README.md

---
license: cc-by-4.0
task_categories:
  - image-segmentation
  - image-classification
tags:
  - image
  - medical
  - medical-imaging
  - surgery
  - surgical
  - surgical-data-science
  - laparoscopic
  - semantic-segmentation
  - organ-segmentation
pretty_name: Dresden Surgical Anatomy Dataset (DSAD)
size_categories:
  - 10K<n<100K
configs:
  - config_name: single_organ
    data_files:
      - split: train
        path: single_organ/train.parquet
      - split: validation
        path: single_organ/validation.parquet
      - split: test
        path: single_organ/test.parquet
    default: true
  - config_name: multi_organ
    data_files:
      - split: train
        path: multi_organ/train.parquet
      - split: validation
        path: multi_organ/validation.parquet
      - split: test
        path: multi_organ/test.parquet
  - config_name: single_organ_multi_rater
    data_files:
      - split: train
        path: single_organ_multi_rater/train.parquet
      - split: validation
        path: single_organ_multi_rater/validation.parquet
      - split: test
        path: single_organ_multi_rater/test.parquet
  - config_name: multi_organ_multi_rater
    data_files:
      - split: train
        path: multi_organ_multi_rater/train.parquet
      - split: validation
        path: multi_organ_multi_rater/validation.parquet
      - split: test
        path: multi_organ_multi_rater/test.parquet
---

<p align="center">
  <img src="banner.gif" alt="DSAD Banner" width="100%"/>
</p>

<h1 align="center">Dresden Surgical Anatomy Dataset (DSAD)</h1>

<p align="center">
  <a href="https://doi.org/10.1038/s41597-022-01719-2"><img src="https://img.shields.io/badge/DOI-10.1038%2Fs41597--022--01719--2-blue" alt="Paper DOI"/></a>
  <a href="https://gitlab.com/nct_tso_public/dsad"><img src="https://img.shields.io/badge/GitLab-Dataset_Code-orange?logo=gitlab" alt="Dataset Code"/></a>
  <a href="https://gitlab.com/nct_tso_public/anatomy-recognition-dsad"><img src="https://img.shields.io/badge/GitLab-Baseline_Code-orange?logo=gitlab" alt="Baseline Code"/></a>
</p>

The **Dresden Surgical Anatomy Dataset** provides semantic segmentation masks for
eleven abdominal anatomical structures in laparoscopic images from robot-assisted
rectal resections: abdominal wall, colon, inferior mesenteric artery, intestinal
veins, liver, pancreas, small intestine, spleen, stomach, ureter, and vesicular
glands. It contains 13,195 images from 32 surgeries, each annotated by three
independent annotators and reviewed by an expert physician. Every image also
includes weak presence labels indicating which structures are visible in the
scene (12 labels — the 11 segmented organs plus uterus).

This is the official HuggingFace mirror of the dataset published in
[Nature Scientific Data](https://doi.org/10.1038/s41597-022-01719-2) (Carstens
et al., 2023). The `single_organ` and `multi_organ` configs contain the final
expert-reviewed masks. Per-annotator masks and STAPLE-merged masks are available
in separate configs (`single_organ_multi_rater`, `multi_organ_multi_rater`) for
inter-rater agreement analysis.

## Quick start

```python
from datasets import load_dataset

# Load single-organ segmentation (default)
ds = load_dataset("nct-tso/dsad")

# Load multi-organ segmentation
ds_multi = load_dataset("nct-tso/dsad", "multi_organ")

# Load only a specific organ (downloads only matching rows)
ds_liver = load_dataset("nct-tso/dsad", split="train", filters=[("organ", "==", "liver")])

# Load per-annotator masks for inter-rater analysis
ds_raters = load_dataset("nct-tso/dsad", "single_organ_multi_rater")

# Stream without downloading
ds_stream = load_dataset("nct-tso/dsad", split="train", streaming=True)
```

## `single_organ` (default)

13,195 images across 11 anatomical structures. Each row has one image and one
binary segmentation mask for a single organ, plus weak presence labels for the
11 segmented organs + uterus.

| Column | Type | Description |
|---|---|---|
| `surgery_id` | string | Surgery identifier (01–32) |
| `frame_id` | string | Frame number within the surgery |
| `organ` | string | Segmented organ |
| `image` | Image | Laparoscopic frame (1280x1024 RGB) |
| `mask` | Image | Binary segmentation mask (0=background, 255=organ) |
| `abdominal_wall` ... `vesicular_glands` | bool | Weak presence labels indicating which structures are visible in the frame (12 columns — 11 organs + uterus) |

## `multi_organ`

1,430 images from the stomach subset, with segmentation masks for up to 7 organs
per image. These images were selected because stomach views frequently show
multiple anatomical structures. Not all organs are visible in every frame —
masks for non-visible organs are empty. Masks may overlap.

| Column | Type | Description |
|---|---|---|
| `surgery_id` | string | Surgery identifier |
| `frame_id` | string | Frame number within the surgery |
| `image` | Image | Laparoscopic frame (1280x1024 RGB) |
| `mask_abdominal_wall` | Image | Binary mask for abdominal wall |
| `mask_colon` | Image | Binary mask for colon |
| `mask_liver` | Image | Binary mask for liver |
| `mask_pancreas` | Image | Binary mask for pancreas |
| `mask_small_intestine` | Image | Binary mask for small intestine |
| `mask_spleen` | Image | Binary mask for spleen |
| `mask_stomach` | Image | Binary mask for stomach |

## `single_organ_multi_rater`

Per-annotator masks and STAPLE-merged masks for inter-rater agreement analysis.
Same rows as `single_organ` - join on `surgery_id` + `frame_id` + `organ`.
Does not include images to avoid duplication.

| Column | Type | Description |
|---|---|---|
| `surgery_id` | string | Surgery identifier |
| `frame_id` | string | Frame number within the surgery |
| `organ` | string | Segmented organ |
| `mask_anno_1` | Image | Annotator 1 mask |
| `mask_anno_2` | Image | Annotator 2 mask |
| `mask_anno_3` | Image | Annotator 3 mask |
| `mask_staple` | Image | STAPLE-merged mask (before expert review) |

Rows are aligned with `single_organ` — combine by concatenating columns:

```python
from datasets import load_dataset, concatenate_datasets

ds = load_dataset("nct-tso/dsad", split="train")
ds_raters = load_dataset("nct-tso/dsad", "single_organ_multi_rater", split="train")
ds_raters = ds_raters.remove_columns(["surgery_id", "frame_id", "organ"])
combined = concatenate_datasets([ds, ds_raters], axis=1)
```

## `multi_organ_multi_rater`

Per-annotator masks and STAPLE-merged masks for the multi-organ subset.
Same rows as `multi_organ` - join on `surgery_id` + `frame_id`.
Does not include images to avoid duplication.

| Column | Type | Description |
|---|---|---|
| `surgery_id` | string | Surgery identifier |
| `frame_id` | string | Frame number within the surgery |
| `mask_anno_{1,2,3}_{organ}` | Image | Per-annotator masks (3 × 7 = 21 columns) |
| `mask_staple_{organ}` | Image | STAPLE-merged masks (7 columns) |

Where `{organ}` is one of: `abdominal_wall`, `colon`, `liver`, `pancreas`,
`small_intestine`, `spleen`, `stomach`.

## Splits

Split by surgery ID following
[Kolbinger et al. 2024](https://doi.org/10.1097/JS9.0000000000000595) to
prevent patient leakage:

| Split | Surgery IDs | single_organ | multi_organ |
|---|---|---|---|
| train | 01, 04–06, 08–10, 12, 15–17, 19, 22–25, 27–31 | 7,889 | 863 |
| validation | 03, 21, 26 | 1,978 | 202 |
| test | 02, 07, 11, 13–14, 18, 20, 32 | 3,328 | 365 |

## Citation

```bibtex
@article{carstens2023dresden,
  title={The dresden surgical anatomy dataset for abdominal organ segmentation in surgical data science},
  author={Carstens, Matthias and Rinner, Franziska M and Bodenstedt, Sebastian and Jenke, Alexander C and Weitz, J{\"u}rgen and Distler, Marius and Speidel, Stefanie and Kolbinger, Fiona R},
  journal={Scientific Data},
  volume={10},
  number={1},
  pages={3},
  year={2023},
  publisher={Nature Publishing Group UK London}
}
```