Radix reads your panoramic radiographs, proposes a mask at every tooth and finding, and puts the images worth labeling next in front of your annotators first — without diluting the rare pathology across healthy dentition.
The order is the pipeline: score first, prompt where it matters, label what's uncertain, reuse everything you've already computed.
Every image is scored at native resolution, so lesion-scale structures don't fall through the feature floor. Peaks become prompts.
SAM is prompted at score peaks and at each numbered tooth. Annotators correct proposals instead of drawing from scratch.
An active-learning loop ranks the pool by what the model is least sure of, split across exploitation, coverage, and exploration.
Fix the mask, assign the class, move on. Tooth annotations carry their FDI position; findings carry their type.
Embeddings and score maps are cached by image and encoder version, so a new round only touches what changed.
Masks and boxes leave as COCO or per-structure crops, deduplicated by pixel content and tagged with their source.