Introduction
Capturing the cellular dynamics of a beating heart in three dimensions remains a significant challenge in biological imaging. High spatial resolution is required to resolve individual cells, while high temporal resolution is necessary to track rapid contractions. Traditional scanning methods, such as confocal or two-photon microscopy, are often too slow. While Light-Sheet Microscopy (LSM) offers increased speed, achieving high-speed volumetric imaging frequently necessitates intense light doses, leading to phototoxicity and photobleaching.
To overcome these limitations, researchers have developed a high-performance computational framework that integrates Compressive Sensing (CS) with LSM. This approach, termed CS-LSM, allows for rapid, low-phototoxic 3D imaging by reconstructing high-fidelity volumetric data from sparse, compressed measurements.
The CS-LSM Framework
The CS-LSM framework is built upon two primary components: hardware-based encoding and a flexible reconstruction algorithm.
Hardware-Based Encoding
The system utilizes a Digital Micromirror Device (DMD) to apply random binary masks to various z-planes (slices) of the heart. Multiple slices are multiplexed and summed into a single 2D camera exposure. This compression significantly reduces the amount of data that needs to be captured, thereby lowering the required light dose.
PnP-ADMM Reconstruction
To reconstruct the 3D volume from overlapping signals, the framework employs the Alternating Direction Method of Multipliers (ADMM) within a Plug-and-Play (PnP) architecture. The PnP approach decouples the physical forward model of the microscope from the denoising process. This modularity allows researchers to "plug in" various denoising algorithms—such as Tikhonov, Total Variation (TV), or BM3D—as priors without altering the core optimization logic.
A critical innovation in this framework is the introduction of temporal regularization along the z-axis. In high-speed acquisition, adjacent slices in a stack exhibit high structural continuity. By enforcing smoothness across these slices, the algorithm effectively reduces artifacts introduced during the compression process.
Results and Discussion
The framework was evaluated using a synthesized 3D dataset of zebrafish cardiac dynamics, simulating real-world imaging conditions including Gaussian noise.
The researchers compared several denoising priors, finding that Temporal-BM3D provided the highest quality reconstructions. In noise-free simulations, it achieved a Peak Signal-to-Noise Ratio (PSNR) of 42.68 dB and a Structural Similarity Index (SSIM) of 0.9923. In noisy scenarios, the temporal coupling proved essential; for BM3D, adding temporal regularization improved the SSIM from 0.41 to 0.77.
While BM3D offers superior image quality, it comes with a higher computational cost, taking nearly an hour per volume compared to less than a second for simpler Tikhonov reconstructions. This creates a trade-off between real-time previewing and high-quality final results.
Conclusion
The CS-LSM framework provides a flexible and powerful solution for high-speed volumetric imaging of dynamic biological processes. By leveraging the PnP-ADMM architecture and temporal regularization, it achieves high-fidelity reconstructions with reduced phototoxicity. Future work will focus on integrating deep-learning-based denoisers to improve reconstruction speed and extending the framework to full 4D (3D + time) cardiac cycle reconstruction.
Summary updated