Introduction

Recently, life sciences have focused on comprehending the working principles of multicellular systems, spanning from basic biology to medical applications (1–5). Multicellular systems are generally complex, comprising heterogeneous cells rather than homogeneous assemblies. To understand the mechanism by which numerous cells cooperate to express the function of the entire system, it is ideal to observe all individual components within the system simultaneously during the time span of a whole phenomenon. In particular, in scenarios where a small fraction of cells in a large multicellular system may influence the fate of the population, or when distant cells and tissues operate synchronously, deducing the entire system from observations of sub-populations becomes challenging. Therefore, the trans-scale measurement of collective dynamics of all individual cells that constitute the system of interest is desired.

In response to this demand, researchers have recently reported the development of imaging methods with large field-of-view (FOV) (6–11). Whereas most of these methods are focused on observing brain activity in neuroscience using two-photon excited fluorescence (7–10), our efforts have been dedicated to developing techniques as versatile tools for studying multicellular systems science and developmental biology. In our previous work, we proposed a fluorescence imaging system for the visible wavelength region, capable of spatially resolving individual cells within a centimeter FOV, which we named AMATERAS-1 (A Multi-scale/modal Analytical Tool for Every Rare Activity in Singularity), enabling simultaneous observation of dynamics in the range of 10⁵–10⁶ cells (11). Additionally, we successfully detected less than 1% of cells with unique roles in multicellular systems and revealed how these rare cells trigger a drastic transition from unicellular to multicellular behavior in Dictyostelium discoideum (12). However, the imaging system had a numerical aperture (NA) of 0.12, which is insufficient for realizing volumetric observation owing to its broad depth of field (DOF). Overcoming this limitation and enabling volumetric imaging for diverse three-dimensional (3D) tissues and small organisms has been a significant challenge.

In this study, we developed AMATERAS-2, a volumetric optical imaging system with approximately 15 × 10 mm² FOV, equivalent to that of AMATERAS-1 (11). The key component for the enhancement is the giant 2× lens system with an NA of 0.25, which offers improved 3D spatial resolution and higher sensitivity. By incorporating novel methodologies of optical sectioning and computational sectioning, we successfully added volumetric imaging capabilities. The effectiveness of these advancements was demonstrated in imaging 1.5-mm-thick brain section blocks and conducting a 25-h time-lapse observation of vascular endothelial cells during quail embryogenesis, allowing us to trace the spatiotemporal dynamics of approximately 4.0 × 10⁵ cells in three dimensions.

Result

Optical configuration and performance

AMATERAS-1 utilized a telecentric macro lens with 2× magnification, thus allowing dynamic observation in the centimeter-FOV with subcellular resolution (11). Despite its efficacy in biological studies, its low NA (0.12) limits both its spatial resolution in the z-direction, and system brightness. To overcome these constraints and expand the possibilities for observing cell dynamics in tissues and embryos, we developed a giant lens system with a higher NA, consisting of a pair of objective and tube lenses. The objective lens is composed of 12 optical elements in 7 groups, while the tube lens is composed of 9 lenses in 6 groups (Fig. 1A). Both lenses are infinity-corrected with NAs of 0.25 for the objective and 0.125 for the tube lens, and have a field number of 44 mm (diameter) suitable for large-area image sensors. They are effectively aberration-corrected to the diffraction limit or lower in the visible wavelength range of 436–656 nm. The objective lens and tube lens have sizes of 144 mm × 310 mm and 170 mm × 345 mm, respectively (Fig. 1B). The objective and tube lenses exhibit zero vignetting, enabling observation of peripheral areas of the FOV with high brightness.

Fig. 1.

As the NA of AMATERAS-2 is 0.25, more than double that of AMATERAS-1 (0.12), our system in principle achieves over twice the spatial resolution in the transverse (xy) direction, more than four times the resolution in the longitudinal (z) direction, and over four times the brightness of AMATERAS-1.

Using this new lens system, we constructed a wide-field epi-illumination fluorescence imaging system, termed AMATERAS-2w, with a magnification of 2×, an NA of 0.25, and a field number of 44 mm (Fig. 1C, Materials and Methods). We utilized either a 120-megapixel CMOS camera or a 250-megapixel CMOS camera, depending on the research purpose. Both the cameras have an approximately 35 mm diagonal, thus providing an observation FOV of 17.8 mm and 17.4 mm diagonal at 2× magnification, respectively. The pixel sizes for these cameras are 2.2 and 1.5 µm, with sampling intervals of 1.1 and 0.75 µm at 2× magnification, respectively.

We installed a high-brightness LED light sources for fluorescence excitation, directing it into the objective lens through a custom-made long-pass dichroic mirror designed for GFP imaging with a cut-off wavelength of 497 nm. To ensure uniform illumination across the entire FOV, we employed a lens-let array pair. Additionally, we incorporated a 2-inch band-pass fluorescence filter between the tube lens and the camera.

We experimentally obtained the optical point spread function (PSF) using green fluorescent beads (center wavelength: 515 nm) with a 0.2-µm diameter, employing both the 120-megapixel CMOS camera (Fig. 2A) and the 250-megapixel CMOS camera (Fig. 2B). The left panels in Figs. 2A and B display the PSFs in the xy and xz planes, respectively. The single-point spatial resolution was evaluated by the full-width of half-maximum (FWHM) of the line profiles determined through Gaussian-function fit. The transverse resolutions (xy-direction) obtained with the 120-and 250-megapixel cameras were 1.24 ± 0.12 (N = 100) and 1.12 ± 0.072 (N = 100) (Figs. 2A-B, right top), both of which are slightly downgraded from the theoretical FWHM,

Fig. 2.

1.05 µm, given by 0.51 λ_em/NA (λ_em = 515 nm, NA = 0.25) (13,14). This can be attributed to the coarse sampling interval (1.1 µm and 0.75 µm, respectively). In particular, in the 120-megapixel case, the apparent single-point resolution varies depending on the relative position of the bead within the pixel (Figs. S1A-B). The coarse sampling also introduces uncertainty in the estimated position of the bead’s center coordinate (Fig. S1E). Although this may pose some challenges in quantitatively evaluating fine shapes in cells, it is not a practical issue in imaging at cell resolution. By contrast, the 250-megapixel camera provides a nearly constant resolution (Figs. S1C-D). Thus, the 250-megapixel camera is suitable when high-resolution images are required, whereas the 120-megapixel camera is suitable for other cases where increasing the number of photons per pixel or suppressing data size is necessary.

In the depth direction (z-direction), we evaluated the DOF by fitting Gaussian curves to the line profiles (Figs. 2A-B, right bottom). The DOF was measured to be 15.6 ± 1.3 (N = 100) and 14.8 ± 1.1 (N = 100) for the 120-and 250-megapixel cameras, respectively. This difference in DOF is attributed to the distinct pixel sizes of the cameras (15). The DOF value for the 250-megapixel camera was found close to the theoretical FWHM, 14.3 µm, given by 1.78 λ_em/NA² (λ_em = 515 nm, NA = 0.25) (14).

The spatial resolution in the transverse direction has been significantly improved, approximately two times compared with that of AMATERAS-1. Figures 2C and D display fluorescence images of cardiomyocytes derived from human induced pluripotent stem cells (hiPSCs) stained with Rhodamine phalloidin (C) across the entire FOV and (D) within a central local region indicated by a magenta square in (C). These images were captured using the 250-megapixel camera, and the FOV size was 14.7 × 9.4 mm². For comparison, Fig. 2E shows an image of the same area observed with AMATERAS-1. Evidently, AMATERAS-2w delivers more detailed images with finer spatial structures. This is further confirmed by the line profiles depicted in Fig. 2F, where the filamentous actin stained with Rhodamine exhibits a sharper spatial resolution with AMATERAS-2w. This result indicated the capability of observing subcellular structures, such as myofibrils, in cell sheets with a large area, such as artificial myocardial sheets, which would enable us to simultaneously investigate microscale structures and macroscale multicellular dynamics.

Spinning pinhole-array disk provides optical sectioning capability in wide-FOV

This imaging system utilizes wide-field illumination and detection, which currently lacks the capability to selectively acquire images of specific z-planes for volumetric imaging. To extend the system’s applicability, we developed two methods to enable the volumetric imaging, namely, optical sectioning and computational sectioning.

For optical sectioning, various options of optical techniques exist for achieving 3D fluorescence imaging with a large FOV in the visible wavelength region. These include scanning confocal microscopy (6,13), light-sheet microscopy (4,16), light-field microscopy (17), two-photon excitation microscopy (7–10,18), and spatiotemporal focusing (19). Among these options, we selected the confocal imaging method for this study. This is because the light-sheet microscopy struggles with uniform illumination over the centimeter scale FOV of AMATERAS (16,20), and the light-field microscopy involves a trade-off between spatial resolution and the ability to resolve three dimensions, which is incompatible with the AMATERAS’s concept (21). In addition, two-photon excitation microscopy, including that using spatiotemporal focusing, is crucial for deep tissue imaging, but still requires further innovations in pulsed laser power and scanning methods to be effectively applied to the wide FOV of AMATERAS.

We considered the confocal method to be the most compatible with our imaging system. Considering the vast FOV, we adopted the multipoint-scanning confocal system. The commonly used method involving a combination of a microlens array and a pinhole array disk (22) could not be employed in AMATERAS owing to the large NA of the tube lens (NA = 0.125) compared with that of standard microscopes. Matching the NA requires shortening the distance between the two disks or enlarging the microlenses. However, shortening the inter-disk distance leads to a loss of space for inserting a dichroic mirror, and enlarging the microlenses reduces the number of foci. While confocal reflection (non-fluorescence) microscopy does not require a beam splitter, allowing for close microlens-pinhole spacing (23), the confocal fluorescence microscopy requires a beam splitter due to the essential separation of fluorescence and excitation light. Therefore, designing microlens-pinhole arrays suitable for AMATERAS-2 presents a significant challenge. This challenge can be circumvented by using an optical configuration where excitation light and fluorescence pass through synchronized pinhole arrays on separate paths, as initially proposed in the early development of multipoint confocal microscopy (24). However, this solution would require duplicating the optics of the giant lens, which is currently unfeasible. Moreover, even if possible, the added complexity of the system would render this approach impractical.

To address this challenge, we opted to use pinhole arrays without microlenses, which is rather a more traditional configuration of multipoint confocal microscopy (13,25,26). Although this method is less light-efficient compared with the one using microlenses, it aligns well with our imaging system’s high NA and low magnification. Here, we designed the pinhole array disk for fluorescence imaging in green (e.g., GFP, YFP) and red (e.g. RFP, mCherry) wavelengths. Commercial confocal systems using pinhole-array disks typically feature large pinhole sizes (e.g., 50 µm) and wide spacing between pinholes (e.g., 200 µm), designed for general microscopes. For AMATERAS-2, however, with its 2× magnification and NA of 0.25, the pinhole size must be much smaller. The diameters of the Airy disk for green and red wavelengths are about 5 µm and 6 µm, respectively, at the focal plane of the tube lens (NA = 0.125). As a prototype, we set the pinhole size to 6 µm (Fig. 3A), matching the Airy disk diameter for red wavelength. A more detailed discussion on the impact of pinhole diameter dependence on 3D resolution is available in the Supplementary File (Note 1, Fig. S2). The pinhole spacing was set as 24 µm, striking a balance between throughput and crosstalk (Fig. 3A). To implement this, we custom-made a pinhole-array disk and incorporated it into a rotary machine (CrestOptics, Rome, Italy).

Fig. 3.

We positioned the pinhole array disk at the image plane of the tube lens and constructed a relay lens system to transfer the plane to the camera. This particular configuration, termed AMATERAS-2c, is illustrated in Fig. 3B (Materials and Methods). To split the light paths of fluorescence excitation light, we placed a short-pass dichroic mirror beneath the disk. For fluorescence excitation, we used a high-brightness LED with a center wavelength of 470 nm, the same as in the wide-field imaging system (Fig. 1C). The relay lens consisted of a telecentric macro lens with 1× magnification, and a 2-inch band-pass fluorescent filter was positioned at the entrance of the macro lens. If necessary, the relay lens can be replaced by a 2× magnification lens to switch the total magnification from 2× to 4×, based on resolution requirements.

Fluorescence images of 0.5 µm fluorescent beads were observed with and without the pinhole-array disk, where fluorescent beads were three-dimensionally dispersed in an agarose gel slab (Fig. 3C). The presence of the pinhole-array disk dramatically suppressed background light from outside the focal plane. This verified that optical sectioning can be achieved by the pinhole-array disk. The variations in the 3D PSF with and without the pinhole-array disk are detailed in the Supplementary File (Fig. S3). Typical 3D PSF was obtained with the fluorescence beads with the relay lens of (D) 1× and (E) 2× magnifications (Figs. 3D and E). As NA of the 1× relay lens (NA = 0.079) is lower than that of the tube lens (NA = 0.125), the transverse spatial resolution (Fig. 3D, right top) is degraded compared to the one obtained by the wide-field configuration without the relay lens (Figs. 2A and B, Note 1). By contrast, the 2× relay lens (NA = 0.12) almost matches with the tube lens, and the reduction in the transverse spatial resolution is relatively small (Fig. 3E). As for the longitudinal resolution, the FWHM of the PSF was approximately 16 µm with a 1× relay lens and approximately 14 µm with a 2× relay lens, both values significantly exceeding the ideal value about 9.5 µm (Note 1). These discrepancies are attributed to the relatively loose focusing of the excitation light and the non-negligible size of the pinhole.

Computational sectioning provides pseudo-depth-resolved imaging capability

In addition to optical sectioning, we developed a computational sectioning method to separate images of the focal plane and out-of-focal planes. During measurement, a z-stack is acquired in the wide-field imaging configuration (Fig. 1C) by scanning the focal position in the z-direction and capturing images at each step. The raw image in a 3D volume results from the superposition of in-focus and out-of-focus plane images. We estimated the contribution from out-of-focus planes as the baseline by iteratively applying low-pass filtering under the assumption that the image’s spatial frequency in the focal plane is higher than that in out-of-focus planes. The cut-off frequency of the low-pass filter was optimized using the principles of independent component analysis (Materials and Methods, Note 2, Fig. S4-S5).

To demonstrate computational sectioning, we observed myocardial organoids derived from hiPSCs (27). This 3D dome structure of a cavity chamber is extensively studied in developmental biology and regenerative medicine as a model for human myocardial tissue development (28,29). We fabricated a large area of organoids across the entire FOV, chemically fixed and immunostained them for cardiac troponin T (cTnT) with Alexa488, and stained the nuclei with Hoechst33342 (Materials and Methods). The z-stack was obtained within a z-range involving the organoids (approximately 250 µm) with 4-µm intervals. Figures 4A and B display two-color overlaid images of the island-formed organoids across the entire FOV, without and with computational sectioning, respectively (both are maximum-intensity projection (MIP) images of the z-stacks). Figures 4C and D present images of six layers in the z-direction of the yellow square region of Figs. 4A and B, before and after computational sectioning, respectively. After the sectioning, the distribution of cTnT and nuclei in each z-layer is clearly visualized (Fig. 4D): The island areas are composed of multilayered cells, and the inter-island spaces are covered with a single layer of cells, as shown in Fig. 4D, z = 20 µm. A 3D isosurface representation (Fig. 4E) shows that the hollow oval structure is formed by a thin layer of cont.-positive cells (cardiomyocytes) whereas the underlying cell layer is composed of cTnT-negative cells, consistent with previous literature (27). This demonstration validates the technique for wide-field imaging. Video files of Figs. 4C–E are available in the Supplementary files (Movies S1 and S2).

Fig. 4.

In principle, the signal component can be extracted as long as the signal intensity (in-focus component) is significantly higher than the statistical noise of the background intensity (out-of-focus component). This method is effective in cellular imaging when fluorescent molecules are localized within the cell, like in the nucleus, or when they exhibit a filament-like structure, thereby enabling cell recognition and dynamic tracking. It relies on the spatial frequency of the fluorescence image in the focal plane being uniform and clearly higher than that outside the focal plane.

Observing a 1.2-cm wide and 1.5-mm thick volume of mouse brain section

To demonstrate the potential for brain imaging, we performed imaging of a 1.5-mm thick mouse brain section in the coronal plane (Fig. 5A). The brain was chemically cleared using CUBIC (30,31), and the cell nuclei were stained with SYTOX-Green (Materials and Methods). The 1× relay lens was primarily employed (total magnification = 2×) to cover the coronal plane in the FOV. Each layer’s exposure time was 1 s, and a total of 378 layers were acquired to cover the entire 1.5-mm-thick volume in steps of 4 µm in the z-direction. Optical and computational sectioning were both applied. Figure 5B shows a fluorescence image at a single z-layer (z = 700 µm), revealing the 12 mm × 8 mm wide section with single cell resolution. Figure 5C displays enlarged views of the xy-plane (coronal plane) in the dotted square region in Fig. 5B, along with xz (transverse plane) and yz (sagittal plane) cross sections, essentially demonstrating successful z-direction sectioning. Figure 5D shows a 3D representation of the same volume of raw data obtained by optical sectioning alone, prior to computational sectioning. However, owing to intense background light overlapping, individual cells are challenging to distinguish, particularly in regions of high cell density. By utilizing computational sectioning to eliminate background light, clear images were obtained even in areas with strong background light. Supplementary files contain videos of the z-scan for the entire FOV and local regions (Movies S3 and S4).

Fig. 5.

Figure 5C visualizes the characteristic 3D structures of several regions, including hippocampal dentate gyrus, medial habenula, and choroid plexus, which are known to be associated with memory formation (32), depression (33), and regulation of cerebrospinal fluid (34), respectively. Close-up views of the hippocampal region (indicated with the dashed square in Fig. 5B) in the xy and xz planes are shown in Figs. 5E and F, respectively. A comparison of the images at 2×and 4× magnification in Figs. 5E and F reveals that the higher magnification (4×) allows for improved spatial resolution and separation of individual cells in both the xy and xz planes.

For cell detection in the 3D imaging data of the mouse brain section, we used a recently developed interactive platform for cell detection and tracking, called ELEPHANT (Efficient LEarning using sParse Human Annotations for Nuclear Tracking) (35). This platform seamlessly integrates manual annotation with deep learning and allows for result proofreading (Materials and Methods). Using this advanced image analysis technique, cell nuclei were detected within the 3D volume. Figure 5G illustrates the cell detection results on specific xy and xz planes within the cortex. As the cell density in the cortex is relatively low, the ellipsoid-shaped nuclei image does not overlap with neighboring cells, resulting in the successful detection of nearly all cells (as clearly shown in the z-scan video, Movie S5). The precision and recall of the cell detection were estimated to be 99.4 % and 97.6 %, respectively (Materials and Methods). Based on these results, the cell density in the cortex was calculated to be 1.18 × 10⁵ cells/mm³, which closely matches previous reports (36). This technique was also applied to two other brain regions, the medial habenula and choroid plexus (Figs 5H-I). In these regions, cell density is higher than in the cortex and the signal-to-noise ratio is lower due to strong background intensity, making it difficult for traditional methods to detect cells accurately. However, with the assistance of ELEPHANT, we successfully detected the majority of cells.

Dynamics of over a 4.0 x 10⁵ cells were captured over 24-hours development of a quail embryo

We employed the imaging methods described for a time-lapse observation of cell migration during quail development. The non-confocal optical configuration (AMATERAS-2w, Fig. 1C) and utilizing computational sectioning alone were sufficient for this 250-µm thick sample. We used a tie1:H2B-eYFP transgenic quail embryo (37) in which enhanced yellow fluorescent protein (eYFP) visualizes the nuclei of vascular endothelial cells (Materials and Methods). The embryo was cultured ex ovo on a 35 mm glass-bottom dish (Fig. 6A). Time-lapse fluorescence imaging began at 36 h (HH10) after the start of egg incubation and captured eYFP signals in the developing embryo over 24 h. For 3D imaging, we acquired z-stacks every 7.5 min, with each z-stack containing 21 layers spaced 12 µm apart. This resulted in a total of 4200 image layers and 200 time-points, with a data size of approximately 500 GB. During the time-lapse observation, we stabilized the relative distance between the lens and the sample using a self-developed autofocus system (Materials and Methods).

Fig. 6.

Representative images of cell distribution based on the nuclear eYFP signals at four time-points (t = 0, 8, 16, and 24 h) are shown in Figs. 6B and C. These images were reconstructed using the HSB color model to represent the 3D distribution after the baseline was removed by computational sectioning (Fig. S6). The brightness indicates the intensity of the MIP image in the z-direction, whereas the hue represents the z-position with the maximum value in the MIP process (Materials and Methods). As observed in Figs. 6B and C, the cell nuclei distribution undergoes significant changes over time, thus resulting in the formation of organs such as the ventral aortae, heart, and dorsal aortae (t = 24 h, Fig. 6C). The heart region appears blurred owing to its oscillating shape caused by beating. Supplementary files include videos of the HSB-color images in the entire FOV and magnified local regions (Movies S6–S8). Figure 6D shows enlarged views of the ventral aortae (indicated by the left dashed square in Fig. 6C) at the four time-points. Additionally, a 3D isosurface of the z-stack of the dorsal aorta region (the right dashed square in Fig. 6C) was calculated (Fig. 6E, and Movie S9), clearly showing the developmental process of the two tube structures of the dorsal aortae (38). Note that the upper vessel wall is partially missing in the second half of the movie because the fluorescent image was disturbed by the blood flow that began around t = 15 h.

At the current spatiotemporal resolution, we successfully traced the movement of individual cells. Segmentation of cell nuclei allowed us to detect approximately 4.5 × 10⁶ cells across all 200 frames, and by linking cells in close proximity over consecutive frames, we tracked about 4.0 × 10⁵ cells. Figures 6F and G show cellular movement along with a feature parameter related to cell movement at t = 8 h and 16 h. Both left panels are magnified views of the square regions in the right panels, with trajectories drawn on MIP images. By analyzing the cell segmentation and tracking results, various feature parameters related to cell nucleus morphology and dynamics were computed, including nuclear size, brightness, aspect ratio, velocity, acceleration, mean square displacement, and meandering index (39,40). Here, we used the meandering index, a measure representing the straightness of cellular movement, to color the trajectories in Figs. 6F and G with the rainbow color table. We found that the distribution of high meandering index at t = 8 h and 16 h was significantly changed. Further detailed analysis and biological interpretation employing the cell-tracking results will be discussed in a future paper.

Discussion

In this study, we presented AMATERAS-2w, a large FOV fluorescence imaging system, along with its variant AMATERAS-2c. The system is built around a newly designed giant lens with an NA of 0.25, magnification of 2×, and field number of 44 mm. This innovative lens configuration allows for imaging with an impressive spatial resolution of approximately 1.1 μm within a centimeter-scale FOV (15 × 10 mm²). Notably, this FOV-to-spatial resolution ratio surpasses those of previously reported large-FOV imaging methods. The key metric "optical invariant”, derived from the product of NA and FOV, has a value of 2.75 (Materials and Methods, 41), thus indicating excellent performance of the lens system. Furthermore, the "space-bandwidth product”, which signifies the FOV to spatial resolution ratio achieved under real measurement conditions (considering image sensor size and wavelength), reaches 4.3 × 10⁸ (Materials and Methods, 42). Both of these metrics significantly outperform those reported in earlier studies on large FOV imaging methods (6–10,41). Although these metrics are essential indicators of our system’s performance, they do not solely determine its superiority over other large-FOV imaging systems. Variations in biological targets and additional optical performance measures (such as frame rate and imaging depth) must also be considered. Our primary focus in designing this system was to expand the FOV, i.e., maximize the number of observable cells at any instant, even at the cost of intracellular spatial resolution. Thus, in this aspect, our system excels as an instrument for such purposes.

The spatial resolution (FWHM) of AMATERAS-2w and -2c was measured to be approximately 1.1–1.8 μm in the transverse direction (xy) and 13–16 μm in the longitudinal direction (z) through PSF measurements (Figs. 2A, 2B, 3D, and 3E). These values represent a significant enhancement compared with the previous AMATERAS version (2.3 and 64 µm, respectively) (11). The improved transverse resolution enables clearer visualization of intracellular filamentous structures (Fig. 2F). Moreover, the refinement in longitudinal resolution was more than four times greater, thus demonstrating that the expected effect was achieved with more than double NA. Depending on the object of observation, it may be debatable whether this anisotropic spatial resolution elongated along the optical axis qualifies "cellular resolution”. Although this may be insufficient to resolve dense cell populations, it proves adequate for cells with only stained nuclei or those sparsely distributed, as observed in Figs. 5F-I, and 6E. Using modern cell detection methods compatible with 3D big data, such as ELEPHANT employed in this study (35), we can achieve accurate cell detection, counting and tracking to some extent, even without clear spatial separation in three dimensions. It goes without saying that the higher spatial resolution in raw image data would enhance the performance of subsequent image analysis. The next step in optical development involves further enhancing the NA of the lens system, including the relay lens, to further improve volumetric resolution and expand the range of observable biological objects.

We successfully achieved volumetric imaging of a 1.5-mm-thick mouse brain coronal section using the multipoint-scanning confocal imaging method, along with computational sectioning and tissue-clearing (Fig. 5). This thickness corresponds to about one-tenth of the whole brain. The exposure time per image was 1 s, and data acquisition took less than 10 min. The ultimate goal is to apply this approach to enable whole-brain imaging. Unlike previous methods such as light-sheet imaging (30) and block-face serial microscopy tomography (43), which require half a day or more for whole-brain data acquisition owing to tiling images from a narrow FOV, our method can significantly reduce the acquisition time when extended to the whole brain. This feature is effective not only for the brain imaging but also for imaging of other organs and whole body of an organism. We also plan to apply it to live imaging of highly transparent tissues such as zebrafish (44) and organoids (45). Nevertheless, as the thickness increases, challenges such as increased background light and larger spherical aberration arise, thus resulting in weaker focal plane images. To address these issues and realize whole-brain imaging, we propose the incorporation of a mechanism for compensating spherical aberrations in the future. Additionally, we aim to improve the pinhole array pattern design and light source intensity to enhance transmittance and increase fluorescence intensity from samples, broadening the scope of applications.

We have shown that AMATERAS-2 can perform time-lapse observation of the centimeter-sized vascular networks in the quail embryo. It is a promising tool for understanding the multicellular behavior, especially the mechanisms of tissue formation during development (1–3). In regard to the quail embryo’s case, the data obtained here are the first demonstration of the dynamics of the extensive organization of the vascular network throughout the quail embryo at the single-cell level. Conventional methods of research employing traditional microscopy have selected a limited area for observation and thus had a difficulty in determining the distribution and synchronicity of cellular dynamics across the entire system. By analyzing such a huge four-dimensional data (xyz and t), we can gain insight into when and where each organ is formed, in what order, and in what interrelationships. In addition, employing this method allows us to break away from the conventional hypothesis-driven research manner and obtain new findings in a data-driven manner. This advantage can be realized not only for quail embryo but also for a variety of other species undergoing widespread and dynamic cellular events, making it a powerful tool in the study of multicellular organisms, especially in the study of tissue formation.

To further advance multicellular systems research, information on individual cell states is required in addition to cellular dynamics, therefore multiplex measurement is an essential technological challenge. Although most of the observations presented in this paper were made with monochromatic fluorescence imaging, multiplex imaging will lead to a more profound understanding if information on the distribution of cellular states can also be obtained by employing multiple fluorescent probes. This requires technological upgrades in optical filters and light sources for multicolor imaging. In addition, spatial transcriptomics has been rapidly developing in recent years and has become one of the most crucial tools for multicellular system study (5). In the near future, dynamics imaging and spatial transcriptomics for the same sample will be a very important technological integration to advance an integrative understanding of self-organization in tissue formation. Rather than simply using AMATERAS sequentially with existing spatial transcriptomics, it could be employed as an optical detection system for spatialtranscriptomics to achieve extremely high throughput in the same wide FOV of cellular dynamics imaging.

Finally, let us discuss a practical challenge we encountered with our instrument, primarily related to the handling of extensive image data. In our studies, the raw data for 3D imaging of the mouse brain section and time-lapse 3D imaging of the quail embryo amounted to approximately 100 and 500 GB, respectively. However, after multiple image processing and analysis steps, a single experimental measurement resulted in several terabyte of image data. Managing such vast data proves challenging for standard computers in terms of both software and hardware capabilities. To address software limitations, we opted not to employ commercial software for data analysis. Instead, we utilized our originally developed programs, implementing certain optimizations to reduce analysis time, minimize write/read cycles, and prevent memory overflows. As for hardware, we established a data server with petabyte storage, thereby enabling efficient data sharing among collaborators from diverse research institutions physically distant from one another. The existence of this infrastructure considerably facilitated the smooth progress of collaborative research within this study. Considering the increasing significance of handling image big data, we expect methodologies for image computation to become even more critical in the future. Ideally, a comprehensive system should be developed, encompassing not only data storage and sharing but also an integrated analysis solution in the cloud. We would like to actively promote the development of such a system, essentially encompassing advancements in imaging techniques and data-handling strategies.

Materials and methods

Wide-field imaging system (AMATERAS-2w) configuration

We aimed for AMATERAS to achieve a larger FOV-to-resolution ratio than conventional microscopes, which required a giant imaging lens system. We realized this design concept by collaborating with SIGMAKOKI CO., LTD. Inc. (Tokyo, Japan) to manufacture the giant objective and tube lenses as shown in Fig. 1. The characteristics of these lens are described in the main text. Designing such large lenses presents significant challenges in selecting optical glass materials compared to regular-sized lenses. Although there are over 200 types of optical glass available, our options are limited because the thickness or diameter of several types of glass may be unprocessable, or their transparency may degrade due to their size. Therefore, we iteratively explored the optimal design, considering the feasibility of material procurement and processing in conjunction with the optical design.

By combining these lenses, we constructed an imaging system with a 2× magnification, NA of 0.25, and a 44 mm field number. To facilitate both water-immersion and dry observation, we incorporated a glass plate whose thickness can be adjusted at the lens tip. The objective lens has a focal length of 120 mm, whereas its working distance is 14 mm owing to the presence of the attachment at the tip. The pupil positions of the objective lens and the tube lens are set externally to the lens bodies. This arrangement allows for the placement of a spatial filter or other devices at the pupil position, facilitating future enhancements in functionality.

For image acquisition, a 120-megapixel CMOS camera (VCC-120CXP1M, CIS, Tokyo, Japan) and a 250-megapixel CMOS camera (VCC-250CXP1M, CIS, Tokyo, Japan) were used, out of which we selected one depending on research purpose. Both the cameras have a chip size of 35 mm (diagonal). The pixel sizes of these two cameras are 2.2 µm and 1.5 µm, respectively. The sampling intervals are 1.1 µm and 0.75 µm at 2× magnification. Image data captured by the CMOS cameras are loaded into a workstation via a CoaXpress frame grabber board (APX-3664G3, Aval Data, Tokyo, Japan). After the imaging experiments, they were transferred to a network server with large storage capacity for long-term storage and data sharing among project members.

By use of the lens system and CMOS cameras, a wide-field epi-illumination fluorescence imaging system was constructed. The imaging system chassis was designed to accommodate both inverted and upright microscope geometries. For this study, all experiments were conducted using the inverted microscope arrangement. Epi-illumination was achieved using a high-brightness LED sources (SOLIS-470C and SOLIS-385C, Thorlabs, Newton, NJ), with an excitation filter (#87-800, Edmund Optics, Barrington, NJ) placed immediately after it. The illumination light was homogenized using a pair of lenslet arrays and projected onto the sample surface. Additionally, a similar LED light source (SOLIS-525C, Thorlabs, Newton, NJ) with an illumination homogenizer was also mounted as transmitted illumination light for bright-field observation. To split the light path, we used a custom-made long-pass dichroic mirror measuring 158 mm × 120 mm × 10 mm (BK7), which was designed for fluorescence imaging of GFP with a cut-off wavelength of 497 nm. A fluorescent filter (#86-992, Edmund Optics, Barrington, NJ) was positioned in front of the camera, which was firmly attached to the imaging lens using an F-mount.

The sample stage comprises a three-axis translational movable stage and a two-axis tilt stage. Only the z-axis stage is motorized, utilizing an electric actuator (SOM-C13E, SIGMAKOKI CO., LTD., Tokyo, Japan). For time-lapse observation under controlled conditions, we mounted a stage-top incubator (SV-141A, BLAST, Kawasaki, Japan) on the five-axis stage, providing a stable environment at 37 °C with CO2 control. To ensure precise measurements, we enclosed the entire sample space within a dark box, shielding it from external light and temperature fluctuations.

Multipoint confocal system (AMATERAS-2c) configuration

In the multipoint confocal imaging system, we utilized a custom-made pinhole array disk with specific dimensions (pinhole size: 6 µm, spacing: 24 µm). This disk was mounted on a high-speed rotating machine (CrestOptics, Rome, Italy) and placed precisely on the image plane of the imaging lens. Alignment throughout the FOV was ensured using a translation stage (XR25P/M, Thorlabs, Newton, NJ) and tilt stage (AIS-1016B, SIGMAKOKI CO., LTD., Tokyo, Japan) to adjust height and tilt. Beneath the disk, we positioned a short-pass dichroic mirror (DMSP490L, Edmund Optics, Barrington, NJ) to allow excitation light to be incident from below. The excitation light source included a high-brightness LED (SOLIS-470C, Thorlabs, Newton, NJ), an excitation filter (#87-800, Edmund Optics, Barrington, NJ), and a pair of lenslet arrays (#63-231, Edmund Optics, Barrington, NJ), together with a 75-mm lens (f = 200 mm, LA1353-A, Thorlabs, Newton, NJ) to ensure uniform illumination onto the disk. The fluorescent image was reflected by the dichroic mirror, transmitted through a fluorescent filter (#86-992, Edmund Optics), and projected onto a 120-megapixel camera using either a 1× or 2× relay lens (LSTL10H-F and LSTL20H-F, Myutron, Tokyo, Japan). The relay lens and camera were attached with an F-mount and secured on a three-axis translational stage (PT3/M, Thorlabs, Newton, NJ). To adjust the focus, the stage for the optical axis direction was motorized using an electric actuator (SOM-C25E, SIGMAKOKI CO., LTD., Tokyo, Japan).

Measurement of 3D point spread functions

To evaluate the optical performance of the imaging systems, the 3D point spread function was measured using green fluorescent beads. Fluorescent beads with a 0.2 µm diameter (FluoSphere F8811, Invitrogen, ThermoFisher Scientific, Waltham, MA) were employed in the evaluation of AMATERAS-2w (Figs. 2A-B) and those with a 0.5 µm diameter (FluoSphere F8813, Invitrogen, ThermoFisher Scientific, Waltham, MA) were employed in the evaluation of AMATERAS-2c (Figs. 3D-E). The beads were dispersed three-dimensionally in agarose gel on a glass-bottom dish, and z-stacks were obtained by moving the sample in the z-direction in 1 or 2 µm steps.

Computational sectioning

In wide-field fluorescence imaging of a 3D volume, the superimposition of a fluorescent image from the focal plane and background light from outside the focal plane is a fundamental issue. This problem is not limited to wide-field imaging but also arises in multipoint confocal imaging of thick or high-density samples. To overcome this challenge, we employed computational sectioning, an image processing technique aimed at removing the background light component. Various algorithms have been proposed (46,47), and Leica microscopes’ standard software includes this functionality. In this study, we developed an original algorithm as a simple and robust method for data analysis.

The critical step in our approach involves estimating the background light component at each layer of the z-stack data. To achieve this, we leverage the assumption that the background light exhibits a low spatial frequency, whereas the focal plane image demonstrates a high spatial frequency. Consequently, we employ an iterative low-pass filtering technique to estimate the background light component. Specifically, as shown in Figs. S6A and B, a two-dimensional (2D) low-pass filter is applied to the raw data (f₀(x,y)) to obtain a smooth surface (L₀(x,y)). To obtain f₁(x,y), f₀(x,y) and L₀(x,y) are compared and the smaller one is chosen to f₁(x,y) at every position (x,y), as expressed by Eq. (1).

where min(a,b) returns the smaller value of a and b. Next, the low-pass filter is applied to f₁(x,y) to obtain L₁(x,y); this process is repeated to make L_j(x,y) closer to the baseline of the original image. The iteration stops when the standard deviation of the difference of f_j(x,y) and L_j(x,y) reaches a preset value (ε).

Among various methods available as 2D low-pass filters, we sequentially applied one-dimensional (1D) infinite impulse response (IIR) filter in two directions (x,y). This method is faster than other methods such as a 2D Fourier transform (48). This speed difference is especially significant when the number of pixels in an image is massive, as in the present case. We adopted the Butterworth type of the IIR low-pass filter.

When the focal plane is empty and there is a bright object in the background, the baseline estimation tends to be skewed by the intensity fluctuations of the background. To mitigate this, we applied a 3×3 moving average filter to the entire image as a pre-processing step before initiating the baseline estimation algorithm. A critical parameter in our algorithm is the cutoff frequency of the low-pass IIR filter. If the cutoff frequency is set too high, the focal plane component would be included in the background; if it is set to low, background light would remain in the focal plane. Thus, it is crucial to find an optimal middle ground for the cutoff frequency. During this optimization, we did not account for cell size or optical system performance. Indeed, we employed a user-friendly blind separation method based on independent component analysis (ICA) to enhance usability (49). Similar to ICA, we assumed that the fluorescence image in the focal plane deviates from the Gaussian (normal) distribution and that the superposition of images from multiple planes, including both in-focus and out-of-focus planes, results in a distribution closer to the Gaussian distribution. To quantify the non-Gaussian nature of the distribution, we considered several measures, including kurtosis, skewness, negentropy, and mutual information (49). Among these measures, we found skewness to be the most robust metric for our image dataset and incorporated it into our algorithm. The cutoff frequency was adjusted to maximize the skewness of the estimated in-focus image. The optimization of the cut-off frequency was performed on a subset of the data before applying it to the entire dataset. An example of the selection of non-Gaussianity measure and the optimization of the cut-off frequency is found in the Supplementary File (Fig. S4-S5, Note 2). The typical cut-off frequency was set at 0.06 in normalized frequency unit of 1/pixel, and the value was applied across all data.

Culture of hiPSCs and differentiation into cardiac organoid with cavity chamber structure

A hiPSC line 201B7 was routinely maintained as previously described (50) on iMatrix-511 (Matrixome, 89292) coated culture dish in StemFit medium (Ajinomoto, AK02N), and 10 µM Y-27632 (Fujifilm Wako, 030-24021) was added for the first 24 h after passage. Four days before inducing differentiation, hiPSCs were dissociated with TrypLE Select Enzyme (Thermo Fisher Scientific, A1285901) and 1.9 × 10⁵ cells were passed into iMatrix-511 coated 12 well plate in StemFit medium supplemented with 10 µM Y-27632. After 24 h, the medium was changed to StemFit without Y-27632 and was exchanged daily for 3 days. Differentiation was initiated with the CDM3 medium (51) composed of RPMI-1640 with HEPES (Fujifilm Wako, 189-02145), 0.5 mg/ml human recombinant albumin (Sigma Aldrich, A9731) and 0.2 mg/ml L-ascorbic acid 2-phosphate (Sigma Aldrich, A8960) supplemented with 3 µM CHIR99021 (Fujifilm Wako, 038-23101) for 48 h. On day 2 of differentiation, the medium was changed to CDM3 supplemented with 5 µM IWP-2 (Fujifilm Wako, 034-24301). From day 4 to day 8, the cells were cultured in CDM3 and the medium was changed every other day. From day 8, the medium was changed to RPMI-1640 with HEPES supplemented with 2% B27 (Thermo Fisher Scientific, 17504044) and was changed every other day. The cells were fixed in 4% paraformaldehyde (PFA; Thermo Fisher Scientific, 43368) on day 15 for immunocytochemical analysis.

Immunocytochemical staining of cardiac organoid

Cells were fixed with 4% PFA for 15 min at room temperature (RT ∼23 °C), permeabilized with 0.2% Triton X-100 (Fujifilm Wako, 807423) for 20 min at RT and incubated with blocking solution composed of D-PBS with 5% BSA and 10 % FBS for 30 min at RT. Subsequently, the cells were incubated with primary antibody (1:200 rabbit IgG anti-cTnT, Bioss, bs-10648R) at 4 °C overnight. After washing three times, cells were incubated with secondary antibody (1:500 alexa488 conjugated goat IgG anti-rabbit IgG, Abcam, ab150077) for 2 h at RT. The cells were washed three times in D-PBS with 1 µg/ml Hoechst33342 (Dojin, H342). Subsequently, the cells were incubated with D-PBS and used for fluorescent observation.

Color representation of the 3D image

Figures 6B–D show the 3D positions of the cells using the HSB (also known as HSV) color model. After applying the method described earlier to remove background light from the z-stack image, we conducted a maximum value intensity projection along the z-direction. The z-position of the maximum value at each xy-position was then recorded (Fig. S6C). The color image was reconstructed by associating the brightness and hue of the HSB table with the value of the maximum intensity and the maximum value position, respectively (Fig. S6D). For the hue table, we assigned red to blue to z-layer numbers ranging from 0 (z = 0 µm) to 20 (z = 240 µm).

Preparation of mouse brain section

All animal care and handling procedures were conducted with approval from the Animal Care and Use Committee of Osaka University (approval numbers R02-8-7). Our utmost efforts were made to minimize the number of animals used. Experiments involved adult male C57BL/6J mice (SLC, Shizuoka, Japan) aged between 2 and 3 mo. To ensure proper anesthesia, mice were deeply anesthetized through intraperitoneal injection of an anesthetic cocktail containing 0.3 mg/kg medetomidine (Nippon Zenyaku Kogyo, Fukushima, Japan), 4 mg/kg midazolam (Sandoz Pharma, Basel, Switzerland), and 5 mg/kg butorphanol (Meiji Seika Pharma, Tokyo, Japan). Subsequently, anesthetized mice underwent transcardial perfusion with saline, followed by 4% paraformaldehyde (Nacalai Tesque, Kyoto, Japan) dissolved in phosphate-buffered saline (PBS). Brain tissues were then excised and immersed in a 4% paraformaldehyde solution until further use. For tissue preparation, brain tissue blocks were sliced into 1.5-mm thick coronal sections using a vibrating microtome (LinearSlicer Pro7N, Dosaka EM, Kyoto, Japan). Subsequent staining and tissue clearing involved immersion of sections in CUBIC-L (Tokyo Chemical Industry, Tokyo, Japan) for delipidation, SYTOX Green (Thermo Fisher Scientific, Waltham, MA, USA) solution (1:5,000) diluted in 20% (vol/vol) DMSO in PBS for nuclear staining, and CUBIC-R+(M) (Tokyo Chemical Industry, Tokyo, Japan) for refractive index matching, following previously established methods (31,52). After clearing, tissue sections were placed on glass bottom dishes and embedded in 2% (wt/vol) agarose gel prepared with CUBIC-R+(M) for subsequent imaging analysis.

Preparation of quail embryo

Transgenic quail line, tie1:H2B-eYFP (37) was bred in quail breeding facility at Kyushu University. Fertilized eggs were incubated at 38 °C. The staging of quail embryos was based on the Hamburger and Hamilton stages of chicken embryos (53). The animal study was approved by the Institutional Animal Care and Use Committee of Kyushu University (Authorization number: A20-019). Ex ovo culture was performed as previously described (54). Black filter paper was used instead of white filter paper to avoid fluorescence background.

Autofocus in time-lapse imaging

To address z-directional drift in the time-lapse observation of quail, we developed an original autofocus method. During the intervals between fluorescence z-images, we acquired bright-field z-stacks and analyzed the images to identify the most in-focus position. To achieve this, we used a black ink marker located on the substrate edge as the autofocus target instead of the variable sample itself. For evaluating the degree of in-focus, we applied a total variation (TV) filter to the marker image and used the kurtosis of the filtered image as the focus score. The z-position with the maximum focus score was identified as the in-focus location, thus signifying the sharpest image of the marker edge. Although other filter types and statistical moments are available options for the focus score (55), our preliminary experiments demonstrated the suitability of the kurtosis of TV for the marker image. In the actual experiment, we acquired 15 images in 15 µm steps and plotted the focus scores against z-positions. Through quadratic function fitting of the focus scores at three z-positions (including the z-position with the maximum value and adjacent positions), we estimated the best in-focus position with an accuracy of 1 µm.

Deep-learning based cell detection for the mouse brain data

Cell detection was conducted using ELEPHANT, a unified platform that facilitates manual annotation, deep learning and proofreading of results within a single user-friendly GUI (35,56). ELEPHANT serves as an extension to Mastodon (57), an open-source framework for large-scale tracking deployed in Fiji, a widely used image-analysis software (58). ELEPHANT supports incremental deep learning with sparse annotation, incorporating algorithms for detecting cells in 3D and tracking them in time-lapse 3D image datasets. In this paper, we utilized the cell detection function exclusively for analyzing 3D imaging data of mouse brain section (Fig. 5G-I). The incremental learning approach allows cell detection models to be trained progressively on a dataset that begins with sparse annotations and is continuously enhanced through human proofreading. For visualization, image data were displayed using BigDataViewer (59) on Mastodon, accessed via BigDataServer (60). This setup permits the visualization of large-scale image data on client computers while maintaining the data on the server. Deep learning capabilities were provided by the ELEPHANT server, enabling remote GPU access. This setup allows non-experts to perform image analysis on big data using deep learning, effectively overcoming typical challenges such as the need for large amounts of high-quality training data, the absence of an interactive user interface, and limited access to computing resources, including substantial storage and high-end GPUs (61,62). The server computer used in this study is equipped with an Intel(R) Xeon(R) Gold 6132 CPU @ 2.60GHz, runs Ubuntu 20.04, and includes 384 GB DDR4 2,666 MT/s RAM, 2x NVIDIA Tesla V100-PCIE-32GB GPUs, and a network-attached Lustre parallel file system with over 2 PB of storage. The client computer features an Apple M1 Pro CPU, runs Sonoma 14.3, and has 16 GB of LPDDR5-6400 RAM and a 500 GB SSD.

In the analyses presented in this paper (Fig. 5G-I), cell detection models were trained for three regions, including the cortex. In each region, the detection model was refined by repeating the sequence (annotation, training, prediction and proofreading) five to seven times. By annotating approximately 100 cells in total, we were able to establish a model capable of detecting the vast majority of the cells. The training of the detection models was conducted on volumes of 256 × 256 x 16 voxels, which were prepared by preprocessing with a random flip in each dimension. During the label generation step, the center ratio was set to 0.4 and the background threshold was set to zero, meaning that only explicitly annotated voxels were used for training. In the prediction step, volumes cropped to dimensions of 700 × 700 × 100 around the target area were used as input. In the postprocessing step, a threshold for nucleus center probabilities was set to 0.5, and r_min, r_max and d_sup were set to 1 µm, 5 µm, and 3 µm respectively (see details about the parameters in 35).

To quantitatively evaluate the performance of cell detection, we calculated precision and recall, defined as TP/(TP+FP) and TP/(TP+FN), respectively, where TP stands for true-positive, FP for false-positive, and FN for false-negative. In the actual 3D data of the cortical region, the values of TP, FP, and FN were manually counted by comparing the detection results of cells (N ∼ 640) within a defined volume of 310 × 170 × 100, one by one, with the actual images. As a result, we obtained a precision of 99.4 % and a recall of 97.6 %.

Cell segmentation and tracking for the quail embryo data

For the analysis of quail embryo data, cell detection, segmentation and tracking were performed using a custom-made program coded in Python. The observed 3D volume has a transverse area (xy) of more than 1 cm² and a height (z) of 240 µm, which is relatively thin in the z direction compared to the xy plane. The overlap of multiple vascular endothelial cells in the z direction occurs with a very low probability. Therefore, to save computational costs and time, 2D cell detection was conducted on z-projected MIP images instead of full 3D cell detection. To avoid miss-detection of overlapping cells, we divided the z-stack data into three blocks along the z-axis and created MIP images for each block, on which 2D cell detection was performed on the MIP images. The resulting cell lists were compared to identify identical cells detected across multiple blocks. By recognizing double-detected cell pairs as identical, we updated the cell list accordingly. This process was applied to all double-detected cells, thus achieving effective cell detection in the 3D volume and enabling cell tracking throughout the observation period.

Evaluation of optical system with optical invariant and space-bandwidth product

We employed two indices to evaluate the lens system’s scale range and compare it with previous research. The first index, the optical invariant, measures the lens system’s performance and is obtained by multiplying the FOV radius and NA. The second index, the space-bandwidth product, considers the image sensor and wavelength while quantifying the ratio of actual resolution to FOV. These indices are calculated using the following simplified formulas, respectively (41).

where I, FN, M, SBP, R_FOV, and dxy denote optical invariant, field number, magnification, space-bandwidth product, FOV radius, and spatial resolution (FWHM).

Data and materials availability

All data needed to evaluate the conclusions in the paper are present in the paper and/or Supplementary Materials.

Acknowledgements

We would like to thank Prof. K. Fujita of Department of Applied Physics, Osaka University, Japan for his valuable comments on optics design. We are also grateful to Prof. S. Miyagawa, and Prof. Y. Sawa of Graduate School of Medicine, Osaka University, for their support on handling human iPS cells and valuable discussion.

Additional information

Funding

Grant-in-Aid for Scientific Research on Innovative Areas “Singularity Biology (No. 8007)”

21H00431 (YS), 18H05416 (HH), 18H05412 (SO), and 18H05410, 18H05408 (TN)

Grant-in-Aid for Transformative Research Areas (A) “Seeing through Scattering Media (No. 20A207)” 21H05590 and 23H041350 (TI)

the Research Program of "Five-star Alliance" in "NJRC Mater. & Dev." (TN)

Precursory Research for Embryonic Science and Technology (PRESTO) JPMJPR18G2 (TI)

JSPS KAKENHI JP23H00395 and JP24K22022 (HH)

AMED Brain/MINDS JP21dm0207117 (HH) and

BINDS JP23ama121054 and JP23ama121052 (HH)

The Uehara Memorial Foundation (TN) Takeda Science Foundation (TN, HH)

Core Research for Evolutionary Science and Technology (CREST) JPMJCR15N3 (TN),

JPMJCR1926 (KS, SO)

RIKEN Cluster for Science, Technology and Innovation Hub (SO)

JST NBDC Grant Number JPMJND2201 (SO)

Author contributions

Conceptualization: TI, TN

Optics design: TI, HH, TN, YT, SE

Lens design and fabrication: YT

Optical system construction: TI, YT, SE

Development of imaging system software: TI

Sample preparation: TK, YS, KS, HH

Data acquisition: TK, YS, TI

Data analysis and visualization: KI, TI, TK, KS

Setup of deep-learning platform: KS, HI, SO

Construction of data share server: HI, SO

Writing—original draft: TI

Writing—review & editing: TI, TN

Competing interests

T.I, H.H and T.N have patent applications (2021-104163(JP), 2023-077341). K.S. is employed part-time by LPIXEL Inc. All other authors declare no additional conflict of interests.