chrisvoncsefalvay/single-bronchoscopy-depth

Single Bronchoscopy Depth

chrisvoncsefalvay/single-bronchoscopy-depth is a single unified bronchoscopy dataset that collects four major open bronchoscopy and airway endoscopy datasets: BI2K, BM-BronchoLC, UAAL, and BERD.

The dataset provides one common table for bronchoscopy images, source provenance, normalised pathology and anatomy labels, and associated depth. It is intended for bronchoscopy depth estimation, geometry-conditioned image generation, text/depth/normal-conditioned appearance modelling, and source-aware sampling across upper-airway and lower-airway endoscopic imagery.

Dataset Summary

The dataset combines 18,396 images from four source datasets.

Source	Rows	Contribution	Original source files and paths
BI2K	2,900	Bronchoscopy diagnosis images labelled as normal, benign, or malignant.	Kaggle archive containing BI2K_3/<class_id>/<case_or_sequence>/<image>.jpg, for example BI2K_3/0/1/1001.jpg.
BM-BronchoLC	2,606	Lower-airway bronchoscopy images mapped to malignant or normal status, with anatomical landmark / lesion annotations.	Figshare files Lung_cancer.zip and Non_lung_cancer.zip; image paths follow <category>/imgs/<patient_id>/<video_id>/<image_id>.png.
UAAL	6,560	Upper-airway anatomical landmark images from clinical and phantom bronchoscopy / intubation settings.	Figshare files coco_ins_clinical.zip, coco_ins_phantom.zip, and annotation visualization.zip; image paths follow coco_ins_<subset>/<split>/<image>.PNG.
BERD	6,330	Representative bronchoscopy examination images from BERD V4, included as image-only rows with local FOV masks, BREA-Depth maps, and compact MedGemma visible-image reads.	Science Data Bank BERD V4 dataset.zip; image paths follow dataset/images/<image>.png.

Hugging Face subset configs

In addition to the default config, the Hub repository exposes filtered configs for common sampling axes. Each subset config preserves the same 31-column schema and contains a sharded train split. The 53 subset configs are stored as 169 Parquet shards so Hugging Face row previews can avoid large single-file scans.

Axis	Config prefix	Subsets	Rows covered
condition	condition__...	10	18,396
source	source__...	4	18,396
anatomy	anatomy__...	39	18,396

Examples:

from datasets import load_dataset

normal = load_dataset("chrisvoncsefalvay/single-bronchoscopy-depth", "condition__normal", split="train")
berd = load_dataset("chrisvoncsefalvay/single-bronchoscopy-depth", "source__berd", split="train")
main_bronchus = load_dataset(
    "chrisvoncsefalvay/single-bronchoscopy-depth", "anatomy__main_bronchus", split="train"
)

Subset names and counts are listed in subsets/subset_manifest.json.

Unified Schema

Each row represents one image and its unified metadata.

Column	Type	Meaning
source_dataset	string	Normalised source name: bi2k, bm_broncholc, uaal, or berd.
source_image_path	string	Original source archive/member path for the image.
source_descriptor	string	Human-readable description derived from the source labels.
pathological_status	string	Unified pathology/status label.
anatomy	string	Unified anatomy label when available.
image	image	Bronchoscopy or airway endoscopy RGB image after source-specific colour correction and field-of-view masking.
mask	image	Binary field-of-view mask applied to image and depth.
mask_source	string	Rule used to generate the mask.
colour_correction	string	Source-specific pixel-order correction applied before masking.
depth	image	BREA-Depth output resized to source image size and masked with mask.
depth_source	string	Method or model used to produce the depth record.
source_subset	string	Source-specific subset, such as clinical, phantom, Lung_cancer, or Non_lung_cancer.
source_split	string	Source split when available.
source_id	string	Dataset-local source identifier.
patient_id	string	Source patient or subject identifier when derivable.
video_id	string	Source video identifier when derivable.
image_id	string	Source image or frame identifier.
width, height	int32	Image dimensions.
source_labels	list[string]	Original labels preserved from the source dataset.
source_annotation_count	int32	Number of original annotations summarised into the row.

Visual Inference Columns

The columns appearance_summary, quality, airway_level, mucosa, lumen_patency, secretions, bleeding, and overall_scene contain compact visual annotations for each frame. inference_seconds records per-frame inference time and inference_model records the model used. In this sync, all 18,396 rows have non-null inference values from google/medgemma-4b-it.

Nomenclature Unification

The source datasets use different directory names, annotation files, and label vocabularies. This dataset keeps the original labels in source_labels and maps the common training concepts into shared fields.

Colour Correction, Masks, and Depth

The dataset applies source-specific field-of-view handling before depth inference.

BM-BronchoLC source images are preserved byte-for-byte in the original source archives, but decode as blue-dominant under standard RGB. For the released image column, BM-BronchoLC rows are converted with a red/blue channel swap before masking and depth inference. BI2K and UAAL rows are kept in their stored RGB order.

The mask column stores the binary field-of-view mask used for both image and depth. BM-BronchoLC uses a fixed octagonal mask. BI2K and UAAL use the largest non-black border-separated field-of-view component. BREA-Depth is run on the corrected, masked image; its native 128 x 128 output is resized back to the source image dimensions and masked with the same mask.

Source Names

Original source name	Unified source_dataset
BI2K / Bronchoscopic Dataset BI2K 3 Classification	bi2k
BM-BronchoLC	bm_broncholc
UAAL Dataset	uaal
BERD / Fine-Grained Bronchoscopy Examination Report Dataset	berd

Pathological Status

Source convention	Unified pathological_status
BI2K path BI2K_3/0/...	normal
BI2K path BI2K_3/1/...	benign
BI2K path BI2K_3/2/...	malignant
BM-BronchoLC Lung_cancer category	malignant
BM-BronchoLC Non_lung_cancer category	normal
UAAL clinical and phantom images	unknown
BERD MedGemma visible read	normal_appearing, airway_narrowing_or_obstruction, bleeding_dominant, abnormal_mucosa, limited_or_non_diagnostic, or secretion_dominant

Observed normalised status counts:

Status	Rows
normal	2,608
benign	600
malignant	2,298
unknown	6,560
normal_appearing	4,396
airway_narrowing_or_obstruction	1,601
bleeding_dominant	317
abnormal_mucosa	10
limited_or_non_diagnostic	4
secretion_dominant	2

Anatomy

BM-BronchoLC and UAAL both include anatomical annotations, but their label inventories differ because they cover different clinical views.

BM-BronchoLC mainly contributes lower-airway bronchoscopy landmarks and lesion annotations, including labels such as vocal cords, main carina, trachea, right main bronchus, and lobar bronchi.

UAAL contributes upper-airway anatomical landmarks from clinical and phantom data, including structures such as nose, nostrils, epiglottis, glottis, glottic aperture, vocal folds, trachea, and channel.

BI2K contributes diagnosis labels but does not provide an anatomy vocabulary in the released classification archive.

BERD contributes MedGemma visible-read airway-level terms, mostly lower-airway bronchoscopy views.

The unified anatomy field stores source anatomy labels when available. The dataset does not collapse upper-airway and lower-airway anatomy into a single artificial ontology; original labels remain available in source_labels.

Source Provenance

Each row records the original source path in source_image_path. Examples:

BI2K:
BI2K_3/0/1/1001.jpg

BM-BronchoLC:
Lung_cancer/imgs/<patient_id>/<video_id>/<image_id>.png

UAAL:
coco_ins_clinical/train2017/<image>.PNG

Intended Use

Recommended uses:

• training and evaluating bronchoscopy depth-estimation pipelines;
• training image-to-depth or depth-conditioned bronchoscopy models;
• training text + depth/normal conditioned bronchoscopy appearance models;
• sampling across upper-airway and lower-airway endoscopy imagery;
• studying label unification across public bronchoscopy datasets.

Out-of-scope uses:

• clinical diagnosis;
• patient-level decision-making;
• regulatory use;
• treating source pathology labels as definitive image-level clinical diagnoses.

Limitations

• The source datasets were collected for different tasks: disease classification, lower-airway landmark / lesion annotation, and upper-airway anatomical landmark segmentation.
• BI2K numeric folder classes are normalised as 0=normal, 1=benign, and 2=malignant based on the public BI2K class description.
• BM-BronchoLC category labels are mapped as Lung_cancer=malignant and Non_lung_cancer=normal; the original category names remain available in source_labels and should not be interpreted as definitive per-image lesion labels.
• UAAL contributes anatomy labels and upper-airway appearance, not pathology labels.
• BERD contributes image-only rows in this release; source report text and report classification labels are not used as target labels. BERD pathology/status fields come from compact MedGemma visible-image reads and should not be interpreted as clinical diagnoses.
• Source licences and dataset terms still apply. BM-BronchoLC and UAAL are released under CC BY 4.0; BI2K usage should follow the Kaggle dataset terms and the cited paper.

Loading

from datasets import load_dataset

ds = load_dataset("chrisvoncsefalvay/single-bronchoscopy-depth", split="train")
row = ds[0]

print(row["source_dataset"])
print(row["pathological_status"])
print(row["source_image_path"])
print(row["image"])
print(row["depth"])

Source Datasets and Citations

BERD

Source dataset: BERD: Fine-Grained Bronchoscopy Examination Report Dataset.
Source DOI: 10.57760/sciencedb.28018.
Version: V4.
Licence: CC BY 4.0.
Original source file: dataset.zip.
Original source path pattern: dataset/images/<image>.png.

@dataset{luo2026berd,
  title     = {BERD: Fine-Grained Bronchoscopy Examination Report Dataset},
  author    = {Luo, Xingjian},
  publisher = {Science Data Bank},
  year      = {2026},
  version   = {V4},
  doi       = {10.57760/sciencedb.28018}
}

BI2K

Source dataset: Bronchoscopic Dataset BI2K 3 Classification.
Original source path pattern: BI2K_3/<class_id>/<case_or_sequence>/<image>.jpg.

@article{sun2025accurate_bi2k,
  title   = {An accurate prediction for respiratory diseases using deep learning on bronchoscopy diagnosis images},
  author  = {Sun, Weiling and Yan, Pengfei and Li, Minglei and Li, Xiang and Jiang, Yuchen and Luo, Hao and Zhao, Yanbin},
  journal = {Journal of Advanced Research},
  volume  = {76},
  pages   = {423--438},
  year    = {2025},
  doi     = {10.1016/j.jare.2024.11.023}
}

BM-BronchoLC

Source dataset: BM-BronchoLC Figshare dataset.
Source DOI: 10.6084/m9.figshare.24243670.v3.
Original source files: Lung_cancer.zip, Non_lung_cancer.zip.
Original source path pattern: <category>/imgs/<patient_id>/<video_id>/<image_id>.png.

@article{vu2024bm_broncholc,
  title   = {BM-BronchoLC - A rich bronchoscopy dataset for anatomical landmarks and lung cancer lesion recognition},
  author  = {Vu, Van Giap and Hoang, Anh Duc and Phan, Thu Phuong and Nguyen, Ngoc Du and Nguyen, Thanh Thuy and Nguyen, Duc Nghia and Dao, Ngoc Phu and others},
  journal = {Scientific Data},
  volume  = {11},
  pages   = {321},
  year    = {2024},
  doi     = {10.1038/s41597-024-03145-y}
}

@dataset{vu2023bm_broncholc_figshare,
  title     = {BM-BronchoLC},
  author    = {Vu, Van Giap},
  publisher = {figshare},
  year      = {2023},
  doi       = {10.6084/m9.figshare.24243670.v3}
}

UAAL

Source dataset: UAAL Dataset Figshare record.
Source DOI: 10.6084/m9.figshare.26342779.v4.
Original source files: coco_ins_clinical.zip, coco_ins_phantom.zip, annotation visualization.zip.
Original source path pattern: coco_ins_<subset>/<split>/<image>.PNG.

@article{hao2025uaal,
  title   = {Upper Airway Anatomical Landmark Dataset for Automated Bronchoscopy and Intubation},
  author  = {Hao, Ruoyi and Zhang, Yang and Tang, Zhiqing and Zhou, Yang and Seenivasan, Lalithkumar and Chan, Catherine Po Ling and Chan, Jason Ying Kuen and others},
  journal = {Scientific Data},
  volume  = {12},
  pages   = {1907},
  year    = {2025},
  doi     = {10.1038/s41597-025-06169-0}
}

@dataset{hao2024uaal_figshare,
  title     = {UAAL Dataset: Upper Airway Anatomical Landmark Dataset for Automated Bronchoscopy and Intubation},
  author    = {Hao, Ruoyi and Zhang, Yang and Tang, Zhiqing and Zhou, Yang and Seenivasan, Lalithkumar and Chan, Catherine Po Ling and Chan, Jason Ying Kuen and others},
  publisher = {figshare},
  year      = {2024},
  doi       = {10.6084/m9.figshare.26342779.v4}
}