| --- |
| license: cc-by-nc-sa-4.0 |
| task_categories: |
| - image-segmentation |
| tags: |
| - surgical |
| - anatomy |
| - endoscopy |
| - laparoscopic |
| - robot-assisted |
| - minimally-invasive-surgery |
| - semantic-segmentation |
| pretty_name: ATLAS-120k |
| size_categories: |
| - 100K<n<1M |
| --- |
| |
| <div align="center"> |
|
|
| # ATLAS-120k |
|
|
| **Surgical Anatomy Recognition with Context Learning using Foundation Representations** |
|
|
| [](https://arxiv.org/) |
| [](https://github.com/TimJaspers0801/atlas) |
| [](https://huggingface.co/papers/2606.22124) |
| [](https://huggingface.co/rlpddejong/ATLAS-pretraining-weights) |
| [](LICENSE) |
|
|
|
|
| <img src="figures/atlas120k.png" alt="ATLAS-120k dataset examples" width="100%"> |
|
|
| </div> |
|
|
| --- |
|
|
| ATLAS-120k is a large-scale clip-level semantic segmentation dataset for surgical anatomy recognition in minimally invasive surgery (MIS). It comprises over **120,000 annotated frames** from **100 surgical videos** spanning **14 procedures** and **42 anatomical classes**, covering both laparoscopic and robot-assisted surgery. |
|
|
| > **Restricted post-release review period — until 31 July 2026.** During this time, we kindly ask the community to provide feedback to help us increase quality control and make continuous improvements. Please report any issues via the [GitHub repository](https://github.com/TimJaspers0801/ATLAS). |
|
|
| > **This repository contains only the segmentation masks and clip index files.** The corresponding video frames must be downloaded separately from YouTube using the tools provided in the [GitHub repository](https://github.com/TimJaspers0801/ATLAS). |
|
|
| ## Dataset Statistics |
|
|
| | Dataset | Procedures | Classes | Videos | Clips | Frames | Laparoscopic | Robot-assisted | |
| |---|---|---|---|---|---|---|---| |
| | Endoscapes-Seg50 | 1 | 6 | 50 | — | 493 | ✓ | | |
| | CholecSeg8k | 1 | 12 | 17 | — | 8,080 | ✓ | | |
| | DSAD | 1 | 11 | 32 | — | 14,625 | | ✓ | |
| | **ATLAS-120k** | **14** | **42** | **100** | **502** | **121,018** | **✓** | **✓** | |
|
|
| ## Procedures |
|
|
| | | | | | |
| |---|---|---|---| |
| | Adrenalectomy | Appendectomy | Cholecystectomy | Colectomy | |
| | Esophagectomy | Gastric surgery | Gastrojejunostomy | Hemicolectomy | |
| | Laparoscopic anterior resection (LAR) | Liver resection | RARP | Rectopexy | |
| | Sigmoid resection | Splenectomy | | | |
|
|
| ## Anatomical Classes |
|
|
| 42 classes including: liver, gallbladder, cystic duct, hepatic ligament, cystic plate, ductus choledochus, ductus hepaticus, stomach, small intestine, colon/rectum, abdominal wall, diaphragm, omentum, aorta, vena cava, artery (major), vein (major), nerve (major), spleen, pancreas, duodenum, kidney, bladder, ureter, uterus, ovary, prostate, seminal vesicles, adrenal gland, mesocolon, mesenterium, V. azygos, esophagus, pericardium, airway (bronchus/trachea), lung, catheter, and tools/camera. |
|
|
| The full class-to-index mapping is provided in `class_index.json` included in this release. |
|
|
| ## Repository Contents |
|
|
| After downloading and extracting this dataset, the directory structure is: |
|
|
| ``` |
| atlas120k/ |
| class_index.json ← class-to-index mapping |
| train/ val/ test/ |
| └── <procedure>/ |
| └── <video_folder>/ |
| ├── clip_index.json ← frame lists per clip |
| └── <clip>/ |
| └── machine_masks/ ← segmentation masks (this repo) |
| ``` |
|
|
| Segmentation masks are single-channel PNG files where each pixel value is a class index. |
|
|
| ## Getting the Full Dataset (with Images) |
|
|
| This HuggingFace release contains only the annotations. To reconstruct the full dataset with images: |
|
|
| 1. **Download annotations** — extract this release to a local directory, e.g. `atlas120k/` |
| 2. **Download raw videos** — use the download scripts in the [GitHub repository](https://github.com/TimJaspers0801/ATLAS) to fetch the corresponding YouTube videos |
| 3. **Extract frames** — run `process_atlas120k.py` to extract the exact annotated frames into each clip's `images/` subfolder |
|
|
| Full instructions and scripts are available at: [https://github.com/TimJaspers0801/atlas](https://github.com/TimJaspers0801/ATLAS) |
|
|
| > **Note:** 3 videos are no longer publicly available on YouTube. For these videos both the segmentation masks and the processed frames are included directly in this HuggingFace release — no download is required. The raw full videos are not available. These 3 videos will be replaced with new ones in a future dataset release. |
|
|
| ## Clip Index Format |
|
|
| Each video directory contains a `clip_index.json` that maps clips to their exact frame indices in the 15 fps video: |
|
|
| ```json |
| { |
| "folder_name": "J5bg8KTYrw0_ROBOT", |
| "youtube_id": "J5bg8KTYrw0", |
| "procedure": "cholecystectomy", |
| "split": "train", |
| "is_robot": true, |
| "frame_digits": 6, |
| "clips": { |
| "clip_0001": [964, 965, 966, ...], |
| "clip_0002": [1004, 1005, ..., 1264] |
| } |
| } |
| ``` |
|
|
| Frame filenames encode the global frame index in the 15 fps video (e.g. `frame_000964.png`). Frames within a clip may be non-contiguous. |
|
|
| ## Repository Ecosystem |
|
|
| | Repository | Description | |
| |---|---| |
| | [ATLAS](https://github.com/TimJaspers0801/ATLAS) | Dataset download and processing scripts | |
| | [ATLAS-model](https://github.com/rlpddejong/ATLAS-model) | ATLAS model implementation and training code | |
| | [ATLAS-Interactive](https://github.com/rlpddejong/ATLAS-Interactive) | Annotation platform, interactive segmentation tools and annotation models | |
| | [ATLAS-bench](https://github.com/TimJaspers0801/ATLAS-bench-public) | Benchmark experiments and ATLAS model implementations | |
| | [SurgeNetDINO](https://github.com/rlpddejong/SurgeNetDINO) | Pretrained DINOv1/v2/v3 surgical foundation backbones | |
| | [SurgeNet](https://github.com/TimJaspers0801/SurgeNet) | SurgeNet pretraining dataset used for surgical foundation models | |
|
|
| ## Ethics |
|
|
| **Sources.** The videos in ATLAS-120k were sourced from publicly available surgical videos on YouTube, originally collected as part of the [GSViT dataset](https://github.com/SamuelSchmidgall/GSViT). Our data collection prioritized videos from medical institutions and surgeons, aiming to maximize compliance with professional consent standards. |
|
|
| **Curation and identifiability.** The dataset consists exclusively of intraoperative endoscopic footage. Prior to annotation, all clips were manually reviewed to remove out-of-body views, patient-identifiable content (e.g., faces, text overlays), and non-surgical content. Annotations were performed by three surgical research fellows under the supervision of three experienced surgeons (each with more than 10 years of experience), and the first frame of every clip was independently reviewed by at least one experienced surgeon for correctness. To the best of our knowledge, patients are not identifiable by current methods from endoscopic video alone. |
|
|
| **Regulatory.** In many jurisdictions, including the EU, the sharing of non-identifiable data acquired for routine medical purposes does not require explicit patient consent and is consistent with the GDPR. |
|
|
| **Opt-out.** If you are the owner of a source video and wish to have it removed from the dataset, please open an issue in the [GitHub repository](https://github.com/TimJaspers0801/ATLAS) or contact us directly. |
|
|
| ## Acknowledgements |
|
|
| The video sources for ATLAS-120k were drawn from the **[GSViT dataset](https://github.com/SamuelSchmidgall/GSViT)**. We thank the authors for making their collection publicly available. |
|
|
| The download pipeline in this repository was inspired by the tooling from **[SurgLaVi](https://github.com/aperezr20/SurgLaVi)** and **[LEMON](https://github.com/visurg-ai/LEMON)**. |
| We also acknowledge **[YouTube-VOS](https://youtube-vos.org/)** as inspiration for creating a large-scale video dataset setting, adapted here to the surgical domain. |
|
|
| ## License |
|
|
| Dataset annotations: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/) |
|
|
| ## Citation |
|
|
| ```bibtex |
| @misc{dejong2026atlas, |
| title={Surgical Anatomy Recognition with Context Learning using Foundation Representations}, |
| author={Ronald L. P. D. de Jong and Tim J. M. Jaspers and Raf A. H. Vervoort and Aron F. H. A. Bakker and Yiping Li and Jip L. Tolenaar and Jelle P. Ruurda and Willem M. Brinkman and Josien P. W. Pluim and Marcel Breeuwer and Daan de Geus and Fons van der Sommen}, |
| year={2026}, |
| eprint={2606.22124}, |
| archivePrefix={arXiv}, |
| primaryClass={cs.CV}, |
| url={https://arxiv.org/abs/2606.22124}, |
| } |
| ``` |
|
|
