{"success":true,"data":{"author":"Stefan Winzeck, Mauricio Reyes et al. (ISLES organizers); mirrored by MedOtter","bucketAccessType":"public","bucketName":"MedOtter/isles2017","bucketPrefix":"","category":"Neuroscience","cloudProvider":"huggingface","createdAt":"2026-11-21T22:16:41.351Z","description":"Training data of the MICCAI 2017 ISLES (Ischemic Stroke Lesion Segmentation) challenge, whose task is to predict where brain tissue will end up dead after a stroke. Inputs are acute multispectral MRI: an apparent diffusion coefficient (ADC) map, perfusion maps (MTT, rCBF, rCBV, Tmax, TTP, computed with Olea Sphere) and the raw 4D perfusion-weighted series (PWI); the target is the final infarct traced by a neuroradiologist on a follow-up T2 scan about 90 days later and mapped back to the acute grid. There are 43 training cases (IDs up to 48, non-contiguous), eight co-registered NIfTI (.nii) volumes per case including the OT ground-truth mask, in the original VSD/SMIR folder naming, plus a preview Parquet, about 6.5 GB (measured 2026-11-21); read with nibabel. ISLES 2017 contains the ISLES 2016 cohort, so do not benchmark the two as independent. The organizers re-deposited the data on Zenodo (DOI 10.5281/zenodo.17736412) under CC BY 4.0; the original SMIR release carried ODbL. Cite Winzeck et al., Frontiers in Neurology 9:679 (2018).","displayTitle":"ISLES 2017: stroke lesion outcome prediction from acute diffusion and perfusion MRI, 43 training cases with 90-day infarct masks","doi":"10.5281/zenodo.17736412","huggingFaceUrl":"https://huggingface.co/datasets/MedOtter/isles2017","id":"ae56d182-13af-5e92-8c06-5f4a0ed15113","license":"CC-BY-4.0","licenseEvidenceUrl":"https://zenodo.org/records/17736412","objectCountEstimate":824,"probeStatus":"ok","referenceUrl":"https://doi.org/10.3389/fneur.2018.00679","region":"us","schemaCard":{"accessHints":[{"language":"python-datasets","snippet":"from datasets import load_dataset\nds = load_dataset(\"MedOtter/isles2017\")"},{"language":"python-hub","snippet":"from huggingface_hub import snapshot_download\nsnapshot_download(repo_id=\"MedOtter/isles2017\", repo_type=\"dataset\")"}],"caveats":["HuggingFace metadata: author=MedOtter, license:cc-by-4.0, size_categories:n<1K, downloads=600"],"format":"mixed","generatedAt":"2026-11-21T22:32:03.164Z","layout":{"example":"training_1/VSD.Brain.XX.O.MR_4DPWI.127015/VSD.Brain.XX.O.MR_4DPWI.127015.csv","pattern":"{repoId}/{filename}"},"mountPath":"/mnt/archive","sampleHeaders":[{"bytes":"Acquisition Number (0020|0012),Instance Creation Time (0008|0032),Slice Position (0020|1041)\r\n1 ,94820.597999999998,1\r\n1 ,94821.342999999993,2\r\n1 ,94820.671000000002,3\r\n1 ,94821.415999999997,4\r\n1 ,94820.745999999999,5\r\n1 ,94821.490999999995,6\r\n1 ,94820.820999999996,7\r\n1 ,94821.566000000006,8\r\n1 ,94820.895999999993,9\r\n1 ,94821.641000000003,10\r\n1 ,94820.971999999994,11\r\n1 ,94821.716,12\r\n1 ,94821.043000000005,13\r\n1 ,94821.788,14\r\n1 ,94821.118000000002,15\r\n1 ,94821.862999999998,16\r\n1 ,94821.192999999999,17\r\n1 ,","encoding":"utf8","key":"training_1/VSD.Brain.XX.O.MR_4DPWI.127015/VSD.Brain.XX.O.MR_4DPWI.127015.csv","truncatedAt":512}],"sampleObjects":["training_1/VSD.Brain.XX.O.MR_4DPWI.127015/License_ODC_ODBL.txt","training_1/VSD.Brain.XX.O.MR_4DPWI.127015/VSD.Brain.XX.O.MR_4DPWI.127015.csv","training_1/VSD.Brain.XX.O.MR_4DPWI.127015/VSD.Brain.XX.O.MR_4DPWI.127015.nii","training_1/VSD.Brain.XX.O.MR_ADC.128020/License_ODC_ODBL.txt","training_1/VSD.Brain.XX.O.MR_ADC.128020/VSD.Brain.XX.O.MR_ADC.128020.nii"],"sidecars":[{"bytes":"---\nlicense: cc-by-4.0\ntask_categories:\n- image-segmentation\nlanguage:\n- en\ntags:\n- medical-imaging\n- mri\n- brain\n- stroke\n- ischemic-stroke\n- lesion-segmentation\n- perfusion\n- isles\npretty_name: ISLES 2017 (Ischemic Stroke Lesion Segmentation / Outcome Prediction)\nsize_categories:\n- n<1K\n---\n\n# ISLES 2017 — Ischemic Stroke Lesion Segmentation Challenge (Outcome Prediction)\n\nMultispectral brain **MRI** from the **MICCAI 2017** edition of the Ischemic\nStroke Lesion Segmentation (ISLES) challenge. The 2017 task is **lesion outcome\nprediction**: estimate the *final* infarct (segmented on a ~90-day follow-up\nscan) from *acute* multispectral MRI.\n\nThis repository contains the **public training split only (43 cases)**, every\ncase carrying its expert ground-truth lesion mask. The 32-case test split is\n**not** included because its ground-truth masks were withheld by the challenge\norganizers for leaderboard scoring (image-only, not useful for segmentation\nevaluation).\n\n## Contents\n\n- **43 training cases** (`training_1` … `training_48`, **non-contiguous** — IDs\n  3, 17, 25, 29, 34 are absent).\n- **8 NIfTI (`.nii`) volumes per case**, each in its own sub-folder:\n\n  | Token | Description | Dim |\n  |---|---|---|\n  | `MR_4DPWI` | Raw 4-D perfusion-weighted source series (40–80 timepoints) | 4-D / 5-D |\n  | `MR_ADC`   | Apparent diffusion coefficient map | 3-D |\n  | `MR_MTT`   | Mean transit time | 3-D |\n  | `MR_rCBF`  | Relative cerebral blood flow | 3-D |\n  | `MR_rCBV`  | Relative cerebral blood volume | 3-D |\n  | `MR_Tmax`  | Time-to-maximum of the residue function | 3-D |\n  | `MR_TTP`   | Time to peak | 3-D |\n  | `OT`       | **Ground-truth lesion mask** (binary 0/1) | 3-D |\n\n- Perfusion maps were computed by the organizers with Olea Sphere (bSVD\n  deconvolution). All volumes within a case share the same spatial grid and are\n  co-registered to the mask. Spatial grids vary across cases (e.g. 192×192×19,\n  256×256×24, 128×128×25, 192×192×30).\n- Note: the raw `MR_4DPWI` series is stored **5-D** `(X, Y, Z, 1, T)` for cases\n  using the `VSD.*` naming and **4-D** `(X, Y, Z, T)` for cases using the\n  `SMIR.*` naming (squeeze the singleton when present).\n\n## Folder / naming convention\n\n```\ntraining_<N>/<PREFIX>.Brain.XX.O.<TOKEN>.<objectID>/<PREFIX>.Brain.XX.O.<TOKEN>.<objectID>.nii\n```\n\n`<PREFIX>` is **`VSD`** for 30 cases and **`SMIR`** for 13 cases\n(`training_36`–`training_48`). A dataloader must match on `<TOKEN>` (e.g.\n`*.MR_ADC.*`, `*.OT.*`) **regardless of prefix**, and must enumerate the actual\ndirectories rather than assuming a contiguous `1..43` range. The `<objectID>`\nin each folder name is the original **SMIR/VSD object ID** and is preserved\nhere — it is the only cross-year linkage handle (see Overlap below).\n\n## Ground truth\n\nThe `OT` volume is the final infarct delineated on the follow-up T2 image by a\nboard-certified neuroradiologist in 3D Slicer. **Binary** (0 = background,\n1 = lesion). All 43 training masks are non-empty (lesion sizes range from a few\ndozen to ~24 k voxels). This is the single, gold-standard annotation tier; there\nis no alternative rater set in the public release.\n\n## Provenance & license\n\n- **Source:** organizers' Zenodo deposit\n  [DOI 10.5281/zenodo.17736412](https://doi.org/10.5281/zenodo.17736412)\n  (Reyes, de la Rosa, Menze; Nov 2025), released **CC BY 4.0**. The original\n  official host (SMIR / `smir.ch`, `isles-challenge.org/ISLES2017`) is offline\n  as of 2026, so this organizer-authored re-host is the working source.\n- The per-folder `License_ODC_ODBL.txt` files shipped inside the original\n  archive carry an Open Data Commons ODbL notice; the umbrella Zenodo deposit\n  re-licenses the posted files under CC BY 4.0. Attribute the challenge when\n  reusing.\n- Case counts match the benchmark paper exactly (43 train / 32 test = 75).\n\n## Cross-dataset overlap (evaluation-leakage note)\n\n- **ISLES 2017 fully contains ISLES 2016** (it is the 2016 cohort extended by\n  8 train + 13 test cases). Never benchmark 2016 and 2017 as independent sets.\n- **Independent of `Angelou0516/isles2022` and ","encoding":"utf8","key":"README.md","name":"README.md","size":4698,"truncated":true}],"sourceObjectKey":"training_1/VSD.Brain.XX.O.MR_4DPWI.127015/VSD.Brain.XX.O.MR_4DPWI.127015.csv","starterCell":"import pandas as pd\ndf = pd.read_csv(\"hf://datasets/MedOtter/isles2017/training_1/VSD.Brain.XX.O.MR_4DPWI.127015/VSD.Brain.XX.O.MR_4DPWI.127015.csv\")\ndf.head()","starterSamplePath":"training_1/VSD.Brain.XX.O.MR_4DPWI.127015/VSD.Brain.XX.O.MR_4DPWI.127015.csv","totalObjects":824},"summary":"Training data of the MICCAI 2017 ISLES (Ischemic Stroke Lesion Segmentation) challenge, whose task is to predict where brain tissue will end up dead after a stroke.","title":"ISLES 2017: stroke lesion outcome prediction from acute diffusion and perfusion MRI, 43 training cases with 90-day infarct masks","updatedAt":"2026-11-22T00:53:17.099Z","visibility":"public","websiteUrl":"https://www.isles-challenge.org/"}}