ETHER

Built to see what others miss.

What the eye misses, Ether finds.

Open the portal

1 radiologist

per 100,000+ people in many underserved regions.

Ether gives a trustworthy second opinion in minutes.

How It Works

01

Texture Analysis

Extracts texture and structural features from the scan.

02

Visual Patterns

Analyzes visual patterns with a deep learning model.

03

Single Finding

Combines both signals into a single finding.

04

Show the Why

Shows exactly which part of the brain drove the conclusion.

We know when not to trust ourselves.

And we tell you.

  • ✓Scan quality check before any prediction is attempted.
  • ✓Confidence threshold enforced — low-confidence cases are flagged, not force-classified.
  • ✓Cases below threshold are marked for mandatory human review.
  • ✓The system never outputs a definitive diagnosis — only decision-support language.

See it in action

DICOM viewer with Grad-CAM overlay — demo recording goes here

This only gets better with more hands on it.

Radiology support shouldn't be a black box owned by one team. Ether is open source so clinicians, researchers, and engineers anywhere can inspect it, challenge it, and push it forward.

View the repository
ETHER

For the technical judges

Explainability by design

Grad-CAM isn't a feature checkbox — it's the trust mechanism at the core of every result.

Data sovereignty

Runs on an on-premise, open-source LLM instead of a locked-in cloud API — built for clinics without reliable cloud access.

Hybrid ML architecture

Classical radiomics plus deep learning generalize better on small, heterogeneous datasets.

Try the Ether demo.

Open the portalGitHub repositoryHow we built Ether

German Joel Ramos

Juan Pablo Botero

Pete Paul Narvaez

Esteban Barreras

Ether