Intensity Variance‐Guided Automated Robotic Optical Coherence Tomography System for Ophthalmic Applications

ABSTRACT

Optical coherence tomography (OCT) enables non‐invasive volumetric retinal imaging. Conventional tabletop and handheld systems rely on skilled operators and patient cooperation. Mobile robot‐assisted OCT (RAOCT) systems can address these challenges, but many depend on complex visual servo modules for low‐latency eye tracking. We present an intensity variance‐guided RAOCT system integrated on a wheeled mobile platform. The system uses one depth camera for coarse eye localization and a single pupil camera for real‐time tracking. By calibrating a lookup table between image variance and distance within the near‐eye region, the pupil camera provides indirect depth estimation for robotic servoing. A motorized reference arm, an electrically tunable lens, and an automated polarization controller further enable image‐quality optimization. Experiments demonstrated 103.50 μm axial and 20.46 μm lateral tracking accuracy. The pupil‐camera response time was 10.53 ms. Automated retinal OCT imaging was achieved, demonstrating the system's potential for point‐of‐care diagnostics in resource‐limited environments.

We developed an intensity variance‐guided robot‐assisted OCT system that enables low‐latency, accurate eye tracking with only one depth camera and one pupil camera. Automated alignment and image‐quality optimization supported rapid, high‐quality retinal OCT and OCTA imaging, highlighting its potential for accessible point‐of‐care ophthalmic diagnostics.

Abbreviations

BPD

balanced photodetector

ETL

electrically tunable lens

FOV

field of view

LUT

lookup table

MAE

mean absolute error

OCT

optical coherence tomography

OCTA

optical coherence tomography angiography

OPD

optical path difference

PC

polarization controller

R ²

coefficient of determination

RAOCT

robot‐assisted optical coherence tomography

ROI

region of interest

SNR

signal‐to‐noise ratio

SS‐OCT

swept‐source optical coherence tomography

VAR

variance

WD

working distance

1. Introduction

Optical coherence tomography (OCT) is a non‐invasive, high‐speed, three‐dimensional optical imaging modality. By leveraging interferometric signals resulting from optical path difference (OPD) between reference and sample arms, OCT can rapidly provide depth‐resolved information of specimens with micrometer‐scale resolution [1]. Since its invention in the early 1990s [2], OCT has become widely used in medical diagnostics, especially in ophthalmology [3, 4, 5]. However, current ophthalmic OCT systems are predominantly bulky, table‐mounted devices that require physicians to manually position the sample arm probe near the patient. Limited automation increases reliance on operator expertise and patient cooperation. These challenges are acute in neonatal intensive care units, where premature infants cannot maintain fixation or remain still. They are also substantial in operating rooms, where supine positioning, restricted ocular access, and workflow constraints complicate probe alignment and stable image acquisition [6, 7, 8]. These challenges increase the likelihood of imaging failures. Consequently, conventional tabletop OCT systems are largely confined to specialized hospital departments, limiting their utility in broader application scenarios, such as emergency rooms or community clinics. While handheld OCT devices offer improved accessibility through portability and bedside use [9, 10, 11], they further exacerbate the reliance on manual operation. Skilled users must maintain probe alignment and track eye motion in real time. Hand tremor, reaction latency, and operator fatigue can significantly degrade image quality [12, 13]. Therefore, fully automated OCT systems are crucial for achieving robust imaging across diverse clinical settings.

Recently, robot‐assisted OCT (RAOCT) system has attracted growing interest from both the robotics and biomedical communities [12, 14, 15, 16, 17, 18, 19, 20]. It has been deployed across diverse applications [21, 22, 23, 24, 25, 26, 27]. In the OCT domain, robot‐mounted systems have demonstrated adaptability for various imaging tasks, including ophthalmic imaging of human eyes [13, 14, 17, 18, 28, 29, 30], large‐field tissue scanning [31, 32, 33, 34, 35], and security screening [36, 37, 38]. For instance, Draelos et al. developed an automated swept‐source OCT (SS‐OCT) system with a robotically‐aligned scanner [14]. They later enhanced its performance through in vivo anterior segment and retinal imaging [28], alignment degrees of freedom [29], and motion correction [30]. Zeng et al. proposed a high‐responsiveness and high‐precision robotic OCT system for rapid pupil tracking and alignment [13]. Ophthalmic RAOCT systems typically utilize a range of visual servo sensors, such as depth and complementary metal‐oxide‐semiconductor (CMOS) cameras, to accurately track the test subject. Ocular motion generally exhibits an amplitude–frequency tradeoff, with larger displacements occurring predominantly at lower frequencies and smaller amplitudes at higher frequencies. Tracker design must therefore balance correction range and update rate. Higher tracking rates reduce control delay and transient misalignment, particularly for patients with involuntary tremor, such as those with Parkinsonian symptoms. They are also valuable during prolonged volumetric OCT or OCT angiography (OCTA) acquisition, where even brief motion can compromise data consistency. However, the data stream from three‐dimensional (3D) depth cameras suffers from high latency, causing severe overshoot in real‐time subject tracking. A single two‐dimensional (2D) CMOS camera can provide low‐latency, high‐resolution pupil detection, but it cannot measure depth directly. Therefore, multiple 2D cameras are commonly used for triangulation [13, 14, 28, 29]. This configuration increases system cost and complexity, limiting the deployment of RAOCT systems in resource‐constrained or point‐of‐care settings.

To simplify this, we propose an intensity variance‐guided RAOCT system using one depth camera and a single inline pupil camera. By leveraging intensity variance from pupil images under controlled lighting, we created a lookup table (LUT) correlating intensity variance to distance, allowing the 2D camera to estimate axial depth indirectly. This approach effectively endows the single 2D CMOS camera with indirect axial depth information, enabling low‐latency pupil coordinate estimation for real‐time robotic tracking and alignment. We have also integrated several automated components, including a custom motorized adjustable reference arm, an electrically tunable lens (ETL) in the sample arm probe, and a custom motorized polarization controller (PC) within the interferometer. Using retinal OCT B‐scan image intensity as the optimization metric, we automatically adjust these components to enhance image quality. We have developed control algorithms and integrated software to orchestrate the entire system operation, successfully acquiring high‐quality retinal OCT and OCTA images. Our work presents a novel pathway to advancing RAOCT development, improving automation while minimizing dependence on complex hardware. This advancement holds potential to broaden the clinical applicability of OCT.

2. Methods

2.1. Robotic Scanner Head Design

We designed a high‐speed SS‐OCT system with the sample arm scanner head mounted on the robotic arm's end‐effector, as shown in Figure 1a. The scanner head has a weight of approximately 5 kg and overall dimensions of 27 × 25 × 17 cm. The system utilizes a collaborative 6‐degree‐of‐freedom robotic arm (UR5, Universal Robots) to precisely track and align the subject during imaging. The detailed optical layout is illustrated in Figure 1b. The sample arm features a large‐diameter 4F telescopic system, providing a subject‐friendly working distance (WD) of 103 mm. Light emitted from the OCT fiber output first passes through a collimator (F240APC‐1064, Thorlabs) and an ETL (EL‐3‐10, Optotune). It is then deflected by a two‐axis galvanometer scanner (S8107, Sunny Tech.), reshaped by the 4F lens system, and finally directed into the pupil region of the eye. The maximum scanning angle of the galvanometer mirrors is set to ±7°, which corresponds to an 11‐mm lateral coverage on the retina. Figure 1c presents the Zemax simulation results of the optical design. Across the ±7° field of view (FOV), the spot diagrams at three wavelengths remain largely confined within the Airy disk despite the presence of geometric aberrations, indicating near‐diffraction‐limited performance. Our visual servo module consists of only two camera sensors: a stereo camera (RealSense D405, Intel) for estimating the 3D position of the human eye, and a monochrome 2D CMOS camera (MV‐CB013‐A0UM‐S, Hikvision) for providing accurate coordinates of pupils and capturing pupil features. The 3D camera is positioned off‐axis relative to the main OCT optical path, while the 2D pupil camera is aligned concentrically with the OCT beam and shares a common objective lens. Four 850‐nm light‐emitting diodes (LEDs) are mounted at the corners surrounding the objective to illuminate the pupil for 2D imaging. To prevent interference between camera illumination and the OCT laser beam, a long‐pass dichroic mirror, with a 950‐nm cut‐on wavelength, is placed at a 45° angle between the objective and the relay lens. Additional fold mirrors are integrated to achieve a compact scanner design.

FIGURE 1.

OCT system design. (a) Photograph of the ophthalmic robot‐assisted OCT system. DOF: degree of freedom. (b) Schematic of the OCT sample arm scanner. DM: dichroic mirror; ETL: electrically tunable lens; WD: working distance. (c) Zemax simulation results: Spot diagrams across a ±7° scan range for three wavelengths.

2.2. Intensity Variance‐Guided Pupil Distance Estimation

We employed an inline 2D camera to capture high‐fidelity pupil images and extracted relevant features to support accurate system localization. To decouple motions in different directions and enable accurate displacement records, we built a separate pupil imaging setup, as illustrated in Figure 2a. We put the visual servo module on a stable optical bench. To replicate the pupil detection process, we employed a mannequin head with a model eye as a phantom. In the captured pupil images, the most prominent feature is the reflection of the LEDs on the corneal surface. As the distance between the scanner and the subject changes, the pupil image alternates between blurred (defocused) and sharp (focused) states, accompanied by variations in the intensity of the LED reflections. Figure 2b displays the detected pupil region of interest (ROI) at various axial depths, highlighting changes in image sharpness and reflection intensity. Inspired by prior autofocus imaging studies [39], we adopted image variance (VAR) as a metric to characterize the variation of image blur with depth. Our goal was to establish a quantitative relationship between VAR and axial depth. By measuring VAR across a range of depths, we observed a consistent trend in the VAR of the pupil ROI, as shown in Figure 2c. The corresponding distance values were obtained using the depth camera. Here, VAR is defined as the sum of squared deviations of grayscale pixel intensities from their mean. The calculation is given by Equation (1):

$$\mathrm{VAR}=\sum _{\left(x,y\right)\in \mathrm{ROI}}{\left[I\left(x,y\right)-\overline{I}\right]}^2$$ (1)

where $\left(x,y\right)$ is the coordinate of a pixel in the detected pupil image, $I\left(x,y\right)$ is the intensity value at the pixel, and $\overline{I}$ is the average pixel intensity over the entire ROI. When the VAR values are plotted as a function of axial distance, the WD point appears on a descending segment of the curve. Within this region, we can identify local extrema (a maximum followed by a minimum) and perform reliable curve fitting. This fitting forms the basis for constructing a LUT that maps VAR values to the axial offset from the WD point. In the near‐eye region (e.g., within ±2 mm of the WD), a measured VAR can be directly translated into an estimate of axial depth. To accurately calibrate the relationship between VAR and distance, we placed the phantom on a translational stage for stepwise axial movement. The optical scanner was fixed on an optical table directly in front of the phantom. At intervals of 100 μm, we recorded both the pupil ROI VAR from the 2D camera and the corresponding axial distance measured by the depth camera. Figure 2d shows the recorded VAR versus depth data alongside several polynomial fitting results in the near‐eye region. For polynomial fits of different orders, Table 1 summarizes the coefficient of determination (R ²) values in the near‐eye region and the mean absolute error (MAE) at the WD point. Across all fits, the maximum R ² reached 0.9976, indicating excellent goodness of fit.

FIGURE 2.

Relationship between axial depth and pupil image intensity variance. (a) Experimental setup for measuring the variance of pupil camera image intensity as a function of axial depth. The OCT scanner head is fixed, while a head‐with‐eye phantom is translated axially. (b) Pupil images captured by the pupil camera at different axial depths, with the detected pupil region highlighted (green square). Left: depth greater than WD; center: depth equal to WD; right: depth less than WD. (c) Plot of intensity variance within the pupil region versus offset from the WD point. (d) Curve fit of the data in the near‐eye region (±2 mm around the WD point).

TABLE 1.

Comparison of curve fitting results with depth camera measurements.

Polynomial fitting order	R 2 in the near‐eye region	Depth estimation MAE at WD point (μm)
1	0.9827	103.62
2	0.9876	12.03
3	0.9908	27.15
4	0.9976	57.34

Abbreviations: MAE, mean absolute error; R ², coefficient of determination; WD, working distance.

After testing on the optical bench setup, we conducted extensive additional pupil tracking tests using the actual robot‐mounted system, to validate the generalization and stability of the VAR‐depth curve fitting in the near‐eye region under different situations. As presented in Figures S1–S3 (Section S1, Supporting Information), the results confirm the robustness of the VAR‐based method.

2.3. SS‐OCT Engine

The schematic of our SS‐OCT is shown in Figure 3. The system employs a 400 kHz swept‐source laser with a center wavelength of 1060 nm and a bandwidth of 90 nm (Axsun Technologies). The laser output is connected to an optical attenuator to adjust the power of the near‐infrared light. Three broadband fiber couplers are used to construct a Mach‐Zehnder interferometer. The resulting interferometric signal is detected by a balanced photodetector (BPD) and digitized using a 1.0 GS/s data acquisition board (ATS9371, Alazar Technologies Inc.). The host computer interfaces with the motors in the reference arm stage, the PC, and the ETL in the sample arm to dynamically control their states and optimize OCT imaging performance. The backscattered power from the reference arm is adjusted to about 0.5 mW before it enters the BPD to prevent signal saturation distortion during digitization. We have calibrated the key imaging parameters of the system. The OCT axial resolution, shown in Figure 3b, is 6.96 μm in air. Imaging a 1951 USAF resolution target (Figure 3c) yielded a measured lateral resolution of 6.20 μm in air. The axial imaging depth at the retina is 3.54 mm.

FIGURE 3.

Robotic OCT engine. (a) System diagram. Red lines: optical signal; Black lines: electrical signal. OA, optical attenuator; PC, polarization controller; BPD, balanced photodetector; DAQ, data acquisition board. (b) Measured axial resolution. The full width at half maximum of the point spread function from a mirror reflection is 6.96 μm (in air). (c) En face image obtained by scanning a 1951 USAF resolution target. (d) Measured lateral resolution. Group 6–3 line pair of the target is clearly resolved, corresponding to a lateral resolution of 6.20 μm (in air).

2.4. Subsystem Configuration

2.4.1. Pupil Tracking and Alignment Subsystem

In RAOCT testing, the depth camera ensures full coverage of the human eye region throughout the tracking process. Both the depth camera and the pupil camera are equipped with fine‐tuned YOLOv8 models [40] to detect the eye region. Figure 4a shows the flowchart of the tracking subsystem. The depth camera provides an initial estimate of the eye's 3D location and triggers motion of the robotic arm. As the scanner approaches the subject, the pupil camera captures a clear ROI of the pupil, and VAR of this ROI can be computed. The robotic arm then performs a rapid depth sweep (~1 s) over a 30 mm range, recording the VAR of the pupil ROI at 0.1 mm depth intervals. Given that the imaging depth of our OCT system is about 3.54 mm, we define the near‐eye region as ±2 mm around the WD point, which corresponds to an absolute depth range of 101 to 105 mm. Within this region, we perform curve fitting to build a LUT that maps VAR to axial depth. In the current implementation, a third‐order fit is used for this LUT construction. This yields an analytical function relating VAR to depth, enabling low‐latency depth estimation from pupil camera data. During operation in the near‐eye region, the robotic arm adjusts its axial position based on the depth estimated from the pupil camera. The depth camera continues to provide direct depth measurements as a backup but is not used for active servo control. Outside the near‐eye region, the robotic arm is guided solely by the depth camera stream. For lateral alignment, we define the optimal imaging position as the center of the pupil camera's FOV. The pixel offset between the detected pupil center and the center of the image frame is multiplied by a pre‐calibrated magnification factor (0.035 mm per pixel), which corresponds to the digital lateral resolution of the pupil camera at the WD. During actual operation, the robotic arm's servo motion frequency is set to 125 Hz. The 3D camera streams depth data at 30 Hz, while the pupil camera operates at 125 Hz to enable continuous eye detection and tracking. Hand‐eye calibration was performed to transform coordinates from the depth camera to the robotic arm, and a Monte Carlo error propagation analysis was carried out to assess the algorithm's robustness against measurement noise. Both the calibration procedure and the noise evaluation results are documented in Section S2 and Figure S4 of the Supporting Information. The tracking module runs concurrently yet independently from the OCT scanning and imaging module, providing rapid positional corrections to maintain the RAOCT system in its optimal working posture.

FIGURE 4.

Overview of the OCT system workflow. (a) Workflow of the tracking subsystem. VAR, variance; LUT, lookup table. (b) Overview of the automatic control component optimization procedure. ETL, electrically tunable lens; PC, polarization controller. (c) Detailed steps of reference arm optimization. CC, correlation coefficient; S _opt, optimal position.

2.4.2. OCT Scanning and Image Optimization Protocols

Our OCT system supports three scanning modes: single‐ or multi‐line B‐scan, wide‐field volumetric imaging, and OCTA. Raster‐scanning waveforms are generated by a multifunction input/output board (PCIe‐6351, National Instruments) and sent to the galvanometer. We developed a software suite based on the Vortex open‐source library [41] for hardware control, signal processing, and real‐time display, which also integrates modules to manage all automated components within the imaging workflow. Once eye tracking and alignment are complete, the system initiates an automated optimization sequence, as illustrated in Figure 4b. The galvanometer is set to repeated B‐scan mode, and all controllable units are reset. Optimization begins with the reference arm. Its motor moves in a stepwise manner to locate the optimal position for first‐order interference. The detailed procedures are depicted in Figure 4c. The motor's forward orientation is defined as the direction that increases the OPD in the reference arm. During this scan, we record the list of motor step counts S, the list of average axial positions of the A‐line intensity peak across the B‐scan P, the mean B‐scan intensity I _mean, and the average peak A‐line intensity I _max. These values are stored in a sliding window of 50 samples. When the window is full, we compute the Pearson correlation coefficient between S and P, along with the signal‐to‐background ratio (I _max/I _mean), to verify the presence of first‐order interference. If both metrics exceed preset thresholds, a linear fit between P and S is performed, and the reference arm is adjusted to center the interference peak for optimal signal acquisition. Next, we sequentially optimize the ETL drive current and the PC motor angle. For each component, we sweep its tuning range and select the setting that maximizes average B‐scan intensity. After the three components are fine‐tuned, the system switches to the target scanning mode for OCT data acquisition and live imaging.

3. Results

3.1. System Performance Evaluation

3.1.1. Tracking Alignment Calibration

We characterized the quantitative tracking performance of our RAOCT system by evaluating its accuracy and precision in both axial and lateral directions. To avoid the uncertainties and instabilities associated with in vivo human eye testing, we used the head‐with‐eye phantom described in Section 2.2 for all quantitative analyses. The OCT scanner was kept stationary and pre‐aligned with the phantom at the optimal imaging position. For axial characterization, we first translated the phantom forward by 10 mm and defined this location as the zero point. We then moved the phantom backward in 1 mm steps and recorded its relative distance from the zero point. A total of 20 trials were conducted, covering an axial range of ±10 mm around the WD point. Each trial consisted of 45 single‐shot samples. A similar approach was used for lateral characterization. The lateral dataset included 24 trials, each containing 50 samples, spanning a range of ±6 mm centered on the camera image frame. The lateral step size was 0.5 mm. To assess accuracy, we compared the system's recorded depth with the known stage displacements. The MAE was computed across all trials after subtracting the true displacement from each measurement. For precision, we calculated the standard deviation of the measurements within each trial after removing the trial‐specific mean.

Figure 5 presents the resulting accuracy and precision values. The axial accuracy is 103.50 μm with a precision of 33.56 μm. The lateral accuracy is 20.46 μm with a precision of 10.34 μm. Within the axial near‐eye region, we evaluated the accuracy of our VAR‐to‐depth estimation method against ground‐truth stage translation data. Table 2 reports the MAE for different polynomial fitting orders. With reasonable curve fitting, our axial localization method achieves an accuracy of approximately 30 μm at the WD point. Additionally, we evaluated the tracking performance of the depth‐camera‐only setup to compare it with the VAR‐based method, as shown in Figure S5.

FIGURE 5.

Tracking accuracy and precision of the visual module. (a) Axial tracking accuracy. The blue dashed line indicates a mean MAE of 103.50 μm across 20 trials. (b) Lateral tracking accuracy. The blue dashed line represents a mean MAE of 20.46 μm across 24 trials. (c) Axial tracking precision. The red dashed lines indicate a precision (standard deviation of errors after trial‐specific mean subtraction) of 33.56 μm across 20 trials. (d) Lateral tracking precision. The red dashed lines indicate a precision of 10.34 μm across 24 trials.

TABLE 2.

Comparison of curve fitting results with ground‐truth data based on translational stage movements.

Polynomial fitting order	Depth estimation MAE among near‐eye region (μm)	Depth estimation MAE at WD point (μm)
1	102.80	127.62
2	52.74	31.48
3	31.88	31.91
4	87.62	81.34

Abbreviations: MAE, mean absolute error; WD, working distance.

3.1.2. Automatic Component Control Tests

We used a 3D‐printed retinal phantom to evaluate the performance of automatic control components. Figure 6a shows a B‐scan image of the phantom, with dimensions of 512 pixels axially and 1200 pixels laterally. The optimal axial signal position appears near index 256. Figure 6b plots the reference arm motor step count and the corresponding average axial index of the A‐line intensity peak over time. When the motor reached approximately 32000 steps, the criteria for optimal first‐order interference (defined in Figure 4c) were met, terminating the transverse search and triggering subsequent optimization steps. Figure 6c,d display the time‐varying ETL drive current and PC motor angle, respectively, alongside the corresponding B‐scan average intensity. We employed a coarse‐to‐fine search strategy in both optimization procedures. During in vivo OCT experiments, the reference arm motor moves at 2.7 mm/s. The ETL has a current tuning range from −10 to 40 mA. The search step size of current is 10 mA at first, then adjusted to 1 mA later. The PC motor has a rotational angle tuning range from 0 to 160 degrees. The search step size of angle is 10 degrees at first, then adjusted to 5 degrees later. The adaptive grid step size can improve optimization efficiency and enhance overall imaging performance.

FIGURE 6.

Records of the intensity‐based automatic component control and optimization process. (a) OCT B‐scan of a home‐built 3D‐printed retinal phantom. (b) Reference arm motor step count and the corresponding average axial position of the A‐line intensity peak across the B‐scan. (c) ETL drive current and the corresponding average B‐scan intensity. (d) PC motor angle and the corresponding average B‐scan intensity.

3.1.3. Visual Perception Latency Calibration

We calibrated the response time of our visual servo module (see Figure S6). As previously described, tracking target information is initially updated using the depth camera and then switched to the pupil camera once the system enters the near‐eye region. Figure 7 shows the measured response time under visual servo with each camera. We recorded 30 trials for each setup. Processing a depth camera frame involves image filtering (e.g., decimation and hole filling), eye detection and localization, together with target information update, resulting in an average response time of 41.78 ms. In contrast, processing a pupil camera frame includes pupil detection and intensity VAR calculation, together with target update, achieving a much faster response time of 10.53 ms. This represents a 75% reduction in visual perception latency compared to the depth camera. In axial eye tracking, low latency helps reduce motion overshoot and enhances system stability (see Supporting Videos 1 and 2). For example, during motion at 10 mm/s, the pupil camera delay corresponds to approximately 0.11 mm of uncorrected displacement, compared with 0.42 mm over the depth‐camera frame. The higher update rate accelerates system stabilization, rather than attempting to correct every small‐amplitude high‐frequency ocular motion.

FIGURE 7.

Response time calibration of visual servo module. (a) Records of depth camera response time. (b) Records of pupil camera response time.

3.2. Imaging Performance

We conducted in vivo ophthalmic OCT imaging experiments on three human subjects using our automatic RAOCT system. Subject 1 is a 28‐year‐old male with mild myopia (−4.00 D), with his left eye imaged using both B‐scan and volumetric OCT modes. Subject 2 is a 30‐year‐old female with emmetropia. Her left eye was imaged using OCTA mode. Subject 3 is a 26‐year‐old male with mild myopia (−2.00 D). His right eye was imaged using volumetric OCT mode. During the experiments, all subjects were seated without a chinrest and required no manual intervention. The overall testing of Subject 3 was recorded to demonstrate the system's automated tracking and imaging process (see Figure S7 and Supporting Video 3). The total system initialization took approximately 6.0 s, comprising per‐subject VAR‐depth calibration (1.5 s), reference arm optimization (1.8 s), ETL optimization (1.2 s), and PC optimization (1.5 s). To ensure ocular safety, the 1060‐nm OCT output power was set to approximately 1.4 mW, and the 850‐nm LED power at the cornea was 0.15 mW. Both values comply with the ANSI Z136.1 safety standard. Additional safety settings include an emergency stop button, a global robot speed limit of 30%, and a restricted step size of 0.5 mm when the scanner is within 10 cm of the subject. All experiments involving human participants were approved by the Clairvo Medical Technology Institutional Review Board (Protocol No. 2025012). The in vivo imaging study was conducted in accordance with the tenets of the Declaration of Helsinki. Written informed consent was obtained from all participants prior to imaging. Written consent for publication of identifiable images or videos was obtained when applicable.

Figure 8 shows retinal imaging results obtained under various scanning modes. Figure 8a,b show two selected ten‐frame‐averaged cross‐sectional frames acquired in the “5‐line” B‐scan mode, where fine details of the retinal and choroidal layers are distinctly visible. Each averaged B‐scan contains 512 × 1200 pixels, covering a lateral FOV of 5 mm. Figure 8c–f present a large FOV volume scan covering an 8 × 8 mm lateral area, clearly capturing both the optic nerve head and macula in a single unstitched volume. Green and blue arrows denote the fovea centralis and optic disc, respectively. The volume consists of 1200 × 600 × 512 voxels. Figure 8g–i show results from OCTA mode. Figure 8g,h display OCT images of the left (OS) eye. In the en face maximum intensity projections (MIPs), red dashed lines mark the B‐scan location through the macula, and red arrows point to the fovea centralis. The corresponding OCTA MIP is shown in figure 8i. The split‐spectrum amplitude‐decorrelation angiography algorithm [42] was utilized to generate the blood flow signals from three repeated OCT scans. The projection in Figure 8i contains 500 × 500 pixels over a 3 × 3 mm lateral FOV.

FIGURE 8.

Acquired ophthalmic OCT and OCTA results. (a, b) Averaged B‐scan images (10 frames averaged). FOV: 5 mm. Scale bar: 500 μm. (c) Large‐field OCT volume encompassing both the optic nerve head (ONH) and the macula. (d) En face MIP of the volume in (c). FOV: 8 × 8 mm; blue arrow: optic disc position; green arrow: fovea centralis. (e, f) Retinal B‐scans of optic disc and fovea centralis, respectively. Scale bars in (d–f): 1 mm. (g) OS OCT volume. (h) En face MIP of the volume in (g). FOV: 3 × 3 mm; red arrow: fovea centralis. (i) En face OCTA MIP of the region in (h). Scale bars in (h) and (i): 500 μm.

4. Discussion

In our proposed RAOCT system, the visual servo module integrates a depth camera for continuous eye detection and a pupil camera for indirect axial positioning with low latency. We observed a stable trend in the VAR‐depth curve, which is highly consistent with findings from previous autofocus work [39, 43]. In the near‐eye region, we establish a per‐subject LUT to map pupil image VAR to depth. This method utilizes the stable descending segment of the VAR‐depth curve, where image sharpness decreases monotonically as the visual system moves from focus to defocus. We noted a slight offset between the VAR peak value and the WD point. This occurs because our depth metric measures the distance to the pupil center, while the LED reflections used for variance calculation are located near the pupil edges. The curvature of the eyeball creates a small geometric offset between these two points. The system automates image quality optimization using three motorized components, enabling hands‐free retinal imaging without constant expert supervision. The image‐ready time is approximately 6 s based on our records, which is significantly faster than the tens of seconds or minutes required by traditional systems that depend on operator skill and patient cooperation. Table 3 compares our work with other ophthalmic RAOCT studies. Our system uses only two camera sensors, which reduces hardware complexity while achieving competitive lateral tracking accuracy and OCT resolution. Although our axial accuracy is slightly lower than that of triangulation‐based systems, this difference is expected. Our depth estimation inherently relies on a single 3D‐camera baseline, which limits the geometric leverage available for triangulation.

TABLE 3.

Comparison of this work with other ophthalmic RAOCT studies.

References	Number of 3D cameras	Number of 2D cameras	Axial accuracy (μm)	Lateral accuracy (μm)	Axial resolution (μm)	Lateral resolution (μm)
[14]	2	2	170	12	10	43
[28]	2	3	63.6	36.6	5.4	8.5
[29]	2	3	31.5	24.0	5.4	9.0
[13]	0	3	27.09	14.15	5.0	Not available
This work	1	1	103.50	20.46	6.96	6.20

A prerequisite for our variance‐based approach is the consistent visibility of LED reflections within the detected pupil region. Several factors can cause failures during in vivo tests. First, sudden lighting changes may induce pupil constriction, shifting reflections outside the detection area and destabilizing variance calculations. Second, physiological activities such as blinking or yawning can temporarily obscure the pupil, creating gaps in depth data. Third, subject‐specific issues such as eyelash occlusion and severe eye diseases like cataracts can hinder pupil detection or distort VAR values. To ensure data reliability and system robustness, we implemented a fallback mechanism for axial positioning. During operation, an online monitor tracks pupil detection status and evaluates the monotonicity of the VAR‐depth curve. If the pupil is undetected for more than 5 s, or the recorded curve lacks monotonically decreasing behavior in the near‐eye region, the system flags the current calibration as unreliable. The system then automatically switches to depth‐camera‐only guidance and reduces the robotic arm's speed to prevent overshoot while maintaining tracking and OCT imaging capability.

To further enhance system performance and robustness, we plan improvements in both engineering practices and clinical validation. First, we aim to combine our VAR‐depth method with established gaze orientation estimation algorithms [13, 29]. Currently, our LUT construction does not account for gaze angle. Therefore, subjects are required to maintain direct gaze toward the scanner. In the future, we plan to quantitatively estimate the effect of gaze angle on the axial‐depth measurement and OCT imaging, establishing a more robust mapping between visual cues and eye position. Second, we plan to replace independent parameter tuning with co‐optimization across multiple devices. Using advanced strategies such as particle swarm optimization or learning‐based methods should enable faster convergence and better image quality. For clinical validation, we will evaluate the effectiveness of the VAR‐based positioning method and the system's tolerance under diverse imaging conditions, including variations in iris pigmentation, ocular diseases, and surgery‐induced pupil deformation. We will also conduct extensive multi‐session experiments to quantify repeatability and failure rates. In parallel, we will develop appropriate physical models to better understand the underlying optical mechanisms in our VAR‐based method.

5. Conclusion

Here, we present an intensity variance‐guided automated ophthalmic RAOCT system. Our design uses only two camera sensors within the visual servo module, reducing system cost and hardware complexity. A standard 2D CMOS camera is endowed with depth estimation capability by establishing a LUT that maps pupil image VAR to the scanner‐to‐eye distance. This enables accurate, low‐latency eye tracking and supports high‐quality OCT imaging in human subjects. The system further automates image quality optimization through integrated motorized components. Our RAOCT scanner offers rapid, unsupervised pupil tracking and retinal imaging, enhancing convenience for both patients and physicians. We anticipate that our approach will expand the potential of OCT technology, facilitating its deployment in resource‐limited settings for remote or point‐of‐care diagnostics.

Author Contributions

Hang Su: writing – original draft, writing – review and editing, investigation, validation, methodology, conceptualization, data curation, formal analysis, visualization. Jie Zhang: conceptualization, methodology, investigation, validation, writing – review and editing, supervision, project administration, resources. Maoyuan Qu: investigation, validation, software, formal analysis, visualization, writing – review and editing. Xiru Gao: validation, investigation, visualization, software, writing – review and editing, formal analysis. Hongqin Chen: methodology, investigation, validation, visualization, writing – review and editing. Congyu Hu: methodology, investigation, validation. Pengfei Song: supervision, resources, funding acquisition, project administration, writing – review and editing. Xingchen Ji: conceptualization, funding acquisition, supervision, project administration, resources, writing – review and editing, visualization, formal analysis, methodology. Yikai Su: resources, writing – review and editing, project administration, supervision, funding acquisition.

Funding

This work was supported by National Natural Science Foundation of China, 62575169, 62341508, 62405180; NSFC Excellent Young Scholars Fund (Overseas); Optica Foundation Challenge Award.

Conflicts of Interest

Dr. Jie Zhang, Mr. Maoyuan Qu, Mr. Xiru Gao, and Ms. Hongqin Chen are employees of Clairvo Medical Technology Co. Ltd. The other authors declare no conflicts of interest.

Supporting information

Figure S1: In vivo pupil tracking test in a human subject. (a) The subject's eye captured by the pupil camera. (b) Plot of intensity variance within the pupil region versus axial offset from the working distance (WD) point. (c) Curve fit of the data in the near‐eye region (±2 mm around the WD point).

Figure S2: Additional eye model tracking test. (a) Diameter measurements on the components of the eye model. (b) Height measurement of the eye model. (c) Plot of intensity variance within the pupil region versus axial offset from the working distance (WD) point. (d) Curve fit of the data in the near‐eye region (±2 mm around the WD point).

Figure S3: Additional eye model tracking test under different lighting conditions. (a) Plot of pupil region intensity variance versus axial offset from the working distance (WD) point with indoor lights both off and on. (b) Curve fit of the data in the near‐eye region with indoor lights on. (c) Curve fit of the data in the near‐eye region with indoor lights off.

Figure S4: Error distributions of the hand‐eye calibration parameters obtained from the Monte Carlo error propagation analysis. (a) Rotation matrix error distribution, indicating a mean error of 0.16°. (b) Translational‐vector error distribution, indicating a mean error of 0.27 mm.

Figure S5: Eye tracking accuracy and precision of the depth camera setup. (a) Axial tracking accuracy. Blue dashed line: a mean MAE of 106.10 μm across 20 trials. (b) Lateral tracking accuracy. Blue dashed line: a mean MAE of 488.02 μm across 24 trials. (c) Axial tracking precision. Red dashed lines: a precision (standard deviation of errors after trial‐specific mean subtraction) of 44.75 μm across 20 trials. (d) Lateral tracking precision. Red dashed lines: a precision of 184.05 μm across 24 trials.

Figure S6: Photograph of the system tracking stability tests using a model eye.

Figure S7: Screenshot of the real‐time OCT imaging software interface with function panel labels.

Video S1: Overshoot depth camera guidance.

Video S2: VAR pupil camera guidance.

Video S3: RAOCT auto imaging process.

Data Availability Statement

The data that support the findings of this article are not publicly available due to privacy concerns. They are available from the corresponding author upon reasonable request.