A second opinion, built where it's hardest to get one.
Ether started from a simple gap: radiologists in resource-limited environments often read brain MRIs alone, with no second set of eyes and no specialist on call. We set out to build an assistant that earns a seat in that room — one that shows its reasoning instead of just a verdict.
What it's built with
FastAPI
Async Python backend serving the inference API, with automatic OpenAPI docs for the clinical team.
PyTorch + Radiomics
A hybrid pipeline: classical GLCM/LBP/wavelet features alongside a CNN, which generalizes better on small, heterogeneous MRI datasets than deep learning alone.
Next.js + Cornerstone3D
The clinical portal radiologists use day to day, with DICOM visualization built on the medical imaging standard Cornerstone3D.
Llama (open source)
An on-premise LLM instead of a locked-in cloud API — built for clinics without reliable cloud access, and for data sovereignty.
How it's structured
Frontend
Next.js, Feature-SlicedThe clinical portal radiologists use day to day: DICOM visualization, clinical summaries, and an LLM assistant, each isolated in its own feature module.
Backend
FastAPI, Service-ClientThin HTTP routes hand off to a services layer for business logic, with external integrations like storage and the LLM kept in their own client layer.
ML Engine
Standalone pipelineAn importable, testable package: DICOM preprocessing, classical radiomics features, and a CNN with Grad-CAM, orchestrated end to end.