Dataset Card for Endovis2017

Dataset Description

Dataset Summary

The Endovis2017 dataset contains preprocessed data for surgical instrument segmentation in robotic endoscopic procedures. This dataset was part of the MICCAI 2017 EndoVis Challenge for robotic instrument segmentation.

The dataset includes high-resolution images from the da Vinci surgical system along with pixel-level segmentation annotations for surgical instruments. It is designed for training and evaluating computer vision models for surgical scene understanding and instrument tracking.

Supported Tasks

• Image Segmentation: Pixel-level segmentation of surgical instruments in endoscopic images
• Medical Image Analysis: Understanding surgical scenes and instrument types
• Computer-Assisted Surgery: Real-time instrument detection and tracking

Languages

Not applicable (image dataset)

Dataset Structure

Data Instances

Each instance in the dataset contains:

{
    'image': PIL.Image,          # RGB endoscopic image
    'label': PIL.Image,          # Segmentation mask (grayscale)
    'image_id': str,             # Unique identifier
    'file_name': str,            # Original filename
    'split': str,                # 'train' or 'val'
    'relative_path': str,        # Path relative to dataset root
    'sequence_id': int           # Sequence/video ID (0 for train, 1-4 for val)
}

Data Fields

• image: RGB image of size 640×480 or similar (varies by sequence)
• label: Grayscale segmentation mask matching image dimensions
• image_id: Unique string identifier for the image
• file_name: Original filename (e.g., "frame000.png")
• split: Dataset split ("train" or "val")
• relative_path: Path relative to dataset root directory
• sequence_id: Integer identifying the surgical sequence (0 for training, 1-4 for validation sequences)

Data Splits

Split	Examples
train	1,800
val	901
Total	2,701

The training set contains images from multiple surgical procedures, while the validation set is organized into 4 different sequences (val1-val4) representing different surgical scenarios.

Dataset Creation

Source Data

The dataset originates from the 2017 Robotic Instrument Segmentation Challenge held at MICCAI 2017.

Original Source: Zenodo Repository

Data Collection

Images were captured using the da Vinci surgical system during robotic-assisted surgical procedures. The dataset includes various instrument types and surgical scenarios to ensure model generalization.

Annotations

Pixel-level segmentation masks were manually annotated by experts. The annotations include:

• Binary segmentation (instrument vs. background)
• Part-level segmentation (shaft, wrist, claspers)
• Instrument type classification

Personal and Sensitive Information

The dataset contains surgical video frames but does not include patient-identifiable information. All images show only the surgical field and instruments, not patients.

Considerations for Using the Data

Social Impact

This dataset enables research in computer-assisted surgery and robotic surgery, which can potentially:

• Improve surgical outcomes through better instrument tracking
• Enable automated surgical skill assessment
• Advance autonomous surgical robotics

Bias and Limitations

• Limited to da Vinci surgical system (may not generalize to other platforms)
• Contains only certain types of surgical procedures
• Annotation quality may vary across different sequences
• Dataset size is relatively small compared to natural image datasets

Recommendations

Users should:

• Test models on multiple surgical systems if deploying in production
• Consider domain adaptation techniques for different surgical contexts
• Validate performance on institution-specific data before clinical use
• Be aware of potential biases toward specific instrument types and surgical scenarios

Usage

Loading the Dataset

from datasets import load_dataset

# Download and cache the full dataset
dataset = load_dataset("tyluan/Endovis2017")

# Access splits
train_data = dataset['train']
val_data = dataset['val']

# Get a sample
sample = train_data[0]
image = sample['image']  # PIL Image
label = sample['label']  # PIL Image (segmentation mask)

print(f"Image size: {image.size}")
print(f"Label size: {label.size}")

Streaming Mode (No Download)

For quick exploration without downloading the entire dataset:

from datasets import load_dataset

# Stream the dataset
dataset = load_dataset("tyluan/Endovis2017", streaming=True)

# Iterate over samples
for sample in dataset['train']:
    image = sample['image']
    label = sample['label']
    # Process sample...
    break  # Just show first sample

Using with PyTorch

from datasets import load_dataset
from torch.utils.data import DataLoader
from torchvision import transforms

# Load dataset
dataset = load_dataset("tyluan/Endovis2017", split="train")

# Define transforms
transform = transforms.Compose([
    transforms.Resize((256, 256)),
    transforms.ToTensor(),
])

# Apply transforms
def apply_transforms(example):
    example['image'] = transform(example['image'])
    example['label'] = transform(example['label'])
    return example

dataset = dataset.map(apply_transforms)
dataset.set_format(type='torch', columns=['image', 'label'])

# Create DataLoader
dataloader = DataLoader(dataset, batch_size=8, shuffle=True)

# Iterate
for batch in dataloader:
    images = batch['image']  # Shape: [8, 3, 256, 256]
    labels = batch['label']  # Shape: [8, 1, 256, 256]
    # Train your model...
    break

Integration with EasyMedSeg

This dataset is part of the EasyMedSeg framework:

from dataloader.image import Endovis2017Dataset

# Download mode (recommended)
dataset = Endovis2017Dataset(
    mode='download',
    split='train',
    hf_repo_id='tyluan/Endovis2017'
)

# Streaming mode
from dataloader.image import Endovis2017StreamingDataset

streaming_dataset = Endovis2017StreamingDataset(
    split='val',
    shuffle=True
)

Additional Information

Dataset Curators

Original dataset curated by the MICCAI 2017 EndoVis Challenge organizers.

HuggingFace version prepared by the EasyMedSeg team.

Licensing Information

This dataset is licensed under Creative Commons Attribution 4.0 International (CC BY 4.0).

When using this dataset, you must:

• Give appropriate credit to the original authors
• Provide a link to the license: https://creativecommons.org/licenses/by/4.0/
• Indicate if changes were made

Citation Information

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

@article{allan2019endovis,
  title={2017 Robotic Instrument Segmentation Challenge},
  author={Allan, Max and Shvets, Alex and Kurmann, Thomas and Zhang, Zichen and Duggal, Rahul and Su, Yun-Hsuan and Rieke, Nicola and Laina, Iro and Kalavakonda, Niveditha and Bodenstedt, Sebastian and others},
  journal={arXiv preprint arXiv:1902.06426},
  year={2019}
}

Contributions

Thanks to:

• MICCAI 2017 EndoVis Challenge organizers for creating the dataset
• Original annotators for high-quality segmentation masks
• EasyMedSeg team for preparing the HuggingFace version

Contact

For questions or issues with this HuggingFace version, please open an issue in the EasyMedSeg repository.

For questions about the original dataset, refer to the challenge website or the Zenodo repository.

References

• MICCAI 2017 EndoVis Challenge
• Original Paper (arXiv:1902.06426)
• Zenodo Repository
• EasyMedSeg Framework