Powered by MedGemma · Built on llama.cpp · MIT Open Source

The Radiologist's
AI Assistant.

Runs Offline. Always.

Generate structured DICOM reports in under 30 seconds — entirely on your device. No cloud uploads, no PHI exposure, no procurement queue. Just open the scan and let the model work.

Free·No account·No internet required after install
Have an NVIDIA GPU? Get the CUDA build →

HIPAA-Safe· On-Device AI· MedGemma· llama.cpp· Tauri 2· MIT License
The Problem

Radiology AI is broken in three specific ways.

Cloud-first products force a trade-off between capability and compliance. We don't.

01

Cloud AI Exposes PHI

Every cloud-based radiology assistant ships protected health information off-site for inference. That's a compliance liability your CIO can't sign off on without months of BAA review.

→ 0 bytes leave the workstation
02

Enterprise Sales Take 18 Months

By the time procurement, security review, and integration are done, the model you bought is two generations behind. Radiologists wait. Patients wait.

→ Installed in < 5 minutes
03

Reporting Backlogs Are Growing

Volume is up. Turnaround SLAs are tighter. The bottleneck is the report, not the read. Drafting a structured impression for every study is wearing the workforce thin.

→ 30-second first drafts
How It Works

Five steps. No PACS integration required.

Drop a DICOM in, get a signed-off draft out. The whole loop fits inside your existing workstation.

1

Install

Download the .dmg or .exe. Native desktop app — no Docker, no Python, no GPU required.

2

Open Scan

Drag a DICOM, NIfTI, or PNG into the viewer. Multi-frame studies load with native windowing.

3

Draw ROI

Lasso or rectangle around the region of interest. Skip for full-image analysis.

4

Generate Report

MedGemma streams a structured Findings/Impression draft in under 30 seconds.

5

Review & Sign

Edit inline, export to .docx or HL7, or copy into your reporting system.

Features

Built like a clinical tool, not a chatbot.

Every design decision optimizes for one thing: a radiologist's first 30 seconds with a study.

Zero PHI Transmission

Every inference happens on the radiologist's workstation. No data ever leaves the device — no cloud, no telemetry, no analytics.

→ HIPAA-Safe by architecture

<30s Reports

End-to-end from open-scan to streaming draft, including ROI segmentation. Benchmarked on Apple Silicon and modern Windows laptops.

→ Median 24.3s on M-series

Native Desktop App

Built on Tauri 2. Tiny binary, native menus, no browser, no Electron bloat. Installs and updates like any other clinical tool.

→ < 20 MB installer

ROI Segmentation

Constrain inference to a region or feed the full study. The model sees what you draw and nothing else.

→ Pixel-level prompt control

Structured Output

Findings, Impression, recommended follow-up — pre-formatted to your institution's template. Export as .docx, JSON, or HL7.

→ Templates configurable

Model-Agnostic

Ships with MedGemma but accepts any GGUF model. Swap in a fine-tune, a domain-specific model, or your own checkpoint.

→ llama.cpp under the hood
Compare

The only one that ships without a cloud dependency.

Honest comparison against the leading radiology AI products as of 2026.

RadiologyAI Rad AI New Lantern AIdoc
Runs offline / on-device Yes Cloud only Cloud only Cloud only
PHI never leaves device Yes BAA required BAA required BAA required
Open-source (MIT) Yes Proprietary Proprietary Proprietary
Install time < 5 min 3–6 mo procurement 3–6 mo procurement 6–18 mo procurement
Bring your own model Any GGUF No No No
Per-study cost $0 Subscription Per-read Subscription

This is the tool I've been waiting for. Finally AI that works in our environment without a six-month procurement process.

— Radiology Resident, Academic Medical Center

Download Free. No Cloud. No Compromise.

v0.1.0 · macOS (Apple Silicon) · Windows 10+ · MIT licensed