← Back to the course

Fold 0 · 300-epoch baseline

Validation results

Masked MAE after restoring synthetic CT predictions to Hounsfield units. Lower is better.

Cases35held out from weight updates
Mean MAE105.0 HUacross the full fold
Median MAE102.0 HUmiddle validation result
Range67.4–189.1 HUbest to worst case

What this page shows

The model was trained on 140 SynthRAD2025 Task 1 abdomen cases and evaluated on the other 35 cases in fold 0. Case 1ABA101 is used in the course figure because its 102.0 HU error is closest to the cohort median, not because it is the best result.

These are internal one-fold validation results from our masked-MSE PlainConvUNet baseline. They are not challenge-test, five-fold ensemble, external-site, or clinical-performance results.

All held-out cases

Per-case masked MAE

Body-mask voxels only

CaseMasked MAE
1ABA01195.8 HU
1ABA031101.1 HU
1ABA047155.6 HU
1ABA04967.7 HU
1ABA061102.6 HU
1ABA073117.5 HU
1ABA080108.7 HU
1ABA08682.4 HU
1ABA09274.4 HU
1ABA101Representative case102.0 HU
1ABA10576.3 HU
1ABA10867.4 HU
1ABA110101.7 HU
1ABA115111.0 HU
1ABB00487.0 HU
1ABB006108.8 HU
1ABB041133.0 HU
1ABB054107.0 HU
1ABB056107.2 HU
1ABB061114.4 HU
1ABB069105.2 HU
1ABB070153.4 HU
1ABB086130.5 HU
1ABB10291.7 HU
1ABB11684.8 HU
1ABB11789.7 HU
1ABB11891.2 HU
1ABB130141.5 HU
1ABB14786.4 HU
1ABB15584.8 HU
1ABB159121.9 HU
1ABB16185.4 HU
1ABB163189.1 HU
1ABB16486.8 HU
1ABC004110.3 HU

Derived from the public SynthRAD2025 Task 1 training data. The displayed figures and derived measurements remain subject to the dataset's CC BY-NC 4.0 terms. No raw medical volumes are redistributed.