Grayye233/BN4601A-Colonoscopy-Curated

BN4601A curated colonoscopy dataset

An academic curation of Kvasir-SEG and PolypGen, with audited identities, fixed center splits, native binary masks, mask geometry and linked weak-label Examine / Explain / Report examples, plus a small original Kvasir-VQA supplement for visual question answering. This repository is hosted by Grayye233 for the BN4601A project. It does not claim ownership of the original images or create a new independently collected clinical cohort.

Public repository: teammates can view and download it without sharing the owner's credentials. All data remain subject to their source-specific research and education terms. This is not a clinical diagnostic dataset with histology labels. Read source terms before use.

See what the data look like

The orange regions are source annotation masks, not model predictions. These examples illustrate train/validation/test and positive/negative inputs; they do not summarize model accuracy. Source credits: Jha et al. for Kvasir-SEG; Ali et al. for PolypGen. Full selected IDs: preview_samples.json.

Use the Dataset Viewer configuration dropdown to select segmentation, core_sft, or vqa_sft, then select a split. Open the VQA training rows directly. VQA rows show the image, question, original source answer and rendered English sentence target. Segmentation rows show image and mask columns plus source, center, group, positivity, original dimensions and geometry. In core_sft, image_id links to the segmentation row; image bytes are stored once rather than repeated in every task.

Included configurations

Configuration	Train	Validation	Test	Contents
segmentation	5,373	1,287	2,375	9,035 source images, native binary masks and metadata
core_sft	6,987	1,062	None	8,049 weak task records linked to 2,329 train and 354 validation images
vqa_sft	3,339	523	530	4,392 original-source VQA records across 850 selected images (640 train / 105 validation / 105 test), with source labels and English sentence targets

Counts in the split columns are rows: images for segmentation, and question/task records for core_sft and vqa_sft. Multiple VQA questions can refer to the same image. See the Version 2 section below for source attribution, overlap screening and limits.

The original project audited 9,649 source frames, removed two exact decoded-image duplicates, and used 9,647 distinct frames. 612 CVC-ClinicDB frames were external evaluation only and are not included as image/mask bytes here. Their audited identity manifest is supplied as external_cvc_reference.csv. The current author site lists a separate data-use agreement; its complete redistribution conditions could not be verified during release. Obtain CVC through its author channel and comply with applicable terms. This omission does not remove any frame from the project's completed evaluation, and it does not remove any training or validation frame.

Source	Role in this repository
Kvasir-SEG	800 train / 100 validation / 100 test; all positive
PolypGen C1-C4	4,573 training frames including negatives
PolypGen C5	1,187 validation frames including negatives
PolypGen C6	2,275 test frames: 407 positive and 1,868 negative
CVC-ClinicDB	612 external identities only; never used for training

Changes made to the original data

1. Verified that each source image opens, its paired mask has matching dimensions, and center/sequence assignments follow the source publication.
2. Assigned frame positivity from the annotation mask, including empty masks in nominal positive folders. For author negative-only sequences without annotation files, generated an empty mask and retained this provenance in mask_origin.
3. Removed two exact duplicate decoded RGB images; preserved the exclusions and all source identities.
4. Fixed the center-held-out split before model training. No sequence crosses train/validation/test. Still-image patient identifiers are unavailable; near-duplicate independence is not established.
5. Kept image file bytes unchanged. Converted annotation masks to native-size 0/255 binary PNGs using the same >127 threshold as the completed training/evaluation. No offline image resizing, augmentation or new human contour annotation is claimed. Resizing and random augmentation happen in the training loader.
6. Computed coverage, exclusive-right/bottom bounding boxes, normalized centroids, 3-by-3 image-relative location and connected-component counts. These measure pixels, not millimeters, bowel segment, histology or verified polyp count.
7. Generated three English core-task records per selected training/validation image. Sequence sampling is capped at 12 uniformly spaced frames. Answers use mask facts and explicitly synthetic clinician context, not expert-written clinical ground truth. Histology remains unavailable. Actual training rotates one task per image per epoch.

split in row metadata uses val, while the Hugging Face split is named validation; they refer to the same fixed validation set.

Load in Colab

!pip -q install datasets pillow
from datasets import load_dataset
repo = "Grayye233/BN4601A-Colonoscopy-Curated"
frames = load_dataset(repo, "segmentation", split="train", streaming=True)
row = next(iter(frames))
display(row["image"])
display(row["mask"])
print({k: row[k] for k in ["id", "source_dataset", "center", "positive", "coverage_pct"]})
tasks = load_dataset(repo, "core_sft", split="train", streaming=True)
print(next(iter(tasks))["answer_json"])
vqa = load_dataset(repo, "vqa_sft", split="train", streaming=True)
vqa_row = next(iter(vqa))
display(vqa_row["image"])
print({k: v for k, v in vqa_row.items() if k != "image"})

For a reproducible rerun, pin the repository revision shown in the project's publication receipt. 04_Browse_Shared_Dataset.ipynb in the separate Colab handoff supplies browsing and optional local materialization. The dataset repository contains no trained model weights; complete weights remain in that handoff package.

Limits and licensing

Research and education only. Kvasir retains its original terms. PolypGen's archive includes conflicting BY and BY-NC-SA notices; both originals are preserved, with the stricter archive notice followed for this release. Our curation and derived text contributions follow CC BY-NC-SA 4.0 subject to source terms; this does not relicense source images. No patient identification, clinical deployment, diagnostic accuracy, physical measurement or histopathology claim is supported. Synthetic family history/removal context is for controlled task learning, not real patient information.

Original source citations

• Jha et al. (2020), Kvasir-SEG: A Segmented Polyp Dataset, MMM. Source and terms, paper.
• Ali et al. (2023), A multi-centre polyp detection and segmentation dataset for generalisability assessment, Scientific Data 10, 75. Paper, author repository.
• Bernal et al. (2015), WM-DOVA maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. Paper, author database site, current access agreement. CVC pixels are not redistributed here.

Version 2: small original VQA extension

Select vqa_sft in the viewer to see image/question/source-answer/natural-sentence records. This configuration is separate from unchanged segmentation/core_sft splits.

Split	Images	QA records
train	640	3339
val	105	523
test	105	530

Seven direct visual tasks: visible polyp/instrument counts, instruments, text/box overlays, and polyp region/color. Colonoscopy-only source categories include Polyps, Instrument and Ulcerative Colitis; the latter contributes direct visual questions, not disease-diagnosis targets. No size, histology, polyp-type or removal-history targets are included. Original short labels are rendered as fixed English sentences; these are not independently authored clinical explanations. Training questions include two handcrafted paraphrases; the third is reserved for evaluation.

Screening compared 2,944 candidate images against all 9,647 existing core frames using 64-bit pHash distance <=6. 245 candidates overlapping core held-out images were excluded from training. 339 of the 640 training images match core training images; the 70 polyp validation/test images reuse their corresponding core held-out split. Thus these records must not be described as 850 new independent clinical cases. Final VQA splits have no cross-split pHash candidates at the chosen threshold. Source patients/centers are unavailable and patient independence is not established. See vqa_sft/audit.json and exclusions.json.

License: original VQA source materials retain CC BY-NC 4.0, attribution to Gautam et al. and their source-specific conditions. See licenses/Kvasir-VQA_SOURCE_CARD.md and source terms. No x1 generated answers are included. Source: https://huggingface.co/datasets/SimulaMet-HOST/Kvasir-VQA/tree/26df9125b98cbad664e5b1801aac1f70a02689e2.