BronAtlas: Bronchus Segmentation and Classification Dataset

A. Research Motivation

In computer-aided diagnosis of bronchial diseases, "the category and abnormality status of bronchial segments" are critically important.

• Existing bronchus classification datasets lack annotations for "abnormal bronchial segments".
• Current bronchus segmentation datasets [2][47][48] do not provide specific segment category information.
• To advance automated bronchial diagnosis, this work introduces BronAtlas, a novel benchmark dataset that simultaneously covers both segmentation and classification tasks.

Dataset Overview:

• Contains 100 chest CT scans
• Provides voxel-level annotations for 20 bronchial segments:
  • Right lung: 10 segments
  • Left lung: 8 segments
  • Main trachea: 1 class
  • Abnormal bronchus: 1 class

B. Data Collection and Annotation Pipeline

Data Sources

• Public databases: EXACT'09 [47] and LIDC [48] (60 cases)
• Collaborating hospitals: 40 new cases

Annotation Team and Workflow

Each CT scan was annotated by two senior radiologists following a two-step annotation process:

1. Binary airway segmentation: Initial segmentation of the entire airway tree
2. Voxel-level segment labeling: Annotation of 20 bronchial segments based on anatomical structure (see Figure 2)

Segment Definitions

Following the standard in [27], 18 segments plus the main trachea are annotated:

• Right lung: RB1–RB10 (10 segments)
• Left lung: LB1+2, LB3, ..., LB7+8, LB9, LB10 (8 segments)
• Main trachea: MAIN
• Abnormal bronchus: ABN (congenital abnormal branches that do not belong to the above segments)

Quality Control

Key quality control measures:

• Except for abnormal branches, all segments must be connected to the main bronchial trunk
• Avoid mis-segmentation, missing segments, and non-abnormal cross-contamination/color mixing
• Cases that do not conform to standard categories are uniformly labeled as "abnormal bronchus"

Compliance and Ethics

• Data acquisition and research procedures comply with the Declaration of Helsinki [49]
• Approved by the institutional ethics committee

C. Dataset Statistics

Basic Information

• Dataset name: BronAtlas (Bronchus Segmentation and Classification Benchmark)
• Dataset size: 100 chest CT scans
• Annotation type: 20-class voxel-level annotations

Class Distribution

• Right lung segments: RB1–RB10 (10 classes)
• Left lung segments: LB1+2, LB3, ..., LB7+8, LB9, LB10 (8 classes)
• Main trachea: MAIN (1 class)
• Abnormal bronchus: ABN (1 class, indicating congenital abnormal bronchus)

Abnormal Case Distribution

• Training and validation sets: 33 cases with abnormalities
• Test set: 12 cases with abnormalities

Data Split

• Train/Validation/Test = 63/7/30 (mixed from different sources before splitting)

Visualization

The dataset provides visualizations of all 20 bronchial categories:

• Segment definitions follow [27]
• MAIN represents the main trachea
• ABN represents abnormal bronchial segments
• Annotations are based on binary airway segmentation
• Except for abnormal branches, segments must be connected to the main bronchial trunk
• Cases not belonging to predefined categories are labeled as abnormal

Summary

Key contributions of BronAtlas:

• Fills the gap in "bronchus segmentation + segment classification + abnormal segment annotation"
• Provides 100 CT scans with 20-class voxel-level annotations
• Rigorous dual-expert two-step annotation process with ethical compliance
• Offers a high-quality benchmark for automated bronchial diagnosis research

Citation

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

@article{huang2024bcnet,
  title={BCNet: Bronchus Classification via Structure Guided Representation Learning},
  author={Huang, Wenhao and Gong, Haifan and Zhang, Huan and Wang, Yu and Wan, Xiang and Li, Guanbin and Li, Haofeng and Shen, Hong},
  journal={IEEE Transactions on Medical Imaging},
  volume={43},
  pages={1-10},
  year={2024},
  publisher={IEEE},
  doi={10.1109/TMI.2024.3448468}
}

@inproceedings{gong2024intensity,
  author={Gong, Haifan and Huang, Wenhao and Zhang, Huan and Wang, Yu and Wan, Xiang and Shen, Hong and Li, Guanbin and Li, Haofeng},
  title={Intensity Confusion Matters: An Intensity-Distance Guided Loss for Bronchus Segmentation},
  booktitle={Proceedings of the IEEE International Conference on Multimedia and Expo},
  pages={1-6},
  year={2024},
  address={Niagara Falls, Canada},
  organization={IEEE}
}