PRISM Embedder

About

Creator:

clemsg

Contact email:

clement.grisi@radboudumc.nl

Image Version:

9e002f06-902a-4c1b-8bce-fe1350e6829d

Last updated:

June 18, 2025, 7:01 p.m.

Model Version:

96f1ba72-b94f-47f7-a622-fd981c02b9c1

Last updated:

June 18, 2025, 5:31 p.m.

Interfaces

This algorithm implements all of the following input-output combinations:

Inputs Outputs

	Inputs	Outputs
1	Tissue Mask Description Segmentation of the tissue in the slide. 0: Background 1: Tissue Kind Segmentation Read from `/input/images/tissue-mask/<uuid>.mha` or `/input/images/tissue-mask/<uuid>.tif` Tissue Mask Whole Slide Image Description Multi-resolution whole slide image Kind Image Read from `/input/images/whole-slide-image/<uuid>.mha` or `/input/images/whole-slide-image/<uuid>.tif` Whole Slide Image	Image-level neural representation Description Neural representation at image-level for each input scan. For each input scan a single, one-dimensional vector of floats should be given. Kind Anything Write to `/output/image-neural-representation.json` Example [ { "title": "transverse-t2-prostate-mri", "features": [ 0.86, 0.12, 0.75 ] }, { "title": "transverse-adc-prostate-mri", "features": [ 0.47, 1.13 ] }, { "title": "transverse-hbv-prostate-mri", "features": [ 0.01, 1.84, 0.01, 0.99 ] } ] Image-level neural representation Whole Slide Tiling Thumbnail Description Low-resolution PNG visualization showing the result of whole-slide tiling. Each square represents a tile. Kind Thumbnail png Write to `/output/whole-slide-tiling-thumbnail.png` Whole Slide Tiling Thumbnail

Model Facts

Summary

This algorithm extracts slide-level representations from whole-slide images (WSIs) using PRISM [1]. We use the open-source Python package HS2P to unroll the WSI into non-overlapping tiles of size 224×224 pixels at a resolution of 0.50 microns per pixel (± 7%). Tiles with less than 25% tissue are discarded. Remaining tissue tiles are encoded using PRISM’s tile encoder -- Virchow [2]. These tile-level features are then aggregated using PRISM’s slide encoder to generate a single 1280-dimensional representation for the entire slide.

[1] PRISM: A Multi-Modal Generative Foundation Model for Slide-Level Histopathology, Shaikovski et al., 2024, arXiv:2405.10254
[2] A foundation model for clinical-grade computational pathology and rare cancers detection, Vorontsov et al., 2024, Nature Medicine.

Mechanism

Algorithm inputs:

a (multi-resolution) whole-slide image
its corresponding binary tissue segmentation mask (0=background, 1=tissue)

Algorithm outputs:

image-neural-representation.json: 1280-dimensional slide-level embedding
whole-slide-tiling-thumbnail.png: low resolution PNG image showing the tissue tiling layout

Validation and Performance

Left empty by the Algorithm Editors

Uses and Directions

This algorithm was developed for research purposes only.

Warnings

This algorithm requires access to both the PRISM slide encoder and the Virchow tile encoder, which are gated models. It can only be executed by users for whom the algorithm editor has explicitly verified that access has been requested and granted for both models. You can request access on the dedicated model pages:

Common Error Messages

Left empty by the Algorithm Editors

Information on this algorithm has been provided by the Algorithm Editors, following the Model Facts labels guidelines from Sendak, M.P., Gao, M., Brajer, N. et al. Presenting machine learning model information to clinical end users with model facts labels. npj Digit. Med. 3, 41 (2020). 10.1038/s41746-020-0253-3