High-resolution scanning fluorescence imaging through scattering via speckle replica alignment and variance computation

1. Problem Statement

Fluorescence microscopy is a cornerstone of biomedical research, but its application deep within biological tissues is severely hindered by optical scattering. Scattering media (like tissue) distort light, causing aberrations and blurring that limit resolution and contrast.

• Limitations of Existing Methods:
  • Wavefront Shaping: Requires active modulation (e.g., Spatial Light Modulators) and often relies on iterative optimization or guide stars, which can be slow and complex.
  • Autocorrelation/Phase Retrieval: Computational methods that reconstruct images from single speckle patterns often require iterative phase retrieval. These are ill-posed problems, highly sensitive to noise, and struggle to converge for complex, non-sparse objects.
  • Nonlinear Imaging: Existing computational methods for linear fluorescence often fail in nonlinear regimes (e.g., Two-Photon/2P) because the effective Point Spread Function (PSF) changes with the illumination pattern, violating assumptions of fixed PSFs.

The authors aim to develop a robust, high-resolution computational imaging method that works for both linear (1P) and nonlinear (2P) fluorescence modalities under unknown random illuminations, without requiring wavefront shaping or iterative phase retrieval.

2. Methodology

The proposed method, termed Speckle Replica Alignment and Variance Computation, leverages the intrinsic correlations within speckle patterns generated by scattering media. The core principle is that a hidden object creates multiple, shifted "replicas" of itself within a single speckle image. By aligning these replicas and analyzing their temporal fluctuations, the object can be reconstructed.

The reconstruction pipeline consists of six key steps:

1. Data Acquisition & Preprocessing:

   • 1P (Linear) Mode: A series of speckle images is captured. The pixel-wise variance of the raw stack is computed first to create the input dataset. This suppresses static background noise and enhances features.
   • 2P (Nonlinear) Mode: The sample is excited by a scanned speckle field (using a rotating diffuser) and detected by a single-pixel detector (PMT). The raw speckle images themselves serve as the input dataset due to the quadratic intensity dependence of 2P excitation.

2. Subimage Division:

   • The input dataset is divided into a set of partially overlapping subimages. The size of these subimages is estimated based on the object's spatial extent (derived from autocorrelation analysis).

3. Cross-Correlation Based Alignment:

   • One subimage is selected as a reference.
   • An iterative process aligns all other subimages to this reference using normalized cross-correlation.
   • The reference is dynamically updated by summing the aligned subimages, creating a highly accurate, perfectly overlaid dataset of "replicas."

4. Variance Computation (Super-Resolution Step):

   • Once aligned, the pixel-wise variance is calculated across the ensemble of subimages.
   • Mechanism: This step acts similarly to Super-Resolution Optical Fluctuation Imaging (SOFI). Since the dominant object replica is consistent across subimages while background noise and weaker, shifted replicas fluctuate randomly, the variance calculation amplifies the object signal while suppressing uncorrelated noise. This effectively enhances resolution beyond the diffraction limit of the raw speckle.

5. Dual Deconvolution:

   • A final dual-deconvolution is performed between the intermediate variance image and the input dataset.
   • This step aggregates information from multiple replicas to further refine the image, improving resolution and quality, especially when the number of available subimages is limited.

6. Output: The final reconstructed image ($Recon(r)$) reveals the hidden object with high resolution.

3. Key Contributions

• Modality Agnosticism: The method successfully operates in both linear (1P) and nonlinear (2P) fluorescence regimes. This is a significant advancement, as previous computational methods often struggled to generalize to nonlinear excitation where the PSF is illumination-dependent.
• No Phase Retrieval: Unlike traditional speckle-correlation methods, this approach avoids the ill-posed, noise-sensitive iterative phase retrieval algorithms. It relies on direct statistical operations (alignment and variance).
• High Resolution via Variance: The integration of pixel-wise variance computation allows the method to achieve super-resolution, resolving details finer than the speckle grain size, analogous to SOFI but applied to scattering media.
• Robustness to Complexity: Numerical simulations demonstrate that the method is more robust than autocorrelation-based techniques when reconstructing complex, dense objects (low sparsity), where phase retrieval typically fails.

4. Experimental Results

The authors validated the method through two distinct experimental setups:

• 1P Fluorescence Imaging:

  • Setup: Fluorescent beads behind a sand-blasted glass slide, illuminated by a CW laser with random phase patterns generated by an SLM.
  • Result: The method successfully reconstructed the bead distribution, closely matching the ground truth. The resolution was significantly improved compared to simple incoherent summation.

• 2P Fluorescence Imaging:

  • Setup: Fluorescent beads behind a scattering layer, excited by a femtosecond pulsed laser with a rotating diffuser. Detection was performed via a PMT (single-pixel) while scanning the speckle pattern.
  • Result: High-resolution images were recovered despite the nonlinear excitation and single-pixel detection.
  • Comparison: The authors compared their variance-based reconstruction against standard incoherent summation and higher-order SOFI. The results confirmed that the variance computation (acting as a 2nd-order correlation) resolves finer details than incoherent summation, approaching the resolution of higher-order SOFI without the computational complexity of higher-order tensor calculations.

• Numerical Validation: Simulations comparing the proposed method against autocorrelation-based reconstruction showed superior performance for objects with varying sparsity levels, particularly for complex, non-sparse objects.

5. Significance and Impact

• Versatility: This framework provides a unified solution for deep-tissue imaging across different fluorescence modalities, removing the need for separate algorithms for linear and nonlinear regimes.
• Practicality: By eliminating the need for wavefront shaping hardware (SLMs) and complex iterative optimization, the method is more accessible and potentially faster for real-world applications.
• Future Applications: The approach opens new avenues for computational imaging in complex scattering environments (e.g., deep brain imaging, thick tissue samples) using spatially incoherent light. It is readily extendable to other nonlinear modalities, such as Three-Photon (3P) fluorescence.

In summary, this paper presents a powerful, computationally efficient strategy that transforms the "noise" of scattering speckle into a structured resource for high-resolution imaging, overcoming the limitations of both traditional microscopy and existing computational reconstruction techniques.