---
license: other
task_categories:
  - image-segmentation
tags:
  - medical
  - endoscopy
  - colonoscopy
  - polyp
  - segmentation
  - bkai
size_categories:
  - n<10K
---

# BKAI-IGH NeoPolyp (binary, MedCLIPSeg mirror)

Re-hosted mirror of the **BKAI-IGH NeoPolyp "Small"** subset (Lan et al.,
2021), originally released through the
[BKAI-IGH NeoPolyp Kaggle competition](https://www.kaggle.com/c/bkai-igh-neopolyp/),
intended for use with [EasyMedSeg](https://github.com/).

This mirror is rebuilt from the
[`TahaKoleilat/MedCLIPSeg`](https://huggingface.co/datasets/TahaKoleilat/MedCLIPSeg)
HF dataset's `BKAI.zip` (the only freely-fetchable HF rehost we found that
ships the masks). We chose this source because the canonical Kaggle URL
requires Kaggle competition-acceptance + an API token, which is awkward
for downstream automation.

## Composition

| Split   | Images | With polyp |
|---------|-------:|-----------:|
| train   | 799    | (computed) |
| val     | 100    | (computed) |
| test    | 100    | (computed) |
| **All** | **999**|            |

Image dimensions: variable (~1280 × 950–1000 px), heterogeneous endoscopy
frames in JPEG. The original Kaggle release contains 1,000 train (with
masks) + 200 test (held-out masks); this mirror uses the 1,000 train pool
re-split into 799/100/100. The 200-image canonical Kaggle test split with
no public masks is **not** included.

## Mask caveat — binary only

The upstream MedCLIPSeg mirror saved the original
3-channel RGB-coded semantic masks as **JPEG-compressed grayscale**.
JPEG compression introduces boundary noise (we observed pixel values
1–40 and 211–254 in addition to 0/255), and JPEG-on-label-map is
inherently lossy.

This mirror **thresholds at > 127** to recover a clean binary
{0, 255} mask. **The 3-class
(background / non-neoplastic polyp / neoplastic polyp) distinction in
the original Kaggle PNGs is NOT recoverable from this source.**
Use this mirror for **binary polyp segmentation** only. Pull the
canonical Kaggle data directly if you need the 3-class formulation
required to reproduce the NeoUNet / BlazeNeo benchmarks.

## Schema

| Column      | Type     | Description                                 |
|-------------|----------|---------------------------------------------|
| `image`     | `Image`  | Source RGB frame (PNG bytes, variable size) |
| `mask`      | `Image`  | Binary mask (`L` mode, 0/255)               |
| `image_id`  | `string` | 32-char hex stem from the source filename   |
| `split`     | `string` | `train` / `val` / `test`                    |
| `has_polyp` | `bool`   | `True` iff the mask contains any foreground |

## License

The original Kaggle release does not declare a public license; usage is
implicitly governed by Kaggle competition rules ("research / academic use").
The intermediate `TahaKoleilat/MedCLIPSeg` mirror redistributes under
**CC-BY-NC-4.0** (mirror-imposed, not author-confirmed). Treat as
**research / non-commercial only** until BKAI/IGH publishes a formal
license.

## Citation

```bibtex
@inproceedings{lan2021neounet,
  title     = {{NeoUNet}: Towards Accurate Colon Polyp Segmentation
               and Neoplasm Detection},
  author    = {Lan, Phan Ngoc and An, Nguyen Sy and Hang, Dao Viet
               and Long, Dao Van and Trung, Tran Quang
               and Thuy, Nguyen Thi and Sang, Dinh Viet},
  booktitle = {Advances in Visual Computing -- ISVC 2021},
  series    = {Lecture Notes in Computer Science},
  volume    = {13018},
  pages     = {15--28},
  publisher = {Springer},
  year      = {2021},
  doi       = {10.1007/978-3-030-90436-4_2}
}

@article{an2022blazeneo,
  title   = {{BlazeNeo}: Blazing Fast Polyp Segmentation
             and Neoplasm Detection},
  author  = {An, Nguyen Sy and Lan, Phan Ngoc and Hang, Dao Viet
             and Long, Dao Van and Trung, Tran Quang
             and Thuy, Nguyen Thi and Sang, Dinh Viet},
  journal = {IEEE Access},
  volume  = {10},
  pages   = {43669--43684},
  year    = {2022},
  doi     = {10.1109/ACCESS.2022.3168693}
}
```