Open Access
Add to BibSonomy
Abstract
Quantitative analysis of zebrafish cerebral vasculature is essential for the study of vascular development and disease. We developed a method to accurately extract the cerebral vasculature topological parameters of transgenic zebrafish embryos. The intermittent and hollow vascular structures of transgenic zebrafish embryos, obtained from 3D light-sheet imaging, were transformed into continuous solid structures with a filling-enhancement deep learning network. The enhancement enables the extraction of 8 vascular topological parameters accurately. Quantitation of the zebrafish cerebral vasculature vessels with the topological parameters show a developmental pattern transition from 2.5 to 5.5 dpf.
© 2023 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
The zebrafish is one of the most widely used animal models for vascular study due to several advantages such as small size, high fecundity, and multiple vascular transgenic lines [1–4]. The zebrafish model served as a bridge between in-vitro assays and in-vivo mammalian studies. The large amounts of studies on this tiny vertebrate have advanced the understanding of vascular development mechanism, disease biology, and drug toxicity over the past two decades [5,6]. For example, the molecular mechanisms of cerebral vascular development in zebrafish have greatly helped the study of hemorrhagic stroke among vertebrates [7,8].
Moreover, the whole-body transparency of zebrafish makes it ideal for optical imaging and quantification of its morphological changes [9–11]. With the confocal laser scanning microscope (CLSM) or more recent light-sheet microscope, high-quality three-dimensional images of zebrafish cerebral vasculature could be obtianed [3,12,13]. Those imaging techniques usually yield tremendous 3D image data that are far beyond the capabilities of traditional imaging analysis pipelines. Visual assessment is subjective to human bias and hard to identify 3D structures. Manual statistics are time-consuming and impractical for 3D vascular morphological changes. Especially, the complex cerebra-vasculature made it impossible to achieve its topological parameters quantitatively by hand [14,15]. So, there is a strong demand to establish an automatic method of topological parameters extraction for the zebrafish cerebral vasculature.
Angiography is the most widely-used method for vessel labeling in which fluorescent dyes were directly injected into the vessel. Serval methods for vascular segmentation from cerebral angiography have been developed. In 2020, Todorov et al. developed a deep-learning-based framework to quantify and analyze mice brain angiography images, named Vessel Segmentation & Analysis Pipeline (VesSAP) [16]. The pipeline used a convolutional neural network (CNN) with a transfer learning approach for the segmentation of the mouse cerebral vasculature and achieved human-level accuracy. Recently, Tetteh et al. developed DeepVesselNet to extract the vascular topological parameters from 3D angiographic vessel images such as vascular skeleton, branches, and intersections [17].
In angiography, the animal to be imaged has to be fixed first. The new transgenic labeling technique allows the live animal to be imaged. In the past two decades, transgenic zebrafish have revolutionized biomedical research in developmental biology, human disease pathobiology, drug development, and so on. As of October 2022, the Zebrafish Information Network (ZFIN) displays 2131 transgenic zebrafish lines about vessels [18]. Among them, Tg(kdrl: EGFP) and Tg(fli1: EGFP) are the two most widely-used transgenic zebrafish models for vasculature labeling [19–21]. Tg(fli1:EGFP) zebrafish express the green fluorescent protein (GFP) in the vascular endothelium, lymphatic vessels and coronary arteries, while Tg(kdrl: EGFP) zebrafish only express GFP in the vascular endothelium. Different from angiography where the fluorescent dyes are distributed in the vessels evenly, the fluorescence intensity of endothelial cells was notably different even in the same vessel [22–25]. In addition, the monolayer endothelial cells lie along the vessel's inner wall, so the vessels of these transgenic fluorescent zebrafish displayed intermittent, hollow, and inhomogeneous structures.
This structure leads to more difficulties in the segmentation and accurate quantitative statistics of vascular topological parameters, compared to angiographic images. Zudaire et al. developed an enhancement method to extract vessel profiles using a fast multiscale Hessian-based enhancement filter [26]. Kugler et al. developed a zebrafish vascular quantification (ZVQ) tool for automated quantification of zebrafish cerebrovascular architecture [27,28], based on Sato enhancement [29] and Otsu thresholding [30]. However, for vessels with weak signals, the vascular enhancement did not perform well and the segmentation led to the interruption or even missing of thin vessels.
Here, we introduced a deep-learning-based filling-enhancement network, called FE-Unet, to enhance the 3D images of transgenic zebrafish cerebral vasculature, and established an automated pipeline for topological parameter extraction of zebrafish cerebral vasculature. Using the 3D images acquired by custom-built light-sheet microscopy, the FE-Unet was trained to transform the intermittent, hollow, and inhomogeneous vessel structure in transgenic vascular images into solid as in angiographic images. With the enhancement of FE-Unet, accurate skeleton and topological parameters of cerebral vasculuture was extracted. The development of zebrafish cerebral vasculature from 2.5 to 5.5 dpf was imaged and quantitatively analyzed.
2. Materials and methods
2.1 Preparation of transgenic labeled and angiographic zebrafish
The zebrafish preparation processes were carried out following the Guide for the Care and Use of Laboratory Animals (published by U.S. National Academy of Sciences, ISBN 0-309-05377-3), and were approved by the animal ethics committee of Suzhou Institute of Biomedical Engineering and Technology, CAS. A stable transgenic zebrafish line (Tg(kdrl:EGFP)) with GFP-labeled vascular endothelium was used in our experiment. The zebrafish were bred and raised according to the standard procedures on a 14 hours light and 10 hours dark cycle at 28.5 °C. Zebrafish egg and larvae were incubated in Holt buffer, and 0.0045% PTU (1-Phenyl-2-thiourea, Sigma P7629, USA) solution was added to the solution one day after fertilization (dpf) to inhibit pigment production. After four days of development, zebrafish were fed paramecium twice a day. For light-sheet imaging, 1.5% agarose (UltraPure Low Melting Point Agarose, ThermoFisher Inc, USA) was used to restrain the larvae maintaining the stability of the posture of the fish.
For angiography, Evans Blue (EB, CAS:314-13-6, Macklin Inc, China) solution was diluted to 1% (mass fraction) using a phosphate buffer solution. Anesthetized embryos were transferred to a microinjection plate by collecting the embryos with plenty of 2× tricaine in a plastic pipette. 2-3 nL diluted Evans Blue solution was injected into the precardiac sinus of the zebrafish embryos using the microinjection dispense system (PV830, World Precision Instruments Inc, USA).
2.2 Two-channel 3D light-sheet imaging
The vasculatures of zebrafish were imaged with a custom-built bi-directional illumination light-sheet microscopy [31]. 488 nm (MLD, Cobolt Inc, Swedish) and 640 nm (OBIS, Coherent Inc, USA) laser were used to excite the transgenic GFP and the Evans Blue dye respectively. The two collimated laser beams were scanned by a galvanometer (CT6215H, Cambridge Technology, UK) and focused by two 10× illumination objective lenses (NA 0.25, Olympus Inc, Japan) to form digital scanning light sheets with the waist thickness of 8 µm. A two-dimensional piezoelectric positioner (SLC-2445-L, SmartAct, Germany) was used to move the zebrafish sample for 3D imaging. The detection path consisted of a water immersion objective (10×, NA 0.3, Nikon Inc, Japan), a quad-band bandpass filter (89402 m, Chroma Technology, USA), and an sCMOS camera (ORCA-Flash 4.0, Hamamatsu Inc, Japan). With the above device, our system can achieve a 1331.2 × 1331.2 µm2 field-of-view, 1.04 µm lateral resolution and ∼8 µm axial resolution. The image frame rate of the light-sheet system is set 10 frames per second and the piezoelectric positioner sets a stable time of 30 ms for each layer. The final volume acquisition speed of our system is 3.227 × 107 pixels per second. To reduce the error caused by vascular movement, two-color images were collected in each piezo-step.
2.3 FE-Unet network architecture and training
The FE-Unet was constructed based on the Unet framework. It consisted of an encoder module and a decoder module, responsible for vessel characteristics extraction and feature fusion respectively. As shown in Fig. 1(A), on the decoder part, the input of each convolution module was the combination of the output of the last convolution module and the corresponding layer of the encoder. In this way, the features could pass through a certain depth without losing too much vessel network information. After up-sampling on the decoder part, the network output the segmentation result with the same size as the original data. The enlarged figure shows the convolution modules, which consisted of two convolutional layers. And the output of each convolutional layer was activated by a GeLu activation function.
Fig. 1. FE-Unet network architecture and image enhancement. (A) Schematic of FE-Unet architecture; Tg-v image: input transgenic vessel image, FE-out image: FE-Unet output image; (B) The enlarged pictures from the white box region in (A); (C) Cross-section of the white line in Fig. 1(B). Scale bars represent 100 µm and 20 µm for (A) and (B).
The FE-Unet transforms the hollow tubular structure in transgenic endothelial cells image (Tg-v image) into uniform columnar structure, as shown in Fig. 1(B). Figure 1(C) shows the cross-sections of the transgenic and FE-Unet enhancement vessels were bimodal and Gaussian shape, respectively. The distance between the bimodal peaks was the diameter of the vessel, which was consistent with the full width at half maximum (FWHM) of the FE-Unet enhancement vessel.
Pairs of the transgenic vascular images and EB angiographic images of 7.5 dpf zebrafish were used as the training dataset of FE-Unet. To increase the data complexity, the 2D slices (2048*2048 pixels) from 3D image stacks were randomly cropped into partially overlapping blocks (1024*1024 pixels). Computations were carried out on a computer with an AMD EPYC 7402 CPU @ 2.8 GHz and 64-bit Windows 10 system. Eight NVIDIA GeForce RTX 3090 GPUs with CUDA11.0 were used to accelerate calculation. The software is based on the architecture of Tensorflow and Python development environment Pycharm. Three metrics were used to evaluate the quality of the output image: the normalized root-mean-square error (NRMSE), peak signal-to-noise ratio (PSNR), and structural similarity (SSIM) index:
2.4 Topological parameter extraction
The enhanced 3D structures of transgenic zebrafish by FE-Unet were used to extract the topological parameters, as shown in Fig. 2. The images were first segmented with different methods to obtain distance map images, skeleton images (Sk) and 3D convex hull images, respectively.
Among them, the 3D envelope of vascular segmentation images, 3D convex hull (Ch), was obtained to represent the drainage volume (Vch) of cerebral vascular network. Euclidean distance map (EDM) of vascular voxel distance to the nearest background voxel was implemented by Jens Bache-Wiig and Christian Henden [32,33], which calculated distance in three-dimensional Euclidean space. In other words, the distance map value at the vascular skeleton position was the vascular radius (Rn).
Skeleton extraction was performed in segmented images using 3D thinning algorithm [34]. The skeleton image was quantified for extraction of branch points (Bp) (coordinate, branch number (Bn), intersection number (In)) and each vessel length by analyzing the skeleton voxels in the vascular path. The average radius (Rave), vascular total length (L) and vascular total volume (V) could be also calculated.
The length distribution chart visually analyzed the vascular length distribution (Ldis) of the cerebral vasculature. 3D complexity (C) was defined as the ratio of the number of vascular branches to the volume of the drainage volume (C = Bn / Vch), which represented the three-dimensional complexity of vascular architecture. 3D density (ρ) was defined as the ratio of the number of intersections to the volume of the drainage volume (ρ = In / Vch), which represented the degree of directional change in the cerebrovascular blood flow. 3D porosity (P) was defined as the ratio of vascular total volume to the volume of drainage volume (P = V / Vch), which was an important reference for the transport capacity of vascular architecture.
3. Results
3.1 3D imaging of zebrafish cerebral vasculature
As shown in Fig. 3(A), the blood vessels of zebrafish was made of multilayers of different types of cells [35,36]. In transgenic zebrafish Tg(kdrl: EGFP), green fluorescent proteins were expressed in the monolayer of vascular endothelial cells. Therefore, the 3D image of the transgenic vessel (Tg-v) was hollow, as shown in Fig. 3(B). In contrast, during angiography, Evans Blue dye (EB-dye) was injected into zebrafish blood vessels and bound to plasma albumin [16,37]. Therefore, the vessels were demonstrated to be a solid tube in the 3D structure, as shown in Fig. 3(C). In the merged image Fig. 3(D), it is clear that the angiographic image with Evans Blue dye was more continuous and homogeneous, which made it much easier to be segmented and quantitatively analyzed. In order to show a close-up 3D look of the vessel to visually showcase the difference between the hollow and solid vessel structure, we crop the box area vascular image and show in Fig. 3(E-G). Several blood vessels marked in the circle area can be observed that hollow Tg-v surrounds the solid angiographic vascular image. The Tg-v image and EB-dye image did not exactly correspond to each other. The Tg-v image mainly contains brain vessels and very weak signals of eye vessels. In contrast, the angiographic images show vessels in all areas where blood flows, including not only the brain vessels but also the gills and eyes.
Fig. 3. Zebrafish cerebral vasculature obtained by 3D light-sheet imaging. (A) Schematic of vascular structure; (B) 3D image of transgenic zebrafish; (C) 3D angiographic image with microinjection of Evans Blue dye; (D) 3D merged image of Tg-v image and EB-dye image. The white dotted line marked the brain region cropped manually. (E-G) 3D enlarged view of the box area in (B-D). The circle areas visually show the difference between the hollow and solid vessel structure.
3.2 FE-Unet enhanced the transgenetic vessel images into angiography-like
Although angiographic vascular images had better image quality than transgenic vascular images, microinjection was laborious and time-consuming. In addition, the injected dye will be metabolized and disappear from the vessel after some time. Transgenic zebrafish is well suitable for in-vivo developmental studies because their fluorescent protein is expressed as they grow. Therefore, we aimed to develop a deep-learning-based network that enhanced the hollow transgenic vascular image to a columnar vascular image, similar to the angiographic image. FE-Unet was developed for zebrafish vascular enhancement.
We validated the FE-Unet performance by quantitative analysis of the 3D filling effect, robustness, fidelity, and transferability. Figure 4(A) shows the enhancement results of the upper, middle and bottom layers of FE-Unet output images (FE-out image). The solid columnar structure in FE-out image is similar to angiographic images and had a clean background, as in the charts of the white dashed cross-section. The maximum intensity projection (MIP) images also show that the vessels in the FE-out image are filled for the entire vascular network, have continuous structures and less blurred background compared with the input Tg-v image (white oval region), as in Fig. 4(B,C). Moreover, Ultrafine blood vessels were also seen in the enlarged images.
Fig. 4. Performance of FE-Unet network. (A) Images of typical bottom, middle and upper layer of cerebral vasculature for a transgenic zebrafish (input, green) and for FE-Unet enhanced result (output, red); The normalized line intensity profiles were plotted; (B, C) Corresponding MIP images for the transgenic zebrafish and FE-Unet enhanced results; (D) Merged image of Fig. 4(B, C); right shows the enlarged 3D images of the two cyan box regions; (E) Enlarged image of the white box region in Fig. 4(D) and the intensity profile along the white dotted line; (F-H) Quantitative parameters in box & whisker plots evaluate the performance of FE-Unet for five zebrafishes: normalized mean squared (F), peak signal to noise ratio (G) and structural similarity (H). Scale bar: 100 µm.
FE-Unet was robust to vessels of different sizes and signal intensities. Two-channel merged images demonstrated the consistency of two channels cerebrovascular architecture, as shown in Fig. 4(D) (3D render image see Visualization 1). The Pearson correlation coefficient was calculated to be 0.81 suggesting the two images were significantly correlated. In enlarged 3D pictures, thin vessels and thick vessels were shown respectively, illustrating that the network was robust to different vascular intensities and radii.
We further quantitatively examined the fidelity of vascular images filled by FE-Unet. The cross-section along the white line in Fig. 4(E) was plotted, and the normalized gray value curves show the center position and diameter of the three vessels with a distribution close to the standard Gaussian curve. We performed paired t-test on two groups of data and obtained a P value of 0.0729 for the diameter and a P value of 0.4427 for the center position, which suggested that FE-Unet was accurate and high-fidelity. Additionally, we calculated three metrics of the NRMSE, PSNR index and SSIM to evaluate the performance of FE-Unet quantitatively. As shown in Fig. 4(F-H), box & whisker plots show the distribution of all 2D images in each group for three metrics. The results of three metrics for 2D sections along the other two directions and 3D volume are shown in Fig. S1. The statistical analysis quantified that FE-Unet typically achieved well performance in terms of all three metrics.
We also verified the FE-Unet transferability for different state zebrafish, including development periods, postures, transgenic lines and imaging equipment. Fig. S2 illustrated that FE-Unet was universal to different development periods and postures zebrafish. Our method had no special requirement for different imaging equipment. Except for 3D imaging via light-sheet microscope, FE-Unet was also will for vessel imaging via confocal microscope. As shown in the Fig. S3(A), FE-Unet fills vessel image acquired by confocal microscope well. Fig. S3(B) illustrated that our proposed method was also accurate in identifying vessel and did not mis-fill images in non-vessel regions (white arrows) for Tg(fli1:EGFP) line zebrafish.
Recently, Kugler et al. reported the ZVQ tool for zebrafish vessel enhancement and automated quantitative analysis [27]. In ZVQ, a core step is the Tubeness plugin for image enhancement. For 3D image stacks, the plugin used the eigenvalues of the Hessian matrix to fill tubular vessels. Before calculating the Hessian matrix, the entire images was convolved with a Gaussian function with the standard deviation (sigma). Generally, sigma was set as the radius of vessel. Compared with the Tubness plugin, FE-Unet was more robust to vessels of different diameters and different intensities. As shown in Fig. S4(A-D), the architecture of the vessels in the FE-Unet enhanced image was more similar to that in the original transgene image. Most of the vessels enhanced by the Tubeness method are filled, but some of the vessels become intermittent. Although a larger sigma value yielded more continuous vessels, the overall vessel diameters will become larger than the true value. In Fig. S4(E-J), the merged MIP images and segmentation images show the difference between the original transgene images (green) and enhancement images (red). By comparing the enlarged images of the box region, it is clear that FE-Unet enhances the weak signal vessels correctly. While the Tubness method with sigma 4 transforms the thicker vessels as two thin vessels (arrow positions). Increasing the sigma to 8 will fill most vessels well, but many fine blood vessels were overfilled to be much thicker (3D rendering images were shown in Visualization 2).
3.3 Vessel topological parameter extraction from the FE-Unet enhanced image
Extraction of the vascular topological parameters was based on the thresholding segmentation, so the accuracy of the segmentation greatly affects the reliability of vascular parameters. Our FE-Unet enhanced vascular image had a continuous columnar structure and less background, which enables accurate thresholding segmentation and topological parameter extraction.
For comparison, we first manually segmented the vessels of a zebrafish with 3D slicer [38] (Fig. 5(A)). The FE-out image of the same zebrafish was segmented with Otsu- or Huang- thresholding. We found that the Huang thresholding [39] segmentation result was much closer to manual segmentation. The Pearson correlation coefficient (Pearson’s R) was used to assessed the correlation between the manual segmentation image and Huang threshold segmentation image was up to 0.87 (Fig. 5(B)). Therefore, the Huang thresholding was used in the subsequent segmentation in this paper. Moreover, we also compared the accuracy of skeleton extracted by 3D thin algorithm [35].
Fig. 5. Thresholding segmentation and skeletonization extract of the cerebral vasculature. (A) Depth-color-coding image of manual segmentation image; (B) Merged image of manual segmentation (green) and Huang-thresholding segmentation (red); Pearson correlation coefficient 0.87; (C) MIP of manual segmentation skeletonization image. (D-G) MIP skeletonization images with Huang-thresholding, Otsu-thresholding, and Tubeness images with a sigma of 4 or 8. Each Pearson correlation coefficient was compared with the manual segmentation. Red circles mark the region with clear differences. (H, I) Merged MIP image of distance map and skeleton image. Color represents the distance of each voxel. (J) Three views of the 3D convex hull image of the cerebral vasculature. Scale bars represent 100 µm and 10 µm for (A-H) and (I).
The difference in segmentation was clearer in the MIP of the skeletonized images, as shown in Fig. 5(C-G). The Huang thresholding of FE-Unet enhancement images (Fig. 5(D), Pearson’s R 0.81) was the closest to the manual reference standard. Even the skeleton of Otsu thresholding segmentation image was still strongly correlated with the manual reference standard (Fig. 5(E), Pearson’s R 0.7). As shown in Fig. 5(F-G), the intermittent skeleton architecture extracted from Tubness segmentation resulted in a low correlation, with big differences in marked red circles. More detailed comparison were shown in Visualization 3. In the subsequent segmentation of this paper, the Huang thresholding was used for segmentation.
In addition to the skeleton image, the Euclidean distance map (EDM) and 3D convex hull, are also calculated as intermediate results in the pipeline of our algorithm (Fig. 2). The EDM of the threshold segmentation image is a 3D gray image whose gray values represent the distance from each vascular voxel to the nearest vascular edge. The composite image of the EDM and skeleton image and the enlarged image were shown in Fig. 5(H, I). According to the definition of EDM and skeleton, the gray values of the EDM at the skeleton positions are the radii of the vessel. We extracted the radii of the entire cerebral vasculature and showed it in Visualization 4, where the colors represent the radius values. To characterize the feature of the entire cerebrovascular architecture, we approximated the drainage volume of cerebrovascular vessels by the 3D convex hull of the segmentation image (Fig. 5(J)) and defined three parameters: 3D complexity (C), 3D density (ρ) and 3D porosity (P). C and ρ indicate the average density of vascular branch number and intersection number in the zebrafish brain, respectively. 3D porosity, a concept drawn on porous media, is the ratio of the flowing portion volume of the zebrafish brain (vascular volume) to the drainage volume, which reflects the compactness of the zebrafish brain and can indirectly represent the transport properties of the cerebral vasculature.
The topological parameters extracted by our method were quantitatively compared with the results of other methods in Table 1. Significantly, the topological parameters obtained by our method were much closer to the manually labeled reference results. Among them, the average radius, total length and intersection number most directly indicated the accuracy of parameter extraction method. In Tubeness method with sigma 4, the average radius and total length were smaller than the reference values, which indicated that with this method many vessels were lost during segmentation. In addition, the intersection number was much larger than the manually labeled result, which reflected the effect due to the intermittent skeleton.
Table 1. Quantitative comparison of vascular topological parameters extracted by different methods.
3.4 Cerebrovascular development statistics
Quantitative analysis of the vascular network of development zebrafish is important for the study of the mechanism of angiogenesis and vascular diseases. Our method was applied to the quantitative analysis of cerebral vessels in developing zebrafish embryos between 2.5 and 5.5 dpf, and the results were shown in Fig. 6. Using the custom-built light-sheet imaging system, we conducted 3D imaging of cerebral vessels of five Tg(kdrl: EGFP) zebrafishes everday, from the 2.5 dpf to 5.5 dpf. The obtained 3D structures were enhanced with FE-Unet and then segmented with Huang thresholding. The vessel skeletons were shown in Fig. 6(A-D), where the red single-pixel lines represented the skeleton architecture, and the yellow dots marked the positions of the intersections. As expected, the volume and length of the zebrafish cerebrovascular skeleton significantly increase with development, and the vascular network becomes more complex.
Fig. 6. Quantitative analysis of zebrafish cerebrovascular development from 2.5 to 5.5 dpf. (A-D) 3D skeleton of the cerebral vasculature of zebrafish from 2.5 to 5.5 dpf. Yellow dots indicate the intersection points. Scale bar: 100 µm. (E) Vessel length distribution for 5 zebrafish from 2.5 to 5.5 dpf. (F) The average vessel radius of zebrafish was not significantly changed from 2.5 to 5.5 dpf (P = 0.1004). (G) Total length was statistically significantly increased from 2.5 to 5.5 dpf (***P = 0.0007). (H) Vascular volume was significantly increased from 2.5 to 5.5 dpf (****P < 0.0001;). (I) Intersections were statistically significantly increased from 2 to 5 dpf (P = 0.0012). (J) The trifurcate intersection ratio was not significantly increased from 2.5 to 5.5 dpf (P > 0.9999;). (K-M) 3D complexity (**P = 0.0018;) and 3D density (**P = 0.0016;) were significantly increased from 2.5 to 5.5 dpf, but 3D porosity was not significantly changed from 2.5 to 5.5 dpf (P = 0.3087;). Data from F-M are Kruskal–Wallis test and mean ± s.d.
We extracted the topological parameters of cerebral vessels to quantitatively analyze the vessel development process. The path and length of each vessel were extracted, and the length distribution of all cerebral vasculatures was plotted in Fig. 6(E). With the development of zebrafish, the vessel number of various lengths increased obviously, but the distribution did not change significantly. The median, mean and standard deviation were similar.
We next analyzed vascular topology parameters, i.e. the volume, length and intersections of the entire cerebral vasculature. We found that the average radius of cerebral vessels had no significant change from 2.5 to 5.5 dpf (Fig. 6(F); P = 0.1004, Kruskal-Wallis test), but the vascular network length (Fig. 6(G); P = 0.0007; Kruskal-Wallis test), intersections (Fig. 6(I); P = 0.0012; Kruskal-Wallis test) and volume (Fig. 6(H); P < 0.0001; Kruskal-Wallis test), significantly increased from 2.5 to 5.5 dpf. In particular, we found that most of the vascular branches in the cerebrovascular network were Y-shaped bifurcate structures (>95%), with only a few trifurcation structures. The ratio of trifurcate intersections to bifurcate intersections has a slow-growing trend but does not significantly change (P > 0.9999; Kruskal-Wallis test) (Fig. 6(J)). Vascular furcation structure has a significant effect on blood flow properties, so the biological meaning of the ratio change needs to be further studied.
Three topological parameters (C, ρ and P) were also analyzed. As shown in Fig. 6(K-M), the C and ρ grow slowly and then stabilize, and significantly increased from 2.5 to 5.5 dpf. The result indicates that the vascular average density in the zebrafish brain slowly changes with development and eventually reaches a steady state. However, there was no significant change in P, indicating that the volume of cerebral vasculature increased proportionally with the growth of brain volume. The result indicates that the compactness of the zebrafish brain is constant during development.
4. Discussion
The quantitative analysis of vasculature was essential for the zebrafish development study. This paper aims to establish an accurate and automated quantitative analysis tool of zebrafish cerebral vasculature. The hollow structure and sparsity of fluorescence expression of the original transgene vascular image seriously affected the continuity and decreased accuracy of vascular segmentation. With angiographic images as reference, an FE-Unet network was developed to obtain a continuous and columnar vascular structure of transgenic zebrafish line. Thresholding segmentation from images enhanced by our method obtains results closer to manual segmentation. That enables accurate and automated extraction of topological parameters of the cerebrovascular network for the developing zebrafish.
The statistical plots of topological parameters with development show many interesting results. We found that most parameters increase progressively with development, but the average radius of the vessels becomes smaller with development and then becomes constant. The majority of intersection type in zebrafish cerebral vasculature is bifurcated structures, and the ratio of trifurcation to bifurcation increases slowly with development but remain a low value. In addition, three parameters (C, ρ and P) characterizing the overall characteristics of the vascular architecture appear stable after slow growth. Attemptly, we assume that the growth of blood vessels and the natural selection of vascular structures are developing towards lower energy consumption and higher transport efficiency. In this paper, we focus on the establishment of topological parameter extraction methods, and the biological mechanisms of these phenomena will be studied more deeply in the future.
Our method put forward the establishment of a digital vascular network atlas. In principle, our method also can be applied to other specimens than zebrafish, such as transgenic mice. It is expected to be also used for the study of cerebral vascular diseases and screening associated drugs.
Funding
National Natural Science Foundation of China (62205368); Jiangsu Provincial Key Research and Development Program (BE2020664); Strategic Priority Research Program of Chinese Academy of Sciences (XDC07040000).
Disclosures
The authors declare no conflicts of interest related to this article.
Data availability
Data underlying the results presented in this paper are not publicly available at this time but may be obtained from the authors upon reasonable request.
Supplemental document
See Supplement 1 for supporting content.
References
1. P. Gut, S. Reischauer, D. Y. R. Stainier, and R. Arnaout, “Little fish, big data: zebrafish as a model for cardiovascular and metabolic disease,” Physiol. Rev. 97(3), 889–938 (2017). [CrossRef]
2. R. M. White, A. Sessa, C. Burke, T. Bowman, J. LeBlanc, C. Ceol, C. Bourque, M. Dovey, W. Goessling, C. E. Burns, and L. I. Zon, “Transparent adult zebrafish as a tool for in vivo transplantation analysis,” Cell Stem Cell 2(2), 183–189 (2008). [CrossRef]
3. S. Isogai, M. Horiguchi, and B. M. Weinstein, “The Vascular anatomy of the developing zebrafish: an atlas of embryonic and early larval development,” Dev. Biol. 230(2), 278–301 (2001). [CrossRef]
4. M. Ye and Y. Chen, “Zebrafish as an emerging model to study gonad development,” Comput. Struct. Biotechnol. J. 18, 2373–2380 (2020). [CrossRef]
5. E. E. Patton, L. I. Zon, and D. M. Langenau, “Zebrafish disease models in drug discovery: from preclinical modelling to clinical trials,” Nat. Rev. Drug Discovery 20(8), 611–628 (2021). [CrossRef]
6. B. C. Das, L. McCormick, P. Thapa, R. Karki, and T. Evans, “Use of zebrafish in chemical biology and drug discovery,” Future Med. Chem. 5(17), 2103–2116 (2013). [CrossRef]
7. M. G. Butler, A. V. Gore, and B. M. Weinstein, “Zebrafish as a model for hemorrhagic stroke,” Methods Cell Biol. 105, 137–161 (2011). [CrossRef]
8. B. P. Walcott and R. T. Peterson, “Zebrafish models of cerebrovascular disease,” J. Cereb. Blood Flow Metab. 34(4), 571–577 (2014). [CrossRef]
9. J. Chen, J. He, R. Ni, Q. Yang, Y. Zhang, and L. Luo, “Cerebrovascular injuries induce lymphatic invasion into brain parenchyma to guide vascular regeneration in zebrafish,” Dev. Cell 49(5), 697–710.e5 (2019). [CrossRef]
10. J. Xiong, W. Wang, C. Wang, C. Zhong, R. Ruan, Z. Mao, and Z. Liu, “Visualizing peroxynitrite in microvessels of the brain with stroke using an engineered highly specific fluorescent probe,” ACS Sens. 5(10), 3237–3245 (2020). [CrossRef]
11. F. M. Benslimane, Z. Z. Zakaria, S. Shurbaji, M. K. A. Abdelrasool, M. A. H. I. Al-Badr, E. S. K. Al Absi, and H. C. Yalcin, “Cardiac function and blood flow hemodynamics assessment of zebrafish (Danio rerio) using high-speed video microscopy,” Micron 136, 102876 (2020). [CrossRef]
12. S. Daetwyler, U. Günther, C. D. Modes, K. Harrington, and J. Huisken, “Multi-sample SPIM image acquisition, processing and analysis of vascular growth in zebrafish,” Dev. Camb. Engl. 146(6), dev173757 (2019). [CrossRef]
13. G. Yang, L. Wang, X. Qin, X. Chen, Y. Liang, X. Jin, C. Chen, W. Zhang, W. Pan, and H. Li, “Heterogeneities of zebrafish vasculature development studied by a high throughput light-sheet flow imaging system,” Biomed. Opt. Express 13(10), 5344 (2022). [CrossRef]
14. N. E. Buglak, J. Lucitti, P. Ariel, S. Maiocchi, F. J. Miller, and E. S. M. Bahnson, “Light sheet fluorescence microscopy as a new method for unbiased three-dimensional analysis of vascular injury,” Cardiovasc. Res. 117(2), 520–532 (2021). [CrossRef]
15. Z. Chu, J. Lin, C. Gao, C. Xin, Q. Zhang, C.-L. Chen, L. Roisman, G. Gregori, P. J. Rosenfeld, and R. K. Wang, “Quantitative assessment of the retinal microvasculature using optical coherence tomography angiography,” J. Biomed. Opt. 21(6), 066008 (2016). [CrossRef]
16. M. I. Todorov, J. C. Paetzold, O. Schoppe, G. Tetteh, S. Shit, V. Efremov, K. Todorov-Völgyi, M. Düring, M. Dichgans, M. Piraud, B. Menze, and A. Ertürk, “Machine learning analysis of whole mouse brain vasculature,” Nat. Methods 17(4), 442–449 (2020). [CrossRef]
17. G. Tetteh, V. Efremov, N. D. Forkert, M. Schneider, J. Kirschke, B. Weber, C. Zimmer, M. Piraud, and B. H. Menze, “DeepVesselNet: vessel segmentation, centerline prediction, and bifurcation detection in 3-D angiographic volumes,” Front. Neurosci. 14, 592352 (2020). [CrossRef]
18. Y. M. Bradford, C. E. Van Slyke, L. Ruzicka, A. Singer, A. Eagle, D. Fashena, D. G. Howe, K. Frazer, R. Martin, H. Paddock, C. Pich, S. Ramachandran, and M. Westerfield, “Zebrafish information network, the knowledgebase for Danio rerio research,” Genetics 220(4), iyac016 (2022). [CrossRef]
19. N. D. Lawson and B. M. Weinstein, “In vivo imaging of embryonic vascular development using transgenic zebrafish,” Dev. Biol. 248(2), 307–318 (2002). [CrossRef]
20. D. Castranova, B. Samasa, M. Venero Galanternik, H. M. Jung, V. N. Pham, and B. M. Weinstein, “Live Imaging of Intracranial Lymphatics in the Zebrafish,” Circ. Res. 128(1), 42–58 (2021). [CrossRef]
21. W. Liao, B. W. Bisgrove, H. Sawyer, B. Hug, B. Bell, K. Peters, D. J. Grunwald, and D. Y. Stainier, “The zebrafish gene cloche acts upstream of a flk-1 homologue to regulate endothelial cell differentiation,” Dev. Camb. Engl. 124(2), 381–389 (1997). [CrossRef]
22. C. P. Choe, S.-Y. Choi, Y. Kee, M. J. Kim, S.-H. Kim, Y. Lee, H.-C. Park, and H. Ro, “Transgenic fluorescent zebrafish lines that have revolutionized biomedical research,” Lab. Anim. Res. 37(1), 26 (2021). [CrossRef]
23. E. Swanton and N. J. Bulleid, “Protein folding and translocation across the endoplasmic reticulum membrane,” Mol. Membr. Biol. 20(2), 99–104 (2003). [CrossRef]
24. C.-J. Huang, T.-S. Jou, Y.-L. Ho, W.-H. Lee, Y.-T. Jeng, F.-J. Hsieh, and H.-J. Tsai, “Conditional expression of a myocardium-specific transgene in zebrafish transgenic lines,” Dev. Dyn. 233(4), 1294–1303 (2005). [CrossRef]
25. J.-A. Lee and G. J. Cole, “Generation of transgenic zebrafish expressing green fluorescent protein under control of zebrafish amyloid precursor protein gene regulatory elements,” Zebrafish 4(4), 277–286 (2007). [CrossRef]
26. E. Zudaire, L. Gambardella, C. Kurcz, and S. Vermeren, “A computational tool for quantitative analysis of vascular networks,” PLoS One 6(11), e27385 (2011). [CrossRef]
27. E. C. Kugler, J. Frost, V. Silva, K. Plant, K. Chhabria, T. J. A. Chico, and P. A. Armitage, “Zebrafish vascular quantification: a tool for quantification of three-dimensional zebrafish cerebrovascular architecture by automated image analysis,” Dev. Camb. Engl. 149(3), dev199720 (2022). [CrossRef]
28. E. Kugler, K. Plant, T. Chico, and P. Armitage, “Enhancement and segmentation workflow for the developing zebrafish vasculature,” J. Imaging 5(1), 14 (2019). [CrossRef]
29. Y. Sato, S. Nakajima, N. Shiraga, H. Atsumi, S. Yoshida, T. Koller, G. Gerig, and R. Kikinis, “Three-dimensional multi-scale line filter for segmentation and visualization of curvilinear structures in medical images,” Med. Image Anal. 2(2), 143–168 (1998). [CrossRef]
30. N. Otsu, “A Threshold Selection Method from Gray-Level Histograms,” IEEE Trans. Syst. Man Cybern. 9(1), 62–66 (1979). [CrossRef]
31. X. Qin, C. Chen, L. Wang, X. Chen, Y. Liang, X. Jin, W. Pan, Z. Liu, H. Li, and G. Yang, “In-vivo 3D imaging of Zebrafish’s intersegmental vessel development by a bi-directional light-sheet illumination microscope,” Biochem. Biophys. Res. Commun. 557, 8–13 (2021). [CrossRef]
32. F. Leymarie and M. D. Levine, “Fast raster scan distance propagation on the discrete rectangular lattice,” CVGIP Image Underst. 55(1), 84–94 (1992). [CrossRef]
33. G. Borgefors, “On digital distance transforms in three dimensions,” Comput. Vis. Image Underst. 64(3), 368–376 (1996). [CrossRef]
34. T. C. Lee, R. L. Kashyap, and C. N. Chu, “Building skeleton models via 3-d medial surface axis thinning algorithms,” CVGIP Graph. Models Image Process. 56(6), 462–478 (1994). [CrossRef]
35. A. Hasan, A. Paul, A. Memic, and A. Khademhosseini, “A multilayered microfluidic blood vessel-like structure,” Biomed. Microdevices 17(5), 88 (2015). [CrossRef]
36. J. Schöneberg, F. De Lorenzi, B. Theek, A. Blaeser, D. Rommel, A. J. C. Kuehne, F. Kießling, and H. Fischer, “Engineering biofunctional in vitro vessel models using a multilayer bioprinting technique,” Sci. Rep. 8(1), 10430 (2018). [CrossRef]
37. M. P. de, S. Goldim, A. Della Giustina, and F. Petronilho, “Using Evans blue dye to determine blood-brain barrier integrity in rodents,” Curr. Protoc. Immunol. 126(1), e83 (2019). [CrossRef]
38. A. Fedorov, R. Beichel, J. Kalpathy-Cramer, J. Finet, J.-C. Fillion-Robin, S. Pujol, C. Bauer, D. Jennings, F. Fennessy, M. Sonka, J. Buatti, S. Aylward, J. V. Miller, S. Pieper, and R. Kikinis, “3D Slicer as an image computing platform for the Quantitative Imaging Network,” Magn. Reson. Imaging 30(9), 1323–1341 (2012). [CrossRef]
39. L.-K. Huang and M.-J. J. Wang, “Image thresholding by minimizing the measures of fuzziness,” Pattern Recognit. 28(1), 41–51 (1995). [CrossRef]
