README.md

---
license: cc-by-4.0
pretty_name: DRAC22
task_categories:
  - image-segmentation
  - image-classification
tags:
  - medical-imaging
  - retina
  - octa
  - uw-octa
  - diabetic-retinopathy
  - ophthalmology
  - miccai-2022
size_categories:
  - 1K<n<10K
configs:
  - config_name: segmentation
    data_files:
      - split: train
        path: segmentation/train-*
      - split: test_unlabeled
        path: segmentation/test_unlabeled-*
  - config_name: image_quality_assessment
    data_files:
      - split: train
        path: image_quality_assessment/train-*
      - split: test_unlabeled
        path: image_quality_assessment/test_unlabeled-*
  - config_name: dr_grading
    data_files:
      - split: train
        path: dr_grading/train-*
      - split: test_unlabeled
        path: dr_grading/test_unlabeled-*
---

# DRAC22 — Diabetic Retinopathy Analysis Challenge 2022

Ultra-wide optical coherence tomography angiography (**UW-OCTA**) benchmark from
MICCAI 2022. 1,103 images, 12x12 mm en-face projections of the **inner retinal layer**
centered on the fovea, acquired on a VG200D swept-source OCTA device at Shanghai Sixth
People's Hospital. All images are **1024x1024 8-bit grayscale PNG**.

This is a faithful re-hosting of the complete official Zenodo release
([10.5281/zenodo.10280358](https://doi.org/10.5281/zenodo.10280358)), reorganized into
three parquet configs. No images or labels were dropped, resampled, or re-encoded.

## Configs

| Config | Task | Train | Test (unlabeled) | Ground truth |
|---|---|---|---|---|
| `segmentation` | Task 1 — DR lesion segmentation | **109** | 65 | **pixel masks** |
| `image_quality_assessment` | Task 2 — image quality | 665 | 438 | 3-class label |
| `dr_grading` | Task 3 — DR grading | 611 | 386 | 3-class label |

`image_quality_assessment` covers the full 1,103-image pool; the other two tasks are
subsets of it. Only `segmentation` carries pixel-level ground truth.

## `segmentation` — read this before evaluating

The three lesion classes are **three independent binary masks, not one multi-class
label map**, and they are **per-class subsets** of the 109 training images:

| Class | Column | Train masks | Median lesion area |
|---|---|---|---|
| 1 — Intraretinal Microvascular Abnormalities (IRMA) | `mask_irma` | 86 | 0.47 % |
| 2 — Nonperfusion Areas (NPA) | `mask_npa` | 106 | 15.0 % |
| 3 — Neovascularization (NV) | `mask_nv` | 35 | 0.60 % |
| | | **227 total** | |

Co-occurrence: IRMA&NPA 83 · IRMA&NV 33 · NPA&NV 33 · **all three 31**. Every training
image has at least one mask, but only 31 of 109 are annotated for all three classes.

> **An absent mask means the class is genuinely absent from that image — it is not an
> all-background label.** Absent classes are stored as **NULL**, never as zero masks.
> Use the `has_irma` / `has_npa` / `has_nv` booleans to select which images are scorable
> for a given class. Treating a NULL as empty ground truth injects 100%-empty references
> into per-class DSC and silently corrupts the metric.

The classes also **overlap spatially** — IRMA and NPA intersect in 73 of the 83 images
carrying both (215,065 px); NPA and NV in 22 of 33 — so flattening them into a single
integer label map is lossy.

Verified over all 227 masks: each is 1024x1024, strictly binary `{0, 255}`, non-empty,
and its id exists in the training image set.

### Columns (`segmentation`)

| Column | Type | Description |
|---|---|---|
| `image` | Image | UW-OCTA en-face image, 1024x1024 grayscale |
| `mask_irma` | Image \| null | Binary IRMA mask (0/255), NULL if class absent |
| `mask_npa` | Image \| null | Binary nonperfusion-area mask (0/255), NULL if class absent |
| `mask_nv` | Image \| null | Binary neovascularization mask (0/255), NULL if class absent |
| `has_irma`, `has_npa`, `has_nv` | bool | Whether that class is annotated for this image |
| `num_classes_annotated` | int | 1–3 (train); 0 for `test_unlabeled` |
| `image_id` | string | Global id, e.g. `082` |
| `image_quality_level` | int \| null | Cross-ref from Task 2: 0 Poor / 1 Good / 2 Excellent |
| `dr_grade` | int \| null | Cross-ref from Task 3: 0 Non-DR / 1 NPDR / 2 PDR |
| `split` | string | `train` or `test_unlabeled` |

Classification configs share `image`, `image_id`, `label` (that task's own label),
`image_quality_level`, `dr_grade`, `has_segmentation`, and `split`.

## Splits and withheld test labels

Test ground truth for **all three tasks** is withheld on the
[Grand-Challenge leaderboard](https://drac22.grand-challenge.org/), which remains open
for post-challenge submission. The test images are included here for inference and
submission, but under the split name **`test_unlabeled`** so they cannot be mistaken for
a scorable split. Every label/mask column is NULL there.

**The usable segmentation set is the 109 training images.**

## Ground truth

Two ophthalmologists annotated every image independently; disagreements were arbitrated
by two additional, more senior ophthalmologists. Only this **consensus** is released —
no per-rater masks exist, so no inter-rater analysis is possible. The challenge metric is
per-class DSC, with IoU as tie-break.

## Cross-task linkage

Filenames form a **globally unique id space** across the release (verified: zero id
collisions between the train and test pools). Task 1's 174 images are byte-identical to
their Task 2 counterparts (verified by md5 over all 174), and the train/test partition is
consistent across tasks. Each config therefore carries the other tasks' labels as
cross-reference columns — e.g. filter segmentation samples by `dr_grade`.

## Caveats

- **No patient metadata exists.** The release ships no age, gender, eye, or history
  fields; the authors state they cannot report the number of eyes or patients. Images
  cannot be grouped by patient, so **patient-level splitting is impossible** and any
  train/val split of the 109 images may leak at the eye or patient level.
- **IRMA and NV are very small targets** (median 0.47 % and 0.60 % of the image).
  Expect low scores from promptless or box-prompted general-purpose segmenters.
- No overlap with OCTA-500 (Nanjing, 3/6 mm, 304 px) or ROSE (Ningbo CAS), and none with
  the color-fundus datasets DRIVE / IDRiD / PAPILA / RAVIR. DeepDRiD shares senior
  authors and the same hospital clinic but is UWF color fundus with no masks, so there is
  no segmentation-benchmark contamination.

## Source

- **Paper**: Qian B, Chen H, Wang X, et al. "DRAC 2022: A public benchmark for diabetic
  retinopathy analysis on ultra-wide optical coherence tomography angiography images."
  *Patterns* 5(3), 100929 (2024). doi:[10.1016/j.patter.2024.100929](https://doi.org/10.1016/j.patter.2024.100929)
- **Data**: Zenodo [10.5281/zenodo.10280358](https://doi.org/10.5281/zenodo.10280358)
- **Challenge**: https://drac22.grand-challenge.org/

## License

**CC BY 4.0**, as published on the Zenodo deposit. Redistribution with attribution is
permitted; please cite the Patterns paper above.