license: cc-by-nc-4.0

SonoVQA

Dataset Summary

SonoVQA is an ultrasound visual question answering (VQA) benchmark with 14,905 expert-annotated QA pairs derived from the UltrasoundCases educational platform. It covers 18 anatomical organs, 6 reasoning levels, and 3 question types across a broad range of ultrasound imaging scenarios.

This is an annotation-only release. The cc-by-nc-4.0 license applies to the SonoVQA annotations, QA pairs, taxonomy labels, splits, and metadata. It does NOT apply to the source ultrasound images, which are not redistributed.

Dataset Contents

• 14,905 QA pairs across train/validation/test splits
• 18 anatomical organs
• 7 anatomical systems
• 6 reasoning levels: protocol, system, organ, sonographic_feature, lesion_category, disease_diagnosis
• 7 reasoning sublevels: imaging_mode_recognition, scanning_protocol_recognition, anatomical_system_identification, target_organ_localization, tumor_non_tumor_classification, fine_grained_lesion_category_classification, fine_grained_disease_diagnosis, sonographic_feature_interpretation
• 3 question types: mcq (10,521), open_ended (4,384)
• Source images from the UltrasoundCases educational platform

Data Files

data/
├── train.jsonl         (10,111 samples)
├── validation.jsonl     (1,462 samples)
└── test.jsonl           (3,332 samples)

Data Fields

Each JSONL line contains one QA sample with the following fields in order:

Field	Type	Description
id	string	Unique sample ID (e.g., sonovqa_train_000000)
split	string	train, validation, or test
source	string	Source platform (UltrasoundCases)
task_type	string	Task type (vqa)
question_type	string	mcq or open_ended
question	string	Cleaned question text
options	dict or null	Option dict for mcq (e.g., {"A": "B-mode", "B": "CDFI"}), null otherwise
answer	string	Answer in A: B-mode format
answer_key	string or null	Option key for mcq (e.g., A), null for open_ended
answer_text	string	Answer text without option key
prompt	string	Original raw prompt text for reproducibility
source_case	string	Case identifier (e.g., 3.2.4 Malignant ovarian lesions/case_016)
image_reference	string	Relative image path within source data
image_filename	string	Original image filename
image_redistributed	boolean	Always false — images are not included
reasoning_level	string	High-level reasoning category
reasoning_sublevel	string	Fine-grained reasoning task
protocol	string	Scanning protocol (e.g., Abdominal, OB-GYN)
system	string	Anatomical system
organ	string	Target organ
lesion_category	string	Lesion category (tumor/non-tumor)
disease_group	string	Disease group (e.g., Benign tumors, Malignant tumors)
diagnosis	string	Specific diagnosis short name

Reasoning Levels

Level	Count	Description
sonographic_feature	4,306	Echogenicity, margin, vascularity, posterior features
lesion_category	3,083	Tumor vs non-tumor classification
protocol	3,007	Scanning protocol recognition
system	1,503	Anatomical system identification
organ	1,503	Target organ localization
disease_diagnosis	1,503	Fine-grained disease diagnosis

Organs

Breast (2,224), liver (1,872), veins (1,157), Pancreas (1,130), Hand (968), SG (952), Kidney (887), Bile ducts (773), Muscle (739), Adnexa (737), Testis (733), Thyroid (704), Biliary (477), Gallbladder (426), Urinary tract (353), Bladder (307), Uterus (300), Nerve (166).

Data Splits

Split	Samples
train	10,111
validation	1,462
test	3,332
Total	14,905

Splits are stratified by anatomical system and diagnosis to ensure balanced coverage across all splits.

Annotation-only Release and Image Copyright

Source ultrasound images are NOT redistributed in this release due to third-party copyright restrictions.

The cc-by-nc-4.0 license applies solely to the SonoVQA annotations, QA pairs, taxonomy labels, data splits, and metadata. It does NOT cover the source ultrasound images, which remain under their original copyright terms from the UltrasoundCases educational platform.

Each sample includes source_case, image_reference, and image_filename as provenance references. Users who need the images should obtain them directly from the original source where permitted, and must comply with the source website's terms of use.

Intended Use

• Ultrasound VQA model training and evaluation
• Multi-modal reasoning benchmark for medical imaging
• Ultrasound education and knowledge assessment
• Fine-grained visual recognition of sonographic features

Limitations

• Image redistribution is prohibited; users must obtain images independently
• QA pairs are expert-annotated but may contain inherent annotator variability
• The dataset covers 18 organs but is not exhaustive for all ultrasound applications
• Some organ categories have fewer samples (e.g., Uterus: 300, Nerve: 166, Bladder: 307)

Loading Example

from datasets import load_dataset

ds = load_dataset(
    "json",
    data_files={
        "train": "data/train.jsonl",
        "validation": "data/validation.jsonl",
        "test": "data/test.jsonl"
    }
)
print(ds["train"][0])

Citation

If you use SonoVQA in your research, please cite:

@misc{sonovqa2025,
  title        = {{SonoVQA}: Ultrasound Visual Question Answering Benchmark},
  author       = {{SonoReasoner Contributors}},
  year         = {2025},
  note         = {Annotation-only release. Images not redistributed.},
}

Contact

For questions, please contact the dataset maintainers.