1. Introduction

The retina, at the posterior aspect of the eye, is a neurovascular network responsible for photon capturing and preliminary visual information processing [1–3]. As an extension of the central nervous system, the retina offers a unique opportunity for noninvasive examination of the body’s microvasculature [4–6] which can potentially reveal signs of systemic and retinal diseases [7–12]. To assess the health of these blood vessels, one measurement used is the wall-to-lumen ratio (WLR), i.e., the ratio of the blood vessel wall thickness (WTN) to lumen diameter (LD) [13]. Increased WLR has been observed in patients with hypertensive retinopathy [14] and diabetic retinopathy [15,16].

To examine WLR, quantitative LD and WTN analysis is a prerequisite. Traditional imaging modalities, such as fundus photography [17,18], fluorescein angiography [19,20], and indocyanine green angiography [21–23], cannot provide sufficient image resolution and contrast for reliable quantification of WTN and LD. By providing improved resolution, adaptive optics (AO) imagers allow better visualization of blood vessels [24–27], and have been recently used for quantitative measurement and analysis of retinal blood vessels [28,29]. Nevertheless, the widespread adoption of AO imaging remains challenging due to its high equipment cost and system operation complications [25].

By providing three-dimensional imaging capability, optical coherence tomography (OCT) has extensively been used for the clinical management of eye diseases [30–35]. Recently, axial projection OCT has been used to quantify retinal blood vessel diameter and WLR [36–38]. However, this axial projection OCT approach has some limitations. A study on OCT angiography found that angular orientation affects the visibility of microvessels, making vertically oriented vessels more difficult to visualize [39]. This finding suggests that the orientation of the scanning light beam relative to the blood vessel may have implications for the reliability and accuracy of vessel wall visualization, particularly for venous vessels [31,32]. Depth projection OCT, i.e., en face OCT, has also been explored for quantitative analysis of blood vessels [40]. Kim et al., reported a decrease in LD in 5xFAD mice compared to wild-type (WT) controls. However, the study was unable to quantify WLR due to the limited visibility of the vessel wall. We speculate that the limited visibility of the vessel wall structure in en face OCT is due to the signal integration in the whole depth projection image. In principle, individual blood vessels only occupy a small depth range of the retinal thickness, and better depth-resolved segmentation may allow better visualization of the blood vessel wall. One such approach is adaptive segmentation, which tailors the segmentation process to a certain depth range for each blood vessel, excluding unnecessary background noise for high contrast observation of the vessel wall and lumen structures. In this study, we test the feasibility of adaptive segmentation for WLR quantification and to validate its use for detecting WLR abnormalities in 5xFAD mice.

2. Materials and methods

2.1 Animal preparation

All associated animal care and experimental procedures were observed in adherence to the principles outlined by the Association of Research in Vision and Ophthalmology concerning the ethical use of animals in ophthalmic and visual science research. The Animal Care Committee of the University of Illinois Chicago (UIC) validated all experimental protocols implemented. The study employed a sample of six-month-old female 5xFAD mice (N = 8) and WT B6SJLF1/J mice (N = 8). The 5xFAD mice were acquired from Mutant Mouse Resource and Research Center (MMRRC Stock No. 34840-JAX; Jackson Laboratory, Bar Harbor, Maine) and housed within UIC’s Biology Resource Lab. The 5xFAD mice were genetically engineered to overexpress mutated forms of amyloid precursor protein and presenilin-1, resulting in the formation of Aβ plagues in the brain and retina. These mice are commonly used in Alzheimer’s disease (AD) research, as they display various pathological characteristics such as behavioral deficits, neuronal death, capillary stalling, and abnormal tissue viscoelasticity in brain subregions, mirroring some of the key features of the disease. The mice were anesthetized and a drop of 1% tropicamide ophthalmic solution was applied to facilitate imaging. A cover glass accompanied by a lubricant ocular gel (GenTeal, Alcon Laboratories, Fort Worth, TX, USA) was applied to the cornea for the dual purpose of preventing dryness and serving as a contact concave lens to enhance the image quality. The head of the mouse was then immobilized using an animal holder that allowed for five degrees of freedom position controls.

2.2 Imaging

A custom-designed OCT system, as described in our prior work [40–42], was used in this study. A near-infrared superluminescent diode (SLD; D-810-HP, Superlum, Ireland) with center wavelength λ = 810 nm and bandwidth Δλ = 100 nm was selected as the light source. A line CCD camera with 2048 pixels (AViiVA EM4; e2v Technologies, Chelmsford, UK) was used to construct the OCT spectrometer. The frame speed of the camera was set to a 50 kHz A-scan rate. The axial and lateral pixel resolutions were 1.8 µm and 0.7 µm, respectively. Each volumetric OCT consisted of 600 B-scans, with each B-scan comprising of 600 A-scans. Four repeated B-scans were averaged for each frame, resulting in a total of 2400 frames for each volumetric OCT. A power of 1 mW was used for OCT imaging.

2.3 Data analysis

2.3.1 OCT reconstruction and preprocessing

Figure 1(A) illustrates the basic procedure of OCT reconstruction and preprocessing. The reconstruction of volumetric OCT B-scans commenced with the acquisition of OCT interferograms, followed by a series of transformations. The Fast Fourier Transformation (FFT) was applied to the interferograms to generate the initial volumetric OCT data. Resizing assisting registration (RAR) was employed for accurate image registration [43]. Briefly, the volumetric image was resized using a built-in function in MATLAB R2021a (Mathworks, Natick, MA, USA) with “nearest” as the interpolation mode, aiming to make the ratio of the lateral and axial pixel resolution equal to 1. All the subsequent image analysis procedures were performed using Fiji software [44] unless otherwise specified. The volumetric OCT was registered using a rigid registration method. Following registration, the lateral and axial pixel resolution was restored to its initial ratio. Subsequently, tri-linear interpolation was performed across the X, Y, and Z axes, effectively doubling the sampling density in each dimension. This process facilitated better vessel quantification by enhancing the resolution within the volumetric data.

 

Fig. 1. (A) A flowchart of OCT image reconstruction and preprocessing. (B1) En face OCT image. Yellow circles represent radial diameter of 250 µm to 400 µm from ONH. The red window represents the location of vessel tracing. (B2) Full-depth en face vessel image. The yellow boundary represents the location of representative OCT B-scans in (C1). (B3) Filtered full-depth en face vessel image of (B2). (C1) OCT B-scans from the yellow boundary in (B2). (C2) Averaged OCT B-scan of panel (C1). The white arrows indicate upper boundaries used as reference points for straightening. The yellow arrows indicate lumen to IPL range. The orange arrows indicate the full-depth projection range. The green boundaries indicate the straightening area (C3) Straightened OCT B-scans of (C1). (D1) Vessel cross-sectional profile image. The yellow arrows indicate vessel wall projection shedding structure. (D2) Lumen-IPL en face vessel image. (D3) Filtered lumen-IPL en face vessel image. (E). Averaged line profile of (D3). A₁, left wall peak; A₂, right wall peak; B₁, left wall trough; B₂, right wall trough; W₁, left wall thickness; W₂, right wall thickness; L_D, lumen diameter; O_D, outer diameter; ONH, optic nerve head; IPL, inner plexiform layer. Scale bars: 50 µm.

Download Full Size | PDF

2.3.2 Full-depth projection OCT and lateral vessel straightening

The oversampled volumetric OCT was averaged to obtain an en face OCT image (Fig. 1(B1)). Individual vessels were selected between 250 µm to 400 µm (yellow circles in Fig. 1(B1)) from the ONH for the following analysis. Vessel tracing on the en face OCT image (red window in Fig. 1(B1)) was adapted to each vessel’s geometry to cover the vessel and wall region. The selected points were transferred to the oversampled volumetric OCT, and the vessel area was straightened at each depth. The straightened oversampled volumetric OCT was averaged to obtain a full-depth en face vessel image (Fig. 1(B2)). A mean filter was applied to the straightened oversampled volumetric OCT before averaging to obtain Fig. 1(B3). Specifically, the filter was implemented in the horizontal direction (X-direction), following the direction of the vessel leaving the ONH, using a neighborhood size of 10 pixels and 3 pixels in the vertical direction (Y-direction) for better visualization of the lumen and wall structures.

2.3.3 Axial projection OCT and vessel straightening

After the lateral straightening procedure, the vessel volume was resliced to obtain the axial projection OCT B-scans. Figure 1(C1) illustrates the OCT B-scans corresponding to the yellow boundary in Fig. 1(B2). To obtain the axial projection vessel image, the OCT B-scans corresponding to the center of the vessel (yellow window in Fig. 1(B2)) were averaged producing an averaged OCT B-scan (Fig. 1(C2)). The upper vessel boundary (white arrows in Fig. 1(C2)) was used as a reference point for axial vessel straightening. The selected points were transferred to the OCT B-scans and then straightened to obtain Fig. 1(C3). Moreover, artery and vein vessels could be differentiated by observing the vascular morphology on the averaged vessel center B-scan [32].

2.3.4 Adaptive segmentation

Following the axial vessel straightening procedure, a vessel cross-sectional profile image could be constructed (Fig. 1(D1)) by averaging the straightened OCT B-scans (Fig. 1(C3)) along the vessel direction. As shown in Fig. 1(D1), the vessel wall structure was observed around the vessel lumen. The vessel wall projection shedding structure was also observed below the vessel within the inner plexiform layer (IPL). It is essential to note that this segmentation process is adaptive, tailoring to the lateral and axial positions of the blood vessel. Therefore, the lumen to the IPL range (yellow arrows in Fig. 1(C2)) could be selected to produce the depth adaptive segmentation of the en face vessel image (Fig. 1(D2)). The selected region was then applied to the corresponding OCT B-scans, resliced, and averaged to obtain Fig. 1(D2). For better visibility, a mean filter was applied to obtain Fig. 1(D3) which provided better vascular feature visibility compared to the full depth en face vessel image (Fig. 1(B2) and 1B3)

2.3.5 Quantitative analysis of WLR

Upon the application of the adaptive segmentation to obtain the en face vessel image, the WLR was quantified for both arteries and veins. To minimize measurement bias due to the heterogenous vessel wall signal, three rules were applied prior to quantification: 1) regions of strong reflection were avoided, 2) the left and right vessel wall must be present, and 3) the left and right vessel wall thickness must be identical. En face vessel images that didn’t adhere to these rules were discarded while those that followed the rules were quantified. Subsequently, the ­en face vessel image was averaged in the horizontal direction returning an averaged intensity profile. A custom-developed MATLAB function facilitated the analysis of the averaged intensity profile to quantify the WLR. The function used an approach based on the full width of half maximum (FWHM) concept. The function accepted four input arguments, representing the vessel wall peaks and troughs, respectively. It computed the midpoint between each peak and trough and returned the corresponding intensity value collinear to it. The distance between the two midpoints with respect to their corresponding intensity values denoted the left WTN (LWTN) as W₁ and the right WTN (RWTN) as W₂ (Fig. 1(E)). The distance between the two midpoints constituted LD. The total WTN (TWTN) was calculated as the sum of LWTN and RWTN:

The outer diameter (OD) was determined as the sum of the TWTN and LD:

The wall-to-lumen ratio (WLR) was obtained by dividing TWTN by LD:

For statistical analysis, the Mann-Whitney U test was used for the comparison between arteries and veins in WT and 5xFAD mice. P-value < 0.05 was considered statistically significant. Statistical analysis was performed on OriginPro, Version 2023 (OriginLab Corporation, Northampton, MA, USA). In the following section, we will further characterize the difference between the full-depth projection and depth adaptive segmented en face OCT vessel images.

3. Results

3.1 Comparative analysis of full-depth projection and depth adaptive segmented OCT vessel images

Figure 2(A1) shows a representative en face OCT of the WT mouse retina. Figure 2(A2) represents the enlarged view of the en face OCT (yellow window in Fig. 2(A1)) showing arterial (red window) and venous vessel (blue window). In order to provide axial insights, Fig. 2(A3) offers profile along representative arterial vessel (corresponding to the red window in Fig. 2(A2)), while Fig. 2(A4) does the same for representative venous vessel (associated with the blue window in Fig. 2(A2)). Figure 2(B) shows full-depth projection en face vessel images of representative arteries (Fig. 2(B1) and Fig. 2(B2)) and veins (Fig. 2(B3) and Fig. 2(B4)). Figure 2(C) shows corresponding adaptive segmented projection en face vessel images of representative arteries (Fig. 2(C1)) and Fig. 2(C2)) and veins (Figs. 2(C3) and 2(C4)). It was observed that the full-depth en face projection, which involves averaging full retina thickness, does not effectively reveal the distinct structure of the vessel walls. In contrast, the adaptive segmented en face vessel images remarkably elucidate the wall structure clearly. In an effort to gain a deeper understanding of the intensity distribution within these images, averaged A-line profiles across the vessels were acquired. These profiles were computed by averaging pixel intensities across the designated image area. The resulting line profiles were then superimposed onto the images, providing a visual representation of intensity variations within the images. It becomes discernible from this comparative examination that the symmetric wall peak, which is evidently present in lumen-IPL en face vessel images (Fig. 2(C1), Fig. 2(C2), Fig. 2(C3), and Fig. 2(C4)) is absent in full-depth en face vessel images (Fig. 2(B1), Fig. 2(B2), Fig. 2(B3), and Fig. 2(B4)).

 

Fig. 2. Comparison of artery and vein visualization using full-depth and lumen-IPL projection. (A1) Representative en face OCT image. (A2) Enlarged illustration of the yellow window in (A1). The red and blue window represents the artery and vein which is used to obtain the axial profile of (A3) artery and (A4) vein. (B) Representative en face vessel images of arteries (B1, B2) and veins (B3, B4) using full-depth projection. (C) Representative en face vessel images of arteries (C1, C2) and veins (C3, C4) using adaptive segmentation. Red and blue lines in (B and C) represent intensity profiles of arterial and venous vessels. Yellow arrows in C indicate clear vessel wall boundaries. Scale bars: 20 µm.

Download Full Size | PDF

3.2 Comparative analysis of vascular anatomical features using adaptive segmentation in control and 5xFAD mice

After the technical demonstration of using the adaptive segmented ­en face profile analysis for WLR quantification in WT mice, we validated its application for a comparative study of WT and 5xFAD mice. A total of 51 blood vessels were obtained from WT mice, and 55 blood vessels were obtained from 5xFAD mice. In total, 106 blood vessels were used for comparative analysis. Figure 3 shows a comparative study between control and 5xFAD mice. Our comparison of OD revealed no statistically significant changes in either arterial or venous vessels. For arterial vessels, the average OD for control mice was 35.29 ± 2.96 µm compared to 35.08 ± 3.96 µm in 5xFAD mice. Similarly, the OD for veins was 37.02 ± 9.08 µm in control mice versus 37.50 ± 7.71 µm in 5xFAD mice. Both comparisons yielded a p-value greater than 0.05 (Fig. 3(A)). A contrasting outcome was observed when we examined LD. The arterial lumen of 5xFAD mice was significantly thinner than those of the control mice (28.81 ± 2.49 µm versus 26.21 ± 2.88 µm; P < 0.05) (Fig. 3(B)). However, no such statistically significant difference was observed in the vein lumen between the control and 5xFAD mice (32.54 ± 8.40 µm versus 31.10 ± 7.21 µm; P > 0.05). Our analysis of TWTN, depicted in Fig. 3(C), unveiled a highly statistically significant increase caused by AD was observed in both arterial and venous vessels. The arterial WTN increased from 6.48 ± 1.22 µm in control mice to 8.76 ± 1.72 µm in 5xFAD mice (P < 0.01). A similar trend was evident in the veins, where TWTN ramped up from 4.48 ± 1.23 µm in control mice to 6.40 ± 1.00 µm in 5xFAD mice (P < 0.01). Lastly, we evaluated WLR as shown in Fig. 3(D). Corresponding WLR increase was observed in both arterial and venous vessels, confirming WLR as the most sensitive marker of the AD condition. Specifically, the artery WLR increased from 0.23 ± 0.04 in control to 0.34 ± 0.07 in 5xFAD mice (P < 0.001). Similarly, the vein WLR increased from 0.14 ± 0.04 in control to 0.22 ± 0.05 in 5xFAD mice (P < 0.001). In summary, although OD remained relatively constant between the control and 5xFAD mice, the arterial LD in 5xFAD mice was reduced. Additionally, the 5xFAD mice exhibited an increase in both WTN and WLR for arteries and veins, with these differences proving to be significant.

 

Fig. 3. Comparative analysis of outer diameter (A), lumen diameter (B), wall thickness (C), and wall-to-lumen ratio (D) of arteries and veins in control and 5xFAD mice. Statistical significance is indicated in each plot: ns, not significant; * P < 0.05; ** P < 0.01; *** P < 0.001.

Download Full Size | PDF

4. Discussion

In this study, we demonstrated adaptive vessel tracing and segmentation for WLR quantification in OCT. Comparative analysis of WT and 5xFAD mice confirmed AD-induced WLR abnormalities.

In our previous research, we highlighted the challenge in accurately quantifying WLR of veins in mouse retinas using axial projection due to thin upper boundary and hyperreflective bottom lumen [32,40]. This specific challenge has not been reported elsewhere, underscoring the novelty of our observation. Despite this challenge, we successfully visualized distinct wall boundaries in arteries and veins by adaptive segmentation. Moreover, when compared with full-depth projection, our method demonstrates a more prominent wall structure. The improved contrast for wall structure visibility can be attributed to two factors.

First, during the data processing, we observed a shedding pattern below the left and right blood vessel walls (Fig. 1(D1)), which shared the same location and thickness as them. A possible explanation for this phenomenon is that the blood vessel wall might act like a waveguide. This is due to its refractive index, which is different from the blood fluid inside the lumen and surrounding tissues outside of the vessel, which guide partial of the incident light into the vessel wall and follow its orientation as shown in Fig. 4. As a result, these OCT A-lines partially show blood vessel wall structure. As the blood vessel walls have stronger reflectance than the IPL layer, these A-lines showed higher intensities than the surrounding IPL structure. During the data processing, we take advantage of this shedding effect to enhance the overall contrast of the OCT image anatomical features, which aids the differentiation between the WTN and the LD.

 

Fig. 4. Illustration of shedding effect of vessel wall structure.

Download Full Size | PDF

The second factor is managing the effects of blood vessel shadow artifacts and hyperreflectivity of the outer retina layer in OCT. Given that OCT employs light to generate images, any obstruction to this light, such as blood within the vessels, creates a shadow, obscuring the tissue visibility in full-depth projection. While these shadow artifacts can affect overall tissue visibility, they do not significantly impede the visualization of the vessel walls in en face projection. As seen in Fig. 1(D), the shadow artifact falls below the lumen and away from the walls of the vessels. However, the hyperreflectivity of the outer retina layer can interfere with the clear visualization of the vessel walls. Therefore, to preserve the visibility of the vessel walls and lumen, we adaptively segmented out the outer retina layer and selected the lumen to IPL region in the axial B-scans. This selection was made to ensure we are capturing the region where the shedding effect of vessel wall structure is most prominent across all blood vessels, thereby minimizing intralayer contrast. Prior studies have highlighted that correcting blood vessel shadow artifacts by minimizing intralayer contrast and maximizing interlayer contrast, tissue visibility can be enhanced. Cheong et al. proposed a novel deep learning algorithm that was used to remove retinal blood vessel shadows in OCT images of the ONH thereby improving tissue visibility [45]. Girard et al. developed two algorithms that significantly minimize intralayer contrast and maximize interlayer contrast thereby improving the quality of OCT images [46].

Subsequently from our technical demonstration of adaptive segmentation in WT mice, we observed arteries to have larger WTN, narrower LD, and greater WLR when compared to veins, aligned with the results of previous histological studies [47,48]. These differences can be attributed to the fact that arteries receive oxygenated blood at greater pressure, necessitating thicker walls and narrower LD to maintain the pressure of blood moving through the retina. Conversely, veins which withstand a much lower pressure from the deoxygenated blood that flows through them have walls that are considerably thinner and lumens that are correspondingly larger in diameter allowing more blood flow with less resistance [49].

When quantifying the WLR in control and 5xFAD mice, we observed no change in outer diameter of both arteries and veins, a phenomenon we postulate may be due to sensitivity limitations with lateral resolution. However, we noticed a reduction in the LD of arteries unlike veins. Additionally, WTN increase was observed in both arterial and venous vessels with a corresponding increase in WLR. Prior research has established that decreased cerebral blood flow is an early sign of AD [50–52]. We speculate that this reduction in blood flow may result in a concomitant narrowing of the LD. This observation in the 5xFAD mice is intriguing, considering that hypoxia typically induces vessel dilation in the retina. However, the complex interplay between oxidative stress, a known contributor to cell death in AD [53,54], and other disease-specific factors may lead to a unique response. Oxidative stress may be engendered by hypoxia, and arteries typically transport oxygenated blood. Therefore, cell death could be a consequence of oxygen deprivation due to the narrowing of the arterial LD. However, veins which transport deoxygenated blood may experience a slower decrease in LD. Moreover, there is increasing evidence of enhanced atherosclerosis in the vessels in AD [51]. It is conceivable that these accumulations, which frequently occur on the vessel wall, may lead to an increase in WTN, as noted in this study. Simultaneously, it has been reported that Aβ plaques in AD contribute to the thickening of the basement membrane in the vessel wall of both arteries and veins, concomitantly leading to reduced blood flow [55,56]. In light of these findings, the future application of deep learning for automatic adaptive segmentation could expedite the quantification process, thereby allowing for better disease prognosis monitoring and effective treatment assessment.

5. Conclusion

We have demonstrated the feasibility of using adaptive vessel tracing and segmentation in OCT for robust quantification of retinal blood vessel WLR. Comparative analysis of WT and 5xFAD mice validated WLR as a sensitive marker of the AD condition.