OCT ILM/BM Segmentation with nnU-Net
Two pretrained nnU-Net models for segmenting the internal limiting membrane (ILM) and Bruch's membrane (BM) in retinal OCT images.
GitHub repository and documentation
The models were trained on the 6 × 6 mm OCT subset of OCTA-500, using ILM and BM reference annotations. The full OCTA-500 dataset is not included here.
| Model package | Input | Dataset ID |
|---|---|---|
Dataset001_OCT-Layer-2d-TR.zip |
PNG B-scans | 001 |
Dataset002_OCT-Layer-3d2d-TR.zip |
NIfTI volume (.nii.gz) |
002 |
Both packages use configuration 2d, fold 0, and trainer nnUNetTrainer_250epochs.
Use
Install nnU-Net v2 following the official instructions, then install one of the downloaded model packages:
nnUNetv2_install_pretrained_model_from_zip MODEL.zip
Run inference with dataset ID 001 or 002:
nnUNetv2_predict -i INPUT_FOLDER -o OUTPUT_FOLDER -d DATASET_ID -c 2d -f 0 -tr nnUNetTrainer_250epochs
Use only the matching examples/<model>/images/ directory as input. Reference labels are provided under labels/. See examples/README.md for the sample layout.
For detailed setup, sample usage, and result inspection, see the GitHub README.
Associated study
Both models were applied without fine-tuning to 6 × 6 mm macular OCT scans from Topcon Maestro2 and Triton in AI-READI version 2.0.0. After post-processing, the two model-derived retinal thickness maps were visually compared with a third map derived from the segmentation provided by the OCT device. The highest-quality map among the three was used in the study:
Citation
Takahashi, N., Gadiraju, N., Kim, J. E., & Wang, J.-K. (2026). Deep learning strategies for estimating retinal thickness from fundus images: a comparative study with multi-device data. Scientific Reports, 16, 22301. https://doi.org/10.1038/s41598-026-51728-z
License
The two model packages, Dataset001_OCT-Layer-2d-TR.zip and Dataset002_OCT-Layer-3d2d-TR.zip, are licensed under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0). You may share and adapt them for noncommercial purposes with appropriate credit, a license link, and an indication of changes, without imposing additional restrictions. See the full legal code. Please identify this model release and cite the associated study when using the models.