prov-gigapath/prov-gigapath

Prov-GigaPath: A whole-slide foundation model for digital pathology from real-world data

What it solves

Prov-GigaPath addresses the challenge of analyzing whole-slide images (WSI) in digital pathology. These images are typically too massive for standard AI models to process as a single input, requiring a way to extract meaningful features from both local tissue patterns and the overall slide-level context.

How it works

The system uses a two-stage encoding architecture:

  1. Tile Encoder: Extracts local patterns and features from individual image patches (tiles) of a slide.
  2. Slide Encoder: Takes the embeddings from the tile encoder along with their spatial coordinates to generate a comprehensive representation of the entire slide.

To improve efficiency, the project provides GigaPath-Flash, a lightweight variant that uses a LongNet backbone and a DINOv2-small tile encoder to reduce memory usage and increase inference speed.

Who it’s for

It is designed for AI researchers working on pathology foundation models, digital pathology, and the encoding of whole-slide image data.

Highlights

  • Foundation Model: A large-scale model trained on real-world digital pathology data.
  • Two-Tier Architecture: Separates local patch-level feature extraction from global slide-level aggregation.
  • Efficiency Variants: Includes "Flash" versions for faster inference and lower memory requirements.
  • Research-Ready: Provides pre-extracted embeddings for datasets like PCam and PANDA to facilitate fine-tuning experiments.

Related

  • Project
  • Project
  • Project
  • Project
  • Project