MedicalNarratives is a large-scale, multimodal medical vision–language corpus built from pedagogical videos and scientific articles. It is designed to unify semantic (classification, retrieval, captioning) and dense (detection, segmentation, grounding) tasks across clinical imaging.

Key properties:

4.7M image–text pairs
≈1M samples with cursor-trace or bounding-box spatial grounding
12 imaging domains including X-ray, CT, MRI, ultrasound, histopathology, dermatology, etc.
UMLS grounding for medical and ROI text
Sources:
- 74k narrated YouTube pedagogy videos (~4,526 hours)
- ≈273k PubMed Central OA articles with ≈1.03M figure–caption pairs

Project page: https://medical-narratives.github.io

Supported Tasks

Vision–language pretraining
Zero-shot & supervised classification
Image–text retrieval
Grounded captioning / visual grounding
Open-vocabulary detection and segmentation

Dataset Structure

General Fields

uid: str — Unique ID {VIDEO_ID}_{IDX} or {PMCID}
index: int — Global dataset index
image: str — Image filepath
text: List[str] — Denoised captions
clip_name: str — Filename of video clip corresponding to image
clip_name: str — Video clip identifier
med_num: int — Number of captions
video: bool — Video-derived sample
article: bool — Article-derived sample
sample_index: int — Index within source video
domains: str — Imaging domain list as string
sample_uids: Dict[str, int] — Mapping of sibling samples in same video
sample_count: int — Total samples from this video
multidomain: bool — True if >1 modality
multisample: bool — True if sample_count >1

Video-Specific Fields

video_id: str — YouTube ID
video_link: str — YouTube URL
video_channel: str — Channel name
duration: float — Seconds
fps: float — Frames per second
spw: float — Seconds per word (ASR alignment)
pair_chunk_time: float — Target chunk duration
chunk_times: List[List[float]] — Stable-scene intervals
roi_chunk_times: List[List[float]] — ROI-activity intervals
width: int, height: int — Frame dimensions
chunk_id: int — Chunk index
chunk_start_time: float, chunk_end_time: float — Timestamps
chunk_medical_text: List[str] — Chunk-level medical text
chunk_roi_text: List[str] — ROI text
chunk_med_umls: List — UMLS for medical text
chunk_roi_umls: List — UMLS for ROI text
noisy_timed_captions: List[Dict] — Raw ASR transcript
asr_correction_dict: Dict[str, str] — ASR corrections
umls: List[List[Dict]] — UMLS for each caption in text
magnification — Video magnification (usually null)
img_time — Processing field (ignore)
roi — ROI spatial annotations (boxes/masks)
traces — Cursor traces (x, y, t)
subdomain: List[str] — Fine-grained subdomains

Article-Specific Keys

PMCID: str — PubMed Central article ID
pubmed_link: str — Article URL (https://www.ncbi.nlm.nih.gov/pmc/articles/{PMCID}/)
img_xref_id: str — Figure or subfigure cross-reference ID from the article
pubmed_caption: str — Cleaned full figure caption
pubmed_caption_raw: str — Raw figure caption before processing
img_inline_sentences: List[str] — Article sentences referencing the figure
subcaptions: List[Dict] — Subfigure caption entries (with image filenames normalized)
subfigures: Dict[str, Any] — Mapping from subfigure image filename → metadata

Data Properties

Mean image resolution: ~1521 × 903 px
Cursor-trace subset: >100M points, ~546k bounding boxes
Mean caption: 22.4 words, 1.9 sentences
~4M UMLS mentions, ~300k unique concepts

Citation

If you use MedicalNarratives, please cite:

@misc{ikezogwo2025medicalnarrativesconnectingmedicalvision,
  title        = {MedicalNarratives: Connecting Medical Vision and Language with Localized Narratives},
  author       = {Ikezogwo, Wisdom O. and Zhang, Kevin and Seyfioglu, Mehmet Saygin and Ghezloo, Fatemeh and Shapiro, Linda and Krishna, Ranjay},
  year         = {2025},
  eprint       = {2501.04184},
  archivePrefix= {arXiv},
  primaryClass = {cs.CV},
  url          = {https://arxiv.org/abs/2501.04184}
}

Downloads last month: 9