Lessons from training a segmentation model
What actually moved the needle when segmenting thin, rare structures — and what didn't.
- computer vision
- segmentation
- deep learning
Placeholder note — edit freely.
A few things I’d tell my past self before starting a hard segmentation project:
- Loss function beats architecture for extreme class imbalance. A combined Dice + focal loss did more than swapping backbones.
- Augment aggressively, but realistically. Augmentations that don’t reflect real variation just add noise.
- Post-processing is underrated. Connected-components filtering removed a lot of false positives for almost no cost.
- Visualize predictions early and often. Metrics hide failure modes that a quick overlay makes obvious.
- 2D vs 3D is a budget decision. 3D context helps but is memory-bound; patch it or accept slower inference.
The meta-lesson: on low-signal problems, spend your time on data, loss, and evaluation before reaching for a bigger model.