Browse Teach Curate

Connectivity-Preserving Segmentation via a Supervoxel Simple-Point Loss

A topology-aware, differentiable loss for instance segmentation of thin filamentous structures that generalizes the digital-topology simple point to connected voxel sets (supervoxels), penalizing split and merge errors so a neural network preserves object connectivity, with training cost that scales linearly rather than nonlinearly like persistent-homology losses.

KITT Text

Efficient Connectivity-Preserving Instance Segmentation with Supervoxel-Based Loss Function Anna

Presents an efficient topology-aware loss function for instance segmentation of thin, filamentous structures (neurons, blood vessels, roads) where preserving connectivity is critical. It extends the …