2026 · Computer vision
U-Net
PyTorch replication of Ronneberger et al. across four medical datasets, plus a skip-connection ablation.
0.971Dice score on MSD Spleen CT
- My role
- Reimplemented U-Net in PyTorch and ran the skip-connection ablation.
- Result
- Dice of 0.864 (ISBI), 0.864 (BUSI), 0.971 (Spleen) and 0.909 (Hippocampus).
- Built with
- PyTorch, U-Net, Medical imaging
Problem
Does the original U-Net hold up beyond the electron-microscopy data it was published on, and which parts of the architecture actually matter?
My role
Reimplemented U-Net in PyTorch and ran the skip-connection ablation.
What I built
- One config switch across EM, ultrasound, CT, and MRI
- Dice 0.971 on MSD Spleen, 0.909 on Hippocampus
- Skip-connection ablation: −15.9% tumor Dice without skips
Outcome
Dice of 0.864 (ISBI), 0.864 (BUSI), 0.971 (Spleen) and 0.909 (Hippocampus). Removing skip connections cut tumour Dice by 15.9%.