Improve representativeness sampling - Githubissues

HealthML / active-segmentation

ActiveSegmentation: A Simulation Framework for Benchmarking Active Learning Strategies for 3D Medical Image Segmentation

GNU Affero General Public License v3.0

16 stars 1 forks source link

Improve representativeness sampling #126

Closed josafatburmeister closed 2 years ago

josafatburmeister commented 2 years ago

This PR adds several extensions to the current implementation of representativeness sampling:

an option to ignore outliers when using the mean shift algorithm for clustering
an option to use the k-means clustering with fixed number of clusters for representativeness sampling
an option to use the plain images as feature vectors for representativeness sampling
a"scan-based" representativeness strategy that evenly samples slices from all scans

In addition, this PR fixes a bug in the computation of the cluster sizes in the cluster-based representativeness strategies.