REVIEW 3 major objections 6 minor 5 references
Rapid Reconstruction of Extremely Accelerated Liver 4D MRI via Chained Iterative Refinement
T0 review · 3 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash
Pith's one-line read CIRNet, a chained iterative refinement network, claims to keep free-breathing liver 4D MRI usable at 30x undersampling, outperforming compressed sensing and Re-Con-GAN at every tested acceleration rate.
desk verdict Solid diffusion-based 4D MRI reconstruction with strong retrospective results, but the 30x clinical acceleration claim outruns the evidence. 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 mechanism is the chained iterative refinement loop: a forward Markovian diffusion process that adds Gaussian noise to the target image sequence over 800 timesteps, paired with a reverse process parameterized by CIRNet, a U-Net adapted from SR3 with BigGAN residual blocks and rescaled skip connections. In training, the network learns to estimate the noise added at each timestep from a random timestep schedule, minimizing a mean squared error loss derived from the variational lower bound. In inference, it denoises a noise sample step by step, conditioned on the undersampled input, so reconstruction quality builds up through repeated refinement instead of a single forward pass. This is what lets the model preserve fine tissue texture at high undersampling, according to the authors.
What would settle it
Run a genuinely 30x accelerated free-breathing liver 4D MRI acquisition on the same scanner, roughly 100 golden-angle spokes collected continuously, reconstruct it with CIRNet, and compare to a separately acquired fully sampled reference; if the PSNR and SSIM gap is much larger than the retrospective 22.35 dB and 0.11 1-SSIM reported here, the central claim fails.
Extended reading notes
Core claim
In the paper's own terms, CIRNet treats accelerated 4D MRI reconstruction as a stochastic iterative denoising problem rather than a direct regression. During training, a forward Markovian diffusion process gradually adds Gaussian noise to the fully sampled ground-truth sequence, and the network is optimized to reverse that process, conditioned on the undersampled input, by minimizing the mean squared error between estimated and true noise. At inference, the reverse process alone recovers the image from noise, using the undersampled nuFFT reconstruction as conditioning. The network processes 4D data as 2D+t temporal slices, with the eight respiratory phases as channels of a modified SR3 U-Net. On the held-out test set, CIRNet consistently outperforms CS and Re-Con-GAN in PSNR, 1-SSIM, and RMSE at all tested acceleration rates, and the authors claim it retains useable image quality at 30x acceleration.
Load-bearing premise
The paper's load-bearing premise is that retrospectively choosing 100 random spokes from a fully acquired 3000-spoke scan is equivalent to a true 30x accelerated acquisition, and that the 3000-spoke reconstruction is a fully sampled ground truth despite averaging only about 375 spokes per respiratory bin.
Editorial extensions
If this is right
- At 30x acceleration, CIRNet reports PSNR of 22.35 dB versus 13.27 dB for CS and 15.89 dB for Re-Con-GAN, so the improvement is not marginal in signal terms.
- Because the network processes 2D+t slices with eight respiratory-phase channels, it reconstructs a whole 4D volume in about 11 seconds, roughly ten times faster than the CS baseline's 120 seconds.
- If the retrospective results hold, the acquisition time for a liver 4D MRI could drop from 8-10 minutes to around 20 seconds at 30x acceleration, reducing motion artifacts and patient burden.
- The authors claim CIRNet preserves subtle tissue textures better than GAN-based reconstruction, which they attribute to modeling noise iteratively rather than regressing the mean.
Reading between the lines
- Beyond the paper: the decisive test is a prospective 30x accelerated acquisition, because randomly selecting 100 spokes from a fully sampled 3000-spoke scan may not reproduce the self-gating and motion-binning statistics of a true continuous short scan.
- Beyond the paper: the stated per-bin average of about 375 spokes at 3000 spokes is below the paper's own Nyquist count of 452 spokes, so the 'fully sampled' reference is mildly undersampled; quality metrics may partly reward smoothness rather than true anatomical fidelity.
- Beyond the paper: feeding raw k-space or multi-coil data directly into the diffusion model would remove the nuFFT preprocessing step and its potential artifact propagation, which the authors acknowledge as future work.
- Beyond the paper: clinical usefulness will depend on whether CIRNet reconstructions change tumor contouring or internal-target-volume margins, not only on PSNR and SSIM.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes CIRNet, a denoising-diffusion-based U-Net that reconstructs respiratory-binned 2D+t liver 4D MRI from undersampled radial k-space data. The method is trained on 37 patients and evaluated on 11 held-out patients, with retrospective random decimation of a fully sampled 3000-spoke acquisition to simulate 3x, 6x, 10x, 20x, and 30x acceleration. CIRNet is compared against compressed sensing and Re-Con-GAN using PSNR, 1-SSIM, and RMSE, and the paper reports consistently better quantitative performance with an 11 s per-volume inference time. The central claim is that CIRNet maintains clinically usable image quality up to 30x acceleration, corresponding to roughly a 20 s acquisition.
Significance. If the reported retrospective results transfer to prospectively accelerated acquisitions, the paper would make a meaningful contribution by extending diffusion-based reconstruction of 4D liver MRI well beyond the 10x acceleration achieved by prior methods, while retaining a clinically practical reconstruction time. Strengths of the study include a comparatively large patient cohort, a patient-level train/test split, evaluation across five acceleration factors, and comparison against both a conventional CS baseline and a published deep-learning baseline. The quantitative advantage of CIRNet is stable and large at high acceleration factors. However, the clinical acceleration claim is currently supported only by retrospective random decimation, and the ground-truth reference is itself not fully sampled by the Nyquist criterion; these issues are load-bearing for the external validity of the conclusions.
major comments (3)
- [Section 2.1] The 3x–30x acceleration experiments are all retrospective random decimations of a fully sampled 3000-spoke acquisition, and the paper equates 100 selected spokes with a true 30x scan. A real accelerated free-breathing acquisition would collect roughly 100 consecutive golden-angle spokes over about 20 s, derive the self-gating/binning signal from those same sparse spokes, and distribute the spokes over only a few respiratory cycles; none of these properties are reproduced by randomly deleting spokes after binning on the complete data. The resulting per-bin count of about 12.5 spokes and the artifact distribution at inference therefore differ from any clinical deployment, so the abstract and Section 4 claim that CIRNet maintains usable image quality for acceleration up to 30 times is not supported by the current experiments. A pseudo-prospective simulation that preserves the golden-angle view order and computes self-gating from the decimated data, or a prospective accelerated acquisition, is needed.
- [Section 2.1] The 'fully sampled' reference is stated to contain on average 375 spokes per respiratory bin, below the 452 spokes required by the Nyquist criterion for the 288x288 matrix. The quantitative metrics are therefore computed against an undersampled anchor, and the absolute statement that CIRNet 'maintains usable image quality' inherits this limitation. The authors should either acquire a truly fully sampled reference for a subset of patients or explicitly quantify and discuss the residual aliasing in the RV-3000 ground truth, since the reported PSNR, SSIM, and RMSE values are relative to that imperfect reference.
- [Sections 3 and 4] The conclusion that CIRNet maintains 'clinically deployable quality' is asserted from PSNR/SSIM/RMSE and visual inspection, but no clinical task evaluation (for example, tumor delineation or ITV generation) or reader study is performed. Since the Discussion invokes clinical deployability and reduced patient burden, the manuscript should either add task-based evaluation or restrict the claim to quantitative reconstruction performance on the retrospective benchmark.
minor comments (6)
- [Abstract and Table 1] At 30x acceleration, the abstract reports Re-Con-GAN PSNR as 13.27±3.89 dB, the same value as CS, whereas Table 1 lists Re-Con-GAN as 15.89±3.65 dB; please correct this inconsistency.
- [Section 2.2] The notation T is used both for the number of diffusion timesteps (T=800) and for the number of motion bins (C=T=8); rename one of these quantities to avoid confusion, and clarify the input shape 8x256x256x1 in the 2D+t setting.
- [Abstract and Section 2.4] The abstract contains the typo 'PNSR' for PSNR, and Section 2.4 says 'close-sourced' where 'closed-source' is meant.
- [Section 4] The text states that a 30x acquisition takes about 20 s and then says this 'approaches the duration of a breathing cycle'; since a typical respiratory cycle is about 3–5 s, 20 s spans several cycles and the wording should be corrected.
- [Table 1] The table caption says that the best and worst scores are bolded and wavy underlined, respectively, but the rendered table does not make this formatting unambiguous; please mark the entries clearly.
- [Tables 1] Only means and standard deviations are reported for the 11 test patients; paired significance tests across patients would substantiate the claim of 'consistently superior performance' compared with the baselines.
Circularity Check
No significant circularity: CIRNet is trained on held-out patients and scored against external baselines; the retrospective undersampling protocol is an external-validity limitation, not a construction-level circular step.
full rationale
The paper's reconstruction pipeline is a conditional denoising diffusion model (Sec. 2.2) optimized with the standard L2 noise-prediction loss (Eq. 2). The network weights are learned from training patients and evaluated on a held-out test split with CS and Re-Con-GAN as external baselines; no test-set parameter is fitted and the reported PSNR/SSIM/RMSE values are not equal to the training loss by construction. The one self-citation (Re-Con-GAN, Ref. 20) is used as a benchmark, not as a load-bearing premise, and its code is executed independently, so it does not import an unverified uniqueness claim. The main concern in Sec. 2.1 is that '30x acceleration' is operationalized as random deletion of spokes from the fully sampled 3000-spoke acquisition, and the RV-3000 reference itself has only ~375 spokes per bin versus the stated 452-spoke Nyquist criterion; thus the clinical claim of a ~20 s acquisition with preserved quality is not directly validated by this retrospective protocol. That is an external-validity / correctness limitation, not a circular derivation: the network's output is not identical to its input by construction, and the comparative evaluation remains meaningful under the stated protocol. The Discussion candidly lists other limitations (2D+t processing, image-domain operation, interpretability), further supporting that the central empirical finding is not an artifact of a definitional loop.
Assumptions & free parameters
free parameters (4)
- Diffusion timesteps T =
800
- Noise schedule gamma =
Piecewise distribution, uniform over timesteps
- Breathing regularity exclusion threshold =
20%
- Training iterations and learning rate =
1,000,000 iterations, Adam lr=1e-4 after 10k warmup
assumptions (4)
- standard math The conditional diffusion variational bound and simplified L2 denoising objective in Equations (1) and (2) are valid.
- domain assumption Retrospective random undersampling from the 3000-spoke acquisition faithfully represents a prospectively accelerated 30x acquisition.
- domain assumption RV-3000 reconstructions serve as ground truth despite per-bin spokes (~375) being below the cited Nyquist requirement (~452).
- domain assumption Temporal 2D+t slices preserve the 4D reconstruction problem; 3D+t volume consistency is not required.
Cite this review
Pith. "Pith review of Rapid Reconstruction of Extremely Accelerated Liver 4D MRI via Chained Iterative Refinement." pith.science (2026). https://pith.science/paper/ZTAJEGNU
@misc{pith2026241210629,
author = {Pith},
title = {Pith review of: Rapid Reconstruction of Extremely Accelerated Liver 4D MRI via Chained Iterative Refinement},
year = {2026},
howpublished = {\url{https://pith.science/paper/ZTAJEGNU}},
note = {Machine review of arXiv:2412.10629}
}
read the original abstract
Abstract Purpose: High-quality 4D MRI requires an impractically long scanning time for dense k-space signal acquisition covering all respiratory phases. Accelerated sparse sampling followed by reconstruction enhancement is desired but often results in degraded image quality and long reconstruction time. We hereby propose the chained iterative reconstruction network (CIRNet) for efficient sparse-sampling reconstruction while maintaining clinically deployable quality. Methods: CIRNet adopts the denoising diffusion probabilistic framework to condition the image reconstruction through a stochastic iterative denoising process. During training, a forward Markovian diffusion process is designed to gradually add Gaussian noise to the densely sampled ground truth (GT), while CIRNet is optimized to iteratively reverse the Markovian process from the forward outputs. At the inference stage, CIRNet performs the reverse process solely to recover signals from noise, conditioned upon the undersampled input. CIRNet processed the 4D data (3D+t) as temporal slices (2D+t). The proposed framework is evaluated on a data cohort consisting of 48 patients (12332 temporal slices) who underwent free-breathing liver 4D MRI. 3-, 6-, 10-, 20- and 30-times acceleration were examined with a retrospective random undersampling scheme. Compressed sensing (CS) reconstruction with a spatiotemporal constraint and a recently proposed deep network, Re-Con-GAN, are selected as baselines. Results: CIRNet consistently achieved superior performance compared to CS and Re-Con-GAN. The inference time of CIRNet, CS, and Re-Con-GAN are 11s, 120s, and 0.15s. Conclusion: A novel framework, CIRNet, is presented. CIRNet maintains useable image quality for acceleration up to 30 times, significantly reducing the burden of 4DMRI.
Reference graph
Works this paper leans on
-
[1]
Introduc+on Owing to its superior so] Bssue contrast compared to computed tomography, magneBc resonance imaging (MRI) has been increasingly adopted for image-guided liver radiaBon therapy (RT)1–3. 4D MRI, a volumetric imaging technique for respiratory-resolved images, is suited for characterizing tumor morphologies and moBons4–6. During the planning for f...
work page 2020
-
[11]
Griswold, M. A. et al. Generalized autocalibraBng parBally parallel acquisiBons (GRAPPA). Magn Reson Med 47, 1202–1210 (2002). 12. Feng, L. et al. Golden-angle radial sparse parallel MRI: combinaBon of compressed sensing, parallel imaging, and golden-angle radial sampling for fast and flexible dynamic volumetric MRI. Magn Reson Med 72, 707–717 (2014). 13. ...
work page 2002
-
[22]
Murray, V. et al. Movienet: Deep space–Bme-coil reconstrucBon network without k-space data consistency for fast moBon-resolved 4D MRI. Magne?c Resonance in Med 91, 600–614 (2024). 23. Huang, Q. et al. Dynamic MRI reconstrucBon with end-to-end moBon-guided network. Medical Image Analysis 68, 101901 (2021). 24. Liu, Z. et al. Video Swin Transformer. (2021) ...
-
[34]
Korkmaz, Y ., Cukur, T. & Patel, V. M. Self-supervised MRI ReconstrucBon with Unrolled Diffusion Models. in Medical Image Compu?ng and Computer Assisted Interven?on – MICCAI 2023 (eds. Greenspan, H. et al.) vol. 14229 491–501 (Springer Nature Switzerland, Cham, 2023). 35. Zhao, X. et al. DiffGAN: An adversarial diffusion model with local transformer for MRI ...
arXiv 2024
-
[45]
Vinogradova, K., Dibrov, A. & Myers, G. Towards Interpretable SemanBc SegmentaBon via Gradient-Weighted Class AcBvaBon Mapping (Student Abstract). AAAI 34, 13943–13944 (2020). 46. Adebayo, J. et al. Sanity checks for saliency maps. Advances in neural informa?on processing systems 31, (2018). 47. Zhou, Y ., Booth, S., Ribeiro, M. T. & Shah, J. Do Feature A...
work page 2020
Reviewed August 11, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.