We analyzed 6-mm-wide anterior segment OCT scans from the Optovue Avanti platform (Visionix). Patients were classified either as (1) healthy volunteers and LASIK candidates with no abnormal corneal findings confirmed by comprehensive slit-lamp examinations, or (2) those with clinically identified corneal opacities observed on slit-lamp examination. Opacity status was based on the clinical slit-lamp determination that corneal opacity was present and no prespecified threshold for visual significance, central location, or minimum severity was applied.

Calibration was performed once for the OCT device and followed the procedure described in our previous publication, using standard diffuse reflectance targets (12%, 25%, 50%) to map the OCT signal magnitude to absolute reflectance. A linear regression model was used to estimate the 100% reflectance level, and OCT intensities were then normalized to this value to ensure consistent reflectance measurements across imaging sessions and calibrated Avanti devices.

As in our previous work, eight radial B-scan images (640 × 1024 pixels, scan depth of 2 mm) were extracted from each OCT scan, maintaining their raw, unprocessed, and non-dewarped format so that each axial scan (vertical column of signal values in the OCT B-scan image array) corresponds to the path of the OCT beam through the cornea. A signal roll-off compensation and a low-pass mean filter (rectangular, 3 × 5 pixels) were applied to each image to reduce speckle noise. Axial scans with central flare artifacts were excluded from subsequent analyses.

Incident angles were computed using the 3-dimensional (3D) approach described in our previous study. This approach derives the surface normal at each point on the anterior surface from the anterior elevation map and assigns it to the corresponding axial scan. It accounts for the incidence angle of the OCT beam relative to the corneal anterior surface normal both in the plane parallel to the B-scan and in the plane perpendicular to the B-scan.

To capture layer-specific optical behavior, we performed binned analyses as previously outlined, dividing the epithelium, stroma, and endothelium into uniform sections based on corneal depth and incidence angle. For the stroma, we used 4% depth increments and 0.005-radian angular bins, while the epithelium and endothelium were divided using 10% and 33.3% depth increments, respectively. Pixels from OCT images from multiple patients were then aggregated into bins with matching depth and angle parameters.

Because the distribution of reflectance within the bins is non-Gaussian and right-skewed, mean ± standard-deviation rules would not robustly characterize the upper tail relevant to opacity detection. Therefore we adopted a percentile-based approach. Within each bin, we calculated the 97th and 99.9th percentile reflectance values to define thresholds for “suspect” and “definite” opacities, respectively. These upper-tail cut-offs serve as empirical analogues to τ-quantiles commonly used in quantile regression. ^, The selected percentile thresholds were finalized during method development and then held fixed for all reported analyses.

Reflectance values at the 97th and 99.9th percentiles were fitted using exponential functions to model reflectance trends as a function of incidence angle and tissue depth. We applied the same functional form and optimization constraints as previously reported:

Where:

• ○

  $Ik\left(\theta \right)$ represents reflectance in layer $k$ at incidence angle $\theta .$

• ○

  $A1k$ , $A2k$ , and $A3k$ are the amplitude coefficients for the exponential terms.

• ○

  $B1k$ , $B2k$ , and $B3k$ are the decay constants.

• ○

  $C$ is the baseline reflectance term.

Pixels were labeled as “suspect opacity” if their reflectance exceeded the 97th-percentile threshold $T_S\left(x,y\right)$ , and as “definite opacity” if they exceeded the 99.9th-percentile threshold $T_D\left(x,y\right)$ , based on their specific depth–angle bin. Here, $\left(x,y\right)$ denotes the pixel coordinates in the OCT B-scan. We generated the initial binary masks using:

A mask merging technique was then applied to combine “suspect” and “definite” opacity pixels into a single mask using morphological operations. Both masks were first smoothed using an ellipse-like structuring element (3 × 5 pixels). The final fused mask was then generated by expanding the processed “definite” opacity mask through 4 sequential dilations with the same structuring element and intersecting it with the processed “suspect” opacity mask:

where $D(.)$ and $E(.)$ represent the dilation and erosion operators, respectively. Minimum-size filtering (100px) was then applied to suppress trivial isolated regions. Figure 2 provides a step-by-step illustration of the thresholding and mask fusion workflow. The resulting final mask was used for subsequent B-scan- and eye-level classification. An eye was classified as opacity-positive if the final post-processed algorithm mask identified opacity on any reviewed B-scan. Otherwise, it was classified as opacity-negative. This post-processing parameter tuning was finalized during method development on a small number of representative B-scans from normal patients (∼10) from our training cohort and then held fixed for the validation step. All data analysis, image processing, and computations were performed using MATLAB R2023b (MathWorks Inc.).