README.md

---
license: cc-by-nc-4.0
task_categories:
- image-segmentation
tags:
- medical
- surgical
- microsurgery
---

# MISAW-Seg: Pixel-level Surgical Tool Segmentation in Microsurgical Anastomosis

This dataset was presented in the paper: [Microsurgical Instrument Segmentation for Robot-Assisted Surgery](https://huggingface.co/papers/2509.11727).

## Dataset Overview

*MISAW-Seg* is a surgical segmentation dataset that extends the original [MISAW dataset](https://www.synapse.org/Synapse:syn21776936/files/) by introducing segmentation annotations for microsurgical tools involved in artificial vessel anastomosis tasks. The original dataset included kinematic data, workflow annotations, and stereo video recordings, but lacked pixel-wise annotations for surgical tools. 

This dataset provides new segmentation masks created using the Roboflow platform, enabling segmentation research in microsurgical environments. *MISAW-Seg* is released under the license CC BY-NC 4.0.

## Data Preparation

The *MISAW-Seg* dataset is constructed by extending the original MISAW dataset, which consists of microsurgical training sessions involving artificial vessel anastomosis. 
<!--From this dataset, we extracted image frames and manually created semantic segmentation masks for surgical tools.-->

The original MISAW dataset includes kinematic data, stereo video, and workflow annotations. These additional components can be accessed separately at:

[Download MISAW dataset](https://www.synapse.org/Synapse:syn21776936/files/)

In this dataset, we focus on providing:

- Extracted image frames (460×540 px) from stereo microscope video
- Corresponding semantic segmentation masks in both PNG and COCO format

<!--Surgical tools were manually annotated on each frame using the Roboflow platform, enabling segmentation tasks in microsurgical environments.-->

## Data Details

Each directory in MISAW-Seg stores raw image data, segmentation masks, and annotation files.

- **Directory Structure**

    ```
    MISAW-Seg/
    ├── images/
    │   ├── 1_1/
    │   │   ├── 1_1_000000.png
    │   │   └── ...
    │   ├── ...
    ├── masks/
    │   ├── 1_1_000000.png
    │   └── ...
    ├── _annotations.coco.json
    ├── fig/
    └── README.md
    ```

- **Image Format**

    - Extracted from stereo-microscope videos (original resolution: 960×540 px)
    - Left side was cropped to generate 460×540 px frames
    - Frame rate: 30 fps

- **Annotation Formats**

    1. **COCO Format** (`_annotations.coco.json`)
    
       The COCO-style annotation file contains polygon-based segmentations, bounding boxes, and category IDs for each object.
    
    2. **PNG Format** (`masks/`)
    
       Pixel-level segmentation masks are provided in PNG format. Each mask shares the same filename as the corresponding image (e.g., `1_1_000030.png`) and stores class IDs as pixel values.

- **Segmentation Classes**

    The dataset contains the following classes:
  
    <center>
      
    | Class Number   | Class Name              | RGB Color        |
    |----------------|-------------------------|------------------|
    | 0              | Left artificial vessel  | (0, 255, 0)      |
    | 1              | Left needle holder      | (255, 255, 0)    |
    | 2              | Needle                  | (0, 255, 255)    |
    | 3              | Right artificial vessel | (0, 0, 255)      |
    | 4              | Right needle holder     | (255, 0, 0)      |
    | 5              | Wire                    | (255, 0, 255)    |  

    </center>

<!--
## Annotation Method

Segmentation masks were manually created using the Roboflow platform. Annotators performed frame-level labeling of surgical instruments and maintained consistency across temporal frames.

![Annotation Process](./fig/annotation_screenshot.png)

The annotation process included the following steps:

- Video frames were extracted from stereo-microscope recordings at 30 fps.
- Annotators used a **Smart Polygon** tool in Roboflow to generate initial object masks based on object contours.
- The automatically generated masks were then manually refined by adjusting vertices to closely fit the actual surgical tool boundaries.
- Each frame was labeled with objects belonging to six predefined surgical tool categories.
- Final masks were stored in COCO JSON (polygon-based) formats.

The annotators were two non-medical professionals. They followed a consistent labeling guide and used tool appearance and continuity across frames to ensure annotation quality.
-->

## Examples of Labeled Data

Figures 1, 2, and 3 show examples of the dataset with segmentation labels.

<center>
  
Figure 1 | Figure 2 | Figure 3
:-------------------------:|:-------------------------:|:-------------------------:
![](./fig/ex1.png)  |  ![](./fig/ex2.png) | ![](./fig/ex3.png)

</center>

<!--
<figure>
    <img src='./fig/ex1.png' width="460" height="540"/>
    <figcaption>Figure 1: Example of Segmentation Mask of Image 1</figcaption>
</figure>

<figure>
    <img src='./fig/ex2.png' width="460" height="540"/>
    <figcaption>Figure 2: Example of Segmentation Mask of Image 2</figcaption>
</figure>

<figure>
    <img src='./fig/ex3.png' width="460" height="540"/>
    <figcaption>Figure 3: Example of Segmentation Mask of Image 3</figcaption>
</figure>
-->

## Citation

If you use this dataset in your research, please cite:

```bibtex
@misc{jeong2025microsurgicalinstrumentsegmentationrobotassisted,
      title={Microsurgical Instrument Segmentation for Robot-Assisted Surgery}, 
      author={Tae Kyeong Jeong and Garam Kim and Juyoun Park},
      year={2025},
      eprint={2509.11727},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2509.11727}, 
}
```