Foundation-model pathology

What is essential is invisible to the eye.

A pathologist estimates tumor purity by eye. EnsoPurity reads it from the slide — ~2× more accurate than the human eye, against absolute genomic ground truth.

real slides · model heatmaps

The problem

Purity is guessed by eye.

Estimating tumor content from a slide is genuinely hard — and the numbers show it. No fault of the pathologist; the truth is simply invisible to the eye.

0.37
how well the human eye
tracks sequencing (ρ)
+16pts
average overestimate
of tumor content
H&E pathology slide
A real H&E — what fraction is tumor? Hard to call by eye.
The method

EnsoPurity reads it.

Validated against absolute genomic ground truth across 32 cancer types — about double the correlation, half the error.

Each point = one slide · model vs pathologist · x-axis = genomic truth
See it live

Draw a region. Read the numbers.

Outline tissue on a real case — purity, nuclei, and DNA yield per region. The unseen, made tangible.

Loading ROI data...
Validated vs genomic truth

Better in every indication.

Pathologist (amber) → EnsoPurity (teal). The gap is the product — grouped the way a pathologist reads, by organ system.

Loading validation…
Method & honesty

What the numbers mean

Held-out TCGA + CPTAC slides with paired genomic purity — every indication shown, including the weaker ones. Ground truth is matched sequencing, not another reader's opinion. EnsoCellularity is built and tested too; held back until its validation set is final.