Acto3D: an open-source user-friendly volume rendering software for high-resolution 3D fluorescence imaging in biology

With the advancement of tissue-clearing technology and fluorescence microscopy systems, it is now possible to label and identify deep structures of large tissues, organs and embryos with the use of fluorescence probes (Kolesová et al., 2021). However, it has remained a challenge to generate readily understandable 3D images. Volume rendering has been a standard method of 3D imaging for visualisation of internal to surface structures of an object based on definition of the intensity and opacity for each spatial coordinate (Levoy, 1988). In practice, 3D images obtained from ultrasonography, computed tomography (CT), or magnetic resonance imaging (MRI) via volume rendering provide much information in a visually and easily understandable manner (Spanaki et al., 2022).

Volume rendering requires a powerful computer, more processing power and larger memory capacity as the space size increases. Specifically, images acquired by fluorescence microscopy often consist of a higher number of pixels in the xyz dimension (incorporating multiple channels) compared with CT and MRI images in the clinics (Fig. S1). Image size inevitably increases as the number of z-slices increases. Reconstruction of 3D images requires a graphics processing unit (GPU), which excels in fast parallel computations. Insufficient dedicated memory in the GPU concerning the size of the imaged data can result in latency during rendering. This latency arises from the necessity to retrieve additional data from storage or system memory, causing delays that impede the attainment of seamless operational performance. Mitigation measures, such as downsizing or compression, may be necessitated under such circumstances. GPUs that possess a sufficiently large, dedicated memory tend to be extremely expensive, however. In addition, volume rendering requires specialised software. Although some open-source software applications are available that are equipped with functions for 3D observation of datasets and advanced analysis functions (Ahlers et al., 2023; Peng et al., 2014, 2010; Royer et al., 2015; Schmid et al., 2010), they pose challenges, such as loading multichannel raw data before compositing, adjusting display settings for individual channels, dynamically modifying sample cross-section orientation and observation direction, managing loading times and processing speeds, and potentially necessitating programming proficiency. There are commercially available software solutions, but these are invariably expensive and exceptionally demanding for high-specification hardware, including dedicated GPUs (dGPUs) (Table S1). These factors impose a substantial barrier to adoption of such technology by researchers in their scientific activities. The ability to easily construct and observe high-resolution 3D images without losing any of the four-channel fluorescence information with low latency, and at low cost, would therefore be of great benefit to the research community.

We here propose a pipeline of observation methods that can overcome these limitations. We have thus newly developed Acto3D (taken from ‘adaptive volume rendering software that allows users to freely adjust and customise colour tones and transfer functions and provides for rich observation in 3D) as an open-source, volume-rendering software application for multichannel fluorescence microscopic imaging. Acto3D leverages the unified memory architecture of Apple Silicon (Apple, Cupertino, CA, USA), in which both the central processing unit (CPU) and the GPU access a shared memory with low latency (Kenyon and Capano, 2022), thereby realising the concept of storing entire image datasets in areas accessible to the GPU – a challenging feat for conventional dGPUs with limited dedicated memory. This architecture, regarded as a type of integrated GPU (iGPU), provides access to a significantly larger memory area than is typically available to dGPUs in personal computers, thus offering a substantial advantage for high-resolution volume rendering without the need for expensive workstations or the larger memory size designated for high-end dGPUs. Furthermore, Acto3D provides a pipeline for extraction of spatial lumen structures and generation of mask images. In the complex process of tissue morphogenesis, a key mechanism involves the formation of fluid-containing lumens, which can manifest as either cavities or interconnected tubular networks, such as the amniotic cavity, blood vessels, neural tube and intra-embryonic body cavity. In this study, we demonstrate that Acto3D can successfully generate 3D images of mouse embryonic pharyngeal arch arteries. With validation of the usefulness and power of Acto3D in imaging of embryonic structures, we propose its adoption as a cost-effective and efficient tool.

We developed Acto3D, using Swift and Metal Shading Language, as part of an open-source project and designed it to be a stand-alone application independent of external packages. This design concept eliminates the need for any specialised environment setup, making the application more accessible to a broad range of users without specialised knowledge. The primary focus of Acto3D is to provide a simple way of viewing fluorescence microscopy images in 3D, supporting free exploration and visualisation in 3D space, without the need for special knowledge of programming languages, such as Java or Python. In volume rendering, the transfer function is a crucial component that determines how scalar pixel values are interpreted and represented in 3D space. It allows users to map these values to specific colours and opacity levels, enabling the precise and flexible visualisation of both external surfaces and internal structures. In Acto3D, users can finely and flexibly adjust the transfer function for each channel in a graphical user interface (GUI). It is also possible to directly edit the transfer functions through programming code in a shader file for advanced drawing. Of note, a high level of observation can be achieved even on a laptop machine by producing high-quality images that are superior to those generated by existing available software (Fig. 1A, Fig. S2; Movie 1). Although it does not require expensive workstations or dGPUs, one limitation is that it is currently a macOS application.

Additions and modifications can be introduced without the need for recompiling the entire program from the source code. Shaders for multiplanar reconstruction (MPR) and maximum-intensity projection (MIP) were included as part of the initial set.

To validate Acto3D, we constructed 3D models of embryonic day (E) 10.0 (32-somite stage) whole mouse embryos from images captured with Zeiss Lightsheet 7 light-sheet microscope, following tissue clearing with CUBIC-R+ (Susaki et al., 2020; Tainaka et al., 2018). Nuclear staining offered a comprehensive visualisation of the entire shape of each embryo (Fig. 3A). The surface structure was observed more readily as the overall opacity was increased, allowing somites and the pharyngeal pouches to be clearly distinguished by free rotation and sectioning (Fig. 3B,C). Conversely, as the opacity and intensity of the nuclear signals were reduced, internal structures became more readily visualised (Fig. 3D-F; Fig. S3A-C). The detailed structure of the intersomitic vessel network was visible at increased magnification (Fig. 3E). Moreover, by customising the transfer function to consider the gradient of the pixel value in the shader, we were able to highlight edges and thereby to identify the pharyngeal arch arteries (PAAs) and the dorsal aorta (Fig. 3F; Fig. S3D). MPR images provided virtual slices from various perspectives that differed from the original imaging direction (Fig. 3G-I; Fig. S4). Acto3D thus revealed readily comprehensible anatomical structures of organs and tissues in 3D.

We next observed development of the mouse embryonic heart (Anderson et al., 2014, 2019; Captur et al., 2016) (Figs 4 and 5; Fig. S5; Movies 2-5). At E9.5, the presumptive left and right ventricles were discernible (Fig. 4A) and the trabeculae were well developed in the ventricles (Fig. 4B,C). At E10.5, the onset of formation of the muscular ventricular and primary atrial septum were apparent (Anderson et al., 2002, 2014) (Fig. 4D,F; Movie 2). The atrioventricular valves still retained the superior and inferior atrioventricular cushions, connecting only to the left ventricle (Fig. 4D,E; Movie 2). In Fig. 4E, the image of the atrioventricular cushion, not composed of myocardium, was revealed through nuclear staining. However, in Fig. 4F, this nuclear staining is obscured, resulting in the removal of the atrioventricular cushion image and revealing the cardiac troponin I (TNNI3)-positive primary atrial septum image located behind it. At E11.5 or later, the rightward shift of the atrioventricular canal responsible for its connection to both ventricles was seen (Fig. 4G,J; Movie 3). The primary atrial septum had further developed, and a secondary foramen had begun to form (Fig. 4G,H; Movie 3). The outflow tract still originated only from the right ventricle and was not separated into pulmonary and aortic trunks (Fig. 4I). By E12.5, the atrioventricular cushions were separated into the presumptive mitral and tricuspid valves (Fig. 4J; Movie 4). The outflow tract was separated into pulmonary and aortic trunks, and its stem was shifted leftward, leading to positioning of the aortic valve over the muscular ventricular septum (Fig. 4K; Movie 4). The opening of the pulmonary vein was observed near the dorsal rim of the atrial septum on the left atrial side (Fig. 4K; Movie 4). Presumptive papillary muscles were also clearly identified at this stage (Fig. 4L; Movie 4), which shows the utility of Acto3D because this had not been confirmed by a conventional method to date (Captur et al., 2016). All these observations were consistent with those of previous studies using episcopic microscopy (Anderson et al., 2002, 2014, 2019; Captur et al., 2016). In these studies, episcopic microscopy methods have demonstrated a remarkable ability to image authentic organ morphology. However, they have encountered hurdles in accurately tracking specific cell populations and/or visualising the 3D distribution of a specific cell population based on protein expression. Moreover, 3D observation facilitates the generation of diverse rendered images through adjustments in viewpoints, and transfer functions. The creation of animations highlighting and emphasising the target structures substantially amplifies the persuasive impact of these observations (Movies 2-5).

Our research efforts then shifted to exploring the feasibility of this approach by focusing on the localisation of HCN4, a marker indicative of the early electrical conduction system, which includes the sinoatrial node (Garcia-Frigola et al., 2003; Stieber et al., 2003; Wiese et al., 2009). HCN4 was initially localised to the sinus venosus (the inflow tract) at E9.5, and gradually it extended to the region where the right and left common cardinal veins are located, forming a V shape around the trachea, and the dorsal wall of the right atrium (Fig. 5; Movie 5). The expression of HCN4 subsequently increased in the common cardinal veins in a left-right asymmetric manner. Consistent with previous observations (Wiese et al., 2009), the common cardinal vein was also found to be positive for TNNI3 (Fig. S5).

These various observations thus indicated that Acto3D facilitates the observation of structures of interest by rendering them more discernible. More importantly, it allowed sectioning of the heart at various angles and thereby provided a vivid depiction of the morphological changes in the internal structure of the embryonic heart.

During embryogenesis, the morphology of PAAs changes markedly through remodelling (Hiruma et al., 2002; Yashiro et al., 2007). We next aimed to observe PAAs by challenging reconstruction from mask images obtained by filling in the vascular lumen edged with fluorescently labelled endothelium. Imaging by fluorescence microscopy usually needs to overcome two key problems: variation in staining intensity among specimens, and the attenuation of emission light intensity that is either caused directly by the presence of material between the target object and the objective lens or due to a reduction in excitation light intensity reaching the target during its passage through intervening tissue (Kervrann et al., 2004). It is therefore not feasible to automatically recognise the vascular lumen by setting a specific threshold for intensity values and filling in the entire lumen.

To address this issue, we iteratively applied one-dimensional k-means clustering (Lloyd, 1982) spatially to single-slice images (Fig. 6A). We initially pre-processed the data by applying a 3D Gaussian blur to remove noise, setting the kernel size to 7×7×7 with a sigma of 1.8, and cropped the local vascular area of interest. We next applied the k-means++ algorithm (Arthur and Vassilvitskii, 2007) to the first slice of this area in order to obtain the initial centres, which were then used to obtain initial centroids of cluster classification by standard k-means clustering. Given that the selection of initial centres by k-means++, which is a type of unsupervised machine learning to determine where to divide the histogram, relies on a weighted probability distribution, the results of the histogram division vary with each computation. In our work on separating vascular lumens on the histogram, the optimal separation would distinguish between the high pixel value endothelial areas of vessels, labelled by fluorescence, and the low pixel value vascular lumen, with soft tissue or autofluorescence occupying the intermediate range. Therefore, we empirically conducted most analyses with a baseline of three clusters. Nevertheless, histograms may not always separate ideally; thus, if the initial division does not lead to the desired image segmentation, either re-running the calculation or increasing the number of clusters can improve the histogram's separation capacity, potentially leading to better results (Fig. 6; Fig. S6A). If imaging was performed with thin z intervals, adjacent slices had structurally similar images and similar intensities as revealed by the structural similarity index measure (SSIM) (Wang et al., 2004) (Fig. S6B). As expected, adjacent images had similar classifications after specifying the obtained centroids in one slice as the initial centres for the next slice (Fig. 5B). The continuous slices obtained in this manner showed a smoother transition with reduced computational costs compared with the application of k-means++ to each independent slice one by one (Fig. S6C-F). Importantly, the vascular structure constructed from the mask images obtained by this iterative method represented well the structure of the vascular lumen with high resolution (Fig. 6C). Furthermore, segmentation in Acto3D allows tracing from any viewpoint and combining multiple mask images. This approach leads to fewer omissions in mask generation compared with tracing from a fixed direction (Fig. S7A-C). Using this method, we were able to construct a 3D model of PAAs in a more accurate formation (Fig. S7D,E), additionally incorporating the surrounding tissues, including the embryonic pharyngeal pouches (Fig. 7A-C; Movie 6).

We observed the chronological 3D model of PAA development (Hiruma et al., 2002; Yashiro et al., 2007) (Fig. 7D-K). At E9.5 (26-somite stage), the first, second, third and fourth PAAs were seen (Fig. 7D). By E10.0 (32-somite stage), the first and second PAAs had lost their connection with the dorsal aorta, and formation of the sixth PAA appeared to be underway in the 3D model (Fig. 7E). However, observation of the acquired images used for construction of this 3D model revealed that the endothelium of the presumptive sixth PAA, although lacking a visible lumen, was continuously connected between the aortic sac and dorsal aorta (Fig. S8A-C). To investigate this finding further, we applied Acto3D to an hourglass-shaped model (Fig. S8D). Theoretically, Acto3D enables a space as small as 1×1 px to be masked. However, if the blur is applied during the pre-processing step, the lumen area may collapse in the case of extremely small lumens. When using the 7×7×7 kernel applied in the same way as when generating PAAs images, we found that the minimum required radius was 2.5, corresponding to a lumen diameter of 5 px. At E10.5 (36-somite stage), formation of the sixth PAA was complete (Fig. 7F). At the 40-somite stage, an aortic sac horn had formed as a common duct from the aortic sac to the third and fourth PAAs, resulting in a slight anterior shift in the positions of these arteries away from the heart (Fig. 7G). At E11.5 (43-somite stage), the aortic sac had separated into the aortic and pulmonary trunks, and the right sixth PAA had started to regress (Fig. 7H). Between E11.5 and E12.5, the base of the outflow tract underwent a rotation, resulting in movement of the aortic trunk posteriorly and relocation of the pulmonary trunk anteriorly (Phillips et al., 2019) (Fig. 7H,I). Around E12.5, the right dorsal aorta of the periphery started to regress below the level of the seventh intersegmental artery (Fig. 7I,J). At E13.5, this portion of the aorta had completely regressed and the morphology of the great artery system was similar to that of an adult (Fig. 7K). All our observations of 3D images of PAAs were thus consistent with previous findings (Bamforth et al., 2013; Hiruma et al., 2002; Yashiro et al., 2007).

In some embryos, we were able to confirm the presence of dorsal collateral vessels between the fourth and sixth PAAs (Fig. 7F,H). It is proposed that persistent fifth aortic arch is not a remnant of the fifth pharyngeal arch artery itself, but a morphological abnormality caused by these collaterals (Anderson et al., 2020; Bamforth et al., 2013; Gupta et al., 2016; Rana et al., 2014).

We here describe the development of software, Acto3D, that allows easy 3D reconstruction for flexible observations from multichannel z-stack fluorescence images. Acto3D is a robust platform designed to allow fine-tuning of opacity and highlighting of target structures, providing sophisticated and readily understandable 3D anatomical images for viewing from multiple perspectives.

Traditionally, the scope of volume rendering has been limited by the constrained dedicated memory of dGPUs. The conventional approach entails the initial rendering of a low-resolution image within the GPU's memory, followed by the retrieval of supplemental high-resolution data from either external storage or system memory. This sequential process can introduce noticeable latency, especially during alterations in viewing angles or slices, thereby impeding workflow efficiency. Therefore, as long as this method is used, a dGPU with a lot of dedicated memory and supporting high computing power, as well as dedicated and expensive software are required to reliably perform volume rendering without stress. Such requirements are often a significant barrier, particularly for research labs operating on limited budgets.

Acto3D, by contrast, takes advantage of Apple Silicon. The architecture of Apple Silicon gives the GPU fast access to a large amount of system memory. By adopting the concept of loading all image data into memory, Acto3D enables advanced volume rendering with a lightweight framework. Acto3D is a unique software that allows up to four channels to be effortlessly visualised from different angles and slices while maintaining accurate dimensions, without complicated operations (Fig. S2). Additionally, it provides an option to programmatically modify shaders for display customisation, allowing detailed display control using the 3D visualisation framework that Acto3D provides (Table S1). We have validated the benefit of this software for detailed characterisation of anatomical structures in 3D. It thus provided sophisticated 3D models both of the mouse embryonic heart and, with the support of an algorithm for tracing vessel lumens, of PAA development. Like other volume-rendering software, Acto3D also allows the creation of animations. It enables users to smoothly interpolate between multiple specified observation points to create videos. Acto3D offers a level of rendering freedom that surpasses commercial software, leading to more compelling visual representations (Movies 2-5). This powerful tool can theoretically be applied not only to embryonic cardiovascular structures but to other targets with appropriate probe labelling and tissue-clearing technology. With its ability to observe 3D morphology as well as to visualise internal structures, Acto3D should prove to be of great benefit in various life science fields.

Previous studies have presented 3D interpretations of the mouse embryonic heart obtained with the use of various techniques, including episcopic microscopy (Anderson et al., 2002, 2014, 2019; Captur et al., 2016) and micro-CT (Li-Villarreal et al., 2023). These studies focused primarily on depicting general morphology without observation of specific structures derived from specific cell lineages. 3D models generated by Acto3D from acquired fluorescence images allow general morphological observations based on nuclear staining only. More importantly, it is possible to study a target structure or the distribution of target cells revealed by specific immunostaining within the general anatomical structure with the use of multichannel 3D images. This task can be smoothly handled even on a laptop computer (Movie 1).

Compared with previous studies of the 3D morphology of PAAs (Bamforth et al., 2013; Hiruma et al., 2002), the application of Acto3D to obtain 3D models of these arteries had several advantages, including technical simplicity, low cost, high resolution, the ability to observe cross-sections and from different viewing directions, and the provision of information on surrounding structures. One limitation was that vascular lumens could not be observed if the vessels were not patent or were too thin (Fig. S8). The use of confocal microscopy instead of light-sheet microscopy may be more suitable for further detailed local observations, as previously suggested (Ramirez and Astrof, 2020). Notably, lumen formation is one of the fundamental processes during embryogenesis. Therefore, the visualisation of 3D lumen structures during embryonic development is a crucial tool that significantly enhances our understanding of morphogenesis and organ formation. The ability to visualise fluid-containing lumens, whether they are cavities or tubular networks, provides researchers with a profound insight into the intricate processes that underlie embryonic development. It provides a holistic view of how these lumens form, expand and morph over time. Acto3D's system of using iterative k-means++/k-means approaches to abstract the shape of large vessels is fully applicable to any lumen structure, as long as the inside and outside of the lumen can be identified with appropriate fluorescent staining. We believe that the functionality of Acto3D, which allows for the convenient and clear visualisation of the 3D structure of lumens filled with fluid in embryos, will be greatly beneficial to many developmental biologists' research.

In general, microscopy images have a lower resolution in the z direction than in the xy direction. Even if the voxel size is isotropic, it is therefore impossible for all resolutions to be equal in space. When examining cross-sections, it may be necessary to optimise specimen placement and align the plane of the light sheet to achieve maximum resolution. Although positioning small embryos can often be challenging, the effect of specimen orientation is negligible in 3D observations, and was also found to have minimal impact in MPR imaging (Fig. S4). Acto3D allows for instantaneous switching between 3D views, MPR, and other display methods while observing from any angle. This ensures that, regardless of the direction of sample placement, 3D understanding can be enhanced with Acto3D while maintaining a consistent resolution.

The use of Apple Silicon played a substantial role in the development of our pipeline, largely as a result of its good fit for our dual objectives of deploying the entirety of image data in GPU-accessible memory and allowing meticulous observations of fine details at any desired time. However, the consequence of this reliance on Apple Silicon is that Acto3D is currently limited to macOS. There is the potential for cross-platform applicability of Acto3D in the future if similar high-performance iGPUs are developed or for computers with large-capacity dGPUs considering the advantages Acto3D offers in facilitating 3D observation.

ICR mice were obtained from Japan SLC. All animal procedures were approved and performed according to guidelines specified by the Animal Experimentation Committee of Kyoto Prefectural University of Medicine (license numbers M2022-172 and M2022-180). The morning of the day on which a vaginal plug was detected after breeding was determined as E0.5. Embryos were dissected at times from E9.5 to E13.5. The impact of blood autofluorescence was minimised by flushing out of the vessels of an embryo by injection of 4% paraformaldehyde in PBS with the use of a glass needle connected to a 1-ml syringe (Terumo) and an injection holder (IM-H1, Narishige Scientific Instrument Lab). The glass needles were prepared with the use of a puller (PC-10, Narishige Scientific Instrument Lab) set at 52°C from borosilicate glass capillaries with outer and inner diameters of 1.0 and 0.58 mm, respectively (GC100F-10, Harvard Apparatus). For observation of the heart, the abdomen of the embryo was punctured and 4% paraformaldehyde in PBS was gradually injected to identify the appropriate injection site for flushing out of blood cells from the heart and great vessels. After visual confirmation with a stereomicroscope that blood cells had been adequately removed from the ventricles, any unwanted tissue was trimmed. For 3D reconstruction of vessels, the ventricles were punctured and blood cells were flushed out by manual injection until no further such cells flowed out from the umbilical cord. In both cases, after perfusion with 4% paraformaldehyde, the specimens were incubated overnight at 4°C in the same fixative.

Immunostaining was performed as previously described (Susaki et al., 2020), with some modifications. After fixation, specimens were washed three times for 30 min each wash with PBS at room temperature. Delipidation was performed with 10% N-butyldiethanolamine (B0725, TCI) and 10% Triton X-100 in deionised distilled water (DDW) (CUBIC-L) at 37°C for 24 to 48 h, until the CUBIC-L no longer changed colour. The samples were then washed three times for 2 h each wash with PBS. For nuclear staining, samples were incubated with SYTOX Green (S7020, Thermo Fisher Scientific) at 1:2500 in a solution of 10% Triton X-100, 5% N,N,N′,N′-tetrakis(2-hydroxypropyl)ethylenediamine (Quadrol, T0781, TCI), 10% urea (35904-45, Nacalai Tesque) and 500 mM NaCl in DDW (ScaleCUBIC-1A with 500 mM NaCl). They were then washed three times for 2 h each wash with 10 mM HEPES-NaOH (pH 7.5). To prevent non-specific binding of probes, the specimens were incubated in a solution containing 10 mM HEPES-NaOH, 10% Triton X-100, 200 mM NaCl, 0.5% casein (030-01505, Fujifilm Wako), 2.5% N,N,N′,N′-tetrakis(2-hydroxypropyl)ethylenediamine and 1 M urea at pH 7.5 (HEPES-TSC buffer with additives) for 90 min at 32°C. In the specific case of HCN4 detection, prior to incubation in HEPES-TSC buffer with additive, the samples underwent an additional 12 h incubation at room temperature in a blocking solution containing 6% donkey serum (S30-100ML, Millipore), 10% dimethyl sulfoxide (13407-45, Nacalai Tesque) and 0.2% Triton X-100 in PBS. Subsequently, the specimens were incubated for 3 days at 32°C with an appropriate combination of primary antibodies to TNNI3 (ab56357, Abcam; diluted 1:200 to a concentration of 5 µg/ml), DyLight 594-conjugated tomato lectin (DL1177, Vector Laboratories; 1:150), and/or antibodies to HCN4 (AB5808, Millipore; 1:250) in HEPES-TSC buffer with additives. The samples were washed first for 2 h at 32°C with 0.1 M phosphate buffer at pH 7.5 (PB) containing 10% Triton X-100 and then three times for 2 h each time at room temperature with PB. After incubation for 90 min at 32°C with HEPES-TSC buffer with additives, the samples were exposed for 2 days at 32°C to secondary antibodies diluted in HEPES-TSC buffer with additives including an appropriate combination of Alexa Fluor 555-conjugated donkey antibodies to rabbit immunoglobulin G (A-31572, Invitrogen; 1:250) for HCN4 and/or Alexa Fluor 633-conjugated donkey antibodies to goat immunoglobulin G (A-21082, Invitrogen; 1:200) for TNNI3. Samples were washed once for 2 h at 32°C with PB containing 10% Triton X-100 and then three times for 2 h each wash at room temperature with PB. The stained embryos were then incubated with 1% formaldehyde (16223-55, Nacalai Tesque) in PB first for 60 min at 4°C and then for 45 min at 37°C. They were then washed three times for 60 min each wash at room temperature with PB. Refractive index matching was performed by incubation of the samples first overnight in a solution of 45% 2,3-dimethyl-1-phenyl-5-pyrazolone (Antipyrine, D1876, TCI) and 30% N-methylnicotinamide (M0374, TCI) in DDW (CUBIC-R+) (Susaki et al., 2020; Tainaka et al., 2018) diluted to 0.5× with DDW, and then for at least 16 h in CUBIC-R+ at room temperature.

Images were acquired with a Zeiss Lightsheet 7 fluorescence microscope. An EC Plan-NEOFLURA 5×/0.16 foc objective lens was used for detection, and an LSFM 5×/0.1 foc objective lens for illumination. The laser setup was applied only on the left side. Each specimen was embedded in a glass tube with an inner diameter of 2.2 mm (701910, BRAND) or 1.5 mm (701908, BRAND). If an appropriate-sized glass tube was not available, the specimen was embedded in 2% agarose (02468-95, Nacalai Tesque) in CUBIC-R+. With regard to acquisition parameters, the zoom was set between 0.36 and 1.3 to ensure that the area of interest fitted within a frame of 1920×1920 pixels; the centre thickness was adjusted to maintain an optimised z interval/xy pixel ratio of <2; the laser output for each wavelength was set within the range of 10% to 50% in order to avoid colour saturation; and the exposure time was fixed at 99.87 ms. In general, image stacks consisted of 600 to 1200 z-slices and were saved in 16-bit depth.

For volume visualisation with the use of Acto3D, the development environment consisted of an Apple MacBook Pro (model MacBookPro18,4) equipped with an Apple M1 Max processor featuring 64 GB of memory and running on OS version 12.5.1. In addition, an Apple M1 Pro processor with 32 GB of memory, and Apple M1 processor with 16 GB of memory, and an Apple M2 processor with 16 GB of memory were used for verification. Although not officially declared by Apple Inc., there is a limitation on the total amount of memory that can be secured by the iGPU at one time for a 3D image. This maximum amount is not clearly stated for each computer model, but it is generally around half of the computer's total memory. Specifically, for the M1 Max with 64 GB of memory, the maximum buffer size was 36 GB; for the M1 Pro with 32 GB of memory, it was 16 GB; and for the M1 and M2 with 16 GB of memory, it was 8 GB. Given these limitations, it is necessary to adjust the size of each image to avoid exceeding these specific constraints. If an image exceeds this limit, it is automatically resized within the software to fit the allowable memory capacity.

We first loaded the Zeiss czi format using Fiji (Schindelin et al., 2012) (version 2.3.0/1.53s). The import options were set to ‘Grayscale’ colour mode, with ‘Virtual Stack’ turned on and ‘Split Channel’ turned off. We then converted the file to multipage TIFF format using the [File]→[Save As]→[Tiff…] function. At this stage, crucial metadata, such as display ranges and voxel size, were stored in the TIFF file. To achieve 3D visualisation, we loaded this TIFF file into Acto3D according to the following procedure: Acto3D→[Open Images]→[Open ImageJ/Fiji TIFF]. We then set the display range for each channel through [Image Option…]. Given that pixel values in fluorescence microscopy can vary as a result of factors such as laser output, staining conditions, and xyz position, there are no universally applicable recommended parameters. Nonetheless, we generally adjusted the display range to avoid intensity saturation and to minimise background noise. The [View in 3D] button was used to execute the volume visualisation. For each channel, the opacity level and intensity could be adjusted through the GUI, as detailed in Figs S2 and S3. Of note, the final image resulting from volume rendering may exhibit variation as a result of the number of overlapping images, even with the same settings. No absolute parameters can therefore be considered universally perfect for all cases. However, we generally achieved good rendering images by setting the opacity to 0 near pixel value of 0, assigning low opacity to low-pixel value areas, and adopting medium to high opacity for high-pixel value areas. This configuration served as a basis for fine-tuning the opacity for each pixel value in each channel, allowing us to obtain optimal images. Several settings can be found in our GitHub repository: https://github.com/Acto3D/Acto3D.

Within Acto3D, users can interact with the displayed 3D image by placing a white sphere at desired coordinates by a simple click. This feature facilitates distance measurements between two specified points and enables the creation of sectional views via a plane defined by three specified points. For precise selection of regions within the 3D image, users are advised to utilise the MPR view, accessible by pressing the ‘Z’ key. In Fig. 4H and Movie 3, this functionality is leveraged to enhance visualisation of anatomical structures, such as the secondary foramen and the primary septum, with the dorsal edge of the latter being demarcated.

Images of specimens labelled with tomato lectin were loaded into Acto3D with the same approach as in the previous section. The display range was set low enough to capture the signal value of endothelial cells of the blood vessels. A dedicated screen was opened by selecting [Segment]→[3D Segment], where we selected the channel that imaged tomato lectin, and [Preprocess] was executed. This step involved application of a Gaussian blur to the 3D data with a kernel size of 7×7×7. From the displayed MPR images, we selected the blood vessels constituting the PAAs and chose the first and last slices for construction using [Set]. We next set the number of k-means++ clusters (usually, three or four) and performed the initial clustering with the reload button. Given that k-means++ clustering relies on weighted probability, the calculation results may vary with each execution. After the necessary cluster classification was obtained, we selected the cluster area representing the blood vessel lumen and pressed the [Start] or [Resume] button to track the same cluster according to the aforementioned algorithm. This process was repeated for multiple blood vessels constituting the PAAs, and their data were finally merged and incorporated into the visualisation by pressing [Apply To Main]. Detailed instructions for these operations can also be found in our GitHub repository: https://github.com/Acto3D/Acto3D.

To compare the performance of volume rendering, the model used in Fig. S2 consists of a sphere with a radius of 256 px, enclosed within a cube the sides of which measure 700 px each. The pixel value for the cube was set at 200, and the sphere's pixel value was 128. As depicted in Fig. S2, the model was quartered, with a central gap of 8 px. Additionally, a Gaussian blur with a kernel size of 7×7 was applied to each cross-section. These quarter segments are designated as channels 1-4. These multichannel images were used as the input for Acto3D and Imaris. This image was saved as a multichannel TIFF using Fiji, and used as input for both Acto3D and Imaris software (Oxford Instruments, version 10.0.0). Regarding Vaa3D, the four-channel image was saved as a .v3draw file in Fiji by selecting [File]→[Save As]→[V3Draw…] and open in Vaa3D (Vaa3D-x.1.0.10).

Regarding all these models, the Python source code used for their creation is accessible from our GitHub repository at https://github.com/Acto3D/Acto3D.

Furthermore, identical models can be created within Acto3D by selecting [File]→[Demo Model] in the Acto3D menu bar.

For the segmentation of the tori model and PAAs in Imaris software (Oxford Instruments, version 10.0.0), we adapted a previously described methodology (Ramirez and Astrof, 2020). Unlike the original approach, which utilised the ‘distance drawing’ mode for contour tracing, we used the ‘magic wand’ mode for the tori model and ‘isoline’ mode for PAAs. For the tori model, mask images were generated by tracing at intervals of every 1, 2, 5, 10, 15, 20 and 30 slices. For PAAs, mask images were traced every ten slices.

In Acto3D, it is possible to create smooth animations between specified views while observing 3D images. Most of the display settings available in Acto3D, such as slice number, rotation angle, opacity and colour, can be automatically interpolated. In order to achieve this, after opening the 3D image, the [Save Parameters] button on the first view was pressed to save the display settings. Then, the image display was adjusted again, and [Save Parameters] pressed to save the new settings. In the ‘Animation’ tab, an animation queue was created by clicking [Add motion] or [Add rotation]. The display setting from which and to which the transition should occur and the duration of the transition time were specified. To link several animations, these steps were repeated. Finally, [Create] was clicked to create the video.

The source code, compiled binaries, and comprehensive documentation for the application, including guidelines on generating custom shaders, along with a dedicated package for interactive integration with Python, can be accessed through the GitHub repository (https://github.com/Acto3D/Acto3D and https://github.com/Acto3D/Acto3D_py).

We thank the Kyoto Prefectural University of Medicine Laboratory for Experimental Animals for their support with the animal experiments, the Kyoto University Live Imaging Center (KULIC) for the acquisition of imaging data with the Zeiss Lightsheet 7 microscope and using Imaris software, as well as Michiyuki Mastuda and Etsuo Susaki for technical advice and helpful discussion.

The peer review history is available online at https://journals.biologists.com/dev/lookup/doi/10.1242/dev.202550.reviewer-comments.pdf