Cancer3.AIAI in OncologyAI Models › CHIEF
Foundation model Open weights Research use only pathology

CHIEF

Clinical Histopathology Imaging Evaluation Foundation model: trained on 60,530 whole-slide images across 19 anatomical sites and validated on 19,491 slides from 24 hospitals for cancer detection, tumour-origin prediction, genomic profiling and survival.

At a glance

DeveloperYu Lab, Harvard Medical School
Released2024-09-04
Licencenon-commercial (see repository)
AvailabilityOpen weights
KindFoundation model
Regulatory statusResearch use only

What it does

Slide-level representations built with weak supervision on top of tile features, plus a text embedding of the anatomical site; one model then serves detection, origin, molecular-profile and prognosis heads.

Notable uses in oncology

Cancer detection across 11 cancer types with an area under the curve near 0.94 on external validation in the paper; survival prediction across cancer types.

Tasks, data types and cancers

CancerPan-cancer
InputWSI tiles + anatomical site
Outputslide-level embedding / task logits

Architecture

Familyweakly supervised slide-level model over tile features
BackboneCTransPath tile encoder + attention aggregation + site text embedding
Pre-trainingself-supervised tile pretraining + weakly supervised slide-level training on 60,530 WSIs

Training data

60,530 WSIs from 19 anatomical sites (public cohorts including TCGA, PAIP and others plus institutional data); validation on 19,491 slides from 24 hospitals worldwide.

Training set size60,530 WSIs (training)
InstitutionsHarvard Medical School and partner hospitals; public cohorts

Linked datasets

Evaluation

Benchmark / datasetMetricValueExternal validationSource
cancer detection, 11 cancer types, external validation AUROC ~0.94 (paper) yes Source
survival prediction across cancer types c-index / log-rank reported in paper yes Source

How to run

LibraryPyTorch (repo code)

Weights are distributed from the GitHub repository under non-commercial terms; the pipeline expects CTransPath tile features.

Regulatory status and intended use

Regulatory statusResearch use only
Intended useResearch.

Source →

Regulatory status is quoted from the source linked above and can change. Research-use-only models must not be used for clinical decisions.

Limitations and bias

  • Tile encoder (CTransPath) is weaker than newer ViT-L/H foundation encoders; CHIEF's strength is the slide-level training.
  • TCGA is in the training set — external evaluation must avoid it.

Sources

  1. Wang X et al. A pathology foundation model for cancer diagnosis and prognosis prediction. Nature 2024
  2. GitHub — hms-dbmi/CHIEF

This page is educational — it is not medical advice and does not replace consultation with an oncologist. Diagnostic and treatment decisions are made solely by specialist physicians.