m2caiSeg - Surgical Scene Segmentation Dataset

Multi-class semantic segmentation dataset for surgical instrument and anatomy segmentation from the m2cai16 challenge.

Dataset Preview

Dataset Details

Property	Value
Source	m2cai16 Challenge
Modality	Endoscopy (RGB)
Task	Multi-class Semantic Segmentation
Classes	19 (1 background + 17 foreground + 1 unknown)
Train	245 images
Test	62 images
Image Format	JPEG (variable resolution)
Mask Format	PNG (RGB color-coded)

Class Mapping (19 Classes)

This dataset uses preprocessed masks. The color-to-class mapping is:

Class ID	Class Name	RGB Color
0	background	(255, 255, 255)
1	grasper	(0, 85, 170)
2	bipolar	(0, 85, 255)
3	hook	(0, 170, 85)
4	scissors	(0, 255, 85)
5	clipper	(0, 255, 170)
6	irrigator	(85, 0, 170)
7	specimen-bag	(85, 0, 255)
8	trocars	(170, 85, 85)
9	clip	(170, 170, 170)
10	liver	(85, 170, 0)
11	gall-bladder	(85, 170, 255)
12	fat	(85, 255, 0)
13	upperwall	(85, 255, 170)
14	artery	(170, 0, 255)
15	intestine	(255, 0, 255)
16	bile	(255, 255, 0)
17	blood	(255, 0, 0)
18	unknown	(0, 0, 0)

The full mapping is also available as class_mapping.json.

Usage

from datasets import load_dataset

# Load from HuggingFace
dataset = load_dataset("Angelou0516/m2caiSeg")

# Access a sample
sample = dataset["train"][0]
image = sample["image"]       # PIL Image (RGB)
mask = sample["mask"]          # PIL Image (RGB color-coded mask)
image_id = sample["image_id"] # e.g., "0"

# Convert mask to class labels
import numpy as np
from json import load

with open("class_mapping.json") as f:
    mapping = load(f)

mask_rgb = np.array(mask)
# Use color_to_class mapping to convert RGB -> class IDs

Reference

Jin, Y., et al. "Tool Detection and Operative Skill Assessment in Surgical Videos Using Region-Based Convolutional Neural Networks." IEEE Winter Conference on Applications of Computer Vision (WACV), 2018.

License

Research use only.