Menu

Model release policy

A trained model moves from research to a clinic only through these steps, in order.

  1. Research. Trained and checked on public data only, with a written dataset licence check.
  2. Validation. Scored on validation subjects the model never trained on. Settings are chosen there and written down.
  3. Frozen test. Scored once on held-back subjects, after every choice is fixed. It is never repeated or tuned.
  4. Promotion gate. It must beat the current reader on the written gate items.
  5. Shadow mode. It runs beside the current reader on clinic scans, with the clinic's written permission, without changing any report.
  6. Release. Only then can it reach a clinic, as an update, with its model card.
  7. Rollback. Any release can be withdrawn. The clinic keeps the previous reader.

FollicleNet has passed research and validation. The frozen test has not been run. It is not in the product.