PathVG (Pathology Visual Grounding) is the first large-scale benchmark designed for linking natural language expressions to precise regions in pathology images. It enables research at the intersection of computational pathology, computer vision, and natural language processing.

The dataset bridges the gap between image-level question answering and region-level localization, providing a standardized benchmark for visual grounding in pathology.

The full paper is available on arXiv: https://arxiv.org/abs/2502.20869

📂 Dataset Statistics

Images: 27,610 pathology images
Annotations: 33,500 expert-verified expressions with bounding boxes
Magnifications:
- 20×: Focus on cell arrangement and interactions
- 40×: Emphasize cellular structure and tumor infiltration
Annotation Quality: All expressions are created and validated by pathology experts via multi-stage review

📊 Data Format

Each sample contains:

image_id: Image filename (unique identifier)
height: Image height in pixels
width: Image width in pixels
bbox: Bounding box coordinates in format [x_min, y_min, x_max, y_max]
bbox_id: Unique bounding box identifier within the image
expression: List of natural language descriptions associated with the bounding box

Example:

{
  "bbox": [1680, 603, 1818, 756],
  "image_id": "8006.jpg",
  "height": 2048,
  "width": 2048,
  "bbox_id": 1,
  "expression": [
    "Upper-right image shows distinct tumor cell cluster spreading."
  ]
}

Downloads last month: 89