Physics-based generation of multilayer corneal OCT data via Gaussian modeling and MCML for AI-driven diagnostic and surgical guidance applications

Training deep learning models for corneal optical coherence tomography (OCT) imaging is limited by the availability of large, well-annotated datasets. We present a configurable Monte Carlo simulation framework that generates synthetic corneal B-scan optical OCT images with pixel-level five-layer segmentation labels derived directly from the simulation geometry. A five-layer corneal model with Gaussian surfaces captures curvature and thickness variability in healthy and keratoconic eyes. Each layer is assigned optical properties from the literature and light transport is simulated using Monte Carlo modeling of light transport in multi-layered tissues (MCML), while incorporating system features such as the confocal PSF and sensitivity roll-off. This approach produces over 10,000 high-resolution (1024x1024) image-label pairs and supports customization of geometry, photon count, noise, and system parameters. The resulting dataset enables systematic training, validation, and benchmarking of AI models under controlled, ground-truth conditions, providing a reproducible and scalable resource to support the development of diagnostic and surgical guidance applications in image-guided ophthalmology.

💡 Research Summary

The paper addresses a critical bottleneck in corneal optical coherence tomography (OCT) research: the scarcity of large, accurately annotated datasets required for training deep learning models. To overcome this, the authors develop a fully configurable physics‑based simulation pipeline that synthesizes high‑resolution (1024 × 1024) corneal B‑scan OCT images together with pixel‑perfect five‑layer segmentation masks and aligned optical‑property maps. The core of the method consists of three tightly coupled components.

First, a five‑layer corneal geometry is generated using a baseline semi‑circular curvature and a parametric deformation that combines low‑ and high‑frequency sinusoidal terms with a localized Gaussian bulge. By sampling the radius, sinusoid amplitudes, frequencies, Gaussian height, width, and lateral offset from physiologically plausible ranges, the framework can produce a wide spectrum of anatomically realistic corneas, ranging from normal eyes to keratoconus‑like steepening. Thickness of each layer (epithelium, Bowman’s, stroma, Descemet’s membrane, endothelium) is independently perturbed by a multiplicative factor, with adaptive constraint enforcement to avoid layer crossing. The geometry is finally scaled to match the typical axial‑to‑lateral aspect ratio of clinical OCT (≈ 1 : 3) and cropped to a clinically relevant lateral window (3–9 mm).

Second, the optical properties of each layer (refractive index n, absorption coefficient μ_a, scattering coefficient μ_s, anisotropy factor g) are assigned based on literature values and assumed piecewise‑homogeneous. From the stacked interfaces a per‑pixel layer index map is derived, which is then used to generate spatially aligned maps of n, μ_a, μ_s, and g. Light transport through this multi‑layer medium is simulated with Monte Carlo Multi‑Layer (MCML) modeling. Photons are propagated using exponential step lengths, scattering angles drawn from the Henyey‑Greenstein phase function, and Fresnel reflection/refraction at each interface. Low‑weight photons are terminated via Russian roulette, and the accumulated photon contributions form depth‑resolved A‑lines.

Third, system‑specific effects are incorporated to make the simulated OCT signals realistic. A Gaussian confocal gating function models the depth‑dependent point‑spread function, while a cosine‑raised‑to‑the‑fourth‑power roll‑off mimics the typical sensitivity decay with depth. The two weighting functions are multiplied with the raw MCML A‑line data, followed by logarithmic compression and contrast normalization to produce the final B‑scan image.

Using this pipeline, the authors generate a synthetic dataset of 10 000 B‑scans: 8 000 representing healthy corneas and 2 000 representing keratoconus‑like eyes. Each sample includes (i) the OCT intensity image, (ii) a five‑layer pixel‑wise segmentation mask, (iii) three optical‑property maps (n, μ_s, g), and (iv) the system‑weighted OCT signal. The dataset is packaged for easy loading and can be re‑parameterized on the fly (e.g., photon count, noise level, PSF width), enabling systematic ablation studies.

To demonstrate the utility of the dataset, two baseline experiments are presented. In Task I, a diffusion‑based inverse model is trained to jointly reconstruct the structural OCT image and the three optical‑property maps from a raw OCT B‑scan, with a physics‑based forward loss enforcing consistency between predicted maps and the reconstructed intensity. Compared with a conventional U‑Net, the diffusion model achieves higher PSNR/SSIM for the structural image and markedly better recovery of scattering‑related parameters (μ_s and g), while performing similarly for refractive‑index reconstruction. In Task II, a simple three‑class segmentation problem (epithelium, Bowman + stroma, Descemet + endothelium) is tackled using a vanilla U‑Net. The network reaches near‑perfect Intersection‑over‑Union for the first two classes but struggles with the thin Descemet‑endothelium class, illustrating how the synthetic data can expose class‑specific challenges.

The discussion highlights several strengths: (1) the physics‑based approach guarantees perfect correspondence between images and labels, eliminating the need for costly manual annotation; (2) the Gaussian parameterization provides a compact yet expressive representation of corneal shape variability; (3) MCML faithfully captures scattering and absorption, offering higher realism than purely analytic OCT simulators; (4) the inclusion of system effects (confocal gating, roll‑off) bridges the gap between simulated and clinical data. Limitations include the computational cost of Monte Carlo simulations, which necessitates GPU parallelization and careful selection of photon numbers, and the need for additional calibration to match specific commercial OCT devices. Future work is suggested to incorporate more complex physiological factors (tear film, post‑surgical remodeling) and to evaluate the synthetic data on a broader range of AI tasks (e.g., disease classification, surgical guidance).

In summary, the authors deliver a scalable, reproducible, and highly configurable synthetic corneal OCT dataset that can serve as a benchmark for training, validating, and stress‑testing deep learning models in ophthalmic imaging. By coupling Gaussian‑based anatomical modeling with MCML light transport and realistic system modeling, the work provides a valuable resource that can accelerate the development of AI‑driven diagnostic and intra‑operative guidance tools in image‑guided ophthalmology.