{"success":true,"data":{"author":"DRAC Challenge organizers; mirrored by MedOtter","bucketAccessType":"public","bucketName":"MedOtter/DRAC22","bucketPrefix":"","category":"Health & Biomedical","cloudProvider":"huggingface","createdAt":"2026-11-21T20:45:56.395Z","description":"DRAC22 (Diabetic Retinopathy Analysis Challenge, MICCAI 2022) is a benchmark of ultra-wide optical coherence tomography angiography (UW-OCTA) images of the retina, which show retinal blood flow without dye. It has 1,103 en-face projections of the inner retinal layer, 12x12 mm centered on the fovea, acquired on a swept-source OCTA device at Shanghai Sixth People's Hospital and stored as 1024x1024 8-bit grayscale PNG. Three tasks are packaged as parquet configs: segmentation (109 labeled training images with separate binary masks for intraretinal microvascular abnormalities (IRMA), nonperfusion areas and neovascularization), image_quality_assessment (665 training images, 3 quality grades) and dr_grading (611 training images, 3 diabetic retinopathy grades), each with unlabeled test images; about 2.2 GB (measured 2026-11-21). A missing mask means the lesion is absent. Read with Hugging Face datasets or pandas plus Pillow. Use for diabetic retinopathy grading, retinal lesion segmentation and OCTA quality control. Mirrored by MedOtter from the organizers' Zenodo release, CC BY 4.0, DOI 10.5281/zenodo.10280358; cite the DRAC challenge paper in Patterns 2024 (doi 10.1016/j.patter.2024.100929).","displayTitle":"DRAC22: Diabetic Retinopathy Analysis Challenge, ultra-wide optical coherence tomography angiography (UW-OCTA) images","doi":"10.5281/zenodo.10280358","huggingFaceUrl":"https://huggingface.co/datasets/MedOtter/DRAC22","id":"447cb88a-a0a6-5475-a719-c690fa3b28de","license":"CC-BY-4.0","licenseEvidenceUrl":"https://zenodo.org/records/10280359","objectCountEstimate":17,"probeStatus":"ok","referenceUrl":"https://doi.org/10.1016/j.patter.2024.100929","region":"us","schemaCard":{"accessHints":[{"language":"python-datasets","snippet":"from datasets import load_dataset\nds = load_dataset(\"MedOtter/DRAC22\")"},{"language":"python-hub","snippet":"from huggingface_hub import snapshot_download\nsnapshot_download(repo_id=\"MedOtter/DRAC22\", repo_type=\"dataset\")"}],"caveats":["HuggingFace metadata: author=MedOtter, license:cc-by-4.0, size_categories:1K<n<10K, format:parquet, downloads=159"],"format":"mixed","generatedAt":"2026-11-21T20:48:42.128Z","layout":{"example":"dr_grading/test_unlabeled-00000-of-00002.parquet","pattern":"{repoId}/{filename}"},"mountPath":"/mnt/archive","sampleHeaders":[{"bytes":"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","encoding":"hex","key":"dr_grading/test_unlabeled-00000-of-00002.parquet","truncatedAt":512}],"sampleObjects":["dr_grading/test_unlabeled-00000-of-00002.parquet","dr_grading/test_unlabeled-00001-of-00002.parquet","dr_grading/train-00000-of-00004.parquet","dr_grading/train-00001-of-00004.parquet","dr_grading/train-00002-of-00004.parquet"],"sidecars":[{"bytes":"---\nlicense: cc-by-4.0\npretty_name: DRAC22\ntask_categories:\n  - image-segmentation\n  - image-classification\ntags:\n  - medical-imaging\n  - retina\n  - octa\n  - uw-octa\n  - diabetic-retinopathy\n  - ophthalmology\n  - miccai-2022\nsize_categories:\n  - 1K<n<10K\nconfigs:\n  - config_name: segmentation\n    data_files:\n      - split: train\n        path: segmentation/train-*\n      - split: test_unlabeled\n        path: segmentation/test_unlabeled-*\n  - config_name: image_quality_assessment\n    data_files:\n      - split: train\n        path: image_quality_assessment/train-*\n      - split: test_unlabeled\n        path: image_quality_assessment/test_unlabeled-*\n  - config_name: dr_grading\n    data_files:\n      - split: train\n        path: dr_grading/train-*\n      - split: test_unlabeled\n        path: dr_grading/test_unlabeled-*\n---\n\n# DRAC22 — Diabetic Retinopathy Analysis Challenge 2022\n\nUltra-wide optical coherence tomography angiography (**UW-OCTA**) benchmark from\nMICCAI 2022. 1,103 images, 12x12 mm en-face projections of the **inner retinal layer**\ncentered on the fovea, acquired on a VG200D swept-source OCTA device at Shanghai Sixth\nPeople's Hospital. All images are **1024x1024 8-bit grayscale PNG**.\n\nThis is a faithful re-hosting of the complete official Zenodo release\n([10.5281/zenodo.10280358](https://doi.org/10.5281/zenodo.10280358)), reorganized into\nthree parquet configs. No images or labels were dropped, resampled, or re-encoded.\n\n## Configs\n\n| Config | Task | Train | Test (unlabeled) | Ground truth |\n|---|---|---|---|---|\n| `segmentation` | Task 1 — DR lesion segmentation | **109** | 65 | **pixel masks** |\n| `image_quality_assessment` | Task 2 — image quality | 665 | 438 | 3-class label |\n| `dr_grading` | Task 3 — DR grading | 611 | 386 | 3-class label |\n\n`image_quality_assessment` covers the full 1,103-image pool; the other two tasks are\nsubsets of it. Only `segmentation` carries pixel-level ground truth.\n\n## `segmentation` — read this before evaluating\n\nThe three lesion classes are **three independent binary masks, not one multi-class\nlabel map**, and they are **per-class subsets** of the 109 training images:\n\n| Class | Column | Train masks | Median lesion area |\n|---|---|---|---|\n| 1 — Intraretinal Microvascular Abnormalities (IRMA) | `mask_irma` | 86 | 0.47 % |\n| 2 — Nonperfusion Areas (NPA) | `mask_npa` | 106 | 15.0 % |\n| 3 — Neovascularization (NV) | `mask_nv` | 35 | 0.60 % |\n| | | **227 total** | |\n\nCo-occurrence: IRMA&NPA 83 · IRMA&NV 33 · NPA&NV 33 · **all three 31**. Every training\nimage has at least one mask, but only 31 of 109 are annotated for all three classes.\n\n> **An absent mask means the class is genuinely absent from that image — it is not an\n> all-background label.** Absent classes are stored as **NULL**, never as zero masks.\n> Use the `has_irma` / `has_npa` / `has_nv` booleans to select which images are scorable\n> for a given class. Treating a NULL as empty ground truth injects 100%-empty references\n> into per-class DSC and silently corrupts the metric.\n\nThe classes also **overlap spatially** — IRMA and NPA intersect in 73 of the 83 images\ncarrying both (215,065 px); NPA and NV in 22 of 33 — so flattening them into a single\ninteger label map is lossy.\n\nVerified over all 227 masks: each is 1024x1024, strictly binary `{0, 255}`, non-empty,\nand its id exists in the training image set.\n\n### Columns (`segmentation`)\n\n| Column | Type | Description |\n|---|---|---|\n| `image` | Image | UW-OCTA en-face image, 1024x1024 grayscale |\n| `mask_irma` | Image \\| null | Binary IRMA mask (0/255), NULL if class absent |\n| `mask_npa` | Image \\| null | Binary nonperfusion-area mask (0/255), NULL if class absent |\n| `mask_nv` | Image \\| null | Binary neovascularization mask (0/255), NULL if class absent |\n| `has_irma`, `has_npa`, `has_nv` | bool | Whether that class is annotated for this image |\n| `num_classes_annotated` | int | 1–3 (train); 0 for `test_unlabeled` |\n| `image_id` | string | Global id, e.g. `082` |\n| `image_quality_level` | int \\| null | Cross-ref from Task 2: 0 Poor / 1 Good / 2 Excell","encoding":"utf8","key":"README.md","name":"README.md","size":7025,"truncated":true}],"sourceObjectKey":"dr_grading/test_unlabeled-00000-of-00002.parquet","starterCell":"from datasets import load_dataset\nds = load_dataset(\"MedOtter/DRAC22\")\nprint(ds)","starterSamplePath":"dr_grading/test_unlabeled-00000-of-00002.parquet","totalObjects":17},"summary":"DRAC22 (Diabetic Retinopathy Analysis Challenge, MICCAI 2022) is a benchmark of ultra-wide optical coherence tomography angiography (UW-OCTA) images of the retina, which show…","title":"DRAC22: Diabetic Retinopathy Analysis Challenge, ultra-wide optical coherence tomography angiography (UW-OCTA) images","updatedAt":"2026-11-21T22:04:25.269Z","visibility":"public","websiteUrl":"https://drac22.grand-challenge.org/"}}