Results

Comparisons across training methods, inference steps, and checkpoint selection.

Ordinary versus rollout training

Choose the checkpoint comparison

Each method gets its best validation checkpoint. This is the overview’s more conservative comparison. All three ordinary runs select the original parent, before additional training.

Ordinary + correctionStep-specific rollout + correction

Means across 3 seeds and 16 scans. Select a step count below for paired uncertainty intervals.

19.2%lower Foreground NRMSE95% CI 18.2–20.1%
0.1246 → 0.1007
lower is better

36.0%lower Full-image LPIPS95% CI 33.7–38.3%
0.0827 → 0.0529
lower is better

+0.0456increase in Foreground SSIM95% CI 0.0437–0.0477
0.8156 → 0.8613
higher is better

Full-image metrics and 10% foreground sensitivity

Metric	Ordinary	Rollout	Difference [95% CI]
Full NRMSE ↓	0.1490	0.1304	-0.0186 [-0.0204, -0.0166]
Full SSIM ↑	0.6808	0.7083	0.0274 [0.0128, 0.0440]
10% foreground NRMSE ↓	0.1232	0.0994	-0.0238 [-0.0251, -0.0226]
10% foreground SSIM ↑	0.8070	0.8549	0.0479 [0.0458, 0.0502]

Difference = rollout − ordinary. Negative is better for error metrics; positive is better for SSIM.

Training for different inference step counts

Two-step-specific training improves two-step inference. At eight-step inference, the four-step-trained model performs better than the eight-step-trained model.

Final training checkpoints, with measurement correction enabled. Lower is better.
Training method	Inference steps
Training method	2 steps	4 steps	8 steps
Ordinary SelfDB	0.1306	0.1422	0.1469
2-step rollout	0.1023	0.1012	0.1019
4-step rollout	0.1044	0.1007	0.0996
8-step rollout	0.1071	0.1030	0.1016

Bold marks the best result in each inference column. Shading marks matching training and inference step counts.

Step-specific training versus four-step training, at the same inference step count. Terminal checkpoints; exploratory contrasts.
Inference	Foreground NRMSE	LPIPS	Foreground SSIM	Full SSIM
2 steps	2.07% lower	6.52% lower	+0.00508	−0.01085
8 steps	1.96% higher	6.49% higher	−0.00567	−0.00268

The paired 95% intervals support improvement at two steps and worsening at eight on all three foreground/perceptual metrics. Full SSIM worsens at two steps; the eight-step full-SSIM interval crosses zero.

Evaluation details

Metrics and uncertainty

NRMSE measures normalized pixel error; lower is better. LPIPS measures feature-space distance; lower is better, but it is not a clinical quality score. SSIM measures local structural similarity; higher is better. Foreground NRMSE and interior-window SSIM use a 5% reference-intensity mask. LPIPS is full-image.

We average slices within each scan, then give scans and seeds equal weight. The paired two-way bootstrap resamples the three seeds and sixteen scans, with 10,000 draws. Displayed intervals are pointwise 95%; the analysis also applies a Bonferroni correction for three step counts. Three seeds still give limited uncertainty coverage.

The test set was held out from model training.

All 21 final configurations, including uncorrected ordinary inference

Checkpoint / inference	FG NRMSE ↓	LPIPS ↓	FG SSIM ↑	Full SSIM ↑
ordinary k2 correction1	0.13064	0.07641	0.79722	0.69506
ordinary k4 correction1	0.14219	0.08605	0.78193	0.68522
ordinary k8 correction1	0.14690	0.09031	0.77387	0.67912
ordinary k2 correction0	0.16518	0.10017	0.69513	0.61900
ordinary k4 correction0	0.17764	0.10139	0.66610	0.60082
ordinary k8 correction0	0.19292	0.10581	0.63491	0.58175
rollout2 k2 correction1	0.10227	0.05270	0.85820	0.69583
rollout2 k4 correction1	0.10117	0.05080	0.86074	0.69373
rollout2 k8 correction1	0.10194	0.05116	0.85949	0.69349
rollout4 k2 correction1	0.10443	0.05637	0.85311	0.70668
rollout4 k4 correction1	0.10069	0.05292	0.86126	0.70826
rollout4 k8 correction1	0.09960	0.05176	0.86398	0.70913
rollout8 k2 correction1	0.10708	0.06031	0.84641	0.70170
rollout8 k4 correction1	0.10302	0.05656	0.85493	0.70520
rollout8 k8 correction1	0.10155	0.05512	0.85832	0.70645
Best validation: ordinary k2 correction1	0.12004	0.07953	0.82011	0.68659
Best validation: ordinary k4 correction1	0.12456	0.08273	0.81564	0.68084
Best validation: ordinary k8 correction1	0.12681	0.08438	0.81159	0.67732
Best validation: rollout2 k2 correction1	0.10198	0.05303	0.85895	0.69711
Best validation: rollout4 k4 correction1	0.10069	0.05292	0.86126	0.70826
Best validation: rollout8 k8 correction1	0.10155	0.05512	0.85832	0.70645

k = inference steps; correction1 = on; correction0 = off. Unprefixed rows use terminal checkpoints.

Earlier saved-model sweep: all configurations

Configuration	FG NRMSE ↓	LPIPS ↓	FG SSIM ↑	Full SSIM ↑
bridge k2 correction1	0.11929	0.06869	0.82990	0.74421
bridge k4 correction1	0.12982	0.07406	0.81937	0.73710
bridge k8 correction1	0.13467	0.07657	0.81191	0.73170
bridge k2 correction0	0.16143	0.10699	0.70663	0.65113
bridge k4 correction0	0.17562	0.10839	0.67350	0.63171
bridge k8 correction0	0.19339	0.11158	0.63780	0.61260
rollout k2 correction1	0.10338	0.05271	0.85708	0.70212
rollout k4 correction1	0.09910	0.04978	0.86586	0.70663
rollout k8 correction1	0.09799	0.04913	0.86857	0.70865
rollout k2 correction0	0.19127	0.18679	0.65256	0.54376
rollout k4 correction0	0.22074	0.23204	0.59560	0.50505
rollout k8 correction0	0.24078	0.25903	0.55801	0.48067

Earlier best-validation models, not the twelve newly trained step-study models. This correction-off diagnostic does not isolate the effect of training without correction.