Pathology slide processing with 22+ foundation models built in

What it does

Trident is an end-to-end toolkit for processing whole-slide pathology images at scale. It takes raw slides—read via OpenSlide, CuCIM, OME-Zarr, Zeiss CZI, and plain images—and runs them through tissue segmentation, patch coordinate extraction, and feature embedding using a wide catalog of patch and slide encoders. The output is a directory of embeddings and machine-readable run reports, ready for downstream analysis.

The interesting bit

The README spends as much space on operational hygiene as on model support, which tells you who the audience is. Per-slide job tracking, .lock files, and a producer/consumer cache pipeline for slow network storage mean this is built for clusters where large batch jobs run for days and can hit flaky storage. That kind of defensive engineering is usually the first thing researchers omit.

Key highlights

• Supports 22+ patch encoders (UNI, Virchow, H-Optimus-0, CONCHv1.5, etc.) and several slide-level encoders (Titan, GigaPath, PRISM, CHIEF, Madeleine, Feather)
• End-to-end pipeline from tissue segmentation through patching to embeddings, runnable as --task all or stage-by-stage
• Multi-GPU distribution with per-slide resume logic and automatic cleanup of stale locks
• WSI cache pipeline that stages slides locally via a producer/consumer pattern for slow or network-backed storage
• Run reports in Markdown and JSON, plus per-slide state files for debugging and accountability

Caveats

• Some models require manual setup (e.g., local CHIEF repository path) or HuggingFace gated access approvals before they will load

Verdict

Grab this if you are a pathology lab sitting on whole-slide images and want to run foundation-model embeddings without maintaining your own preprocessing stack. Everyone else can safely keep scrolling.