mahmoodlab/TRIDENT

What it solves

Trident simplifies the complex process of large-scale whole-slide image (WSI) processing for AI pathology. It provides a unified toolkit to handle the end-to-end pipeline of converting images, segmenting tissue, extracting patch coordinates, and generating embeddings using a wide variety of state-of-the-art foundation models.

How it works

Trident operates as a multi-stage pipeline that can be run as a single command or in discrete steps:

1. Tissue Segmentation: Uses models like HEST or GrandQC (or Otsu's method) to separate tissue from background and remove artifacts.
2. Tissue Patching: Extracts coordinates for small image patches from the segmented tissue regions at specific magnifications.
3. Feature Extraction:
   • Patch-level: Uses one of 33 supported patch encoders (e.g., UNI, CONCH, Virchow) to create embeddings for individual patches.
   • Slide-level: Uses one of 12 slide encoders (e.g., Titan, GigaPath, PRISM) to aggregate patch data into a single embedding for the entire slide.

To handle the scale of pathology data, it includes a WSI cache pipeline for slow storage, multi-GPU distribution, and a smart resume system to skip already processed slides.

Who it’s for

It is designed for researchers and practitioners in AI pathology who need to process massive histological datasets and generate high-dimensional features for downstream analysis using various foundation models.

Highlights

• Extensive Model Support: Integrates 33 patch encoders and 12 slide encoders.
• End-to-End Workflow: Single command execution for the entire pipeline from segmentation to embedding.
• Robust Infrastructure: Features multi-GPU support, smart resume capabilities, and a producer/consumer cache for network storage.
• Broad Format Support: Compatible with OpenSlide, CuCIM, OME-Zarr, Zeiss CZI, and pyramidal TIFFs.