less than 1 minute read

Yesterday I presented our paper Seeded iterative clustering for histology region identification at the NeurIPS 2022 Learning Meaningul Representations of Life (LMRL) workshop as a contributed talk. We propose a latent feature clustering approach to suggest a patch-based dense segmentation at WSI level requiring as little as one click per class from the pathologist. Some key contributions:

  1. By decoupling feature extraction and interaction, we have no need of a GPU at run-time while still getting results in seconds.
  2. By not retraining at any point, we can compare different feature extraction spaces in a less biased way.
  3. By performing the clustering per-slide, the method is more robust to domain shift artifacts.

This was just our proof of concept, but we are already improving it and turning it into a full interactive annotation tool!

See the complete talk here: