OCTseg-Zeiss

This is a U-Net–based method for retinal layer segmentation from OCT B-scans, including code for training, testing, a pretrained model, and RNFL thickness map computation.

• The OCT b-scan images are assumed to be acquired with a Cirrus HD-OCT (Zeiss).
  • Each OCT volume covers a 6 × 6 × 2 mm³ area of the retina and is stored as a 200 × 200 × 1024 (horizontal × vertical × depth) data cube. Selected B-scans were then extracted to form the dataset.

• This package was tested with:
  • Python 3.8.16 and PyTorch 1.12.1
  • CUDA 11.3.1 and cuDNN 8.2.0.53-11.3 (for GPU training)

---

Train the Model

An example dataset is provided here to only illustrate the required folder structure. The original dataset is proprietary and cannot be shared. To train the model on your own data, you must follow the same directory structure.

Training command:

python main_nyupitt.py --lr 0.001 --batch-size 1 --epoch 50 --data-dir ./my-example-datasets/onh-3subsets --log_path ./logs --test-name nyu-segmenter

After running this command, the trained model and test outputs will be saved in ./logs/OCTseg-Zeiss/onh-3subsets/ under two subdirectories: train and test. Each subdirectory contains an output folder named nyu-segmenter_0.001, where the folder name is derived from the --test-name and --lr arguments.

• The trained model (best and last checkpoints) will be stored at ./logs/OCTseg-Zeiss/onh-3subsets/train/nyu-segmenter_0.001/model.
• Test outputs will be saved at ./logs/OCTseg-Zeiss/onh-3subsets/test/nyu-segmenter_0.001.

Test the Pretrained Model

You can use the pretrained model in two ways: with or without ground-truth masks. A ready-to-use pretrained model is provided at ./my-pretrained-model/model.

With Ground-Truth Segmentation Masks

Assuming the test images are located in onh-3subsets/test/img and their corresponding segmentation masks in onh-3subsets/test/mask, you can run inference using the following command, specifying the pretrained model path with --model-path.

Test command:

python main_nyupitt.py --test-name nyu-segmenter --lr 0.001 --batch-size 1 --data-dir ./my-example-datasets/onh-3subsets --log_path ./logs --model-path ./my-pretrained-model/model/model_best.pth.tar

Without Segmentation Masks

When running a pretrained model on OCT B-scans without corresponding segmentation masks, in addition --model-path, you must:

1. Use the --predict flag
2. Ensure the data directory contains a onh-3subsets/predict/img folder where input images (without ground-truth masks) are stored.

Test command:

python main_nyupitt.py --test-name nyu-segmenter --lr 0.001 --batch-size 1 --data-dir ./my-example-datasets/onh-3subsets --log_path ./logs --model-path ./my-pretrained-model/model/model_best.pth.tar --predict

The output results will be saved in ./logs/OCTseg-Zeiss/onh-3subsets/predict/nyu-segmenter_0.001

Benchmarking the model

Since the exact dataset from Reference 1 was not available, a direct comparison is not possible. My results, however, are based on the same cohort. In Table 1 of Reference 1, the average Dice scores have been reported as 0.80 across all layers and 0.88 for the RNFL layer in the fully supervised setting. My results are as follows:

Average Dice	RNFL (Dice_1)	GCL+IPL (Dice_2)	INL (Dice_3)	OPL (Dice_4)	ONL (Dice_5)	IS (Dice_6)	OS (Dice_7)	RPE (Dice_8)
0.84	0.87	0.85	0.82	0.75	0.93	0.82	0.87	0.85

Application to RNFL Thickness Map Computation

The segmentation model can be applied to compute retinal nerve fiber layer thickness maps (RNFLTs) from a folder of OCT volumes.

Requirements:

• A pretrained model checkpoint (e.g., ./my-pretrained-model/model/model_best.pth.tar)
• A dataset directory containing .img OCT volume files. An example is saved in ./my-example-datasets/onh-oct-volumes folder.

RNFLT computation command:

python compute_rnfl_thickness_map_batch.py --model-path ./my-pretrained-model/model/model_best.pth.tar --batch-size 1 --data-dir ./my-example-datasets/onh-oct-volumes --log_path ./logs/

This command performs slice-by-slice segmentation of each OCT volume:

• The resulting segmentation volume is optionally saved as a .npy file.
• A corresponding RNFL thickness map is computed and saved as a .png file.
• All outputs are stored in the directory specified by --log_path.
• Notes:
  • Low contrast maps: RNFL thickness maps appear dark when raw segmentation values are min-max normalized, because the intensity distribution is skewed toward low values and the optic disc/segmentation errors act as outliers. Excluding the disc and applying contrast enhancement improves visualization.
  • Colormap choice: Commercial OCT devices typically use custom colormaps to enhance visual contrast.

Example: The effect of Contrast enhancement on RNFLT maps [UPDATED - Feb. 2026]

• contrast_rnfl_thickness_map: An interactive tool for contrast enhancement of RNFL thickness maps. This tool uses a custom color map (not equivalent to Commercial devices), and it also supports manual optic disc segmentation to demonstrate how disc removal affects the final visualization. Two methods for contrast enhancement have been applied.

  Method 1: Percentile based contrast enhancement

  

  After excluding the optic disc:

  

  Method 2: MAD/IQR-based Contrast Mapping

  

  After excluding the optic disc:

  

• Final note:

  • The contrast-enhanced RNFLT visualizations shown here are intended solely for exploration and research. The resulting images should not be interpreted as clinically calibrated thickness maps, nor do they match the proprietary contrast pipelines used in commercial OCT devices.
  • The example above also illustrate an important point: RNFL thickness maps can appear markedly different depending on the chosen contrast-enhancement method, even when the underlying imaging and/or segmentation is identical.

Citation and Acknowledgement

If you find this repository helpful, please consider starring it or citing our work. This repository was developed in the context of my work supervised by Prof. Hiroshi Ishikawa at Oregon Health & Science University (OHSU). The images were provided through the courtesy of Prof. Joel S. Schuman.

References

The following references were utilized in the development of this repository:

1. The dataset has been prepared based on this paper: Sedai, S., Antony, B., Rai, R., Jones, K., Ishikawa, H., Schuman, J., Wollstein, G., & Garnavi, R. (2019). Uncertainty guided semi-supervised segmentation of retinal layers in OCT images. In International Conference on Medical Image Computing and Computer-Assisted Intervention (pp. 282–290). Springer, Cham. https://doi.org/10.1007/978-3-030-32239-7_32

2. The U-Net architecture is adapted from MGU-Net.