Whole Slide Concepts

Building a whole-slide foundation model normally takes two self-supervised stages: pretrain a patch encoder on millions of unlabeled tiles, freeze it, then train a second model to aggregate those frozen patch features into a slide embedding. Neither stage ever looks at a clinical label — even though every slide out of a diagnostic lab already carries one. Subtype, mutation status, and eventual outcome are all recorded as routine, at slide level rather than tile level.

Whole Slide Concepts (WSC) uses those labels as the training signal. The patch encoder and the slide-level aggregator are optimized jointly, end to end, against slide-level targets from 18 tasks at once — cancer subtyping, overall survival, and gene mutation prediction — so a single shared attention-pooled slide representation has to serve all of them. The team at Fraunhofer MEVIS, with collaborators at Fraunhofer ITEM and Hannover Medical School, trained it for 500 hours on one A100 — about 200 kWh, plus 37 kWh inherited from Tissue Concepts, the group's earlier patch encoder used to initialize the backbone. That total is roughly 16% of the energy Prov-GigaPath reports for its aggregator alone, and 5% of the CTransPath encoder underlying CHIEF.

Posted as a preprint in July 2025 as Tissue Concepts v2 (TCv2) and renamed Whole Slide Concepts in the January 2026 revision, it ships as the WSC-MTL-tiny checkpoint in the MedicalMultitaskModeling (M3) library, under a Fraunhofer license limited to non-commercial scientific research.

#Key Features

• End-to-end supervision on slide labels: Within-bag sampling lets gradients flow from a slide-level target back through attention pooling into the patch encoder, so both halves are learned together, not in sequence.
• One representation, three clinical question types: Subtyping, risk estimation, and mutation prediction share the same pooled slide vector, each solved by a small linear head.
• Attention as a tumor detector: Attention pooling assigns each patch a weight, and on the held-out TUPAC16 cohort those weights concentrate inside expert-annotated tumor regions, with no fine-tuning.
• Frozen-encoder evaluation: Downstream tasks are solved by a linear decision layer over the frozen encoder and aggregator, so results come from a fixed checkpoint, not a per-task retrain.
• Reproducible pretraining corpus: Training used only public cancer cohorts, with the NCI Imaging Data Commons queries for the TCGA and CPTAC portions published in the appendix.

#Technical Details

The backbone is a Swin Transformer V2 tiny, about 28M parameters, initialized from Tissue Concepts weights. Each training bag holds 128–256 augmented patches from one slide; they pass through the shared encoder, are compressed by a squeezer module, and are pooled by eight attention heads into a 768-dimensional slide vector. Losses across the 18 tasks are accumulated before each optimizer step, and the pipeline ran for 200 epochs. Tasks were built from TCGA and PLCO — 22,225 slides for pan-organ subtyping, 1,023 breast slides supplying subtype, survival, and TP53, SPAT1, and CDH1 targets — with CPTAC held out. Three variants were trained: a tiny model on slide labels only, a tiny model on slide plus patch labels, and a small-backbone version.

Across four runs per task, the slide-plus-patch tiny model reaches AUC 0.98 on TCGA→CPTAC NSCLC subtyping against 0.97 for CHIEF and Prov-GigaPath, and AUC 0.89 on PANDA ISUP grading, on par with UNI and Prov-GigaPath and ahead of CHIEF at 0.83. It also leads on prostate survival, a c-index of 0.66 on the LEOPARD cohort against 0.58 for CHIEF, though CHIEF is stronger on CPTAC-LUAD at 0.70 versus 0.64. The slide-labels-only variant takes both out-of-domain tasks: AUC 0.85 on BRACS breast subtyping and 0.84 on classifying collagen burden in precision-cut lung slices. Scaling to the small backbone did not improve subtyping.

#Applications

WSC suits groups holding slide archives with routine outcome labels but lacking the compute for self-supervised pretraining: a slide vector extracts on modest hardware and feeds a linear head for subtyping, mutation screening, or risk stratification. The attention map accompanying each prediction shows pathologists which regions drove it, and the lung fibrosis experiment shows the features transfer to non-oncological readouts such as quantifying collagen in preclinical tissue slices.

#Impact

WSC shows the field's default recipe is not the only workable one: weak clinical labels already sitting in hospital systems are a competitive substitute for large-scale self-supervision. Training only on public cohorts also makes the pretraining distribution inspectable for bias auditing, in a field where flagship models are usually built on proprietary archives. The limits are real: the model is uni-modal, with no alignment to pathology reports of the kind TITAN uses; survival results carry high variance the authors attribute to sampling only 300 patches per slide; the PLCO portion of the training data needs an approved access request; the work remains a preprint; and the license permits scientific non-commercial use only, with diagnostic and therapeutic use excluded.