Neural Active Contours

An exploration of medical image segmentation that predicts an evolving boundary rather than a dense per-pixel mask.

Encoder-decoder segmentation networks label every voxel independently, so anatomical boundaries can come out fragmented or topologically implausible. Classical active contour methods take the opposite view: represent the structure as a curve and evolve it under image and smoothness forces. This project asks what a learned version of that formulation looks like, and where it helps on medical data.

Current direction:

• Represent the target structure as a deformable contour instead of a dense mask
• Learn the evolution forces that drive the contour toward anatomical boundaries
• Keep boundary smoothness and topology as properties of the representation rather than post-processing steps
• Compare against standard U-Net baselines on the same data

This is an early-stage exploration; the repository tracks ongoing experiments rather than a finished benchmark.