DiLFM: an artifact-suppressed and noise-robust light-field microscopy through dictionary learning

The study addresses the challenge of achieving high-speed, high-resolution volumetric imaging of biological dynamics, which often occur on millisecond timescales in 3D. While LFM enables single-shot volumetric acquisition by encoding 4D light fields onto a 2D sensor, conventional reconstructions via iterative Richardson–Lucy (RL) deconvolution suffer from multiple artifacts: a trade-off between contrast and edge ringing, block-wise artifacts near the native image plane (NIP) due to low spatial sampling, depth crosstalk that introduces grid-like patterns across layers, and severe degradation under low signal-to-noise ratio (SNR). The purpose of the study is to analyze these artifacts and propose a reconstruction approach, DiLFM, that suppresses artifacts and improves noise robustness without hardware modifications, thereby enhancing LFM’s utility for in vivo biological imaging under low-light conditions.

Prior volumetric imaging approaches (confocal, multiphoton, selective plane illumination, structured illumination) require axial scanning and are limited by mechanical inertia or phototoxicity. Multiplexed strategies (multiplane/multifocal imaging, scanning temporal focusing, random access microscopy) increase throughput but face heat tolerance and sample density constraints. LFM offers scanning-free volumetric imaging at camera frame rate but reconstruction artifacts persist. Existing mitigation strategies include limiting the imaging volume to one side of the NIP (reducing depth range), shifting the microlens array (sacrificing depth of field), point spread function (PSF) shaping or adding additional views/hardware (increasing system complexity and limiting applications in freely behaving animals). Algorithmic methods such as adding strong blur reduce artifacts but degrade resolution; the phase space reconstruction improves convergence and reduces block artifacts but cannot resolve the contrast-versus-ringing dilemma. All these methods remain vulnerable to noise, especially under low-light conditions.

DiLFM is a reconstruction framework that combines a small, carefully chosen number of RL iterations with a learned dictionary-based patching step. The RL iterations are limited to operate below the onset of edge ringing, providing a baseline 3D reconstruction with reduced ringing but residual artifacts. DiLFM then applies dictionary patching: a pair of dictionaries (low- and high-fidelity) is trained under the LFM forward model using sample priors from general biological images. Low-fidelity patches (snippets) drawn from the RL reconstruction are matched to the learned low-fidelity dictionary, and the corresponding high-fidelity elements are used to replace/update them, leveraging sparse representation in an over-complete dictionary to restore structure, suppress block artifacts near NIP, and mitigate depth crosstalk. Training incorporates noise models representative of LFM acquisitions, including mixed Poisson (shot) and Gaussian (readout) noise, to enhance robustness. The method requires no hardware modifications and exploits domain similarity across biological samples. Performance is evaluated via simulations (USAF targets, spheres, gradual-intensity objects) and experiments on scattering biological specimens (Drosophila embryos and brains) and in vivo zebrafish preparations (blood flow at 100 Hz under low illumination; whole-brain calcium imaging), using metrics such as SSIM, frequency-domain analysis, intensity cross-sections, peak-to-valley ratios, and downstream analyses (blood cell counting via cross-sectional intensity fluctuations and neuron detection via CNMF-E).

- DiLFM balances contrast and artifact suppression: it achieves the contrast of RL with 10 iterations but with significantly reduced edge ringing. - NIP block-wise artifacts and grid-like depth crosstalk patterns are substantially reduced; reconstructions display smoother structures and more natural frequency responses. - Simulation comparisons show DiLFM outperforms RL, RL with anti-aliasing filtering, and the phase space method. Reported SSIMs: DiLFM 0.89 vs phase space 0.81, anti-aliasing RL 0.69, and RL 0.65. - For gradual intensity profiles, DiLFM maintains superior fidelity relative to other methods. - In Drosophila samples, DiLFM reduces block artifacts, restores smooth embryo boundaries and brain sulci, converts erroneously square cell shapes near NIP to round, and improves the peak-to-valley ratio of a brain sulcus by ~1.2× at z = 10 µm without sacrificing resolution. - Noise robustness: Under mixed Poisson and Gaussian noise, DiLFM yields clear backgrounds at PSNR ≈ 23.5 dB and achieves the best SSIM across PSNR 15.8–33.2 dB. When PSNR < 20 dB, other methods degrade markedly while DiLFM remains robust. - Zebrafish blood flow imaging at 100 Hz under very low illumination (0.12 mW mm⁻² vs 6.8 mW mm⁻²): traditional LFM becomes too noisy (e.g., at z = −30 µm), whereas DiLFM recovers clear hollow-core vessels and elliptical blood cells, reduces background and depth crosstalk, and improves blood cell counting via clearer cross-sectional intensity fluctuations. - Zebrafish whole-brain calcium imaging (HUC:H2B-GCaMP6s) at 0.37 mW mm⁻²: DiLFM provides sharper images with reduced artifacts and higher ΔF/F activity contrast; CNMF-E detects 779 neurons in an 800 × 600 × 100 µm³ volume versus 383 with traditional LFM, with more uniform detection across depths.

By integrating limited-iteration RL with dictionary-based patch replacement, DiLFM directly addresses core LFM reconstruction pathologies—edge ringing, NIP block artifacts, and depth crosstalk—while enhancing resilience to realistic mixed noise. The artifact suppression and noise robustness translate into tangible biological benefits: clearer structures, improved contrast and frequency characteristics, and more reliable quantitative analyses (e.g., accurate blood cell counting under low light and enhanced neuron detection and activity inference). The approach avoids hardware changes, preserving LFM’s compactness and suitability for in vivo and potentially mobile or constrained setups. Leveraging learned priors from biological image domains enables high-fidelity recovery even in low-SNR regimes, achieving a favorable balance between resolution preservation and artifact reduction compared with existing algorithmic alternatives.

The study introduces DiLFM, a dictionary learning-based reconstruction framework for light-field microscopy that couples a few RL iterations with dictionary patching to substantially suppress artifacts and boost noise robustness without hardware modifications. Across simulations and diverse biological experiments, DiLFM delivers higher-fidelity 3D reconstructions, improved contrast with reduced ringing, diminished NIP and crosstalk artifacts, and superior performance under low illumination. These gains enable reliable high-speed volumetric analyses, including robust blood cell counting and increased neuron yield in whole-brain calcium imaging. Potential future directions include further optimization of dictionary training across broader sample types and real-time implementations for online volumetric imaging.