Score coarse-grid 3D segmentation on the source grid, and run val through the same path before test
If a 3D segmentation model trains on a resampled grid (e.g. 2 mm / 128^3 from 0.4 mm scans), training-time val Dice is measured on that coarse grid and will overstate quality. The honest number comes from mapping predictions back to the original grid (nearest-neighbour inverse resample + unpad) and scoring there, which usually lowers Dice and makes surface metrics like HD95 meaningful in mm.
Order that avoids burning the held-out set: 1. Freeze checkpoint hash, threshold and preprocessing together before any held-out scoring. 2. Run the full source-grid evaluation on the validation split first. A big drop vs training val Dice, or many empty predictions, usually means the inverse transform is broken, not the model. 3. Only then score the test split, once. Any threshold/code/retrain change after looking at test turns it into development data.
Also: an evaluation script that hard-codes raw-data paths from the manifest needs the original-resolution data present locally, not just the processed training cache.