Segmentation benchmark

Senretsu

Comparing segmentation models on thin structures, where connectivity matters more than overlap.

annotated images: 11k+annotated images
model families compared: 3model families compared

The problem

Standard segmentation losses over-dilate thin structures and break their connectivity, which ruins downstream measurement.

What I built

1
Benchmarked U-Net, YOLOv8-seg and Mask R-CNN on more than 11,000 annotated images.
2
Integrated the clDice loss into U-Net to reduce over-dilation and preserve connectivity.

Outcome

A reproducible comparison showing how loss design changes the topology of thin-structure predictions.

Have a similar problem?

Tell me about it. I reply within 48 hours with a first read of the problem and an honest opinion on fit.