Modeling the Foveal Pit from Radial OCT

Master's Project (Engineering cycle, ISEP Paris) — Computer Vision.
Goal: segment retinal layers in radial OCT B-scans, register the slices, reconstruct the 3D foveal surface, and fit a 2D Gaussian model to extract biomarkers (depth, width, centering).

Context

Optical Coherence Tomography (OCT) provides high‑resolution cross‑sections of the retina. Around the fovea, radial acquisitions (0–360°) enable 3D reconstruction of the foveal pit and derivation of clinically useful parameters.

Pipeline

1. Pre‑processing: angle sorting, cropping, intensity normalization (percentiles), masking of instrument artefacts.
2. Multilayer segmentation (ILM, Hyper‑HRC, Ext‑HRC):
   • vertical gradient (Sobel) + Canny fused into a cost map;
   • dynamic programming shortest‑path search per column;
   • polynomial smoothing;
   • quality control (local variance, luminance) and optional active contour refinement.
3. Inter‑slice registration: affine alignment (from segmented curves) followed by phase correlation (FFT) for residual translation.
4. 3D reconstruction: polar coordinates ((r,\theta)) → interpolation → Cartesian grid ((x,y,z)).
5. Mathematical modeling: 2D Gaussian fit (A,\sigma_x,\sigma_y,x_0,y_0,C) via non‑linear least squares.
6. Evaluation & visualization: overlays, registration RGB checks, 3D surfaces vs. model, metrics (Recall, Precision, Accuracy, F1).

Key Results

• ILM → Ext‑HRC surface (robust zone): mean Accuracy 99.17%, F1 97.44% across 8 series.
• Hyper‑HRC → Ext‑HRC surface (thinner zone): Accuracy 99.17%, F1 87.66%.
• Registration: typical RMSE < 0.2 after affine + phase correlation.

Interpretation: ILM→Ext‑HRC is the most stable and precise; inner HRC is more sensitive to noise/artefacts.

Data

• 8 radial OCT series centered on the fovea (regular angles), grayscale 2D slices.
• Three interfaces segmented: ILM, Hyper‑HRC, Ext‑HRC.
• Ground truth available for quantitative evaluation.

Installation

# Python 3.10+ recommended
python -m venv .venv && source .venv/bin/activate  # (Windows: .venv\Scripts\activate)
pip install -r requirements.txt

**requirements.txt **

numpy
scipy
scikit-image
opencv-python
matplotlib
scikit-learn
tqdm
imageio

Technical Highlights

• Cost map: weighted fusion of Canny edges and Sobel vertical gradient for robust boundaries.
• Dynamic programming ensures continuous, anatomically plausible interfaces.
• Automatic QC (local variance, luminance) with fallback Otsu + active contour.
• Hybrid registration (curve‑driven affine + phase correlation) for angular consistency.
• 2D Gaussian model yields interpretable parameters (A, σx, σy, x0, y0, C).

Validation

• Metrics: Recall, Precision, Accuracy, F1 (per series and per surface), registration RMSE.
• Qualitative review: residual peripheral fringes and their impact on reconstruction.

Roadmap

• Replace affine + phase correlation with a learning‑based registration (e.g., self‑supervised descriptors) to reduce local errors.
• Add a lightweight CNN refinement after DP for challenging HRC interfaces.
• Extend modeling beyond a single Gaussian (ellipsoidal paraboloid, mixtures).

Authors

Hugo Anselme, Baptiste Aubrée — Engineering cycle (ISEP Paris).