README.md

---
license: cc0-1.0
task_categories:
  - image-segmentation
tags:
  - medical
  - dermoscopy
  - skin-lesion
  - melanoma
  - lesion-segmentation
  - isic
pretty_name: ISIC 2016 (ISBI 2016) — Part 1 lesion segmentation
size_categories:
  - 1K<n<10K
---

# ISIC 2016 — Part 1: Lesion Segmentation

**1,279 dermoscopy images** (900 train / 379 test) of pigmented skin lesions
with **expert binary lesion-boundary masks**, from the ISBI 2016 challenge
*"Skin Lesion Analysis toward Melanoma Detection"* hosted by the International
Skin Imaging Collaboration (ISIC). A snapshot of the ISIC Archive.

- **Modality:** dermoscopy (2D RGB JPEG), variable 0.5–12 MP resolution
- **Organ:** skin (pigmented lesions — melanoma vs benign)
- **Ground truth:** one binary mask per image (PNG, `0`=background,
  `255`=lesion), traced by an expert clinician via a semi-automated
  (seed + flood-fill) or manual (polyline) process. Single annotation tier —
  every image has exactly one mask.
- **Expert ceiling:** pairwise inter-observer Jaccard on 100 images of this
  data is **≈ 0.786** (Codella et al., IBM J. Res. Dev. 2017; see
  arXiv:1902.03368 §2.1) — treat scores near that as expert-level.

## Scope

The ISBI 2016 challenge had five parts. This repository ships **Part 1
(lesion segmentation) only.** Not included: Part 2 superpixel + dermoscopic-
feature classification JSONs, Part 2B dermoscopic-feature masks (globules /
streaks on an 807-image subset — *not* lesion segmentation ground truth), and
Part 3/3B malignancy classification labels.

## Schema

| column | type | contents |
|---|---|---|
| `image_id` | string | ISIC Archive id, e.g. `ISIC_0000000` — the cross-challenge join key |
| `image` | Image | original challenge JPEG, unmodified |
| `mask` | Image | original challenge PNG, single-channel, values `{0, 255}` |
| `in_isic2018_train` | bool | `True` iff this image is also in ISIC 2018 Task 1 **training** ground truth |

Splits: `train` (900 rows), `test` (379 rows). No validation split was released
for this challenge.

## ⚠️ Overlap with ISIC 2018 (leakage warning)

The ISIC 2016/2017/2018 challenge datasets are successive snapshots of the same
archive and share the `ISIC_<7-digit>` id namespace. Measured against the
official ISIC 2018 Task 1 training ground truth (2,594 ids):

| intersection | count |
|---|---|
| ISIC 2016 **train** ∩ ISIC 2018 train | **806 / 900** (89.6%) |
| ISIC 2016 **test** ∩ ISIC 2018 train | **339 / 379** (89.4%) |
| ISIC 2016 (any) ∩ ISIC 2018 **val/test** | **0** |

Consequences:

- Any model **fine-tuned on ISIC 2018 training data is contaminated** for
  evaluation on ISIC 2016 (both splits). Filter with the
  `in_isic2018_train` column:
  `ds.filter(lambda r: not r["in_isic2018_train"])`.
- Zero-shot evaluation on both challenges' *eval* sets never scores the same
  image twice (2016's test set is disjoint from 2018's val/test).
- For the 1,145 shared images the 2018 masks were slightly **revised**
  (IoU 0.988–0.997 vs the 2016 masks) — do not mix the two years as
  interchangeable label sources.

## Provenance

Built from the four official challenge zips (no registration required),
fetched byte-exact against their Content-Length and matching the challenge
paper's counts (900 train / 379 test) exactly:

```
https://isic-challenge-data.s3.amazonaws.com/2016/ISBI2016_ISIC_Part1_Training_Data.zip
https://isic-challenge-data.s3.amazonaws.com/2016/ISBI2016_ISIC_Part1_Training_GroundTruth.zip
https://isic-challenge-data.s3.amazonaws.com/2016/ISBI2016_ISIC_Part1_Test_Data.zip
https://isic-challenge-data.s3.amazonaws.com/2016/ISBI2016_ISIC_Part1_Test_GroundTruth.zip
```

Images and masks are byte-identical to the originals (no re-encoding, no
resizing). Every mask was verified single-channel with values ⊆ {0, 255},
non-empty, and pixel-dimension-identical to its image.

## Usage

```python
from datasets import load_dataset

ds = load_dataset("MedOtter/ISIC2016")          # train / test
sample = ds["train"][0]
image = sample["image"]                          # PIL RGB
mask = sample["mask"]                            # PIL L, {0, 255}
binary = mask.point(lambda p: p > 0)             # -> {0, 1}

# leakage-safe subset w.r.t. models trained on ISIC 2018:
clean_test = ds["test"].filter(lambda r: not r["in_isic2018_train"])  # 40 rows
```

## License

**CC0 1.0 (public domain)** — as stated for the 2016 challenge on the
[ISIC challenge data page](https://challenge.isic-archive.com/data/).
Attribution is requested: cite the challenge paper below.

## Citation

```bibtex
@article{gutman2016skin,
  title   = {Skin Lesion Analysis toward Melanoma Detection: A Challenge at
             the International Symposium on Biomedical Imaging (ISBI) 2016,
             hosted by the International Skin Imaging Collaboration (ISIC)},
  author  = {Gutman, David and Codella, Noel C. F. and Celebi, Emre and
             Helba, Brian and Marchetti, Michael and Mishra, Nabin and
             Halpern, Allan},
  journal = {arXiv preprint arXiv:1605.01397},
  year    = {2016}
}
```

## Related

- [`MedOtter/ISIC2018`](https://huggingface.co/datasets/MedOtter/ISIC2018) —
  ISIC 2018 Task 1 (2,594 / 100 / 1,000). See the overlap table above before
  using both.