paper | poster
This repository implements a connectivity-preserving loss function designed to improve instance segmentation of curvilinear structures. The paradigm shift here is to evaluate segmentation quality at the “structure-level” as opposed to the voxel-level. The loss is computed by detecting supervoxels in the false positive and false negative masks during training, then assigning higher penalties to supervoxels that introduce connectivity errors.
Figure: Visualization of loss computation, see Method section for description of each step.
To do...
To use the software, in the root directory, run
pip install -e .supervoxel-loss is licensed under the MIT License.