{"id":"aa1bee43-1f25-4209-8201-a21cefbeaf2b","arxiv_id":"2505.24136","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SPIC-SSDU adds a sparse parallel-imaging consistency term to the MM-SSDU loss, improving artifact and noise reduction in self-supervised MRI reconstruction at acceleration rates 6 and 8.","lead":"A new self-supervised training loss for MRI reconstruction adds a sparse-domain consistency term that compares a network's response to added perturbations with the perturbations themselves, reducing aliasing artifacts at high acceleration. The method improves over several baselines on fastMRI knee and brain data, though gains are small and no error bars are reported.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1 contradicts the central claim: CC-SSDU has higher PSNR than SPIC-SSDU in two of six settings, so 'consistently outperforms' in Sec. 4.3 is false.","rationale":"The reader's weakest_assumption focuses on the linearity assumption behind Eq. (7). That is a legitimate theoretical concern, but the more immediate and decisive issue is that the paper's own quantitative results do not support the stated central claim. The abstract and Sec. 4.3 both say SPIC-SSDU 'consistently outperforms' all self-supervised baselines in PSNR and SSIM, yet Table 1 shows CC-SSDU wins on PSNR in two settings. This is a factual contradiction, not a matter of statistical interpretation alone. Since the entire paper is framed around this outperformance claim, it is the most load-bearing concern. The linearity assumption, by contrast, could be viewed as a heuristic that might still work empirically even if Eq. (7) is not exact; the numerical contradiction cannot be similarly waived. The reader's rationale does mention 'occasionally inconsistent improvements,' but the weakest_assumption field does not identify this as the primary risk. For that reason I disagree with the reader's chosen weakest assumption. Despite the overstated claim, the method still achieves the best SSIM in all settings and the best PSNR in four of six settings, so the paper has merit and a CONDITIONAL verdict is appropriate, requiring the authors to correct or qualify the 'consistently outperforms' claim and ideally report significance tests. The verdict therefore remains UNCHANGED.","tokens_in":7341,"tokens_out":6946,"duration_ms":77169,"concrete_test":"For the two contested cells (cor PD R=8 and ax T2 R=6), rerun the exact trained models on the same test splits using the released code and data, and compute paired PSNR/SSIM differences between SPIC-SSDU and CC-SSDU. If CC-SSDU still exceeds SPIC-SSDU in PSNR in either cell, the 'consistently outperforms' claim must be revised; if the table entries are typos, the corrected numbers would resolve the contradiction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim that SPIC-SSDU 'consistently outperforms other unsupervised and self-supervised methods in terms of PSNR and SSIM values' (Sec. 4.3) is contradicted by its own Table 1. On coronal PD at R=8, CC-SSDU achieves PSNR 35.76 vs. SPIC-SSDU 35.64; on axial T2 at R=6, CC-SSDU (and MM-SSDU) achieve PSNR 35.32 vs. SPIC-SSDU 35.24. Thus SPIC-SSDU is not the PSNR leader in two of the six reported settings. Furthermore, in cells where SPIC-SSDU leads, differences are often tiny (e.g., 0.01-0.05 dB, 0.001-0.002 SSIM), and no error bars, confidence intervals, or significance tests are provided, so it is unclear whether even these favorable comparisons are statistically meaningful. As written, the quantitative claim is unsupported and inconsistent with the reported evidence.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes SPIC-SSDU, a self-supervised training scheme for physics-driven deep learning (PD-DL) MRI reconstruction. It augments the MM-SSDU loss with a sparse-domain consistency term computed by adding designed perturbations to the measured k-space data and comparing the network's response difference against the true perturbation in the dual-tree complex wavelet domain with a reweighted L1 loss. Experiments on fastMRI knee and brain datasets at R=6 and R=8 with equispaced undersampling compare the method with supervised learning, MM-SSDU, ULIM, and CC-SSDU using PSNR/SSIM and example images. The paper claims consistent artifact and noise reduction and quantitative improvements over the unsupervised/self-supervised baselines.","tokens_in":7538,"tokens_out":7277,"duration_ms":79054,"significance":"The proposed idea of adding a sparse-domain parallel-imaging consistency term to MM-SSDU is interesting and, if validated, could improve self-supervised reconstruction at high acceleration factors. Strengths include the use of public fastMRI knee and brain data, comparison against three relevant baselines, and use of the same network architecture for all methods. However, the quantitative support for the central claim is currently weak: the reported improvements are small, the PSNR result is not consistent across all settings, and the key assumption behind Eq. (7) is not tested. The contribution is therefore a promising incremental method that needs further validation before it can be assessed as a reliable improvement.","major_comments":[{"comment":"The claim that SPIC-SSDU 'consistently outperforms other unsupervised and self-supervised methods in terms of PSNR and SSIM values' is contradicted by the reported numbers: on coronal PD at R=8, CC-SSDU has PSNR 35.76 versus 35.64 for SPIC-SSDU, and on axial T2 at R=6, MM-SSDU has PSNR 35.32 versus 35.24 for SPIC-SSDU. Moreover, the favorable differences when SPIC-SSDU does lead are often 0.01-0.05 dB, and no error bars, confidence intervals, or significance tests are reported. Please revise the claim, add per-slice or repeated-run statistics, or both.","section":"Section 4.3 / Table 1"},{"comment":"The definition of the sparse parallel-imaging consistency term in Eq. (8) relies on the equality p = f(y+q, E; theta) - f(y, E; theta) stated in Eq. (7). This equality is exact for a linear reconstruction operator that is a right inverse of the encoding operator, but the unrolled PD-DL network is nonlinear due to the CNN regularizer, and the paper provides neither a theoretical argument nor an empirical test that the equality holds approximately for the chosen perturbations. Because the entire proposed loss is built on this relation, this assumption is load-bearing. Please provide such a validation, e.g., measuring the residual f(y+q)-f(y)-p on validation data for the actual trained models.","section":"Section 3, Eq. (7)"},{"comment":"The construction of the perturbations p_k is not specified. The text says '3 perturbations with 3 k-space masks' and describes them only as 'well-designed perturbations' that prevent fold-overs, but it does not state how each perturbation is generated, how it is paired with a mask, or how the design guarantees recoverability by parallel imaging. Without this information the method cannot be reproduced and the validity of Eq. (7) cannot be assessed. Please provide the explicit construction algorithm.","section":"Section 4.1"}],"minor_comments":[{"comment":"The sentence that SPIC-SSDU achieves performance 'matching supervised learning' while reducing its artifacts at R=6 across all datasets is not supported by Table 1, since supervised learning has higher PSNR than SPIC-SSDU on coronal PD-FS at R=6 (34.45 vs. 34.05) and on axial T2 at R=6 (35.42 vs. 35.24). Please qualify this statement.","section":"Section 4.3"},{"comment":"The choice beta = 5e-3 and the number of perturbations/masks are reported without a sensitivity study; given the importance of the trade-off between the MM-SSDU term and the sparse PIC term, a small hyperparameter sweep would strengthen the paper.","section":"Section 4.1"},{"comment":"The paper does not report the hyperparameters used for the ULIM and CC-SSDU baselines beyond their loss formulations; for a fair and reproducible comparison, state whether all methods used the same training schedule, loss weighting, and mask settings.","section":"Section 4.2"},{"comment":"Figure 4 demonstrates the benefit of the sparse-domain loss over the L2 spatial-domain loss with a single qualitative example; provide quantitative metrics or a small dataset-wide comparison to support this claim.","section":"Figure 4"},{"comment":"There are minor typographical errors: 'A schmeatic' (Section 2.2) should be 'A schematic', and 'demonstrates slights artifacts' (Section 4.3) should be 'demonstrates slight artifacts'.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be a short conference-style paper being considered for a full journal venue; the evaluation is thin for the strength of the claims. Two of the four baselines (MM-SSDU and CC-SSDU) are from the same group, and the comparisons rely on single-number metrics without uncertainty quantification. For a journal-level decision, I would require the additional statistical validation and complete perturbation-construction details listed in the major comments. The proposed method itself is not circular, and the underlying idea is worth developing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Alçalar and Akçakaya propose a new self-supervised loss for accelerated MRI reconstruction. The core idea: add carefully designed perturbations to k-space during training and enforce that the network's output change in the sparse domain matches the perturbation, using a reweighted ℓ1 penalty. That specific combination—parallel-imaging-motivated perturbations plus sparse-domain consistency—is new, as far as I can tell. The paper is clearly written, the baselines include strong recent self-supervised methods (MM-SSDU, ULIM, CC-SSDU), and the ablation comparing their sparse loss to a spatial ℓ2 loss is a nice touch. The visual results are believable and the quantitative results mostly favor the proposed method.\n\nHowever, the paper's central claim is not supported by its own Table 1. The authors write that SPIC-SSDU 'consistently outperforms' other unsupervised and self-supervised methods in PSNR and SSIM. But in two of six settings, CC-SSDU has higher PSNR (cor PD R=8: 35.76 vs 35.64; ax T2 R=6: 35.32 vs 35.24). In those same settings SPIC-SSDU wins on SSIM, so the table is mixed rather than consistently better. Beyond that, the margins when SPIC-SSDU leads are sometimes 0.01–0.05 dB with no error bars or significance tests, so even the favorable comparisons are hard to interpret as real. This is a paper-internal inconsistency, not a critique from outside.\n\nAnother soft spot is Eq. (7), which defines the estimated perturbation as f(y+q,E)-f(y,E). The paper says PD-DL models are 'expected to satisfy' p equals that difference. That is an additivity/linearity assumption on a network that is otherwise nonlinear due to the CNN regularizer. The loss may still work if it is treated as a soft constraint, but the paper does not test or discuss the validity of the equality. It would be good to see a footnote or a simple experiment showing that the difference image actually looks like the perturbation.\n\nFinally, the method does not ship code or a precise description of how the perturbations are constructed beyond avoiding aliasing overlap, which makes reproduction harder than it needs to be.\n\nWho is this for? Researchers working on self-supervised MRI reconstruction will find the method plausible and probably worth trying, especially at high acceleration rates. But the quantitative support is weaker than the prose suggests. A serious referee should ask for statistical validation and a more careful wording of the claims.\n\nRecommendation: send it to peer review but require revisions: fix the overclaim, add error bars or at least a paired comparison, and address the linearity assumption. The core idea is sound enough.","headline":"A genuinely new self-supervised training loss for MRI reconstruction, but the paper's own Table 1 contradicts the 'consistently outperforms' claim and the core linearity assumption is untested.","tokens_in":8115,"tokens_out":2900,"would_cite":true,"duration_ms":33601,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Adding a sparse parallel-imaging consistency term to self-supervised training lets MRI reconstruction networks suppress aliasing and noise at acceleration factors 6 and 8 without fully-sampled references.","keywords":["self-supervised MRI reconstruction","parallel imaging","sparse-domain consistency","data undersampling","physics-driven deep learning","aliasing artifacts","dual-tree complex wavelet transform","high acceleration rates"],"falsifier":"Train the same PD-DL network on the same knee data with and without the sparse consistency term (Eq. 8), holding all other settings fixed, and compare test PSNR and SSIM at acceleration rate 8; if the difference is within the run-to-run variability of the self-supervised baselines, the central claim would be undercut, since the paper attributes its gains specifically to this term.","tokens_in":7119,"feed_emoji":"🧲","tokens_out":9353,"duration_ms":87284,"temperature":0.7,"pith_summary":"This paper proposes a way to train MRI reconstruction networks when no fully-sampled reference images are available, by adding a consistency term to the standard multi-mask self-supervised loss. The term feeds the network the same frequency-domain (k-space) data with and without a small designed perturbation and trains the difference of the two outputs to match the true perturbation in a sparse wavelet domain. On knee and brain MRI data undersampled at acceleration factors 6 and 8, the authors report that this term reduces aliasing artifacts and noise more effectively than three prior self-supervised and unsupervised methods, bringing quality close to supervised training. If it holds up, the method would make higher acceleration factors practical in clinical MRI without the usual loss of fidelity.","feed_headline":"Sparse-consistency loss cleans up self-supervised MRI at high rates","feed_subtitle":"A term that checks perturbation recovery cuts aliasing and noise at R=6 and R=8 without fully-sampled references.","key_machinery":"The load-bearing object is the sparse parallel imaging consistency term $\\mathcal{L}_{\\mathrm{spic}}$ (Eq. 8), computed as the reweighted $\\ell^1$ distance between the true perturbation $p$ and the network's predicted perturbation $p_{\\mathrm{est}} = f(y_\\Omega+q_\\Omega,E_\\Omega;\\theta)-f(y_\\Omega,E_\\Omega;\\theta)$, after applying the dual-tree complex wavelet transform (a shift-sensitive multi-scale transform that sparsifies MR images) as the sparsifying operator $W$ and normalizing by $|(Wp_{\\mathrm{true}})_n|+\\epsilon$. The perturbations are designed so that, at acceleration rate $R$, their aliased copies do not fold over one another, making them recoverable by parallel imaging; the sparse-domain weighting follows reweighted $\\ell^1$ minimization. The role of this term is to couple the network's behavior to classical parallel-imaging consistency during training, which is what the paper credits for artifact and noise suppression at high acceleration.","core_discovery":"The central claim is that a sparsity-driven parallel imaging consistency term, appended to the MM-SSDU objective, is what lets a self-supervised physics-driven deep learning network suppress coherent aliasing and noise at high acceleration. The paper defines perturbations $p$ whose undersampled k-space signatures $q_\\Omega = E_\\Omega p$ alias without overlapping in the field of view, so a parallel-imaging reconstruction can in principle recover them. It then asserts that a well-trained network should satisfy $p = f(y_\\Omega+q_\\Omega,E_\\Omega;\\theta)-f(y_\\Omega,E_\\Omega;\\theta)$, and enforces the estimated perturbation in a dual-tree complex wavelet domain with a reweighted $\\ell^1$ loss. The result, SPIC-SSDU, is reported to achieve the best PSNR and SSIM among MM-SSDU, ULIM, and CC-SSDU on coronal PD, coronal PD-FS knee, and axial T2 brain data at acceleration rates 6 and 8, approaching supervised learning and even removing some residual artifacts seen in supervised reconstructions.","pith_inferences":["The paper does not test the additivity assumption behind Eq. (7); measuring $p_{\\mathrm{est}}$ against $p_{\\mathrm{true}}$ for a range of inputs would show how much bias this introduces.","Because the sparse consistency term targets a different failure mode than cyclic-consistency methods, it could plausibly be combined with them for further gains, a combination the paper does not explore.","The perturbation design is tailored to coherent, equidistant undersampling, so the largest benefits would be expected in that clinically relevant setting; random undersampling may show smaller improvements."],"forward_implications":["SPIC-SSDU outperforms MM-SSDU, ULIM, and CC-SSDU in PSNR and SSIM on all three tested datasets at both acceleration rates 6 and 8.","Sparse-domain perturbation consistency beats a plain $\\ell^2$ spatial-domain perturbation loss, yielding sharper images with fewer artifacts.","Self-supervised training with SPIC-SSDU approaches supervised reconstruction quality, and in some cases removes residual aliasing that supervised training leaves behind.","The same framework is claimed to be applicable to random Cartesian and radial undersampling patterns, though only equidistant sampling is evaluated."],"supporting_citations":[{"why":"Supplies the multi-mask SSDU baseline loss that SPIC-SSDU augments, along with the recommended mask split ratio.","marker":"[7]"},{"why":"Cycle-consistent SSDU baseline included in the comparisons; its remaining artifacts motivate the new consistency term.","marker":"[11]"},{"why":"ULIM baseline that the paper compares against, representing cyclic measurement consistency without the MM-SSDU loss.","marker":"[14]"},{"why":"Provides the multi-coil knee and brain datasets used for training and testing.","marker":"[13]"},{"why":"Defines the unrolled variable-splitting PD-DL network with conjugate-gradient data fidelity used as the reconstruction backbone.","marker":"[3]"},{"why":"Motivates the reweighted $\\ell^1$ weighting inside the sparse consistency loss.","marker":"[20]"},{"why":"Supplies the dual-tree complex wavelet transform used as the sparsifying domain for the consistency comparison.","marker":"[23]"},{"why":"SENSE parallel imaging principle behind designing perturbations whose aliased copies do not overlap.","marker":"[15]"}],"fun_headline_variants":["Sparse consistency check sharpens self-supervised MRI at high rates","Self-supervised MRI cleans up via sparse consistency loss","Sparse consistency term reduces MRI artifacts without fully-sampled data","Perturbation-recovery consistency cleans up accelerated MRI","Sparse-domain consistency check tames aliasing in self-supervised MRI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The training signal assumes that adding a designed perturbation to the network's input shifts the network's output by exactly that perturbation, even though the network is a nonlinear learned map; if this does not hold, the consistency term enforces a relation that may not be true.","fun_headline_variants_meta":{"raw":{"variants":["Sparse consistency check sharpens self-supervised MRI at high rates","Self-supervised MRI cleans up via sparse consistency loss","Sparse consistency term reduces MRI artifacts without fully-sampled data","Perturbation-recovery consistency cleans up accelerated MRI","Sparse-domain consistency check tames aliasing in self-supervised MRI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001287,"raw_usage":{"total_tokens":5250,"prompt_tokens":932,"completion_tokens":4318,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":4231}},"tokens_in":548,"tokens_out":4318,"duration_ms":30558,"temperature":1.0,"reasoning_tokens":4231,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:32:17.513411+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same PD-DL network on the same knee data with and without the sparse consistency term (Eq. 8), holding all other settings fixed, and compare test PSNR and SSIM at acceleration rate 8; if the difference is within the run-to-run variability of the self-supervised baselines, the central claim would be undercut, since the paper attributes its gains specifically to this term.","supporting_citations":[{"cited_title":"A deep cascade of convolutional neural networks for dynamic MR image reconstruction,","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-mask SSDU baseline loss that SPIC-SSDU augments, along with the recommended mask split ratio."},{"cited_title":"ADMM-CSNet: A deep learning approach for image compressive sensing,","cited_arxiv_id":null,"evidence_quote":"Cycle-consistent SSDU baseline included in the comparisons; its remaining artifacts motivate the new consistency term."},{"cited_title":"Equivariant imaging: Learning beyond the range space,","cited_arxiv_id":null,"evidence_quote":"ULIM baseline that the paper compares against, representing cyclic measurement consistency without the MM-SSDU loss."},{"cited_title":"Multi-mask self-supervised learning for physics-guided neural networks in highly accelerated magnetic resonance imaging,","cited_arxiv_id":null,"evidence_quote":"Provides the multi-coil knee and brain datasets used for training and testing."},{"cited_title":"Our key innovation involves augmenting the MM-SSDU loss, as defined in (4), with a novel consistency term, which we refer to as sparse parallel imaging consistency","cited_arxiv_id":null,"evidence_quote":"Defines the unrolled variable-splitting PD-DL network with conjugate-gradient data fidelity used as the reconstruction backbone."},{"cited_title":"Unsuper- vised learning from incomplete measurements for in- verse problems,","cited_arxiv_id":null,"evidence_quote":"Motivates the reweighted $\\ell^1$ weighting inside the sparse consistency loss."},{"cited_title":"Gen- eralized autocalibrating partially parallel acquisitions (GRAPPA),","cited_arxiv_id":null,"evidence_quote":"Supplies the dual-tree complex wavelet transform used as the sparsifying domain for the consistency comparison."},{"cited_title":"Robust compressed sensing MRI with deep generative priors,","cited_arxiv_id":null,"evidence_quote":"SENSE parallel imaging principle behind designing perturbations whose aliased copies do not overlap."}],"review_version":1}