Light Sheet Imaging and Interactive Analysis of the Cardiac Structure in Neonatal Mice

Abstract

Light-sheet microscopy (LSM) enables us to strengthen the understanding of cardiac development, injury, and regeneration in mammalian models. This emerging technique decouples laser illumination and fluorescence detection to investigate cardiac micro-structure and function with a high spatial resolution while minimizing photodamage and maximizing penetration depth. To unravel the potential of volumetric imaging in cardiac development and repair, we sought to integrate our in-house LSM, Adipo-Clear, and virtual reality (VR) with neonatal mouse hearts. We demonstrate the use of Adipo-Clear to render mouse hearts transparent, the development of our in-house LSM to capture the myocardial architecture within the intact heart, and the integration of VR to explore, measure, and assess regions of interests in an interactive manner. Collectively, we have established an innovative and holistic strategy for the image acquisition and interpretation, providing an entry point to assess myocardial micro-architecture throughout the entire mammalian heart for the understanding of cardiac morphogenesis.

Graphical Abstract

Light-sheet microscopy coupled with virtual reality enables the interactive visualization and analysis of micro-architecture of the neonatal mouse heart. We propose a holistic strategy including system construction, control algorithm, sample preparation, image processing, and quantitative analysis to advance the understanding of myocardial structure in cardiac development and regeneration.

Introduction

Heart failure remains the leading cause of mortality worldwide, largely due to the lack of regenerative capacity of mature cardiomyocytes. The advanced understanding of cardiac architecture and function is critical to uncover the underlying mechanism of cardiac morphogenesis and the remodeling process in response to myocardial infarction^1-2. Recent progress has demonstrated that neonatal mice have the capacity to repair cardiomyocytes following cardiac injury³, laying the foundation to investigate the cues to structural and functional abnormalities in mammals. While computed tomography (CT), magnetic resonance imaging (MRI), and ultrasound (US) are capable of 3-dimensional (3D) imaging, the paucity of specific fluorescence labeling techniques and low spatial resolutions limit the in-depth investigation of cardiac micro-architecture in neonatal mice⁴. For this reason, we customized a light-sheet microscope (LSM)⁵ along with tissue clearing methods^6,7 and virtual reality (VR) ⁸ for the study of the intact neonatal mouse heart.

While these methods allow us to advance the understanding of cardiac morphology and function, they also result in unprecedented challenges, such as complicated 3D micro-ultrastructures and volume expansion post tissue clearing. In this context, we have substituted the conventional image analysis methods for interactive operations via VR to enable user-directed investigations of the trabecular network in model systems ^8-9. To enhance interactive capability and establish a straightforward framework for cardiac studies, we sought to create an integrative workflow to provide more insights into cardiac morphogenesis.

Our newly developed system allows for the rapid imaging with a lateral resolution of 2.30 ± 0.37 μm and an axial resolution of 2.91 ± 0.34 μm. In combination with eight-view fusion, our method achieves a nearly isotropic resolution of 1.90 μm along the lateral and axial directions. The implemented image stitching¹⁰ method further enables us to capture images over a field of view (FOV) of 3 – 4 mm, covering the entire neonatal mouse heart. The horizontal layout of the detection module¹¹ simplifies multiview fusion during image acquisition, obviating the misalignment issue resulting from dual-illumination in our previous version⁵. Furthermore, we have adapted the established protocol of Adipo-Clear⁶ to minimize artifacts and maximize imaging depth in the intact neonatal mouse hearts. In contrast to other methods, this approach preserves native morphology and expedites the processing time for adipose tissue, dramatically improving its stiffness and reducing its volume expansion^6,12. To explore the complicated ventricular trabeculation extracted from LSM imaging data, we have customized the interaction between users and digital hearts, demonstrating the benefits of highly free user-directed quantitative analysis via VR models over the conventional pre-defined analysis¹³.

Overall, our customized framework of LSM imaging, tissue clearing, and VR-based analysis of neonatal mouse hearts provides a means to study myocardial architecture across intact hearts, holding the great potential to elucidate the contribution of myocardial microenvironment, cell lineages, and signaling pathways in response to cardiac anomaly and dysfunction.

Results

Customized light-sheet microscope

Our LSM system is composed of a horizontal detection arm¹¹ and a static beam-based illumination module⁵. The light-sheet beam is reshaped by a cylindrical lens, a group of relay lenses, and an illumination objective lens for the imaging of the intact mouse heart. The auto-fluorescence emitting from the heart is captured by the detection objective lens and sCMOS camera (Figure 1A). A four-axis stage is used to translate and rotate the heart along the detection axis while the light-sheet is stationary (Figure 1B), enabling us to capture a 3D image stack of the heart. A customized LabVIEW control algorithm is used to synchronize the propagation of the laser, the stage movement, and sCMOS camera exposure (Figure 1D). During the image acquisition process, the inherent trigger and rolling shutter on the sCMOS are employed to capture sequential 2D images, while the stage continuously translates the heart across the light-sheet at a constant velocity that is determined by the step size and exposure time. In contrast to our previous solution⁵, this customized chamber allows us to adapt the Adipo-Clear tissue clearing protocol for specific cardiac research, minimize refractive index mismatch, align the light-sheet, elongate the depth of field, and capture fluorescence emitting from the sample in a large working distance (Figures 1C-D)^14,15. All raw images were used for volume rendering in the conventional mode, and the subsequent segmentation and VR reconstruction were implemented for interactive analysis.

Figure 1. Schematic of LSM imaging and Adipo-Clear tissue clearing.

(A) LSM system hardware with the representative Gaussian light-sheet captured under the system. (B) Layout of detection lens, illumination lens, and the chamber. The heart is mounted in a holder within the chamber at the intersection of both lenses. (C) Close up of the chamber with the sample attached to the holder. (D) Summary of Adipo-clear protocol used to render the heart transparent. BE: beam expander; CL: cylindrical lens; L2: lens 2; L1: lens 1; IL: illumination lens; DL: detection lens; FW: filter wheel; TL: tube lens, DBE: Dibenzyl ether. Scale bars: 500 μm.

Point spread function calibration

We imaged fluorescent beads with a diameter of 0.53 μm to measure the point spread function (PSF) of our system. To verify the spatial resolution and minimize the potential issue of opacity at different depths, we calibrated the PSF across the whole sample, and demonstrated the results of the full width at half maximum (FWHM) at depths of 3.5 μm, 40.0 μm, 75.0 μm, and 85.0 μm. Representative images of the beads are presented in the ‘Fire’ pseudocolor associated with ImageJ (Figure 2), and all PSF data for the 20 beads measured are included in Supplementary Figure S1. The averaged lateral and axial resolutions indicated by the FWHM are 2.30 μm, and 2.91 μm, respectively. The comparability of the resolution values can be attributed to the comparability of the numerical aperture (NA) of the illumination (NA = 0.30) and detection (NA = 0.25) lenses used. To achieve isotropic resolution, minimize the stripes in the imaging data, and maximize the imaging depth, we further implemented multiview deconvolution using beads¹⁶.

Figure 2. Point spread function (PSF) is used to quantify the lateral and axial resolution of the LSM system.

(A) Images of fluorescent beads in the XY (Scale bar: 50 μm) and YZ planes (Scale bar:10 μm), and the averaged PSF of 20 beads. (B) PSF of representative beads at depths of 3.5 μm, 40.0 μm, 75.0 μm, and 85.0 μm (Scale bars: 5 μm). The values are depicted by full width at half maximum (FWHM).

Multiview fusion and deconvolution of beads

The employed multiview technique involves the use of fusion and deconvolution to accurately reconstruct the overlapped region of interest from images obtained at different angles. We have captured images of beads from eight angles (Figure 3A), fused and deconvolved the raw data using the Efficient Bayesian-based multiview deconvolution approach^17-18. We have summarized the detailed steps to generate the fusion and deconvolution results in Figure 3B. This approach involves the use of the individual PSF’s acquired from each angle to estimate the PSF of the system (Figure 3C). To demonstrate the efficiency of multiview deconvolution, we increased the thickness of our light sheet to generate a larger confocal region for ease of visualization and overlap of beads from multiple perspectives. Based on this new configuration, the raw data of lateral and axial resolution was around 2.70 μm and 15 μm respectively pre-multiview. A reconstructed lateral resolution of 1.89 μm, and an axial resolution of 1.98 μm across the whole sample were obtained post-multiview (Figure 3D). The multiview deconvolution allows to not only capture images from different angles and minimize stripes, but also generate the isotropic resolution due to the conversion of lateral resolution to the axial resolution in different views (Figure 3D and Supplementary Video 1).

Figure 3. Multiview deconvolution enables to obtain images with improved isotropic resolution.

(A) Illustration of the acquisition of 2D images from eight angles. (B) Summary of the steps taken to generate multiview imaging results. (C) Images of the PSF of 0.53 μm fluorescent beads from each angle and the fused PSF for the system from the XY and XZ views. (D) Imaging results indicate an improved isotropic resolution of around 1.90 μm. Scale bars: 5 μm.

Imaging of the myocardial architecture of the mouse heart at postnatal day 1

Our LSM system allows for the rapid investigation of the 3D myocardium in neonatal mouse hearts at postnatal day 1 (P1) without multiview deconvolution and image stitching within 2 minutes (Figure 4). Following the Nyquist sampling criterion, we imaged the entire P1 heart from atria to ventricles to apex with 1600 raw images at a step size of 1 μm. Our results indicated that the ventricular chambers, atria, and valve leaflets were well preserved and discernible after tissue clearing and imaging of the heart (Figures 4A&B). Our raw image stack was projected to three orthogonal perspectives (XY, XZ, and YZ) for the comparison of image contrast and resolution within a single scan. The conventional 3D volume rendering approach further allowed us to resample the digital heart along the coronal, sagittal, and transverse planes regardless of the sampling direction (Figure 4C). By virtue of the rapid scanning of LSM, the 3D digital reconstruction of the P1 heart provides a straightforward way to track cardiac micro-structures such as the myocardial trabeculation and compaction within the ventricles (Figures 4C-G).

Figure 4. Single-perspective LSM images of the neonatal mouse heart at postnatal day 1 and the 3D rendering of a thin section of representative compartments of the heart.

(A) 2D images of the P1 mouse heart at different depths from the XY, YZ, and XZ directions. (B) Trabeculation is captured under the custom-built LSM. (C) Volume rendered images of the P1 mouse heart from the XY, YZ, and XZ directions. Close-up on (D) left atrium, (E) valves, (F) trabeculae within the ventricles, and the (G) right atrium. Scale bars: (A-C) 500 μm and (D-G) 200 μm.

Image stitching for an extended field of view

The confocal region determined by light diffraction usually limits the effective Gaussian light-sheet region. To extend the confocal region across the entire heart, we have implemented BigStitcher¹⁰ to stitch multiple image stacks of the P1 mouse heart for evenly distributed illumination and comparable spatial resolution in different tiles. To demonstrate the feasibility, we fused three representative tiles, indicated in blue, green, and yellow, during the post-image analysis (Figure 5). Rbpms (RNA-binding protein with multiple splicing) is an RNA binding protein that plays an essential role in heart development. One recent study demonstrated that the absence of Rbpms causes severe noncompaction cardiomyopathy in the neonatal mouse heart¹⁹, which is an optimal model to test our LSM imaging technique in visualizing the detailed structure of the myocardium and trabeculae. We have qualitatively compared the ventricular myocardial details of P1 Rbpms knockout (KO) (Figure 5A) and the wild-type littermate mouse hearts (Figure 5B) and observed obviously hyper-trabeculated myocardium in the KO heart (Figure 5A). The results indicate that our LSM imaging, tissue clearing, and post-processing methods perform equally well on both models. In contrast to the unstitched image (blue box in Figure 5), the micro-architecture of the trabeculae was more discernible due to the enhancement of image contrast and cellular resolution in the stitched image (red box in Figure 5); A side-by-side comparison of the unstitched and stitched images is included in Supplementary Figure S2 and Video 2. This further allows us to compare myocardial trabeculation and compaction in the wild-type mouse heart with the noncompaction cardiomyopathy model system in our future work.

Figure 5. Image stitching enables to improve the image quality and reduce the overall blurriness of regions illuminated by the thicker portions of the light sheet.

(A & B) Tiles 1, 2, and 3 represent the higher spatial resolution regions of images obtained with the thinnest portion of the light sheet at three representative positions. Tiles 1, 2, and 3 are stitched to create a less blurry 2D image as compared to the unstitched image. RA: right atrium; LV: left ventricle. Scale bars: 500 μm.

Multiview deconvolution and image stitching of P1 mouse heart

To demonstrate the full capacity of cardiac imaging, we implemented multiview deconvolution and image stitching to investigate the trabecular network across the entire P1 mouse heart. We have suspended the heart, post-clearing, between a transparent resin¹⁸ and a column of agarose gel in a Dibenzyl ether (DBE)-filled glass tube, along with the fluorescent beads embedded in the resin (Figure 6A). Similar to our stitching result (Figure 5), we scanned the heart thrice to cover different sections using the effective light-sheet thickness for each angle (Figure 6B), and then stitched the representative tiles. After capturing images of the heart and beads, simultaneously, from six angles, we fused the stitched results from the different views (Figure 6C) and reconstructed the multiview model (Figure 6D and Supplementary Video 3). The fluorescent beads were used for registration, i.e., to determine the transform function for the deconvolution of the heart images. In contrast to the results from a single perspective, our multiview deconvolution and stitching results allowed us to navigate inside the intricate trabecular network across ventricles and atria with near isotropic resolution in both 2D and 3D (Figure 6D). The detailed myocardial trabeculation and compaction provide a comprehensive map inside the ventricles and holds the great promise for the in-depth analysis of cardiac development, regeneration in response to infarction, and the contribution of different cell lineages to cardiac anomaly and dysfunction.

Figure 6. The combination of multiview deconvolution and image stitching enables to investigate the intricate trabecular network in the whole heart with the isotropic resolution.

(A) Illustration of the set-up for the simultaneous imaging of the P1 mouse heart and fluorescent beads. Tiles 1, 2, and 3 represent the higher spatial resolution regions of images obtained with the thinnest portion of the light sheet at three representative positions for each angle. Tiles 1, 2, and 3 are stitched to improve image quality before multiview deconvolution. (B) Raw images of P1 mouse heart and beads at different angles of 0°, 60°, 120°, 180°, 240°, and 300°. (C) Reconstructed image of P1 mouse heart and fluorescent beads post-multiview deconvolution and stitching. (D) Volume rendered image of the multiview deconvolved P1 mouse heart highlighting the trabeculation within the ventricular cavity. Scale bars for dashed boxes: 100 μm and all other figures: 500 μm.

Quantitative analysis of the ventricular cavity

While the conventional volume rendering enables us to assess 3D myocardial structure in the P1 mouse heart (Figures 7A-B), the in-depth investigation of the trabecular network remains challenging due to the complicated interconnection of the myocardium uncovered by the LSM. To better understand the 3D structure, we sought to provide more freedom for users to investigate the whole architecture in its natural form by obviating the need for pre-defined perspectives and operations. We have introduced user-directed operations inside the ventricle for the interactive analysis of 3D trabeculae. For example, even though we were able to quantify the width of the ventricular cavities, the volume, and the thickness of the myocardium as ~1.35 mm, ~0.33 mm³, and 0.06 - 0.18 mm respectively, this kind of conventional analysis largely depends on the manual annotation of the region of interest through well-established software, such as Amira (Figure 7B). The specific annotation for analysis leads to the generation of different models, being time-consuming and remaining the bottleneck of high-throughput analysis in collaborative projects. In comparison, our own VR method ^8,9,13 allows users to conduct different kinds of analysis and operations on a well-established model such the 3D trabecular network in its natural form (Figure 7C-F). Using this novel method, we are able to interact with the entire left ventricle of the mouse heart and a portion of the trabecular network that we are interested in during our exploration of this digital model (Supplementary Video 4). By viewing the heart in a virtual 3D environment, we can explore the interconnected aspect of the trabecular network. In addition, the platform allows for quantitative measurements such as the distance between two arbitrary trabeculae stems, 0.04 mm (Figure 7D), or the volume of sections of the trabecular network, 2.20 * 10⁻³ mm³ (Figure 7E). Besides visualization and analysis on the original reconstruction, we were also able to manipulate our digital model by arbitrarily cutting the trabeculae for further investigations and measurements (Figure 7F). In this context, the integration of interactive image analysis with volumetric imaging holds the great potential to augment quantitative analysis and user-directed manipulation in cardiac studies.

Figure 7. Segmentation of the P1 mouse heart reveals quantitative information.

(A) Conventional volume rendering of whole P1 mouse heart indicates the atria and valves. (B) Manually segmented ventricular chamber reveals the ventricular cavity and enables us to determine its volume. (C) Customized VR demo for user-directed analysis of the trabecular network. (D) Distance between two trabeculae stems is determined using the distance tool of our VR platform. (E) A section of trabecula is cut using the cutting tool of our VR platform. (F) Volume of a section of trabecula is determined using the volume tool of our VR platform. Scale bars: (D) 10 μm and (E-F) 5 μm.

Methods and Materials

Implementation of imaging system

Our new LSM was optimized based on the previous one for better resolution and post-image processing in cardiac studies⁵. The illumination arm is composed of three continuous-wave diode-pumped solid state lasers of wavelengths 473 nm, 532 nm, and 589 nm (LRS-0473-PFM-00100-05, LRS-0532-PFM-00100-05, LRS-0589-GFF-00100-05, Laserglow Technologies, Canada). The three laser beams are aligned, and then expanded by a 10x achromatic beam expander (GBE10-A, Thorlabs). An iris (ID20/M, Thorlabs) is used to shape the beam, and the resultant beam is focused into one dimension after passing through a plano-convex cylindrical lens (f = 50 mm, ACY254-050-A, Thorlabs). A group of achromatic doublets serve as the relay lens (f₁ = 100 mm, and f₂ = 60 mm) and an objective lens (Plan Fluor 10X/0.30, Nikon) is used to reshape the beam into a vertical light sheet. Inspired by mesoSPIM¹¹, we installed the detection arm horizontally for ease of multiview implementation. The detection module is composed of an objective lens (MV PLAPO 1X/0.25, Olympus), Zoom body (MVX-ZB10, Olympus), a filter wheel (LB10-W32, Sutter) with a set of filters (ET525/30, ET585-40, ET645-75, Chroma), MVX10 tube lens (Olympus), and an sCMOS camera (Flash 4.0 v3, Hamamatsu) installed orthogonal to the illumination plane. The zoom body enables us to capture images with magnifications ranging from 0.63X to 6.3X. Four Physik Instrumente (PI) motors are used to position the heart, and focus the detection module. During imaging, the heart is mounted in a 3D printed chamber filled up with DBE for refractive index matching (Figure 1C). The illumination and detection modules are controlled by a computer with a dedicated Solid State Drive Redundant Array of Independent Disks level 0 (SSD RAID 0) storage for fast data streaming.

Data acquisition and image post-processing

The axial resolution of the LSM system is determined by the waist of the light sheet which is defined as 2ω₀ = 2λf/πω, where λ is the emission wavelength, f is the focal length of the illumination objective lens, and ω denotes half of the width of the illumination beam before focusing. The lateral resolution of the system is determined by the NA of the detection objective lens and the emission wavelength, defined as $\frac{\lambda }{2\ast N{A}_{detection}}$. To digitally sample the PSF, our zoom body allows us to zoom in from 0.63X to 6.3X resolution, corresponding to ~10 – 1 μm digital resolution in our LSM system. The cardiac imaging data presented in this paper were captured within a camera exposure time of 50 ms, and the step size of mechanical scanning was 1 μm which is smaller than one-half of the light-sheet thickness (2.91 μm), consistent with the Nyquist-Shannon sampling theorem. Volume rendering of the acquired images was done using the Amira software.

Sample preparation

Animal protocols, experiments, and housing in this manuscript have been approved and conducted under the oversight of the University of Texas at Dallas and University of Texas Southwestern Institutional Animal Care and Use Committee. The generation of Rbpms knock-out mice was detailed in the other literature¹⁹. Animals were euthanized at P1 for whole heart collection. Freshly isolated hearts were immersed in room temperature 0.2M KCl to arrest in diastole. Hearts were then fixed in 4% PFA in PBS at 4°C with gentle agitation on a rocker for 2-3 days.

Tissue clearing: Adipo-Clear

Adipo-Clear was used to render the heart transparent for imaging (Figure 1D). After fixation, the heart was rinsed thrice with 1x Phosphate Buffered Saline (PBS) for 15 minutes each. Next, the heart was incubated with shaking at 4°C in 20%, 40%, 60%, 80%, and 100% methanol/B1n buffer (Glycine, Triton X-100, deionized water) for an hour each to dehydrate it. After dehydration, the heart was delipidated by incubating it in Dichloromethane (DCM) at 4°C three times for an hour each, followed by incubation in 80%, 60%, 40%, 20% methanol/B1n buffer at 4°C, and 100% B1n buffer at room temperature (RT) to rehydrate the heart. The heart was then washed in PTxwH (10x PBS, Triton X-100, Tween 20, Heparin, deionized water) buffer at RT for 2 hours to make it permeable. After delipidating and permeabilizing the heart, it was embedded in a 1% agarose in 1x PBS solution. Next, it was dehydrated in 25%, 50%, 75%, and 100% methanol at RT for an hour each, followed by incubation in 100% DCM at RT three times for an hour each. Lastly, the heart was incubated in DBE till it turned transparent⁶.

PSF calibration and multiview deconvolution

The fluorescent beads were diluted to a concentration of 1:150,000 using a solution of 0.8% Agarose+ RIMS (Refractive index matching solution: Histodenz+DI water). The diluted beads were mounted in a glass tube and immersed in a glycerin-filled chamber for imaging. All raw images from the multiple views were imported into the BigSticher plugin in ImageJ for deconvolution, registered using the beads as a point of interest, aligned and fused¹⁰. The averaged PSFs of each view were used to estimate the overall PSF of the system. This estimated PSF was then used to deconvolve the dataset into a single stack of images following the Efficient Bayesian-Based Multiview deconvolution approach^17,18. For registration and to preserve fluorescence of beads in the multiview imaging of the mouse heart, we adapted the protocol^18,20 to embed beads in the resin, together with the cleared heart in the same holder. The resin was prepared by mixing D.E.R. 332 (bisphenol-A diglycidyl ether), IPDA (isophorone diamine, 5-amino-1,3,3-trimethylcyclohexanemethylamine), and D.E.R 736 (polypropylene glycol diglycidyl ether) at a volume ratio of 4:1:1. Fluorescent beads were extracted from their storage solution by centrifugation and mixed in with the curing reagent (IPDA) before the addition of the flexibilizer (D.E.R. 736) and epoxy resin (D.E.R. 332). The solution was degassed in a vacuum chamber for 3 hours and poured into a silicone mold before being left to cure for 2-3 days. Our workstation for multiview deconvolution and fusion contains an Intel (R) Xeon (R) Gold 6258R CPU @ 2.69 GHz, 2 processors, a NVIDIA RTX A6000 GPU, and 256 GB of RAM. Multiview deconvolution of the P1 mouse heart presented in Figure 6 required 20-30 hours of processing and consumed about 60GB of memory.

Image stitching

We also used the BigStitcher plugin in ImageJ for image stitching¹⁰. All the raw images were combined and imported into BigStitcher. After pre-processing, the tiles were retrieved and manually annotated, followed by the assessment of image quality using the Fourier Ring Correlation across datasets. To stitch multiple files, we aligned them with a translation model following pairwise shift calculation, filtering pairwise shifts, and global optimization. Once done, all aligned tiles were merged into one image. Our Image stitching workstation contains an Intel (R) Xeon (R) W-2245 CPU @ 3.90 GHz, a NVIDIA Quadro RTX 4000 GPU, and 64 GB of RAM. Image stitching of the P1 mouse heart presented in Figure 5 required 3-4 hours of processing and the resultant data consumed 15 GB of memory.

VR procedure

The LSM imaging data was imported into 3D Slicer, an open-source platform for image analysis and visualization, and segmented using image processing and manual tools, such as threshold, island, paint, draw, and smoothing. A surface mesh was produced from the segmentation data, stored as a .obj file, and imported into Maya 2020 for mesh correction, where we could manipulate the surface normal vectors and triangles of the model. The mesh was duplicated, and its duplicate reversed to create an opposing piece. These opposing meshes were merged to create a model with both outer and interior normal vectors. The resultant model was exported as a .fbx file into our VR platform made with the Unity engine, where we were able to interactively visualize and analyze the object through the Oculus Quest II HMD and its controllers. Operations and manipulations such as cutting a section of trabeculae and measuring the distance between two trabeculae stems were initiated by the controllers¹³. The mouse heart that is visualized as a VR model in Figure 7 was stored as a 1.36 GB .obj file. Importing this model into Unity, the 3D development engine, took approximately 6 minutes. Our VR desktop contains an Intel i9 CPU @ 3.70 GHz and 10 Cores, a NVIDIA Quadro RTX 5000 GPU with 16 GB of VRAM, and 64 GB of RAM. The upper limit of data amount processed by the VR platform depends on the number of polygon meshes of the model, the interactive operations created for the analysis, and the computational power of the workstation.

Data availability

The data sets generated during and/or analyzed during the current study are available from the corresponding author.

Code availability

The computer code generated during and/or analyzed during the current study are available from the corresponding author.

Statistics

Data are presented as mean ± standard deviation.

Conclusion and Discussion

LSM systems have been demonstrated to foster cardiac studies due to the minimal risk of photobleaching and phototoxicity, high spatial resolution, and optical sectioning as opposed to other optical imaging methods^4,21-27. Thus, to better understand the 3D myocardial architecture during cardiac development, injury, and regeneration, we have established a holistic strategy including imaging system construction and control, Adipo-Clear tissue clearing, and subsequent quantitative analysis in both conventional and user-directed VR modes. Our method allows us to investigate the ventricles, atria, aortic arch, and valve leaflets in the intact P1 heart within minutes, bypassing mechanical slicing. The subsequent user-directed analysis using VR further provides an entry point for the user to interact with the digital model with more freedom and unlimited perspectives, allowing for the in-depth investigation of the 3D model in its natural form. The successful implementation of this strategy further paves the way for other studies in characterization of the contribution of multiple highly specialized lineages and numerous signaling pathways to myocardial trabeculation and compaction.

The imaging system is equipped with translational stages to position the sample along three axes to enable quick positioning of the heart within the confocal region of the light sheet. In comparison to other laser-scanning designs^28-29, our scanning method is easier for cleared samples as it obviates the need to synchronize the scanning beam with the detection module by moving the sample instead. The 3D printed sample chamber is made from Nylon which is resistant to the DBE solution used for refractive index matching. Cover glasses with a thickness of 0.13 - 0.17 mm are used as the chamber windows (Figure 1C), minimizing the artifacts and refractive index mismatch. The width of the cover glass, 24 mm, is wide enough to enable the unobstructed capturing of fluorescence emitted from the sample by the zoom body, and to enable us to position the sample at the focus of the detection and illumination lenses. The system also makes use of magnetic mounts for quick sample and chamber exchange. Additionally, we implemented Adipo-Clear tissue clearing technique to render the P1 mouse heart transparent. Adipo-Clear⁶ is a variation of the immunolabeling-enabled imaging of solvent-cleared organs (iDISCO) that ensures complete delipidation while preserving tissue morphology. Originally intended for clearing adipose tissue, this versatile and robust tissue-clearing technique has proven its effectiveness in rendering neonatal mouse hearts transparent. Additionally, the use of Adipo-clear is likely responsible for the expanded ventricular cavities of the cleared heart, revealing the intricate trabecular network within it. The cardiac morphology was also retained contrary to hearts cleared with the CLARITY tissue clearing protocol, which increased in size and fragility.

While the actual lateral and axial resolution of the imaging system is 2.30 μm and 2.91 μm at the thinnest part of light sheet, respectively (Figure 2A-B), the tradeoff between the waist of the Gaussian beam and the confocal region resulted in anisotropic resolution of the P1 mouse heart (Figure 4). Additionally, to capture the whole heart within the FOV, we changed the magnification ratio of the zoom body from 6.3X to 3.2X, leading to the degraded digital resolution from 1 to 2 μm, and therefore the degraded average resolution of 4.68 μm (lateral) and 5.04 μm (axial) for the multiview deconvolution of mouse heart (Figure 6 and Supplementary Figure S3). However, our combination of multiview deconvolution and image stitching allows us to investigate the trabecular network across the entire neonatal mouse heart using three representative tiles (Figure 6). The multiview results of fluorescent beads indicate the feasibility of this system to generate an isotropic resolution of ~1.9 μm over the FOV of hundreds of microns, proving the improvement of axial resolution by 7.5-fold in comparison to the raw data of 15 μm (Figure 3).

To further improve the efficiency and accuracy of interactive analysis, we have created machine learning-based approaches for the assessment of the ventricular myocardium^30-31. Our interpretable neural network model, consisting of a VGG-19 network with gradient-weighted class activation mapping (Grad-CAM) and uniform manifold approximation and projection (UMAP) embedding methods, enables us to accurately extract and visualize features from a region of interest in our imaging data without explicitly defining them. This model can be used for the quantification of the trabeculae within the ventricles to assess the differences between the P1 Rbpms knockout (KO) and the wild-type littermate mouse hearts in our future work.

This system is a starting point for the LSM imaging of injury and regeneration in the neonatal mouse hearts. In mammalian models, our method provides a powerful strategy to trace the distribution of highly specialized myocardial lineages across the intact heart, allowing us to study cardiomyocytes that differentiate from numerous progenitors by taking advantage of the multichannel and large-scale capacities of the LSM system. LSM-generated images of organ-specific stem cell differentiation allow a more detailed understanding of cardiac development and regeneration. When exported into our custom VR platform, these images can be used to develop an interactive model to explore the distribution of cardiomyocytes and their progenitors within the heart. To further optimize our strategy for such studies, the techniques of remote focusing^27,32 and PEGASOS³³ will be implemented for the system hardware and tissue clearing, respectively. Their inclusion will offer the advantages of submicron resolution across an extended FOV and preservation of endogenous fluorescence. The use of remote focusing is complementary to image stitching for the larger specimen, however, the longer acquisition time increases the risk of photodamage. On the other hand, multiview enables us to image thick and scattering samples with balanced high throughput and perspective views, and the combination of image stitching and multiview deconvolution methods allows us to cover the large FOV with an enhanced axial resolution, image contrast, and acquisition rate. However, it is limited by the high computational cost. Due to the advantages and disadvantages of these methods, we seek to establish a pipeline for the imaging of neonatal mouse hearts based on the biological applications, sample size, transparency quality, computational resources, and guidelines³⁴. Overall, our framework will bring advanced imaging, tissue clearing, and interactive visualization and analysis methods to fundamental studies of congenital cardiac anomalies and diseases, being instrumental to uncovering the mechanism of cardiac morphogenesis and addressing controversial issues in heart development and regeneration.