Method
Train through the model’s own predictions, while keeping each step tied to the scanner readings.
Self-supervised training
A scanner collects frequency-space readings, called k-space. Fewer readings can mean a faster scan, but also an incomplete picture.
Imagine we acquired 100 readings. Hide some, reconstruct from the rest, then check the predicted image against the original 100. The hidden readings teach the network. A clean target image is never needed for the training loss.
The numbers are illustrative. SelfDB uses two nested levels of extra masking to build its diffusion bridge.
y is the acquired data. A models coil sensitivities and Fourier encoding. M selects available measurements. The implementation also guards normalization numerically.
Rollout training
Ordinary bridge training
Construct a synthetic intermediate state by mixing degraded reconstructions and adding noise. Train one prediction at a random bridge time.
Rollout training
Run the actual sequence. Later inputs depend on earlier predictions. Backpropagate through every step, so a final error can change the first prediction.
At four steps, the final prediction gets 81.25% of the weight; each earlier prediction gets 6.25%. All steps share the same network weights. This is gradient-based unrolled training, not reinforcement learning.
Measurement consistency correction
After the network proposes an image, simulate what the scanner would measure. Adjust the image to reduce disagreement with the input readings, while staying near the network’s proposal.
z is the proposed image. O contains only the readings allowed as input. Hidden training readings enter the loss, never this correction.
Both the rollout model and the ordinary baseline get the same correction at evaluation. We report the combined system; this study does not separate a newly trained rollout-only model from correction-only training.
Measurement and foreground masks
Measurement masks
Choose which k-space readings the model can see. The historical implementation had non-nested masks. We fixed nesting in every new branch, so the comparison does not isolate that fix.
Foreground masks
Choose where to score an image. Pixels above 5% of the reference maximum define foreground. SSIM uses windows fully inside that mask. This keeps black background from dominating.
Foreground scoring was introduced after seeing earlier results. The final study fixed that rule before its runs, and also reports full-image metrics and a 10% threshold sensitivity. Reference images are used for validation and test scoring, not as training targets.
Experimental setup
| Ingredient | Final step study |
|---|---|
| Network | Repository-default U-Net; 149,580,546 parameters, with a complex-input padding wrapper. |
| Data | Calgary: 5,032 training slices / 37 acquisitions; validation 96 slices / 8 scans; test 192 slices / 16 scans. |
| Splits | Acquisition- and exam-prefix-disjoint. The training acquisitions have 35 exam prefixes; unique participant identities were not independently verified. |
| Sampling | 25% of the full grid within 85% partial-Fourier support; nested training levels 1/6 and 1/8; 24 × 24 calibration region retained. |
| Paired initialization | Three seeds. Every branch starts from its seed’s original 3,000-call pretrained checkpoint. |
| Training budget | 24,000 additional logical network calls per branch; batch size 4. Equal calls do not mean equal FLOPs, updates, data exposures, or time. |
| Schedule | Learning rate 10⁻⁴ → 3 × 10⁻⁵ → 10⁻⁵ at 12k / 18k calls. EMA decay 0.995 per call. |
| Checkpoint selection | Terminal checkpoint is the registered primary endpoint. Secondary: minimum validation foreground NRMSE, checked every 4,000 calls; ordinary selects at four steps, rollout at its native step count. |
Training samples the expanded pool with replacement. These runs neither guarantee every slice was visited nor establish global convergence. Calgary references are coil-combined partial-Fourier acquisitions, not noise-free fully sampled ground truth.
See what improved, and what didn’t →