Effect of topology-aware losses on tracer-histology fiber-bundle segmentation

Determine how topology-aware loss functions affect fiber-bundle segmentation performance in tracer histology, including structural and pixel-overlap quality.

Background

The paper compares BCE–Dice, clDice, Betti matching, and Topograph for segmenting fiber bundles in macaque tracer histology. BCE–Dice emphasizes pixel overlap, whereas clDice, Betti matching, and Topograph introduce structural or topological objectives intended to better preserve elongated, connected fiber-bundle regions. The paper reports different trade-offs among these losses, but the general effect of topology-aware losses on this segmentation task is explicitly identified as unresolved. The question is relevant because automated segmentation must capture both the spatial extent and structural organization of annotated fiber bundles.