Open-source segmentation and biometry dataset using spectrally-multiplexed whole-eye optical coherence tomography

Abstract: Whole-eye optical coherence tomography (WEOCT) has emerged as a transformative imaging modality capable of simultaneously capturing the anterior and posterior segments of the human eye. WEOCT enables comprehensive ocular biometry, which is critical for a wide range of clinical and research applications-from intraocular lens power calculation, myopia progression monitoring, and refractive surgery planning to the precise measurement of the visual and optical axes and the generation of personalized eye models for eye tracking in virtual, augmented and mixed reality(VR/AR/MR). However, existing WEOCT systems often face trade-offs between signal-to-noise ratio, imaging speed, and the ability to capture dynamic processes without motion artifacts. To address these limitations, we present a novel spectrally-multiplexed WEOCT system that utilizes two synchronized 200 kHz swept sources at 1310 nm and 1060 nm. Coupled with an automated end-to-end processing pipeline involving deep learning-based surface segmentation, 3D distortion correction, surface fitting and ray-tracing refraction correction, our system enables anatomically accurate 3D reconstruction of the segmented ocular layers. Through a 300+ participant user study and comprehensive phantom studies, we demonstrate that our system can provide simultaneous accurate measurements of cornea topography and 3D pupil center. While labeled retinal OCT data is abundantly available in open-source repositories, labeled B-scan or volumetric anterior segment data remains significantly limited. Consequently, research groups working in related domains must often acquire their own data using custom imaging systems. To help bridge this gap, we are releasing as open-source a comprehensive dataset comprising 6,621 processed volumes from 276 unique participants with corresponding segmentation and calibrated 3D anterior point clouds.

• The paper introduces a spectrally multiplexed whole-eye OCT system operating at 1310 and 1060 nm that captures anterior-eye and retinal structures simultaneously while reducing motion artifacts.
• The paper validates its automated segmentation, calibration, and refraction-correction pipeline with a 14 µm median corneal-radius error and 22 µm median 3D pupil-center error on precision phantoms.
• The paper releases 6,621 processed volumes from 276 participants, including raw spectral data, segmentation masks, calibrated 3D anterior point clouds, and manual annotations for ocular biometry research.

Overview

This paper presents a spectrally-multiplexed whole-eye optical coherence tomography (WEOCT) system developed at Meta's Reality Labs, together with an automated biometry processing pipeline and a large open-source dataset. The system images the anterior segment and retina simultaneously using two 200 kHz swept sources at 1310 nm and 1060 nm, and the accompanying release comprises 6,621 processed volumes from 276 participants with segmentation labels and calibrated, refraction-corrected 3D anterior point clouds (2605.19191). The authors state this is the first publicly available whole-eye OCT dataset of its kind, and the first OCT dataset to include distortion- and refraction-corrected 3D anterior point clouds with anatomical labels.

System design

The instrument combines two independent Thorlabs swept-source lasers: an anterior channel centered at 1310 nm (60 nm bandwidth) and a retinal channel centered at 1064 nm (102 nm bandwidth), each operating at 200 kHz and digitized via dual-edge sampling. The anterior channel achieves approximately 50 mm imaging depth in air—sufficient to span eyelashes through the posterior crystalline lens—with a measured depth resolution of 12.4 µm; the retinal channel reaches 12.2 mm depth at 8.9 µm resolution. Because the nominally identical sweep rates exhibit residual mismatches up to tens of milliseconds, custom acquisition software synchronizes channels at the volume level.

The wavelength split is motivated by ocular tissue optics: 1310 nm carries a higher maximum permissible exposure (MPE) limit, permitting greater illumination power (6.7 mW at the pupil plane, peak sensitivity 107.2 dB) for anterior imaging, while 1060 nm provides high-SNR retinal imaging with minimal water absorption attenuation (1.8 mW, peak sensitivity 102.5 dB). The anterior channel offers a 35 mm lateral field of view (FOV) with near-telecentric, diffraction-limited performance and a deliberately enlarged 48.8 µm spot size for extended depth of focus across the ~4 mm anterior eye depth. The retinal channel covers a 45° FOV at 8.9 µm diffraction-limited lateral resolution on a model eye, includes an electrically tunable lens providing −10 D to +8 D diopter correction, and a fast steering mirror enabling pupil steering over ±15° gaze (not used in the present data collection). A dichroic mirror combines the channels before a shared objective at ~5 cm working distance, with auxiliary pupil alignment camera and fixation display paths supporting gaze targets up to ±12° eccentricity.

This design addresses a known weakness of prior simultaneous WEOCT approaches—SNR loss from MPE constraints or sample-arm transmission losses—and avoids the motion artifacts inherent to sequential systems that cannot capture the visual and optical axes at the same instant.

Processing pipeline

The automated pipeline follows three stages consistent with prior OCT biometry work by Ortiz et al.: surface segmentation, distortion correction, and surface fitting with ray-tracing refraction correction.

Segmentation uses a coarse-to-fine two-stage DenseUNet combining U-Net skip connections with DenseNet-style dense connectivity, trained on 51,398 annotated B-scans from 552 volumes (30 subjects) with 7,951 B-scans from 79 volumes (8 subjects) held out for validation. Annotations were produced by trained graders assisted by SAM 2, with cornea and sclera labeled as a single class due to their poorly defined boundary in B-scans; separation is deferred to post-processing using the iris boundary. Model selection used best Jaccard index under weighted cross-entropy loss with AdamW optimization.

System calibration maps galvanometer scan voltages to physical 3D rays. A custom pseudo-random dot-pattern target is translated by a hexapod robot to seven Z-positions; per-pixel 3D observations are fit with least-squares lines to recover beam origin and direction, then expressed as third-order 2D polynomial functions of galvo voltages, with optional joint nonlinear refinement including axial pixel size. Rigid-transform registration places both channels in a common coordinate frame.

Refraction correction fits biconic surfaces to the segmented corneal point clouds (RANSAC outlier rejection followed by nonlinear least squares) and applies Snell's law at each interface using fixed population-average refractive indices ($n = 1.376$ cornea, $n = 1.336$ aqueous). The paper concedes two assumptions here: refractive indices are population averages rather than per-subject measurements, and the cornea is treated as homogeneous rather than modeling layered microstructure. Rays encountering total internal reflection are deactivated.

Validation results

Two phantom studies quantify accuracy. Against a 12.5 mm radius Zeiss reference ceramic sphere imaged at four lateral positions, the pipeline achieved a median radius error of 14 µm (0.11%) and p95 error of 19 µm (0.15%), with median and p95 3D center localization errors of approximately 30 µm. These figures establish tens-of-microns-level corneal biometry accuracy attributable primarily to the calibration pipeline.

End-to-end validation used a custom high-precision phantom eye (Thorlabs LE5243 phantom cornea, $n = 1.43$, 3 mm central thickness) imaged across ±5° yaw/pitch rotations and ±5 mm translations (45 volumes). The 3D pupil center error was 22 µm median (p95: 65 µm), with no discernible error trend versus rotation angle, indicating refraction correction robustness across gaze angles. Notably, the authors argue these results are conservative upper bounds for human eyes because the phantom exhibits a far larger index contrast ($\Delta n \approx 0.43$ vs. $\Delta n \approx 0.04$ in vivo) and a roughly 5× thicker cornea, amplifying refraction-induced distortion beyond physiological levels.

Data collection and dataset

Human data were acquired with 500 A-scans × 250 B-scans per volume (~1.6 Hz volumetric rate), 35 mm × 17.5 mm anterior FOV, and 36° × 18° retinal FOV. Each participant completed seven lighting/eye-condition combinations (bright, standard, dark; wide vs. normal eyes) at central fixation plus eight peripheral gaze directions on a 9° ring under bright/wide-eye conditions, yielding up to 84 volumes per participant. Manual quality assurance rejected 33% of volumes for motion artifacts and 8% for poor retinal visibility—retinal visibility rejections concentrated in downward gaze and bright-light conditions with small pupils—leaving 45% acceptance overall. This substantial rejection rate is an honest constraint on effective yield and reflects the difficulty of simultaneous dual-channel acquisition in uncontrolled fixation conditions.

The released cohort (mean age 39.5 ± 13.3 years, 55% female, recruited near Redmond, WA) excludes participants with refraction stronger than ±5 D who do not wear contact lenses—a screening criterion tied to a follow-on study not included here, which limits representation of high ametropes. The dataset ships raw spectral volumes with scan waveforms, processed B-scans with segmentation masks and boundary pixels, calibrated labeled 3D point clouds, and manual annotations from a 34-participant subset.

Limitations and open questions

Several limitations are stated or implicit. Refraction correction relies on population-average indices and a homogeneous cornea model, so per-subject accuracy for structures behind the cornea depends on how well individual anatomy matches those averages—an assumption not directly validated in vivo. Validation was performed on phantoms rather than against an independent gold-standard biometer in human subjects, so human-eye accuracy inherits only the conservative upper-bound argument. The pupil steering capability was not exercised during collection, leaving its contribution untested. Retinal-channel coverage failures in downward gaze and miosis reduced usable yield substantially, and the exclusion of strong uncorrected ametropes narrows demographic generalizability. Open questions include whether the reported phantom accuracies transfer to in vivo corneal topography and visual/optical axis measurement, and whether segmentation trained on 30 subjects generalizes across the full population without additional annotation.

Conclusion

The paper delivers a validated dual-wavelength WEOCT platform achieving sub-25 µm median 3D pupil center and 14 µm corneal radius errors on traceable phantoms, plus the largest publicly available annotated whole-eye OCT dataset with calibrated anterior point clouds. By releasing raw volumes, annotations, and processed geometry together, it lowers the barrier for groups working on ocular biometry, personalized eye models, and eye tracking who previously had to build custom systems and collect their own labeled data.