Hessian-Guided Feature Injection in U-Net for Nucleus Instance Segmentation
- Nucleus instance segmentation in H&E histopathology images using Hessian-derived Morse critical points as structured geometric input features.
About
A U-Net trained end to end on a small dataset like CryoNuSeg (30 H&E images, with the test set held out by organ) learns geometric structure implicitly. That implicit learning runs out of capacity at the hardest case: two nuclei touching with no clear seam between them. This project adds explicit geometric features to the U-Net's input alongside the original RGB image. The features come from a Hessian analysis of the intensity surface: local maxima mark nucleus centers, and saddle points appear at the bridge between two touching nuclei. The final model concatenates the RGB image with six Hessian-derived channels and trains a single binary head with Dice loss. Three controlled ablation scenarios isolate the contribution of the geometric prior, and a finer-grained validation analysis confirms that the saddle channel does what it was designed to do.