# OCTID Processed Dataset

Processed version of the OCTID retinal OCT dataset prepared for
research experiments comparing Swin Transformer architectures.

## Dataset Summary

- Dataset: OCTID
- Total images: 572
- Number of classes: 5
- Image size: 224 x 224
- Image mode: RGB
- Image format: JPEG

## Classes

| Class | Images |
|---|---:|
| Normal | 206 |
| AMD | 55 |
| CSR | 102 |
| DR | 107 |
| MH | 102 |
| Total | 572 |

## Preprocessing

The preprocessing pipeline consists of:

1. Aspect-ratio-preserving resize
2. Center crop to 224 x 224
3. Non-Local Means denoising
4. CLAHE enhancement

ImageNet normalization is intentionally not stored in the images.
It will be applied during the PyTorch training pipeline.

## Experimental Splits

Five repeated stratified splits are provided.

Each split contains approximately:

- 50% training
- 20% validation
- 30% testing

Seeds:

- 42
- 52
- 62
- 72
- 82

The same split files should be used when comparing different
models to ensure a fair experimental comparison.

## Repository Structure

images/
splits/
metadata/
README.md

## Intended Use

This processed dataset is intended for research experiments
involving retinal OCT image classification and comparison of
deep learning models, including Swin Transformer and custom
Swin Transformer architectures.

The provided split files should be used consistently across
models to maintain a fair comparison.

## Image Preprocessing

The saved images have already undergone:

- Aspect-ratio-preserving resize
- Center crop
- Non-Local Means denoising
- CLAHE enhancement

ImageNet normalization is NOT included in the saved JPEG files.

For model training, ImageNet normalization should be applied
after loading the image and converting it to a tensor.

## Reproducibility

The repository includes five stratified train/validation/test
splits and the experiment configuration used during dataset
preparation.

These files are provided so that the same experimental partitions
can be reused across different models.

## Dataset Integrity

- Original images checked: 572
- Corrupted original images: 0
- Original duplicate files: 0
- Processed images: 572
- Missing processed images: 0
- Processed duplicate files: 0
- Processed image size: 224 x 224
- Processed image mode: RGB

## License and Citation

This repository contains processed research data derived from OCTID.

Please consult the original OCTID dataset/source for its licensing,
citation requirements, and redistribution conditions before using
or publishing this dataset.