SonoIntent / README.md
QingboKang's picture
docs: link pretrained models (SonoIntent-Models) in dataset card
5727634 verified
|
Raw History Blame Contribute Delete
11.4 kB
metadata
license: cc-by-4.0
language:
  - en
task_categories:
  - image-segmentation
  - visual-question-answering
pretty_name: SonoIntent
size_categories:
  - 10K<n<100K
tags:
  - ultrasound
  - medical-imaging
  - reasoning-segmentation
  - clinical-objective

SonoIntent

SonoIntent is a benchmark for clinical objective-driven ultrasound reasoning segmentation, built from 30 existing ultrasound datasets. The full benchmark contains 130,928 image-question samples from 83,324 images; every sample pairs an ultrasound image with a clinical question, a reference reasoning that resolves the question to the target structure, and a binary target mask.

Repository: https://huggingface.co/datasets/QingboKang/SonoIntent

This repository releases the HC18, PSFHS, BrEaST, CCAUI, CUBS, FASS and Segthy subset, including the rendered images and binary masks. No other source's pixels are included.

Release status

The released subset contains 16,869 annotations with pixel data for seven sources whose licenses confirm redistribution. No sample ID, public ID, split assignment, question, reasoning text, asset path or mask content was changed.

source license images train val test masks test ood_type
HC18 CC BY 4.0 998 898 50 50 998 ID
PSFHS (FH-PS-AOP) CC BY 4.0 1,358 2,324 130 128 2,582 ID
BrEaST CC BY 4.0 225 202 11 12 225 ID
CCAUI CC BY 4.0 1,005 904 50 51 1,005 ID
CUBS CC BY 4.0 2,176 0 0 2,176 2,176 OOD_intent
FASS CC BY 4.0 1,588 4,109 257 187 4,553 ID
Segthy CC BY 5,330 0 0 5,330 5,330 OOD_intent
released total 12,680 8,437 498 7,934 16,869

The released test split has 428 ID samples and 7,506 OOD_intent samples. It contains no Dataset-OOD, Organ-OOD or target-absent samples. Every image is stored once. PSFHS has 1,224 images shared by two tasks and FASS has 1,588 images shared by multiple structure tasks.

The HC18 and PSFHS rows and assets are unchanged from the earlier release: their public IDs, splits, text, asset paths and SHA-256 values are byte-identical.

All other sources are withheld. EchoNet-Dynamic prohibits redistribution; CAMUS, ACOUSLIC-AI, BUS-UCLM, GIST514-DB and regPro are restricted; the remaining sources lack a confirmed data-level license for derived annotations and pixels. Per-source status, evidence and the release_disposition column are in metadata/source_register.csv; metadata/release_tiers.yaml and metadata/statistics.json carry the release block.

The full benchmark composition below is reported for context. It is not a statement that all referenced records are downloadable here.

Full benchmark composition

split task samples unique image frames
train 102,248 59,329
val 5,810 3,383
test 22,870 20,612
main total 130,928 83,324
target-absent (independent test) 656 656

Test settings (by ood_type): ID 5,385, Dataset-OOD 3,829, Organ-OOD 3,917, Objective-OOD 9,739.

Data format

Only train.jsonl, val.jsonl and test.jsonl are present; the target_absent split is not part of this release. Each line:

field meaning
public_id opaque per-task alias (never changes)
dataset source dataset name
split train / val / test
ood_type ID or OOD_intent for test rows; empty for train/val
problem clinical question
thinking reference reasoning
asset_image / asset_mask relative image and mask paths under assets/
image_sha256 / mask_sha256 SHA-256 of the frozen image / mask
license_tier source tier from the release audit
asset_availability released for every row in this repository
asset_available_now yes for every row in this repository

metadata/image_index.csv carries the same per-task fields for the released subset and joins to the annotations on public_id; all shared fields are equal.

Images and masks

Assets are grouped by source:

assets/hc18/img/<public-image-id>.png
assets/hc18/mask/<public-id>.png
assets/fh_ps_aop_psfhs/img/<public-image-id>.png
assets/fh_ps_aop_psfhs/mask/<public-id>.png
assets/breast/img/<public-image-id>.png
assets/breast/mask/<public-id>.png
assets/ccaui/img/<public-image-id>.png
assets/ccaui/mask/<public-id>.png
assets/cubs/img/<public-image-id>.png
assets/cubs/mask/<public-id>.png
assets/fass/img/<public-image-id>.png
assets/fass/mask/<public-id>.png
assets/segthy/img/<public-image-id>.png
assets/segthy/mask/<public-id>.png
source images masks image format
HC18 998 998 472x320, 8-bit grayscale PNG
PSFHS 1,358 2,582 256x256, 8-bit grayscale PNG
BrEaST 225 225 source resolution (363-1024 x 289-618), RGBA PNG
CCAUI 1,005 1,005 709x749, RGB PNG
CUBS 2,176 2,176 618x464 or 800x652/636, 8-bit grayscale PNG
FASS 1,588 4,553 1024x768, 8-bit grayscale PNG
Segthy 5,330 5,330 source resolution (395-696 x 462-617), 8-bit grayscale PNG

There are 12,680 unique images, 16,869 masks and 1,285,911,223 bytes of asset data in total. Every mask is binary (0/255) and has the same dimensions as its image. All masks are the frozen benchmark targets; no mask content or assignment was changed. BrEaST images retain the source RGBA alpha channel; 29 fan-shaped images have transparent pixels outside the acquisition fan. assets/README.md gives the per-source attribution, version and modification notes.

License and attribution

The SonoIntent-authored question and reasoning text, the index and the public identifiers are offered under the Creative Commons Attribution 4.0 International license; see LICENSE-ANNOTATIONS. Each source's image and mask files remain under that source's own CC BY terms, with the attribution and modification notices below and in assets/README.md. This repository does not relicense any withheld source.

The released assets and annotations are derived from:

  • HC18 - van den Heuvel et al., "Automated measurement of fetal head circumference using 2D ultrasound images", PLoS ONE 13(8):e0200412 (2018), https://doi.org/10.1371/journal.pone.0200412. Official Zenodo record v2: https://zenodo.org/records/1327317. CC BY 4.0. Changes: 998 of the 999 training frames are used; frames are rendered at 472x320; masks are binary fills of the released 1-pixel ellipse outlines; public identifiers and question/reasoning annotations were generated.
  • PSFHS (FH-PS-AOP) - Lu et al., "PSFHS: Intrapartum ultrasound image dataset for AI-based segmentation of pubic symphysis and fetal head", Scientific Data 11:436 (2024), https://doi.org/10.1038/s41597-024-03266-4. Official Zenodo record V1: https://zenodo.org/records/10969427. Only the public 1,358-frame subset is used. CC BY 4.0. Changes: images converted to 8-bit grayscale 256x256; the three-class source labels are split into binary pubic-symphysis and fetal-head masks; public identifiers and question/reasoning annotations were generated.
  • BrEaST - Pawłowska et al., "Curated benchmark dataset for ultrasound based breast lesion analysis", Scientific Data 11:148 (2024), https://doi.org/10.1038/s41597-024-02984-z. Official TCIA collection BREAST-LESIONS-USG, Version 1 (updated 2024-01-08): https://www.cancerimagingarchive.net/collection/breast-lesions-usg/, DOI https://doi.org/10.7937/9WKK-Q141. CC BY 4.0; TCIA marks the collection "Data Citation Required". Changes: 225 images used by the frozen benchmark (the official collection has 522 series from 256 subjects); images retain source resolution, RGBA mode and alpha channel; binary masks are derived from the source tumor-mask alpha channel; public identifiers and question/reasoning annotations were generated. TCIA's Data Usage Policy applies: cite the dataset DOI and the article, and do not attempt re-identification.
  • CCAUI - Momot, "Common Carotid Artery Ultrasound Images", Mendeley Data v1 (2022-11-14), https://data.mendeley.com/datasets/d4xt63mgjm/1, DOI https://doi.org/10.17632/d4xt63mgjm.1. CC BY 4.0. Changes: 1,005 of the 1,100 source images are used by the frozen benchmark; RGB images are retained at 709x749; expert masks are binarized to 0/255; public identifiers and question/reasoning annotations were generated.
  • CUBS - Meiburger et al., "Carotid Ultrasound Boundary Study (CUBS): An Open Multicenter Analysis of Computerized Intima-Media Thickness Measurement Systems and Their Clinical Impact", Ultrasound in Medicine & Biology 47(8):2442-2455 (2021), https://doi.org/10.1016/j.ultrasmedbio.2021.03.022. Mendeley Data v1 (2021-05-04), https://data.mendeley.com/datasets/fpv535fss7/1, DOI https://doi.org/10.17632/fpv535fss7.1. CC BY 4.0. Changes: the 2,176 public images are rendered as 8-bit grayscale PNG at source resolution; the manual A1 intima-media boundaries are filled into a binary task mask; public identifiers and question/reasoning annotations were generated.
  • FASS - Correggio et al., "Fetal Abdominal Structures Segmentation Dataset Using Ultrasonic Images", Mendeley Data v1 (2023-10-18), https://data.mendeley.com/datasets/4gcpm9dsc3/1, DOI https://doi.org/10.17632/4gcpm9dsc3.1. CC BY 4.0. Changes: 1,588 public .npy records are used; each RGB 768x1024 image is converted to an 8-bit grayscale 1024x768 PNG; each annotated fetal structure becomes a separate binary task mask, giving 4,553 tasks; public identifiers and question/reasoning annotations were generated.
  • Segthy - Krönke et al., "Tracked 3D ultrasound and deep neural network-based thyroid segmentation reduce interobserver variability in thyroid volumetry", PLoS ONE 17(7):e0268550 (2022), https://doi.org/10.1371/journal.pone.0268550. Official dataset page: https://www.cs.cit.tum.de/en/camp/publications/segthy-dataset/. The page states the dataset is licensed under a CC BY license, permits commercial use and requires acknowledgement of the creators. Changes: 5,330 slices are selected from the public US volumes; each frame is rendered as 8-bit grayscale PNG at source resolution; the thyroid labels are binarized into task masks; public identifiers and question/reasoning annotations were generated.

PNG text/EXIF metadata is absent from the released assets, and filenames are opaque public aliases. BrEaST images retain the source DPI and ICC color-profile chunks; these carry no patient information. A manual review of sampled images across all seven sources found no readable patient name, ID, date or other identifying burned-in text. The visible overlays are non-identifying machine, caliper or relative-timestamp marks. See NOTICE.md and metadata/source_citations.md for further attribution.

Pretrained models

Checkpoints fine-tuned on SonoIntent are released in a separate model repository: https://huggingface.co/QingboKang/SonoIntent-Models (Sa2VA-14B and VRS-HQ; usage examples and license files are included there).

Citation

TODO: paper citation - do not cite this repository alone.

Contact

Qingbo Kang
kangqingbokb@gmail.com