Dataset Name: LitMedImage – literature-derived Medical vs. Non-Medical Image Dataset
Description:
LitMedImage is a curated dataset of biomedical literature figures labeled as MEDICAL or NON-MEDICAL. The dataset is built from images extracted from PubMed Central Open Access (PMC-OA) articles and includes corresponding captions and parsed image metadata. Labels were generated using a large language model (LLM) following strict imaging definitions. This dataset is intended for research in figure classification, document parsing, and biomedical vision-language models.
Columns:
PMCID: PubMed Central article identifier (e.g., PMC1234567) Image_num: Index of the image within the article Online_file_path: Direct file path to the image under https://ftp.ncbi.nlm.nih.gov/pub/pmc/
Image_info_Cleaned: Parsed metadata describing the image contents Caption_Clean: Cleaned image caption from the original publication label: Binary classification label ("yes" for MEDICAL images, "no" for NON-MEDICAL images)
Label Definitions:
Label = "yes" (MEDICAL) Images that belong to clinical or biomedical imaging modalities, including: Radiology, Echocardiography, Dermoscopy, Histopathology, X-ray (Radiography), CT, MRI, Ultrasound, Nuclear Medicine (PET, SPECT, PET-CT, PET-MRI), Optical Imaging, Thermography, Elastography, Mammography, Digital Breast Tomosynthesis, Fluoroscopy, Clinical Imaging
Label = "no" (NON-MEDICAL) Images that are statistical plots or schematic illustrations, including: Bar charts, Histograms, Line graphs, Scatter or Bubble plots, Pie or Donut charts, Area charts, Heatmaps, Box or Violin plots, Radar or Spider charts, Treemaps, Network graphs, Drawings, Conceptual diagrams
Task Instruction:
Given an image, classify whether it is a MEDICAL image or a NON-MEDICAL image.
Data Source: All images originate from the PubMed Central Open Access Subset via the public FTP archive at: https://ftp.ncbi.nlm.nih.gov/pub/pmc/
Intended Use:
- Binary image classification (medical vs. non-medical)
- Multimodal image + caption classification
- Figure filtering for automated document processing
- Pre-filtering figures for vision-language model training or inference
Citation: [To be added]