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

Three-dimensional (3D) biological dynamics at high spatiotemporal resolution are crucial for understanding cellular processes. While techniques like confocal and multiphoton microscopy offer 3D imaging, their temporal resolution is limited by scanning. Light field microscopy (LFM) emerges as a high-speed alternative, capturing 4D light fields on a single 2D detector. However, LFM faces challenges: trade-offs between contrast and ringing artifacts, block-wise artifacts near the native image plane (NIP), depth crosstalk, and poor performance in low light conditions. Existing solutions, such as adjusting optical components or modifying reconstruction algorithms, have limitations. This paper proposes DiLFM, a novel method using dictionary learning to address these issues without hardware modifications. DiLFM leverages sparse signal representation and combines a few Richardson-Lucy (RL) iterations with dictionary patching to suppress artifacts and improve contrast, even under noisy conditions. The method's effectiveness is demonstrated through simulations and experiments on various biological samples.

The introduction provides a comprehensive review of existing microscopy techniques, highlighting their limitations in achieving high spatiotemporal resolution in 3D. It discusses various approaches to 3D imaging, including confocal, multiphoton, selective plane illumination, and structured illumination microscopy, emphasizing their trade-offs between speed and resolution. The limitations of LFM, such as artifacts and noise sensitivity, are clearly outlined, along with existing attempts to mitigate these issues, such as adjusting the imaging volume, reshaping the point spread function (PSF), adding views, and modifying reconstruction algorithms. The review sets the stage for the proposed DiLFM method, positioning it as a superior alternative that addresses the shortcomings of previous approaches.

DiLFM is a novel light-field microscopy (LFM) technique that combines few Richardson-Lucy (RL) deconvolution iterations with a dictionary learning-based patching process to reconstruct high-quality 3D images from raw LFM measurements. The RL iterations provide an initial, ringing-reduced reconstruction, which is then refined by the dictionary patching approach. A pair of dictionaries, one representing low-fidelity image patches from the RL reconstruction and another representing high-fidelity image patches (representing artifact-free ideal image patches), are trained using a set of representative biological samples. The training process takes into account both Gaussian and Poisson noise, enhancing robustness in real-world scenarios. The algorithm iteratively identifies low-fidelity patches in the initial reconstruction and replaces them with corresponding high-fidelity patches from the learned dictionary, effectively suppressing artifacts and improving contrast. The number of RL iterations is carefully chosen to balance contrast enhancement and ringing suppression. The dictionary training considers a forward model of the LFM imaging system to ensure that the learned dictionaries accurately represent the relationships between the raw measurements and the reconstructed images. The entire process is designed to minimize the introduction of new artifacts while maintaining image resolution and high signal-to-noise ratio (SNR).

DiLFM effectively suppresses various LFM artifacts. Numerical simulations demonstrate DiLFM's superiority over existing methods in reducing block-wise artifacts near the NIP, mitigating depth crosstalk, and improving contrast without sacrificing resolution. The Structural Similarity Index (SSIM) is used to quantitatively assess reconstruction quality, showcasing DiLFM's superior performance over traditional RL deconvolution, RL with anti-aliasing filters, and the phase-space reconstruction approach. Experiments on Drosophila embryos and brains confirm artifact reduction and improved structural clarity. In vivo zebrafish blood flow imaging at 100 Hz demonstrates DiLFM's robustness in low-light conditions, resulting in clearer blood cell visualization and improved accuracy in blood cell counting. Furthermore, DiLFM exhibits significant noise robustness, outperforming other methods under various noise levels. In vivo whole-brain calcium imaging of zebrafish larvae under low illumination power shows that DiLFM reveals significantly more neurons (779 vs 383) compared to traditional LFM. This improved detection results from the reduced background noise and artifacts, enabling more accurate neuron activity inference.

DiLFM's success addresses the long-standing challenges in LFM by providing a computational solution that significantly enhances image quality without requiring expensive or complex hardware modifications. The use of dictionary learning enables the algorithm to generalize to a wide range of biological samples, making it a versatile tool for various applications. The findings demonstrate the potential of DiLFM to significantly improve the accuracy and reliability of quantitative analyses based on LFM data, opening avenues for deeper insights into complex biological processes. The enhanced noise robustness is particularly significant, allowing for high-quality imaging with reduced phototoxicity, which is crucial for long-term in vivo studies. The results suggest that DiLFM could be broadly applicable to other 3D microscopy techniques that suffer from similar artifact and noise issues.

DiLFM offers a significant advancement in light-field microscopy by effectively suppressing artifacts and improving noise robustness through a dictionary learning approach. This method enhances the quality of 3D reconstructions, enabling more reliable and accurate biological imaging in various applications. Future research could explore the development of more sophisticated dictionary learning strategies, optimizing the dictionary training process for specific biological samples or imaging conditions. Expanding the application of DiLFM to other imaging modalities and exploring its potential in combination with other advanced microscopy techniques would also be valuable future directions.

While DiLFM demonstrates significant improvements, certain limitations exist. The performance of DiLFM depends on the quality and representativeness of the training data used to generate the dictionaries. An insufficient or biased training dataset may limit the generalizability of the method to novel samples or imaging conditions. The computational cost associated with dictionary learning and patching might be a factor in high-throughput applications. Further research is needed to optimize the computational efficiency of the algorithm.