REVIEW 4 major objections 4 minor 1 cited by
A Self-supervised Diffusion Bridge for MRI Reconstruction
T0 review · 4 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash
Pith's one-line read SelfDB trains a diffusion bridge for MRI reconstruction without clean references, beating ambient diffusion in few steps.
desk verdict A clever self-supervised extension of diffusion bridges to measurement-only training, but the reverse update is asserted rather than derived and the evaluation is too thin to fully back the performance claim. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the two-operator subsampled-measurement diffusion process $y_t = (1-t)\bar M y + t M' y + \sigma_t \epsilon$, where a diffusion bridge means a diffusion process built to interpolate between two paired distributions. This process replaces a failed direct adaptation in which applying the ambient-diffusion operator to a bridge state gives $y_t \approx M'y$ for every $t$, leaving no actual transition to learn; by inserting an intermediate subsampling level, SelfDB creates endpoints that differ and a reverse process that adds information instead of repeatedly solving the same inverse problem. The reverse update in Algorithm 1, patterned on the diffusion-bridge posterior step, uses the network output $\hat x = f_\theta(y_t,t)$ through $M A \hat x$ as a stand-in for the missing measurement, and the training loss in (12) ties that prediction back to the observed $y$.
What would settle it
Compare one step of Algorithm 1 against the exact conditional expectation $\mathbb{E}[y_{t-\delta}\mid y_t]$ computed by Monte Carlo under the process in Eq. (11) with the same fitted network; if the update deviates systematically from that expectation as $\delta \to 0$, the claimed posterior step is not the one being simulated.
Extended reading notes
Core claim
On its own terms, the paper's discovery is that a diffusion bridge can be trained self-supervised by making the bridge process a transition between two levels of undersampling of the measured data. Given a measurement $y = MAx + e$, SelfDB introduces two extra subsampling operators, an intermediate $\bar M$ and a severe $M'$, and defines $y_t = (1-t)\bar M y + t M' y + \sigma_t \epsilon$. Because the two endpoints are different sub-sampled copies of the same measurement rather than an image and Gaussian noise, reversing the process gradually fills in missing k-space information. The network $f_\theta(y_t,t)$ is trained with the loss $\mathbb{E}\|M A f_\theta(y_t,t)-y\|_2^2$, and inference iterates the posterior-style update $y_{t-\delta} = \frac{\delta}{t} M A \hat x + (1-\frac{\delta}{t}) y_t + (t-\delta)\sqrt{\sigma_{t-\delta}^2-\sigma_t^2}\,\epsilon$ with $\hat x = f_\theta(y_t,t)$. The paper reports that this measurement-only training outperforms a conditional ambient-diffusion model at four inference steps on brain MRI.
Load-bearing premise
The iterative refinement works only if the network's measurement-domain prediction is an accurate enough stand-in for the unmeasured clean data inside the reverse update; the paper does not prove that this substitution is the true conditional expectation.
Editorial extensions
If this is right
- Diffusion bridges no longer require paired clean references, so the method can be trained in settings where only routine undersampled acquisitions are stored.
- The few-step speed advantage claimed for supervised diffusion bridges carries over to self-supervised training: four inference steps outperform the conditional ambient-diffusion baseline on NRMSE, SSIM, and LPIPS.
- The number of inference steps acts as a tunable perception-distortion knob: two steps give NRMSE 0.101 and LPIPS 0.039, while eight steps give NRMSE 0.118 and LPIPS 0.035.
- Because the construction is stated for orthogonal forward operators and nested subsampling, the same argument transfers to other linear inverse problems with that structure.
Reading between the lines
- A natural test the authors did not run is a comparison against a supervised diffusion bridge with the same architecture; the gap would quantify how much accuracy is traded for removing the need for references.
- If the reverse update in Algorithm 1 were replaced by an exactly derived posterior step, SelfDB could in principle extend to non-orthogonal forward operators, such as general CT line-integral models, where the current nested-subsampling identity does not hold.
- The observed perception-distortion trend suggests SelfDB could be used as a tunable sampler, with the step count chosen according to whether the downstream task rewards fidelity or human-reader quality.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. This paper proposes SelfDB, a self-supervised training scheme for diffusion bridges in MRI reconstruction. The method defines a new diffusion process in k-space by applying two additional subsampling operators to the available measurements, so that the forward process interpolates between an intermediately subsampled measurement and a more aggressively subsampled measurement. A network is trained without ground-truth images using the available measurement as its own training target; at inference, iterative refinement is performed in measurement space. Experiments on fastMRI brain data report that SelfDB with four inference steps outperforms a conditional ambient-diffusion baseline and an ambient-diffusion diffusion-bridge baseline on NRMSE, SSIM, and LPIPS. The paper also documents a perception-distortion trade-off controlled by the number of inference steps.
Significance. If the theoretical gaps were resolved, the contribution would be practically significant: it would extend the self-supervised paradigm from denoising diffusion models to diffusion bridges, enabling reconstruction without high-quality references and with few inference steps. The paper explicitly targets CS-MRI, where ground-truth images are often unavailable. The strengths are the clear problem statement, the use of an external benchmark with held-out fastMRI data, and the inclusion of both distortion and perceptual metrics with visual comparisons. However, the significance is currently tempered by an unproven reverse-update step and by empirical evidence from a single split without error bars. The contribution is therefore promising but not yet fully established.
major comments (4)
- [Section 3, Eqs. (11)-(12) and Algorithm 1] There is a load-bearing mismatch between the forward process and the training target. Eq. (11) defines a bridge whose t=0 endpoint is the intermediate measurement \bar{M} y, while Eq. (12) trains M A f_theta(y_t,t) to match the full measurement y. In Algorithm 1, line 11 uses \hat{y} = M A f_theta(y_t,t) as the estimate of the bridge endpoint. Unless the same symbol M in Eq. (12) is intended to denote the intermediate operator rather than the original sampling operator from Eq. (1), the network is trained to predict a richer target than the endpoint of Eq. (11). The paper never specifies this distinction or derives the connection between Eq. (12) and the conditional expectation E[y_{t-\delta}|y_t]. With the current training loss, the iterative update in Algorithm 1 is not reversing Eq. (11). Please either change the forward process to bridge from the full measurement y to M'y, or change the training loss to compare \bar{M} A f_theta(y_t,t) with \bar{M} y, and state the chosen convention explicitly.
- [Section 2.2 and Section 3, Eq. (10)] The simplification M'A x_t = M' y + sigma_t M' A epsilon requires A A^H = I. For the multi-coil forward operator A = F S with ESPIRiT sensitivity maps used in Section 4, A A^H = F S S^H F^H is not the identity, so the last equality in Eq. (10) does not hold. This invalidates the derivation of Ambient-DB and weakens the stated explanation for why Ambient-DB is suboptimal. Please either restrict the claim to settings with orthogonal A or correct the derivation for the multi-coil CS-MRI setting actually used in the experiments.
- [Section 4, Table 1] All metrics are reported on a single train/validation/test split of fastMRI with no error bars, multiple training runs, or significance tests. The NRMSE gap between SelfDB and C-Ambient-DDM is 0.005 (0.108 versus 0.113), which is small relative to typical run-to-run variability in deep-learning reconstruction. To support the claim of superior performance, please report mean plus/minus standard deviation over at least three independent training runs, or provide a paired statistical test on the test set.
- [Section 4 and Algorithm 1] The method cannot be reproduced from the manuscript because the noise schedule sigma_t, the inference step size delta, the network architecture, and the training hyperparameters are not specified. Algorithm 1 depends directly on sigma_t and delta, and the reported perception-distortion trade-off in Table 2 depends on the number of inference steps, so these details are essential. Please provide the missing specifications, and ideally release code.
minor comments (4)
- [Section 4, first paragraph] The sampling rates are listed as '25%, 16.6%, and 12.5% as M, M, and M′', which appears to be a typo: the second operator should be a distinct symbol such as \bar{M} to avoid confusion with the 25% operator.
- [Figure 1 caption] The phrase 'read arrow' should be 'red arrow'.
- [Section 2.2, first paragraph] The sentence 'Once a DNN is trained to reverse the this process' contains a typo and should read 'reverse this process'.
- [Section 4, data description] Reference [16] describes the fastMRI knee dataset, but the experiments use T2-weighted brain data; please cite the appropriate fastMRI brain data source or clarify the relationship.
Circularity Check
No significant circularity: SelfDB's self-supervised training is validated against external ground-truth benchmarks, and no prediction reduces by construction to a fitted input.
full rationale
SelfDB's central claim is that a diffusion bridge can be trained using only raw measurements as targets and still outperform a self-supervised denoising diffusion baseline. The training loss in Eq. (12) does use the available measurement y as the regression target, but the reported performance is evaluated against held-out fully-sampled ground truth (NRMSE/SSIM/LPIPS in Table 1), so the headline result is not forced by the training objective. No parameter is fitted to the test set, and no load-bearing uniqueness or equivalence theorem is imported from the authors' prior work. The citation to the authors' own [12] is contextual support for the general effectiveness of measurement-space self-supervised losses, not the basis of the SelfDB construction. The reverse update in Algorithm 1 line 11 is asserted to be E[y_{t-δ}|y_t], with the network output M A f_θ(y_t,t) used as a stand-in for the intermediate measurement; since Eq. (12) trains toward y while the forward bridge endpoint in Eq. (11) is M y, there is a potential target mismatch. That is an unproven approximation and a correctness risk, but it is not circularity: the paper's performance claims are empirical comparisons against external baselines, not consequences derived by definition from its own equations.
Assumptions & free parameters
free parameters (3)
- intermediate subsampling rate M_bar =
16.6% (mask between 25% and 12.5%)
- noise schedule sigma_t =
not specified
- number of inference steps =
4 (default); 2 and 8 also reported
assumptions (4)
- standard math The Gaussian diffusion bridge formulation from Delbracio-Milanfar [10] and Chung et al. [9] is valid for inverse problems.
- domain assumption The measurement operator A is orthogonal in CS-MRI.
- ad hoc to paper Applying two further subsampling operators to the observed measurement y yields a meaningful diffusion process.
- domain assumption Measurement noise e in y = M A x + e is ignored in the diffusion process and training targets.
Cite this review
Pith. "Pith review of A Self-supervised Diffusion Bridge for MRI Reconstruction." pith.science (2026). https://pith.science/paper/D6WJFL3F
@misc{pith2026250103430,
author = {Pith},
title = {Pith review of: A Self-supervised Diffusion Bridge for MRI Reconstruction},
year = {2026},
howpublished = {\url{https://pith.science/paper/D6WJFL3F}},
note = {Machine review of arXiv:2501.03430}
}
read the original abstract
Diffusion bridges (DBs) are a class of diffusion models that enable faster sampling by interpolating between two paired image distributions. Training traditional DBs for image reconstruction requires high-quality reference images, which limits their applicability to settings where such references are unavailable. We propose SelfDB as a novel self-supervised method for training DBs directly on available noisy measurements without any high-quality reference images. SelfDB formulates the diffusion process by further sub-sampling the available measurements two additional times and training a neural network to reverse the corresponding degradation process by using the available measurements as the training targets. We validate SelfDB on compressed sensing MRI, showing its superior performance compared to the denoising diffusion models.
Figures
Forward citations
Cited by 1 Pith paper
-
Self-Consistent Nested Diffusion Bridge for Accelerated MRI Reconstruction
A nested bidirectional diffusion bridge with a self-consistency loss and contourlet embedding improves magnitude-image MRI reconstruction over prior diffusion baselines.
Reference graph
Works this paper leans on
-
[1]
Sparse MRI: The application of compressed sensing for rapid MR imaging,
M. Lustig, D. Donoho, and J. M Pauly, “Sparse MRI: The application of compressed sensing for rapid MR imaging,” Magn. Reson. Med. , vol. 58, no. 6, pp. 1182–1195, 2007
work page 2007
-
[2]
Deep learning techniques for inverse problems in imaging,
G. Ongie et al., “Deep learning techniques for inverse problems in imaging,” IEEE J. Sel. Areas Inf. Theory, vol. 1, no. 1, pp. 39–56, 2020
work page 2020
-
[3]
MoDL: Model-Based Deep Learning Architecture for Inverse Problems,
H. K. Aggarwal, M. P. Mani, and M. Jacob, “MoDL: Model-Based Deep Learning Architecture for Inverse Problems,” IEEE Transactions on Medical Imaging , vol. 38, no. 2, pp. 394–405, Feb. 2019
work page 2019
-
[4]
A Survey on Diffusion Models for Inverse Problems,
G. Daras et al., “A Survey on Diffusion Models for Inverse Problems,” arXiv:2410.00083, 2024
arXiv 2024
-
[5]
DOLCE: A model-based probabilistic diffusion framework for limited-angle ct reconstruc- tion,
J. Liu et al., “DOLCE: A model-based probabilistic diffusion framework for limited-angle ct reconstruc- tion,” in Proc. IEEE Int. Conf. Comput. Vis. , 2023, pp. 10498–10508
work page 2023
-
[6]
Denoising Diffusion Probabilistic Models,
J. Ho, A. Jain, and P. Abbeel, “Denoising Diffusion Probabilistic Models,” in Proc. Adv. Neural Inf. Process. Syst., 2020, vol. 33, pp. 6840–6851
work page 2020
-
[7]
Diffusion Posterior Sampling for General Noisy Inverse Problems,
H. Chung et al., “Diffusion Posterior Sampling for General Noisy Inverse Problems,” in Proc. Int. Conf. Learn. Represent., Feb. 2023
work page 2023
-
[8]
DisC-Diff: Disentangled Conditional Diffusion Model for Multi-Contrast MRI Super- Resolution,
Y. Mao et al., “DisC-Diff: Disentangled Conditional Diffusion Model for Multi-Contrast MRI Super- Resolution,” in Proc. Med. Image. Comput. Comput. Assist. Intervent , 2023, pp. 387–397
work page 2023
Show all 17 references
-
[9]
Direct Diffusion Bridge using Data Consistency for Inverse Problems,
H. Chung, J. Kim, and J. C. Ye, “Direct Diffusion Bridge using Data Consistency for Inverse Problems,” in Proc. Adv. Neural Inf. Process. Syst. , 2023
2023
-
[10]
Inversion by Direct Iteration: An Alternative to Denoising Diffusion for Image Restoration,
M. Delbracio and P. Milanfar, “Inversion by Direct Iteration: An Alternative to Denoising Diffusion for Image Restoration,” Trans. Mach. Learn., 2024
2024
-
[11]
Self-supervised learning of physics-guided reconstruction neural networks without fully sampled reference data,
B. Yaman et al., “Self-supervised learning of physics-guided reconstruction neural networks without fully sampled reference data,” Magn. Reson. Med. , vol. 84, no. 6, pp. 3172–3191, 2020
2020
-
[12]
Self-Supervised Deep Equilibrium Models With Theoretical Guarantees and Applications to MRI Reconstruction,
W. Gan et al., “Self-Supervised Deep Equilibrium Models With Theoretical Guarantees and Applications to MRI Reconstruction,” IEEE Trans. Comput. Imaging , vol. 9, pp. 796–807, 2023
2023
-
[13]
Equivariant Imaging: Learning Beyond the Range Space,
D. Chen, J. Tachella, and M. E. Davies, “Equivariant Imaging: Learning Beyond the Range Space,” in Proc. IEEE Int. Conf. Comput. Vis. , Aug. 2021, pp. 4379–4388
2021
-
[14]
Ambient diffusion: Learning clean distributions from corrupted data,
G. Daras et al., “Ambient diffusion: Learning clean distributions from corrupted data,” in Proc. Adv. Neural Inf. Process. Syst. , 2023
2023
-
[15]
Image Super-Resolution via Iterative Refinement,
C. Saharia et al., “Image Super-Resolution via Iterative Refinement,” IEEE Trans. Pattern Anal. Mach. Intell., 2022. 6
2022
-
[16]
fastMRI: A publicly available raw k-space and DICOM dataset of knee images for accelerated MR image reconstruction using machine learning,
F. Knoll et al., “fastMRI: A publicly available raw k-space and DICOM dataset of knee images for accelerated MR image reconstruction using machine learning,” Radiol.: Artif. Intell. , vol. 2, no. 1, pp. e190007, 2020
2020
-
[17]
ESPIRiT-an eigenvalue approach to autocalibrating parallel MRI: Where SENSE meets GRAPPA,
M. Uecker et al., “ESPIRiT-an eigenvalue approach to autocalibrating parallel MRI: Where SENSE meets GRAPPA,” Magn. Reson. Med. , vol. 71, no. 3, pp. 990–1001, 2014. 7
2014
Reviewed August 10, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.