Skip to content

Model-assisted annotation

Drawing every annotation by hand is slow. Once you've trained even a rough first model, you can let it do most of the work: test its output on a slice, copy that whole slice into the Annotation tab, and fix only the mistakes by hand. This turns annotating from drawing into editing — much faster. It also means you annotate and test at the same time, and can add training data exactly where the model is still inaccurate.

Copying a model's output into the training annotations, then correcting the mistakes.

How to use it

  1. Train a quick first model — see Training a model. It doesn't have to be good; it only has to be a starting point.
  2. Apply it to a tomogram and play with the threshold — and, optionally, the test-time augmentation multiplicity — until the output looks about right. See Assessing model performance.
  3. Right-click the model's panel and open copy output to annotation, then pick the destination annotation from the list of features.
  4. Ais copies the output into that annotation and switches you to the Annotation tab. Now correct it: erase the false positives, and draw in whatever the model missed.
  5. Add and correct a few boxes, then train again. A useful rule of thumb: annotate about 20 boxes and train a first version; then use that model to assist you up to roughly 50 boxes, export, and train again — and see how much better it gets.

What gets copied

Copy output to annotation transfers only the slice you're looking at, and only the pixels above the current threshold. Take care when copying onto a slice you have already annotated: it overwrites your hand-drawn labels there, though any boxes you placed stay in place.