Vessel Tech: A high-accuracy pipeline for comprehensive mouse retinal vasculature characterization

Abstract

Mouse retinal vasculature is a well-recognized and commonly used animal model for angiogenesis and microvascular remodeling. Morphological features of retinal vasculature reflect the vessel’s biological functions, and are critical in understanding the physiological and pathological process of vascular development and disease. Here we developed a comprehensive software, Vessel Tech, using retinal vasculature images of postnatal mice. This pipeline can automatically process retinal vascular images, reconstruct vessel network with high accuracy and assess global and local vascular characteristics based on the recent machine-learning techniques. The development of Vessel Tech provides a powerful tool for vascular biologists.

Letter to Editor

Blood vessels constitute the largest networks in our body and play important roles in oxygen transportation, nutrient delivery, inflammatory response, wound healing and respond to many stimuli. A subtle micro-environmental variation can result in significant morphological changes of blood vessels, especially small vessels[1]. It has been appreciated that vascular morphology can reflect its biological functions, and thus, analysis of vessel morphology is critical in understanding physiological and pathological processes in vascular development and diseases[2]. The current study aims at developing a comprehensive pipeline, named Vessel Tech, for automatically characterizing vascular morphology in mouse retina, which is a well-recognized animal model for angiogenesis and microvascular remodeling[3]. From postnatal day 0 (P0) to P7, the first vessel in arises from the optic nerve head and spreads radially over the entire retina following a complex vessel network. Two pivotal processes occur during this period: angiogenesis and micro-vascular remodeling[4–6]. Angiogenesis is the process where new blood vessels grow from existing ones by sprouting, and vascular remodeling is usually expressed as structural alterations such as vessel regression, vessel diameter and area change, and artery-vein specification without the formation of new vessels (Supplementary Fig 1A–C).

Our Vessel Tech pipeline relies on the recent development of convolutional neural network (CNN), and can achieve high accuracy in both image segmentation and vessel network reconstruction. A special CNN architecture U-net model is used to perform vessel segmentation. U-net gets its name from its architecture, which when visualized, appears similar to the letter U. It has been successfully applied in pixel-based image segmentation and will be effective even with limited training images[7]. However, its application to the mouse vascular imaging analysis is still limited. As the current state-of-the art vessel analysis tools, such as Angioquant[8], Angiotool[9], RAVE[10], and Autotube[11] are still using a threshold-based segmentation method with a relatively low accuracy in the vessel segmentation and in later morphological analysis.

The workflow of Vessel Tech contains four modules (Fig. 1A): image preprocessing, vessel segmentation, vessel network reconstruction and retinal vascular feature extraction. Both code and data involved in the training and testing of Vessel Tech can be downloaded through https://github.com/zhengwu/Vessel-Tech. The raw mouse retina images from P2-P7 (there are few vessels before P2) were first pre-processed with grayscale transformation, intensity adjustment, illumination correction and denoising to enhance the image quality. The vessel segmentation step separates the vessels from the background using the U-net model. The output image from the U-net model is a probability map, and we employed the Otsu method[12] to convert the probability map into a binary image.

Fig. 1. Workflow of Vessel Tech (A), and evaluation of its vessel segmentation (B) and vessel network reconstruction (C) accuracy.

(B) Left: Illustration of vascular area dividing with rings. The scale bar represents 1mm. Right: Plots of vessel segmentation accuracy in different rings with Vessel Tech and Autotube. Here 200 pixels equal to 0.25mm. (C) Nodes within a patch of retinal vasculature at P5 were detected by different methods and accuracies were calculated. From left to right: manually annotated ground truth image, Vessel Tech results, and Autotube results. The patch size is 150 * 150 pixels, which is 1/400 of the whole retina image (3000 * 3000 pixels). 150 pixels equal to 187.5μm. The detection accuracy is shown on the top right of each plot.

After having a binary image with 1 indicating the vessel area and 0 the background, we reconstructed the vessel network, e.g., to identify the edges (the skeletons) and nodes (branching points) of the vessels while preserving topological structures. Then relying on rich literature on topological data analysis[13] and network data analysis[14], we extracted various morphological features to delineate the vessel network’s shape, connectivity and topology. Vessel Tech pipeline provides vascular morphological features described in Supplementary Fig.1D, and from these basic features, we can derive many other features to quantify the vessel network, such as spatial density of nodes, and other network-related topological features.

To evaluate Vessel Tech’s performance, we focused on two aspects: (1) the vessel segmentation accuracy, and (2) the vessel network reconstruction accuracy. The vessel segmentation accuracy is defined as $ACC=\frac{{\displaystyle \sum _{i=\{0,1\}}{n}_{i,i}}}{{\displaystyle \sum _{i=\{0,1\}}{\displaystyle \sum _{j=\{0,1\}}{n}_{i,j}}}}$, where n_i,j is the number of pixels in class i predicted as class j. Given that the complexities of vessel network structures vary across different locations, we divided the entire developing retinal vessel network with circles, as indicated in the left panel of Fig. 1B, and evaluated segmentation accuracy of Vessel Tech in the rings independently. As most existing tools for vessel network analysis (such as AutoTube[11] and RAVE[10]) used similar threshold-based methods for vessel segmentation, we only compared Vessel Tech with the newest software – AutoTube[11]. Vessel Tech provided greater accuracy for the segmentation of the whole vessel network (Vessel Tech v.s. Autotube, 92.29% v.s. 90.91%) and local rings, especially in the middle rings where the vessel network becomes more complex (Fig. 2B). We also quantified the accuracy of the recovered vessel network, which is done through measuring the accuracy of network node recovery. To find the correctly identified nodes, we first matched nodes in the reconstructed network to the ones in the ground truth network based on distances between nodes in the two networks. For the matched node-pairs that have distances less than a default value (the same value used for merging nodes in the network reconstruction section; refer to the Materials and Methods), we regarded these nodes in the reconstructed network as a correct detection. The default value used for merging nodes was 21.74μm (5 pixels in our data). The accuracy of the node detection is defined as $ACC=\frac{n_{matched}}{n_{all}}$, where n_matched denotes the number of nodes in the ground truth image that has been detected in the prediction images, and n_all is the total number of nodes in the ground truth image. Evaluated on a patch of P5, Vessel Tech correctly recovered 83.3% of the nodes, while Autotube only recovered 64.3% (Fig. 1C). For the whole vascular network, Vessel Tech achieved 85% accuracy, while the Autotube only had 79%. We noticed that in blurred areas or low contrast areas, Vessel Tech was more sensitive and accurate at detecting the vessel network.

Fig. 2. Illustration of global and local features extracted with Vessel Tech from mouse retinal vasculature at P2-P7.

(A-B) Amplified images of retinal vasculature and its network reconstruction. (A) The whole mount staining of mouse retinal vasculature at P5. (B) The corresponding network images generated by Vessel Tech. A: artery. V: vein. All scale bars represent 1mm. (C-I) The vascular features at P2-P7 were analyzed and their corresponding increasing rates were calculated. (C) total number of nodes (branching points) in the whole network, (D) average vessel length, (E) vessel segments, (F) total vessel area, (G) vessel network diameter (from the center to the edge of vessel network), (H) mean diameter and (I) vessel network diameter of the whole retinal vasculature. In (D), the unit is pixel and 10000 pixels equal to 43.48 mm. (J-L) Vessel diameter and tortuosity changes in precise locations of mouse retinal vasculature at P5. (J) Reference image. The whole retina was first divided into ring areas with circles having the same center with varying radius from 100 to 400, and then further divided into 12 pieces counterclockwise with annotations at each piece (0–360 degrees). The feature statistics (average diameter and average tortuosity) were calculated in each small region and plotted in (K) and (L). The two points a, b in (K) and (L) correspond to the red boxes in the reference image (J). In (K), 4 pixels equal to 17.39 μm.

We first applied Vessel Tech to reconstruct vessel networks and analyze global vasculature characteristics and their changes from P2 to P7. To better visualize the details, we enlarged the images of stained retinal vasculature at P5 and its reconstructed network in Fig.2A–B. We extracted global features of the whole retinal vasculature, such as the number of nodes (branching points), average vessel length, vessel segments, vessel area, and distance from vessel center to vessel edge. These angiogenic features in the whole retinal vasculature demonstrated a time dependent increase with the maximum achieved at P7 (Fig. 2C–I). The increasing rates of these global features had a substantial elevation at P3 and were maintained at high levels until P5, indicating that the peak of sprouting was between P3-P5. At P6-P7, on the other hand, the rates sharply decreased to 10–20%, implying a decrease in sprouting angiogenesis and a possible increase of vessel regression. In contrast, the mean vessel diameter and mean vessel tortuosity were comparable at different times.

The mouse retinal vasculature develops fast during P2 to P7, and it is critical to study local vessel features. Thus, we divided the whole retinal vascular network into 5 regions using 5 circles and summarized the vascular features of these regions in Supplementary Fig. 2. Based on the changes of the angiogenic features in every regions, we assigned three zones: the sprouting zone (zone 40–100% where sprouting angiogenesis occurs), the intermedia zone (zone 20–40% where vessels transition from sprouting to mature), and the mature zone (zone 0–20% where vessel maturation has finished). We further analyzed new specific metrics such as the vessel branching pattern and the vessel diameter distribution et al (Supplementary Fig.2–4). The alternations of diameter and tortuosity in certain vessels indicate special physiological or pathological conditions and requires more precise evaluations; therefore, the whole retina was divided into 12 small ring sectors and annotated with degrees and radii (Fig. 2J). The statistics were calculated in each small region and presented in Fig. 2K–L, providing geographical visualization of vessel diameter and tortuosity.

Vessel Tech has several advantages over current vessel network analysis software programs. a. High accuracy. Utilizing advanced imaging preprocessing procedures and deep learning models, Vessel Tech achieves better vessel segmentation and network reconstruction, especially in regions with complex vessel structures. b. High efficiency. Compared to the existing manual quantification[15], using Vessel Tech can substantially shorten the analysis time. c. Comprehensive vessel features characterization. Vessel Tech provides all features of whole retinal vasculature listed in Supplementary Fig.1D as well as specific features in distinct developmental zones and even in a precise location.

Vessel Tech will be a compelling tool and its future applications include a. Identifying genes responsible for various vascular metrics. Vessel Tech can perform unbiased analyses with high accuracy, and automatically process large amount retinal vascular images from different mouse strains to identify specific genes controlling different vascular metrics. b. Discovering subtle changes of vascular metrics in pathological conditions. Prematurity and diabetic retinopathy are common vascular diseases. Application of Vessel Tech in these disease mouse models may capture previously unnoticeable variations during the progression of the disease and reveal novel insights into their pathophysiology. c. Exploring the communications between EC and other cell types in retina. Combination of staining of different cell markers and EC makers will provide merged images with more sophisticated information[16,17]. Analyzing these images with Vessel Tech will reveal novel regulatory mechanisms mediated by no-ECs on vascular morphology. d. Analyzing vasculature in other organs. The current version of Vessel Tech can be updated to analyze 3D vessel networks to study more sophisticated vasculatures in different mouse models or human organs.