to distinguish diffraction-modeling accuracy and phase-unwrapping robustness in ODT schemes (Ayoub et al., 2019).

7. Advantages, Limitations, and Broader Significance

The SRF approach offers several key advantages:

• It obviates the need for physical phantoms or external ground-truth information.
• It applies without modification to arbitrarily complex and dynamic biological samples.
• It is sensitive to both algorithmic errors from phase unwrapping and inaccuracies in scattering/diffraction modeling.

Principal limitations include the computational intensity of solving the full Lippmann–Schwinger equation (notable memory and time demands), as well as residual errors potentially arising from imperfect hologram calibration, absorption losses, or system misalignments. Nonetheless, SRF as formulated and validated by Ayoub et al. constitutes a powerful, quantitative, ground-truth-free metric for benchmarking and optimizing ODT algorithms in real experimental conditions (Ayoub et al., 2019).