Synthetic Medical Vision · OpenMed
synthvision-seeds
A curated collection of 119,137 medical image records serving as seeds for the SynthVision data generation pipeline.
What synthvision-seeds is
synthvision-seeds is a medical imaging dataset published by OpenMed. It brings together 119,137 records from four open medical image repositories to serve as reference imagery and seed prompts for the SynthVision generation pipeline, which produces annotated training corpora for medical vision tasks.
Developing multimodal vision models for healthcare requires accessible reference imagery across different medical domains. This dataset unifies diverse clinical sources spanning radiology, pathology, and mixed modalities alongside their associated textual descriptions, question-and-answer pairs, and labels into a single standardised schema.
How it was made
The corpus aggregates entries from four open-access datasets: eltorio/ROCO-radiology (65,393 records), OpenMed/multicare-images (50,000 records), flaviagiammarino/path-vqa (3,430 records), and flaviagiammarino/vqa-rad (314 records). Duplicate images across these collections are detected and filtered using SHA-256 hashes.
Each entry in the dataset is structured under a consistent schema consisting of a unique identifier, a relative image path, the source dataset identifier, the specific imaging modality, and a metadata dictionary containing any captions, visual question-and-answer pairs, or classification labels provided by the source.
Key facts
- Record count
- 119,137
- Modalities
- Radiology, Pathology, Mixed
- File format
- Parquet
- Task category
- Visual question answering
- Constituent datasets
- ROCO-radiology, multicare-images, path-vqa, vqa-rad
- Deduplication method
- SHA-256 hash
What you can use it for
- Seed generation for synthetic medical datasets
- Supplying prompt text and initial image references to feed the SynthVision pipeline for generating annotated medical computer vision corpora.
- Visual question answering research
- Training or evaluating multimodal algorithms on clinical visual questions and answers derived from collections such as VQA-RAD and Path-VQA.
- Multimodal medical data benchmarking
- Serving as a standardised multi-source index across radiology and pathology images for comparative vision-language analysis.
Explore the data
The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from OpenMed/synthvision-seeds at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 119,137 rows of default/train
id
string_text- min
- 6
- median
- 10
- mean
- 12.20
- max
- 23
image_path
string_text- min
- 75
- median
- 79
- mean
- 83.39
- max
- 95
metadata
string_text- min
- 26
- median
- 26
- mean
- 35.79
- max
- 61
modality
string_label- radiology88%
- pathology11%
- ct_scan1%
seed_labels
list- min
- 0
- median
- 0
- mean
- 0
- max
- 0
seed_qa
string_text- min
- 70
- median
- 102
- mean
- 113.30
- max
- 306
96.9% empty
source
string_label- roco55%
- multicare42%
- pathvqa3%
tier
string_label- B97%
- C3%
Strengths and limits
Good at
- Aggregates 119,137 records from four distinct open medical datasets into a unified schema.
- Removes redundant records across source corpora through cryptographic SHA-256 deduplication.
- Preserves original annotations, including image captions, question-and-answer pairs, and diagnostic labels within metadata fields.
- Spans multiple clinical imaging disciplines, including radiology and pathology.
Watch out for
- Records reference relative image paths rather than embedding raw image pixels directly.
- Metadata content and depth vary depending on the availability of annotations in each contributing source collection.
Questions
Can I look inside synthvision-seeds without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.
What licence is synthvision-seeds under?
Its authors declare apache-2.0. Read the terms on the original page before reusing the data.
Which datasets are included in synthvision-seeds?
It contains 65,393 radiology records from eltorio/ROCO-radiology, 50,000 mixed records from OpenMed/multicare-images, 3,430 pathology records from flaviagiammarino/path-vqa, and 314 radiology records from flaviagiammarino/vqa-rad.
What information does each record contain?
Every record includes a unique identifier, a relative image path, the source dataset name, the imaging modality, and a metadata dictionary containing captions, Q&A pairs, or labels.
How are duplicate images handled?
Images are deduplicated across the aggregated collections using SHA-256 cryptographic hashing.
What role does this dataset play in SynthVision?
It provides the source imagery and seed prompts that feed the SynthVision pipeline to produce annotated medical-imaging training sets.
Written from synthvision-seeds's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.