README.md

---
license: cc-by-nc-4.0
task_categories:
- image-segmentation
tags:
- medical
- ophthalmology
- cataract-surgery
- surgical-video
- intraocular-lens
- pupil
size_categories:
- n<1K
---

# LensID — lens & pupil segmentation

Segmentation subsets of **LensID** (Ghamsarian et al., MICCAI 2021), a cataract-surgery
dataset from ITEC, Alpen-Adria-Universität Klagenfurt and the Department of
Ophthalmology, Klinikum Klagenfurt. Frames are extracted from surgical microscope
video of the anterior segment of the eye.

| | |
|---|---|
| Modality | Cataract surgery microscope video (RGB), annotated on extracted 2D frames |
| Anatomy | Eye, anterior segment |
| Targets | `lens` (intraocular lens implant), `pupil` |
| Images | **401 unique** (train 292 / test 109) |
| `lens_mask` | 401 (every row) |
| `pupil_mask` | 189 (train 141 / test 48) — `null` on the other 212 |
| Resolutions | 1024×768 (299 frames) and 720×720 (102 frames) |
| Mask format | binary PNG, single channel, values `{0, 255}` |
| Official split | preserved as released; **no video appears in both splits** |

## Important: `lens` ⊂ `pupil` — the two targets NEST

The intraocular lens sits inside the pupil aperture. Measured over all 189 frames
that carry both masks:

* **99.59 %** of `lens` pixels fall inside `pupil` (per-frame minimum 94.01 %)
* only 68.19 % of `pupil` pixels fall inside `lens`

They are therefore published as **two independent binary masks**. Do **not** merge
them into a single `{0, 1=pupil, 2=lens}` label map — that would silently reduce
"pupil" to a rim annulus and change what the benchmark measures.

## Important: `pupil` is an annotation layer, not extra images

The 189 images in the upstream `Dataset_pupil.zip` are **byte-identical duplicates**
of 189 images in `Dataset_lens.zip` (md5 match on all 189, same split, same
filename). This mirror stores each image **once** and marks pupil availability with
`has_pupil_mask`. Unique images = 401, not 590.

Pupil annotations cover exactly the 189 **non-`case_`** frames; the 212
`case_3xxx` frames have a lens mask only.

## Naming and grouping

`group_id` is the safe key for group-wise splitting or video assembly — it is never
null. `video_id` is null only for the 10 `t1xxxx` frames, whose video of origin is
not recoverable from the release.

| `source_group` | example | `video_id` | n | size |
|---|---|---|---|---|
| `case_XXXX` | `case_3155_000002` | `case_3155` | 212 | 1024×768 |
| `Vnn_nnn` | `V11_123` | `V11` | 87 | 1024×768 |
| `Vn_6digit` | `V1000006`, `V000139` | `V1`, `V` | 92 | 720×720 |
| `t1xxxx` | `t10001` | `null` | 10 | 720×720 |

The V-series rule is *strip the trailing 6 digits*. Ten files carry no video digit
and share the bare prefix `V`; grouping them as one video is what makes the paper's
counts reconcile exactly — 21 train / 6 test videos for lens, 13 / 3 for pupil,
27 videos in total.

## Deviations from the upstream archives

1. **`Dataset_phase.zip` is not mirrored.** It is 367 GB of `.avi` clips for binary
   Implantation-vs-Rest *classification* and contains no segmentation masks.
2. **One orphan mask dropped**: `lens/test/V000268_31.png` had no matching image
   (402 masks vs 401 images) and was 512×512 RGB where its group is 720×720 L.
   A proper `V000268` image+mask pair is present and unaffected.
3. **Masks normalised to single-channel `L`.** Upstream ships a mix of RGB and L.
   Every RGB mask was verified to have three identical channels, so this is lossless.
   Values remain exactly `{0, 255}`.
4. Images are byte-for-byte the upstream PNGs; no resizing or re-encoding.

## Scope note

"LensID" is the name of the *framework* in the paper. The `lens` target is the
**artificial intraocular lens (IOL) implant after implantation** — not the natural
crystalline lens and not the cataract.

## Overlap with other cataract datasets

No frame overlap with Cataract-101 (`case_269`–`case_934` vs LensID's
`case_3091`–`case_3262`; the ID spaces are disjoint), Cataract-21, IrisPupilSeg,
InSegCat or CatRelDet. CaDIS / CATARACTS-2018 were recorded at Brest University
Hospital, France; LensID is Klagenfurt, Austria. **Cataract-1K** shares the same two
anatomy targets but was recorded 2021–2023, after LensID published — target
duplication, not data duplication. No cross-reference ID column exists upstream.

## License

**CC BY-NC 4.0**, as stated on the official dataset page.

The authors' page adds a further restriction, reproduced verbatim:

> This dataset is exclusively provided for scientific research purposes and as such
> cannot be used commercially or for any other purpose. If any other purpose is
> intended, you may directly contact the originators of the datasets.

This mirror exists for non-commercial scientific research only. Source:
<https://ftp.itec.aau.at/datasets/ovid/LensID/>

## Citation

```bibtex
@inproceedings{ghamsarian2021lensid,
  title     = {LensID: A CNN-RNN-Based Framework Towards Lens Irregularity
               Detection in Cataract Surgery Videos},
  author    = {Ghamsarian, Negin and Taschwer, Mario and
               Putzgruber-Adamitsch, Doris and Sarny, Stephanie and
               El-Shabrawi, Yosuf and Schoeffmann, Klaus},
  booktitle = {MICCAI 2021},
  series    = {LNCS},
  volume    = {12908},
  pages     = {76--86},
  year      = {2021},
  doi       = {10.1007/978-3-030-87237-3_8}
}
```