Single-scan adaptive optics-enabled quantitative optical coherence tomography angiography for absolute three-dimensional retinal blood flow mapping

Abstract

Accurate measurement of blood flow is critical for understanding metabolic function and disease progression, particularly in the retina, where conditions such as diabetic retinopathy, macular degeneration, and glaucoma are closely linked to impairment in microvascular circulation. However, current imaging techniques, including optical coherence tomography angiography (OCTA) and Doppler OCT, do not generally provide direct and absolute blood flow measurements, particularly at the capillary level. Adoption of adaptive optics (AO) and high-speed swept-source lasers in OCT systems has enabled video-rate volumetric acquisition with the ability to resolve individual red blood cells (RBC). Here, we present a quantitative AO-OCTA approach that enables 3D blood flow mapping by integrating OCTA-based vessel morphology with a 3D Radon transform of RBC streaks to measure cell velocity from the raw spatio-temporal OCT data. We describe a complete post-processing pipeline to determine single-cell velocity, vessel diameter, and flow rate using a single scan. We demonstrate depth-resolved flow rates across retinal vessels with diameters from 5 to 120 μm and velocities ranging from 0.5 to 54 mm/s, capturing the wide range of flow dynamics in the human retina. Our methodology is a powerful tool for quantitative blood flow imaging, establishing a robust method for noninvasive, high-resolution microvascular flow quantification with multiple potential applications in biomedical research and clinical diagnostics.

1. INTRODUCTION

Dysfunction of retinal blood flow (RBF) has been increasingly recognized as a key contributor to the pathogenesis of major retinal diseases, including glaucoma [1,2], age-related macular degeneration [3,4], and diabetic retinopathy [5,6]. In these conditions, impaired microvascular perfusion and altered hemodynamics are closely linked to neuronal degeneration, hypoxia, and potentially leading to vision loss, underscoring the critical role of blood flow in retinal health and disease progression [7,8]. Retinal blood flow velocity has been reported to decrease by approximately 17% in diabetic retinopathy [9], owing to increased blood viscosity and capillary dropout, whereas in glaucoma, velocity reductions of 15%–30% are attributed to diminished perfusion pressure [10,11]. A growing body of clinical and experimental evidence supports the association between flow abnormalities and disease severity, highlighting the importance of accurate methods to quantify retinal microcirculation. Precise measurement of RBF is therefore essential to understand the underlying mechanisms of retinal disorders, to advance early disease diagnosis, to monitor disease progression and therapeutic response, and to guide the development of novel interventions.

Various techniques have been developed to quantify RBF, ranging from invasive methods to non-invasive imaging modalities [12,13]. Invasive approaches, such as fluorescein angiography and indocyanine green angiography, have long been used to assess retinal circulation, providing valuable insights into perfusion dynamics. However, these methods rely on contrast agents, which introduce potential risks and limitations, particularly in a clinical setting. As a result, non-invasive imaging techniques are increasingly favored for their safety, convenience, and rapid scanning capabilities. Among these, Doppler-based methods, such as laser Doppler velocimetry [14], laser Doppler flowmetry [15], Doppler optical coherence tomography (OCT) [16], and color Doppler imaging [17], have been widely employed to measure blood velocity through frequency shifts in reflected light. Advanced Doppler approaches include three-beam Doppler OCT for total retinal blood flow measurement [18] and Doppler holography methods that enable blood flow imaging in both anterior and posterior segments [19,20]. While these techniques are non-invasive and enable accurate flow measurements, they are limited by poor spatial resolution, inability to resolve flow in depth, single-point flow measurement (in some cases), and dependency on accurate Doppler angle estimation. Speckle-based approaches, such as laser speckle imaging, estimate relative flow by using speckle correlation times, which inversely relate to mean velocity [21,22]. However, converting the width of the power spectral density in laser Doppler methods or correlation times in laser speckle methods into absolute flow is challenging due to the unknown quantity and orientation of cells within the scattering volume. Moreover, both laser Doppler and speckle-based techniques lack depth discrimination, which hinders precise flow localization within the retinal layers. Collectively, these limitations highlight a critical gap: the absence of a method that is non-invasive, depth-resolved, and capable of directly quantifying absolute blood flow in microvasculature. A direct method offering three-dimensional, spatially resolved, and absolute RBF measurements would significantly enhance experimental and clinical evidence of RBF alterations in retinal diseases.

In parallel, optical coherence tomography angiography (OCTA), a functional extension of OCT, has rapidly emerged as a cornerstone technology in biomedical imaging, enabling clinicians and researchers to visualize microvascular networks in vivo without the need for exogenous contrast agents [23]. Since its inception, OCTA has provided unprecedented insight into retinal vascular plexus by harnessing the intrinsic motion contrast of circulating RBCs. Although OCTA provides impressive mapping of the retinal vessels, it does not directly measure flow velocity. Instead, vessel visualization is accomplished by extraction of OCT signal fluctuations that arise from moving RBCs. Attempts have been made to provide quantitative flow metrics from OCTA images (i.e., a flow index based upon vessel density) that yielded high variability across different devices [24]. To address some of these shortcomings, Choi et al. [25] introduced the variable interscan time analysis (VISTA) OCTA technique, which enables differentiation of blood flow speeds within a specific relative range—from the smallest detectable to the largest distinguishable flux. Temporal autocorrelation methods, while not directly resolving single RBCs without adaptive optics, provide quantitative, repeatable measurements that are sensitive to pulsatility and normal aging [26–28]. Other approaches utilize dynamic forward scattering signals [29] and correlation ratios [30] to link autocorrelation measurements to blood flow speeds, enabling flow quantification at both large vessel and individual capillary levels. However, absolute quantitative OCTA remains elusive, yet essential, to enable objective interpretation of clinical data.

Adaptive optics (AO) has enhanced imaging modalities like confocal scanning laser ophthalmoscopy (SLO), flood illumination, and OCT, enabling in vivo cellular-level spatial resolution by compensating for both lower- and higher-order ocular aberrations [31–33]. The combination of AO with OCTA has been demonstrated to visualize retinal micro-capillaries and choriocapillaris with enhanced resolution [34–37]. Several groups have demonstrated the measurement of absolute RBF measurement with AO-enabled systems [38–40]. For instance, AO-SLO operated in a custom line-scan mode, where the slow axis scanner is paused to produce spatio-temporal images containing RBC motion traces or streaks, has been used to measure the absolute distribution of blood velocity [22]. This technique resolves densely packed RBCs flowing through medium to large vessels and captures single-file flow within capillaries [39,40]. Our group recently demonstrated repeatable RBF measurements with the AO-SLO line-scan method in healthy volunteers [41]. Retinal blood cell velocity has also been quantified using a high-speed flood illumination ophthalmoscope [42], an AO-enabled partially confocal multi-line ophthalmoscope [43], a dual-channel AO-SLO [44], and a high-speed AO-SLO system [45]. Despite their promise, these direct approaches are restricted to two-dimensional (2D) imaging (no depth information) and typically permit measurement of only one vessel at a time within the field of view (FOV). In particular, the AO-SLO line-scan method underestimates velocity when imaging non-planar vessels (e.g., optic nerve head vessels) [46].

A robust and quantitative method for RBF imaging must ideally meet the following key criteria for clinical utility: (1) provide three-dimensional depth-resolved measurements; (2) offer absolute quantification without the need for calibration or indirect inference; (3) capture a wide dynamic range of flow rates and vessel calibers, including single-file flow in capillaries, within the same FOV; (4) simultaneously extract blood cell velocity and vessel morphology from the same dataset for accurate flow measurement; (5) achieve rapid acquisition (<1 s); and (6) be insensitive to vessel orientation within the tissue.

We present an AO-enabled quantitative OCTA (AO-qOCTA) methodology, inspired by advances in both AO and OCTA, that paves the way toward fulfilling all key criteria for quantitative blood flow imaging. The key insight lies in leveraging sufficient spatio-temporal resolution—achieved in the lateral and axial spatial domain, respectively, with AO and OCT and in the temporal domain with an ultrafast laser—thereby enabling visualization and direct quantification of RBC streaks within an AO-OCTA volume. Individual absolute RBC velocity and vessel morphology are measured with a standard OCTA (multiple B-scan) volumetric acquisition. We demonstrate high-fidelity streak detection and vessel segmentation, addressing key limitations of conventional OCTA. We present a comprehensive post-processing pipeline for voxel-wise flow rate computation. Unlike the AO-SLO analog, our method captures both axial and transverse velocity components and enables true depth sectioning. We validated the approach using a flow phantom, line-scan AO-SLO imaging, and the conservation of flow principle in a branching vessel. Additionally, in vivo imaging of healthy volunteers highlights the ability to discriminate overlapping vessels, detect laminar flow profiles across larger vessels, and capture both pulsatile and capillary-level flow. This direct measurement approach not only improves flow quantification accuracy but also facilitates the construction of 3D flow rate maps in absolute volumetric units (μL/min). Crucially, the method integrates seamlessly into current OCTA systems without requiring hardware modifications, assuming sufficient spatio-temporal resolution can be achieved to resolve RBC streaks. By enabling direct, absolute flow rate measurement, our method holds significant potential for transforming early diagnosis and management of retinal vascular diseases via sensitive, quantitative vascular biomarkers.

2. METHODS

A. AO Imaging

Six healthy volunteers (H1–H6; age: 37.3 ± 12.4 years) were enrolled in this study (Table S1). Written informed consent was obtained from all participants after explaining the potential risks. The study protocol was approved by the Institutional Review Board of the U.S. Food and Drug Administration (FDA) and adhered to the tenets of the Declaration of Helsinki.

Details of the Fourier-domain mode-locked (FDML) adaptive optics system used in this study have been reported previously [47]. Briefly, the system includes both AO-OCT and AO-SLO channels where the OCT channel was used for AO-qOCTA data collection, and the SLO channel was used for real-time visualization and to collect validation data in line-scan mode. The AO-OCT system employs a 3.348 MHz FDML laser (NG-FDML-1060-8B-FA, OptoRes GmbH, Munich, Germany) with a center wavelength of 1060 nm and a bandwidth of 76 nm, providing a nominal axial resolution of 8.4 μm in the eye (assuming refractive index n = 1.38). The resonant scanner operates at 3.270 kHz in a bidirectional acquisition mode (512 pixels), yielding an effective OCT B-scan rate of 6.539 kHz (f_fast). The slow-axis galvanometer operates at 12.77 Hz (f_slow) with 512 pixels. All system timing and synchronization are derived from the FDML laser.

The standard AO-SLO mode was used to identify the region of interest (ROI) following pupil dilation with 1% tropicamide. Similar to the 2D analog AO-SLO approach [38], target imaging regions were selected such that the majority of vessel orientations formed a shallow angle (χ < 15°) to the AO-OCT fast scan direction, as shown in Fig. 1. This shallow angle facilitates the generation of discrete RBC streaks in space–time images, enabling more reliable velocity measurement. For AO-OCTA imaging, the acquisition protocol covered a 1.5° × 1.5° FOV with eight repeated B-scans (four forward and four backward) acquired at each slow scan position. The system focus was positioned at the superficial vascular plexus of the inner retina to maximize vascular signal. Streaks embedded within the repeated B-scans (space–time images) represent individual RBC motion traces. Given the known system parameters (i.e., B-scan speed and OCT A-scan spacing), the absolute velocity of each RBC can be computed by analyzing its projections along the three axes (Fig. 1).

Fig. 1.

Illustration of RBC flow through the imaging plane. A raster-scanning AO-OCT beam captures backscattered signals from moving RBCs intersecting the imaging plane (orange vector denotes RBC velocity vector). At each slow-scan position, repeated B-scans record RBC motion as streaks in the space–time domain, reflecting spatial displacement over the elapsed scan time. The streak characteristics depend on the vessel orientation relative to the fast axis, with corresponding projection angles (θ, φ, and χ) measured in the X-Y (space–time), Y-Z (space–time), and X-Z (space–space) planes.

B. Method Validation

To validate the proposed method, we employed a combination of flow phantom experiments and in vivo comparisons. First, a compact linear flow phantom developed by our group was used to verify velocity measurements against known target values [48]. Briefly, the phantom consists of a linear conveyor belt driven by a high-torque DC motor, with ~10 μm diameter beads uniformly deposited in a monolayer to mimic RBCs. In this study, the phantom was positioned at a 30° angle relative to the fast scan axis to simulate vessel inclination. The motor was driven at calibrated voltages corresponding to known velocities. Second, to compare with an established method, we imaged the same region using both the AO-qOCTA approach and the AO-SLO channel operating in line-scan mode to acquire space–time videos. We previously described analysis of AO-SLO video to extract velocity using a custom algorithm developed in MATLAB software (R2018a, Version 9.2, MathWorks, Natick MA, USA) [41]. Last, we used the conservation of flow principle to image branching vessels in a healthy volunteer.

C. OCTA Processing Pipeline

The acquired AO-OCTA volumes were reconstructed, and the resonant scanner nonlinearity was corrected [47]. Eye motion along the fast scan direction was corrected using a cross-correlation function in MATLAB, and the volume is flattened to the outer retina. The pre-processed volume was then input to two separate pipelines (Fig. S1) to derive vessel morphology (OCTA processing pipeline) and velocity (OCT flow processing pipeline).

The OCTA image was generated using an amplitude decorrelation approach [36]. Decorrelation (D) of the OCT amplitude signal (A) was calculated across N repeated scans with

$$D_v=1-\frac{1}{N-1}\sum _{k=1}^{N-1}\frac{A_{v,k}*A_{v,k+1}}{1/2({A_{v,k}}^2+{A_{v,k+1}}^2)},$$ (1)

where k denotes the B-scan index, and υ denotes the voxel indices of the OCT signal. Bulk target motion-induced artifact is minimal owing to the high B-scan rate (6.539 kHz). To extract and quantify vessel morphology from the AO-OCTA volume and quantify vessel diameter a custom algorithm was developed in MATLAB. Prior to the morphological feature extraction step, Hessian-based Frangi vesselness filter [fibermetric()] was applied on the Gaussian filtered (σ = 4−15, spanning capillaries to the largest vessels) OCTA volume for enhancement of tubelike structures and suppression of non-tubular structures and background noise [49], thereby greatly improving the accuracy of OCTA vessel segmentation. The Frangi filter selectively enhances the intensity of individual voxels based on their “vesselness,” which is quantified using the eigenvalues of a local Hessian matrix. MATLAB functions for skeletonization and the Euclidean distance transform [bwskel() and bwdist()] were used on the binarized volume to measure voxel-wise vessel orientation and diameter (Fig. 2).

Fig. 2.

Post-processing pipeline for 3D flow rate quantification using AO-qOCTA. A pre-processed reconstructed OCT volume is processed through two parallel chains: OCTA analysis for vessel morphology and OCT flow analysis. The amplitude decorrelation method visualizes retinal vasculature, which is enhanced with Frangi filtering to generate a 3D vessel mask. Vessel diameter and orientation in 3D are extracted using MATLAB-based centerline tracing and Euclidean distance calculations. Concurrently, RBC motion streaks in the OCT intensity volume (Visualization 1) are denoised and Sobel filtered to suppress vertical static features. A custom 3D Radon transform determines streak orientation, which, combined with vessel orientation, enables voxel-wise velocity measurement. Vessel diameters and velocities are then integrated to compute absolute flow rates.

D. OCT Flow Processing Pipeline

To reduce speckle noise and improve streak detection and angle measurement accuracy, the shearlet domain thresholding method [50] greatly improves SNR while preserving edge details essential for our RBC streak analysis. A square root transformation was first applied to the pre-processed spatio-temporal OCT volume before 3D shearlet transformation. A fixed threshold (2.5) was applied to the computed 3D shearlet coefficients as described by Yang et al. [51]. Further, application of a Sobel filter, similar to a high-pass filter, removes the static structural features that appear as bright vertical lines in the pre-processed volume.

To automatically quantify individual particle velocity within each image, a sliding ROI defined by the longest streak length (15–30 square pixels) is applied to the two space–time projection planes (X-Y and Y-Z). A circular mask on each ROI is implemented to avoid Radon artifacts. The Radon transform is implemented in an iterative manner for each ROI to achieve higher angle precision (step angle ~0.001°) in a shorter time [52]. The maximum variance angle provides the dominant streak orientation (θ in X-Y and φ in Y-Z). Furthermore, we use the OCTA vessel mask to reduce the number of ROIs by restricting analysis to vessel regions, which significantly decreases the computational load.

To compute 3D particle velocity, knowledge of the vessel orientation to the fast axis scan direction is required, as it defines the direction of particle motion. This is derived from the segmented OCTA image by calculating the local gradient of the vessel centerline. The magnitude of the particle velocity can be calculated with

$$v_{\text{particle}}=\sqrt{{{(f*\text{cot}\theta *\text{sec}\chi )}^2}_{XY}+{{(f*\text{cot}\phi *\text{sec}\chi )}^2}_{YZ}},$$ (2)

where υ_particle is the local individual particle velocity (mm/s), f is equal to the B-scan rate (Hz) multiplied by sampling density (μm/pixel), θ and φ are streak angles measured on the X-Y and Y-Z planes relative to the fast axis, and χ is the vessel orientation to the fast axis. In Eq. (2), the X-Z plane projection is neglected as only the X-Y and Y-Z planes form space–time information whereas the X-Z plane only forms space–space information. A conservative SNR threshold of 2.5, defined in Eq. (S1) (Supplement 1), was selected based on visual assessment to determine the reliability of streak angle measurements within the measurable range [41]. Voxel-wise flow rates were computed by combining velocity and lumen diameter assuming a circular cross-sectional area for the retinal vessels:

$$Q=\sum _{i=1}^n{v}_i*\frac{\pi *{\text{Diameter}}_i^2}{4},$$ (3)

where Q is the mean blood flow rate in μL/min. OCT flow processing is the most time-consuming step, with the 3D Radon transform being the primary computational limiting factor. To maximize accuracy, we used a 0.001° angular step size, resulting in a total processing time of ~12 h per OCT volume with N = 8 repeats on a standard CPU.

All statistical analyses were performed in Microsoft Excel using the Analysis ToolPak add-ins. Agreement between ground-truth and AO-qOCTA velocity measurements in the flow phantom was assessed using the intraclass correlation coefficient (ICC) (two-way mixed-effects model). The effects of system parameters were evaluated using repeated-measures ANOVA, with statistical significance set at p < 0.05.

3. RESULTS

A. Method Validation

We validated our method with three distinct experiments. First, we performed a flow phantom experiment with a known ground-truth velocity of 4.16 ± 0.15 mm/s. Processing data through the full AO-qOCTA pipeline yielded a measured velocity of 4.03 ± 0.19 mm/s, demonstrating excellent agreement with the ground-truth velocity values [ICC(3, 1) = 0.973]. Second, we conducted a cross-validation study using AO-SLO line-scan measurements on the same retinal vessel in the macular region, where the microvasculature is predominantly planar (Fig. S2). The velocity measured by AO-SLO was 9.21 mm/s compared to 9.01 mm/s from AO-qOCTA, corresponding to a 2.2% difference. Last, we evaluated conservation of flow at a retinal bifurcation (Fig. S3). The mean flow rate in the parent vessel was 0.309 μL/min, while the sum of the flow rates in the two daughter branches was 0.296 μL/min. This small difference (~4.3%) supports the sensitivity and quantitative accuracy of the AO-qOCTA technique. A summary of the validation results is shown in Table 1. Overall, the method demonstrated robust performance across phantom, cross-modal, and physiological validation scenarios.

Table 1.

AO-qOCTA Validation Summary

Experiment		Diameter (μm)	Mean Velocity (mm/s)	Mean Flow Rate(μL/min)
Flow phantom	Set velocity	–	4.16 ± 0.15	–
	AO-qOCTA	–	4.03 ± 0.19	–
Cross-validation	AO-SLO	–	9.21 ± 0.35	–
	AO-qOCTA	40.13 ± 1.52	9.01 ± 0.21	0.683 ± 0.08
Conservation of flow rates	Parent	26.6 ± 1.3	9.24 ± 0.71	0.309 ± 0.05
	Daughter 1	20.1 ± 1.0	8.39 ± 0.59	0.159 ± 0.03	0.296 ± 0.05
	Daughter 2	16.9 ± 0.8	10.17 ± 0.74	0.137 ± 0.02

B. Depth-Resolved Quantitation

Our method enables volumetric quantification of blood flow with depth specificity, a key advancement over traditional approaches. Figure 3 shows a 3D flow rate map, highlighting spatial variation in flow across vessel branches in depth. A cross-sectional velocity profile extracted from an individual vessel exhibits clear laminar (parabolic) flow, further validating the accuracy of our measurements.

Fig. 3.

Depth-resolved flow quantification using AO imaging. A target vessel identified from a clinical SLO scan was imaged across four single AO scans. AO-OCTA analysis revealed a densely connected 3D vascular network. Simultaneous analysis of AO-OCT intensity data enabled computation of a depth-resolved absolute flow rate map. A representative cross-sectional velocity profile demonstrates the expected parabolic flow within the vessel, which can be extracted at any lateral position along its course. The black mask represents the inner lumen boundary of the measured vessel.

Additionally, our approach enables clear discrimination of overlapping vessels, which often confound conventional 2D modalities such as AO-SLO. As shown in Fig. 4(a), vessel crossovers occurring at different depths are successfully resolved with AO-OCTA, and depth-specific flow rate measurements reveal parabolic velocity profiles in both vessels. Furthermore, detailed analysis of streak orientation indicates that flow direction influences the angle of streaks, appearing in either the first (0° to 90°) or second quadrant (90° to 180°), as illustrated in Fig. 4(b). By incorporating prior anatomical knowledge of vessel origins at the optic nerve head, we were also able to infer vessel type, demonstrating the capability to distinguish arterioles from venules within complex vascular networks. These findings underscore the capability of our method to deliver precise, depth-resolved flow quantification with vessel-level specificity.

Fig. 4.

Discriminating flow in overlapping vessels. (a) AO-OCTA analysis revealed overlapping vessels. (b) Careful inspection of streaks in the AO-OCT intensity volume revealed directionality based on flow direction (yellow arrows) at different depths. Together with the knowledge of branching vasculature from the optic nerve head, we could differentiate vessel type (arteriole or venule).

To demonstrate flow measurement accuracy for a wide variety of retinal vessel sizes and flow directions using the AO-qOCTA approach, we measured flow velocity in an inter-plexus connection in the macula (Fig. 5) and an optic nerve head (ONH) vessel segment with a steep axial angle (γ = 34.3°) (Fig. S4). Flow measurement across the three vascular plexuses, including non-planar inter-plexus connections in the inner retina, is achieved with tightly controlled system focus set to the inner nuclear layer (INL) with the deformable mirror (Fig. 5). The flow velocity in one inter plexus connecting vessel (red arrow) was found to be 6.9 mm/s, where the X-Y and Y-Z components are 5.6 and 4.0 mm/s, respectively. For the ONH vessels (Fig. S4), our approach enables accurate flow velocity measurement even along curved segments, yielding a velocity of 54.3 mm/s. In contrast, the 2D analog AO-SLO technique, which assumes planar geometry, underestimates the velocity by 18.8% (υ_xy = 5.6 = mm/s) and 13.8% (υ_xγ = 46.8 mm/s) for the inter-plexus connecting vessel (Fig. 5) and the ONH vessel (Fig. S4), respectively, due to its inability to resolve the axial velocity component of flow (υ_yz = 4.0 mm/s; υ_γz = 27.6 mm/s).

Fig. 5.

AO-qOCTA depth measuring capability enables absolute flow quantification within the inter-plexus connection. (a) Fine focus control enabled by the AO system allows analysis of the human retinal vasculature across three distinct plexuses: the superficial vascular complex (SVC), the intermediate capillary plexus (ICP), and the deep capillary plexus (DCP). Panels below show representative RBC streaks observed across each plexus in both X-Y and Y-Z. Several streaks in the Y-Z plane from the interconnecting capillary are denoted by the cyan arrows. (b) Flow within the inter-plexus connection (red arrow) linking all three vascular plexuses as visualized in the AO-OCTA B-scan was also quantified. Scale bar is 50 μm.

C. Dynamic Range

To evaluate the versatility of our method across different flow regimes, we assessed the dynamic range of measurable RBC velocities and vessel calibers. As shown in Fig. 6, the AO-qOCTA technique successfully quantified flow in both large retinal vessels near the optic nerve head and in fine capillary networks surrounding the foveal avascular zone (FAZ). The accompanying fundus image delineates the macula, FAZ, and parafoveal regions, where we captured a spectrum of flow profiles. We found that configuration with higher repeated B-scans improved streak visibility and extended the lower measurable velocity to 0.5 mm/s (Fig. S6). Notably, our method was sensitive enough to resolve slow, single-file RBC movement (down to 0.5 mm/s velocity) within narrow capillaries adjacent to the FAZ, while also maintaining accuracy in higher-velocity flow up to 54 mm/s within arterioles and venules. This ability to measure both low and high velocities within vessels of varying diameter highlights the broad dynamic range and sensitivity of the AO-qOCTA technique, enabling detailed characterization of microvascular perfusion across the retina.

Fig. 6.

Dynamic measurement range of velocity and vessel calibers. (a) Representative AO-OCTA visualization of FAZ, and (b) the AO-enabled qOCTA method enables robust measurement of slow, single-file flow within capillaries near the FAZ. (c) The fundus image from a healthy subject highlights the macula, FAZ, and parafoveal region. (d) Blood flow velocities were quantified from five subjects in vessels spanning large optic nerve head vessels to capillaries in the parafoveal region. Error bars denote standard deviation.

D. Effect of System Parameters

Our method enables flexible tuning of spatio-temporal acquisition parameters to capture dynamic flow characteristics such as pulsatility. To investigate this capability further, we modified the system to acquire volumes at rates of 1.6, 3.2, 6.4, and 12.8 vols/s by progressively reducing the number of lateral samples in the slow-scan direction (512 × 512, 256 × 512, 128 × 512, and 64 × 512 pixels, respectively) as shown in Fig. 7(a). This trade-off in spatial resolution provided increased temporal sampling necessary to resolve flow dynamics over the cardiac cycle. By acquiring consecutive volumetric datasets at each volume rate, we were able to visualize temporal variations in velocity within individual vessels. Figure 7(d) shows representative pulsatile flow waveforms, with the best-fit curve capturing the characteristic velocity fluctuations across the cardiac cycle. These results demonstrate that appropriate parameter tuning enables robust sampling of cardiac-driven flow dynamics, while preserving quantitative accuracy across different spatio-temporal configurations. Additionally, we examined the impact of lateral pixel density and the number of repeated B-scans [Figs. 7(b), 7(c), 7(e), and 7(f)]. First, in our flow phantom, we confirmed that measurements across different FOVs (1, 1.5, 2, and 2.5°; 6.21 ± 0.04 mm/s; p = 0.383) and different number of repeated B-scans (2, 4, 8, and 16; 6.16 ± 0.12 mm/s; p = 0.495) were 98.3% accurate to the set velocity of 6.1 mm/s. Next, in human subject imaging, we confirmed that measured mean velocity remained stable across FOV (3.76 mm/s; p = 0.875) and repeated B-scan (5.94 mm/s; p = 0.893) settings. The large variability for in vivo velocity measurements across multiple datasets is attributed to pulsatility in these vessels, further underscoring the importance of collecting flow measurements across the entire cardiac cycle for larger vessels.

Fig. 7.

Tuning of spatio-temporal parameters. Panels (a) and (d) illustrate pulsatile flow measurement in the microvasculature by analyzing consecutive frames of the same vessel acquired with adjusted slow-scan parameters. Temporal resolution was enhanced by reducing spatial sampling along the slow-scan axis, enabling appropriate sampling of the human cardiac cycle (~60 beats/min). (b),(e) Separately, varying the FOV from 1° to 2.5° and (c),(f) the number of repeated B-scans altered lateral pixel density and the dwell time at each slow scan position, without affecting the measured mean velocity at a given vessel segment in the presence of pulsatile variability. Scale bar is 100 μm.

4. DISCUSSION

We developed an AO-enabled qOCTA technique to overcome longstanding limitations in RBF imaging by enabling direct, absolute, and depth-resolved measurements. While prior research has examined quantitative assessments of RBF, these investigations have largely relied on indirect or 2D measurements. Building on the principle of the 2D AO-SLO line-scan method, we recognized that the faster imaging speeds of our AO system allowed 3D RBC motion traces (or streaks) to be captured directly within the volumetric OCTA scans. The 3D flow rates are thus calculated from the combination of streak orientation quantification enabling individual RBC velocity measurement and vessel morphology quantification from the AO-OCTA volume. We have demonstrated the complete methodology, from image acquisition through data processing to absolute flow quantification from a single scan, highlighting its potential for robust and precise RBF assessment.

Our method was validated using a three-pronged approach, employing a flow phantom, cross-validation with line-scan AO-SLO, and conservation of flow analyses, to confirm the accuracy of the AO-enabled qOCTA methodology. One of the most significant advantages of our methodology is the ability to provide depth-resolved quantification of blood flow. By accurately capturing the full three-dimensional geometry from the same single scan, our approach ensures that flow measurements incorporate both axial and transverse components, thus delivering a comprehensive assessment of microvascular perfusion. In contrast, the 2D analog AO-SLO line-scan method measures the blood flow velocity component parallel to the en face plane, thereby limiting access to retinal vessels due to significant errors induced by axial orientation [38,46]. We believe AO-qOCTA provides a more generalizable method, with AO-SLO line-scan representing a special case applicable primarily to planar vasculature.

The depth-sectioning capability of OCT even allows for the differentiation and flow assessment of overlapping vascular structures. As depicted in Fig. 4, our method resolves streaks in vessels that cross over at different depths—a relatively common feature where 2D imaging methods are inadequate. In these regions, quantification of flow is further challenged by vessel curvature. Our analysis revealed that even in segments where a venule passes beneath an arteriole with an axial bend, the AO-qOCTA technique accurately computes flow velocities by accounting for the full 3D orientation. The axial component measurement with our method is further illustrated for a representative ONH vessel in Fig. S4. In contrast, planar techniques underestimate flow due to their inability to resolve out-of-plane components. Our demonstrated measurement precision of <5% (Table 1) provides sufficient sensitivity to detect the 15%–50% velocity reductions commonly reported in diabetic retinopathy [9,53], glaucoma [10,11], and age-related macular degeneration [4,54]. We demonstrate a broad dynamic range for measuring blood flow velocities and vessel calibers across the retina including high flow in large vessels near the optic nerve head, as well as slow, single-file RBC motion in capillaries near the FAZ. This expansive dynamic range, spanning velocities from approximately 0.5 to 54 mm/s, is well suited for comprehensive studies of retinal microcirculation, facilitating investigations into conditions where both macrovascular and microvascular changes occur concurrently.

The ability to tune spatio-temporal resolution is vital for capturing rapid hemodynamic changes. Our experiments showed that by reducing the number of lateral pixels in the slow-scan direction, the acquisition rates can be increased from 1.6 to 12.8 volumes per second, thereby enhancing temporal resolution. This optimization was crucial for visualizing pulsatile flow dynamics corresponding to the human cardiac cycle. Moreover, the observed insensitivity of measured mean velocities to variations in lateral pixel density and number of repeated B-scans, apart from natural cardiac cycle-related fluctuations, affirms that the AO-qOCTA methodology is robust across different RBF domains.

Additional observations included a bead-like appearance within spatio-temporal slices, rather than the expected continuous streaks or vertical lines (Fig. S5). This phenomenon may arise when the RBC displacement is undetectable during the scan, due to steep vessel angles and/or velocities beyond the measurable range, resulting in discrete rather than elongated traces. In our method, the lower limit of measurable velocity is primarily constrained by the B-scan rate, streak detectability, and signal-to-noise ratio, while the upper limit is governed by the vessel’s orientation relative to the fast scan axis and the interscan time (δt). Increasing the number of repeated B-scans above 20 extended the sampling duration (>3 ms) and improved the visibility of streaks from slow-moving blood cells (Fig. S6), enabling reliable measurement of velocities as low as 0.5 mm/s and enhancing sensitivity to slower flows in the single-file capillary range.

Despite its promising performance, our approach has several challenges that warrant further investigation. First, the current implementation of the AO-qOCTA pipeline is computationally intensive, which is an inherited limiting factor from the Radon transform analysis. It has been reported the 2D Radon analysis takes about 4.7 h for a 10 s long AOSLO line-scan video on a CPU [43]. The 3D Radon transform can be configured for more rapid analysis at the expense of accuracy in streak orientation measurement. Porting the post-processing pipeline to a high-performance computing platform could further significantly reduce computation time. In addition, while our validation experiments confirm robust performance in healthy volunteers and a well-controlled phantom, the sample size is limited, and the impact of pathological variations in retinal tissue remains to be fully explored. Furthermore, similar to other AO studies, imaging in patients with significant media opacity or extensive eye motion may present additional challenges due to reduced SNR or motion artifacts.

Although our method enables velocity measurements in capillaries, quantifying flow rates along extended capillary segments remains challenging. This is primarily due to the sparse sampling of RBCs at each slow scan position, a consequence of single-file flow in these narrow vessels. Additionally, individual RBCs occupy a substantial portion of the capillary lumen, and other factors such as cell–cell interactions [55] and variable spacing further complicate accurate velocity estimation in these microvascular regions. During imaging sessions, target vessels were selected to ensure a shallow intersection angle (<15°) with fast scan axis, which limited the number of accessible vessels for measurement. This constraint arose from the fixed orientation of the fast-axis scanner in our system, which prevented rotation of the laser scan to align with varying vessel orientations. Future implementations could overcome this limitation via hardware modifications that enable beam rotation—such as incorporating a cylindrical or Powell lens or mounting the scanner on a rotational stage.

In summary, our AO-enabled qOCTA method addresses critical gaps in current retinal blood flow imaging technologies. By integrating high-speed imaging with advanced 3D processing, it delivers direct, absolute measurements of blood flow from a single scan that are essential for both clinical diagnostics and fundamental research. These advancements hold significant promise for transforming the assessment of retinal hemodynamics and enhancing our understanding of microvascular dysfunction in retinal diseases.

Supplementary Material

Supplemental document. See Supplement 1 for supporting content.

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.