{"id":"c7bc385a-2274-4bde-87d4-3324fbd11d28","arxiv_id":"2412.09998","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A nested bidirectional diffusion bridge with a self-consistency loss and contourlet embedding improves magnitude-image MRI reconstruction over prior diffusion baselines.","lead":"SC-NDB reconstructs sharper MRI magnitude images from under-sampled ones by nesting two diffusion bridges with a self-consistency loss and adding contourlet-based detail features. The paper reports higher PSNR and SSIM than several diffusion baselines on fastMRI and IXI datasets, with only 20 sampling steps.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The nested self-consistency loss can be minimized by θ1/θ2 co-adaptation without improving the inference network; no control separates this from genuine gains.","rationale":"The reader's weakest assumption identifies exactly the load-bearing risk: the nested bridge uses the model's own one-step estimate as a target endpoint, and the paper gives no analysis of the resulting bias. I read the central claim as an empirical SOTA claim for magnitude-image reconstruction, and the experiments, if reproducible, may well support it. However, the novel component is the self-consistency mechanism, and the current evidence does not isolate its effect from co-adaptation artifacts or from the CDEM module. A stop-gradient control is the minimal experiment that would settle whether the circularity is benign. Since this concern does not by itself overturn the empirical comparison, my read leaves the reader's CONDITIONAL verdict unchanged. Secondary issues such as the unfair non-magnitude baseline comparison and missing code are real but less central to the method's internal logic.","tokens_in":14070,"tokens_out":14172,"duration_ms":164188,"concrete_test":"Retrain SC-NDB on the same fastMRI knee 4x protocol with a stop-gradient on ¯x0 and ¯y0 when forming y'_t2 and x'_t2 in Eq. (8), so the nested loss does not backpropagate into θ1 through the predicted endpoints. Compare the resulting PSNR/SSIM to the reported 34.84 dB / 89.73%. If the detached variant matches, the circular co-adaptation is not the source of the gain and the concern is resolved; if it drops materially, the headline result depends on the unanalyzed feedback and the method needs a non-circular justification or a corrected claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Algorithm 1 lines 7-14 and Eqs. (7)-(10) define a training loop in which the endpoint of the second Brownian bridge is the model's own one-step estimate (¯x0, ¯y0) and gradients flow through both denoisers. Because θ2 is asked to recover y0 from y'_t2 = (1 - m_t2)y0 + m_t2¯x0 + √σ_t2 ε, its success depends on the distribution of ¯x0 being close to the clean x0 distribution. Early in training this is false, and nothing in the loss prevents θ1 and θ2 from co-adapting on the training set: the self-consistency term can decrease by θ2 learning to invert θ1's particular errors rather than by θ1 improving its marginal prediction. The paper provides no analysis of the endpoint distribution, no stop-gradient or EMA target, and no control experiment that removes the circularity. The ablation in Table 4 only compares BBDM with the full method; it does not test a clean-target or detached-endpoint version, so the 0.64 dB gain over selfRDB in Table 1 cannot yet be attributed to a valid self-consistency prior.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SC-NDB, a training-time nested diffusion-bridge scheme for undersampled-to-fully-sampled magnitude MRI reconstruction. It trains two denoising networks θ1 and θ2 on the forward and reverse Brownian-bridge losses between under-sampled and fully-sampled magnitude images, and adds a self-consistency loss that treats each network's one-step estimate (¯x0, ¯y0) as the endpoint of a second bridge. A contourlet-decomposition embedding module (CDEM) injects multi-scale directional features into the U-Net backbone. The method is evaluated on fastMRI knee (in-distribution), IXI (in-distribution), and fastMRI brain (out-of-distribution), reporting improved PSNR/SSIM/NMSE relative to BBDM, selfRDB, DDPM, I2SB, and several non-magnitude-based methods, while using only 20 training and sampling steps and a single denoising network at inference.","tokens_in":14355,"tokens_out":8308,"duration_ms":88583,"significance":"The target scenario—magnitude-image-only MRI reconstruction from DICOM data—is clinically relevant and relatively underexplored. The 20-step training and sampling regime with a single inference network is practically appealing, and CDEM is a conceptually sensible way to inject explicit structural and textural features. The self-consistency idea, if it genuinely improves the single-network inference path, would be a useful training-time regularizer. However, the empirical support is currently incomplete: the main comparisons are not matched in step counts, and the self-consistency term is trained against the model's own predictions without a control for co-adaptation. With the additional controlled experiments described below, the paper could make a solid contribution.","major_comments":[{"comment":"The central state-of-the-art claim is not yet supported because the comparison protocol is uncontrolled. For SC-NDB, both training and sampling use T=20; for BBDM, DDPM, and I2SB, training uses T=1000 and sampling uses 200 DDIM steps. Performance differences in Tables 1–3 therefore conflate the proposed nested self-consistency mechanism with the step-count schedule. In particular, the reported 0.64 dB gain over selfRDB (34.84 vs 34.20 dB in Table 1) is of the same order as the sensitivity to the noise-variance multiplier shown in Table 6. Please add matched-step experiments, for example by training BBDM and selfRDB with T=20 and sampling with 20 steps, or by training SC-NDB with T=1000 and sampling with 200 steps, and report both quality and inference time under those matched settings. Also clarify the apparent contradiction between §4.1, which says other diffusion methods use 1000 training and 200 sampling steps, and §4.4, which says the training and sampling steps for SC-NDB and selfRDB were set to 20 each.","section":"§4.1, §4.4, Tables 1–3"},{"comment":"The self-consistency loss can be minimized by co-adaptation between θ1 and θ2 rather than by an improved single-network inference path. In Algorithm 1, the endpoint ¯x0 is produced by θ1 and then used as the target endpoint of the second bridge whose reconstruction ¯¯y0 is supervised by ||y0−¯¯y0||, with gradients flowing into both θ1 and θ2. Early in training, ¯x0 is far from a clean fully-sampled image, so θ2 can learn to invert θ1's particular errors; the paper provides no analysis of this bias. The ablation in Table 4 only toggles the whole self-consistency term on and off and does not distinguish a genuine prior from co-adaptation. I request (a) a clean-endpoint control replacing ¯x0 with ground-truth x0 in Eq. (8), (b) a detached-endpoint control with stop-gradient on ¯x0 and ¯y0, and (c) ideally an EMA target-network variant. Without at least one of these controls, the 0.36 dB self-consistency gain in Table 4 cannot be attributed to the intended mechanism.","section":"Algorithm 1, Eqs. (7)–(11), Table 4"},{"comment":"The derivation implicitly assumes that the one-step estimate ¯x0 is an approximate sample from the fully-sampled magnitude-image distribution, so that the second Brownian bridge y0 → ¯x0 in Eq. (8) has valid marginals. This is not established. Since ¯x0 depends on θ1 and on the same noise ε used in xt1, its distribution is neither independent of the bridge state nor equal to qdata(x0); during early training it is a heavily biased prediction. Because only θ1 is used at inference, the correctness of the whole approach rests on the claim that the training-time nested bridge improves θ1's marginal prediction. Please provide a justification or, at minimum, empirical diagnostics (for example, PSNR of ¯x0 versus training iteration, and the evolution of ||x0−¯x0|| relative to ||x0−¯¯x0||) showing that the endpoint assumption holds in the regime where the self-consistency loss is active.","section":"§3.2, Eqs. (6)–(8)"}],"minor_comments":[{"comment":"The second line should read ¯y0 = yt1 − ϵθ2(yt1, t1); as printed it mixes xt1 with θ2's output and is inconsistent with Algorithm 1.","section":"Eq. (7)"},{"comment":"The step count used for selfRDB is ambiguous: §4.1 appears to include selfRDB among methods using 1000/200, while §4.4 states selfRDB uses 20 steps. Please state explicitly which T was used for each baseline in each table.","section":"§4.1 and §4.4"},{"comment":"The caption says 'Denosing network' instead of 'Denoising network', and the module is called 'Contourlet Decomposition' in §3.3 but 'Contour Decomposition' in the title and abstract; please use one name consistently.","section":"Fig. 5 and §3.3"},{"comment":"The header 't1 = t2 ∼ U t 1 ∼ U, t2 ∼ U' is garbled and should be reformatted; also clarify whether the two timesteps are sampled jointly or independently.","section":"Table 5"},{"comment":"Zhang and Ghanem (2018a) and (2018b) refer to the same ISTA-Net paper; the duplicate should be merged into one reference.","section":"References"},{"comment":"Please specify how inference time was measured (GPU model, batch size, image size) and whether all compared methods used the same hardware and sampling-step settings; without this, the efficiency comparison is difficult to reproduce.","section":"Fig. 11 and §4.4"},{"comment":"The noise-variance multiplier √σt2 is tuned on the validation set; the selection procedure should be described so that the reported result is understood as a tuned configuration rather than a fixed hyperparameter.","section":"§4.5, Table 6"}],"recommendation":"major_revision","confidential_remarks":"The paper is promising, but the state-of-the-art claim needs matched-step comparisons and an endpoint-control ablation before it can be accepted. Please also check the relationship to selfRDB (Arslan et al. 2024): the neighboring term 'self-consistent' appears in both works, and the current text does not clearly explain the differences between its nested self-consistency and the recursive self-consistency of selfRDB."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper has a real clinical motivation (magnitude-only images are what DICOM gives you) and a concrete, trainable idea. The experiments span fastMRI knee, IXI, OOD brain, with ablations and significance tests. That's more than many papers do.\n\nWhat's new: combining two nested Brownian bridges with a self-consistency loss that uses the model's own one-step estimate as the endpoint of the second bridge, plus contourlet features injected into the U-Net. The contourlet module is modest but sensible.\n\nThe soft spots are real. First, the self-consistency loss is exactly the kind of target that can be minimized by θ1/θ2 co-adaptation without improving the inference network, because gradients flow through both and there's no stop-gradient, EMA, or clean-target control. The paper offers no analysis of the endpoint distribution. This isn't fatal, but it means the 0.64 dB over selfRDB in Table 1 is not yet attributable to a valid prior. Second, the evaluation is confounded for most baselines: SC-NDB uses 20 steps while BBDM, I2SB, DDPM use 1000/200. The paper says in Sec. 4.4 that selfRDB also used 20, but Sec. 4.1 seems to say all diffusion baselines used 1000/200. That inconsistency needs to be resolved. The ablation in Table 4 uses BBDM as baseline, and if BBDM ran 200 sampling steps, the +0.36 dB from self-consistency is contaminated.\n\nNovelty is overstated: selfRDB already does self-consistent recursive bridges, and the related work claims self-consistency was overlooked. That needs rewording.\n\nThe math itself is okay: Eq. (6) correctly rearranges the forward process. The variance multiplier tuning on validation is model selection, not circularity.\n\nBottom line: the paper deserves referee time because the task is clinically real and the mechanism is non-trivial. But a serious referee should ask for a control experiment with detached targets or stop-gradient, a step-matched comparison with all baselines, and a clarification of selfRDB's settings. I'd send it to review, with the expectation of a major revision.","headline":"A clinically motivated nested diffusion bridge for magnitude-only MRI, with a plausible but unvalidated self-consistency mechanism; worth reviewing but needs controls.","tokens_in":14877,"tokens_out":3804,"would_cite":false,"duration_ms":38792,"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":"Nested diffusion bridge beats MRI baselines in 20 steps.","keywords":["accelerated MRI reconstruction","magnitude image","diffusion bridge","self-consistency","contourlet decomposition","image translation","fastMRI","IXI dataset"],"falsifier":"Train a variant in which the second bridge is built on the true ground-truth image rather than the model's estimate \\(\\bar{x}_0\\); if this variant achieves the same PSNR and SSIM as SC-NDB, the self-consistency mechanism is not the source of the improvement. A cheaper check is to measure early in training whether the one-step estimates \\(\\bar{x}_0\\) overlap the distribution of real fully-sampled images.","tokens_in":13882,"feed_emoji":"🧲","tokens_out":5512,"duration_ms":52611,"temperature":0.7,"pith_summary":"The paper proposes SC-NDB, a diffusion-bridge framework that reconstructs fully-sampled MRI magnitude images from under-sampled ones, the data format actually stored in clinical DICOM files. It claims that a nested pair of Brownian bridges, trained with a self-consistency loss, produces the best reconstruction quality on the fastMRI knee and IXI benchmarks among both magnitude-based and k-space/image-space diffusion models, for example 34.84 dB PSNR at 4x acceleration on fastMRI knee versus 33.79 dB for the BBDM baseline. The design also cuts inference cost sharply: only one denoising network is used at inference, with 20 training and sampling steps instead of the usual 1000/200. A contourlet decomposition embedding module is added to feed multi-scale directional high-frequency information into the denoiser.","feed_headline":"Nested diffusion bridge beats MRI baselines in 20 steps","feed_subtitle":"Uses only magnitude images and one denoising network, yet raises PSNR by about 1 dB over prior diffusion methods.","key_machinery":"The central object is the nested diffusion bridge: an outer Brownian bridge between under-sampled y0 and fully-sampled x0, and a second bridge started from the model's own one-step estimate \\(\\bar{x}_0\\) (or \\(\\bar{y}_0\\)). A self-consistency loss measures the distance between the original image and the reconstruction produced by running the opposite denoiser on this second bridge, forcing the two networks to encode consistent source priors. The supporting machinery is the Contourlet Decomposition Embedding Module (CDEM), which uses Laplacian pyramids and directional filter banks to extract multi-scale, orientation-selective high-frequency features and injects them into the U-Net denoiser.","core_discovery":"SC-NDB models accelerated MRI reconstruction as a bidirectional image-translation problem between under-sampled and fully-sampled magnitude images, and shows that forcing two nested diffusion bridges to agree with each other during training improves the accuracy of the single-network reverse path used at inference. Concretely, the training objective combines the usual reconstruction losses for x←y and y←x with a self-consistency loss that runs each denoiser on the other's one-step estimate and penalizes the resulting mismatch. On fastMRI knee at 4x acceleration the method reaches 34.84 dB PSNR and 89.73% SSIM, beating the previous magnitude-image diffusion baselines BBDM (33.79 dB / 87.85%) and selfRDB (34.20 dB / 89.19%), and also beating k-space and image-space diffusion models such as MC-DDPM and HFS-SDE. The gains hold at 8x acceleration, on the IXI T1 dataset, and in out-of-distribution fastMRI brain scans, and the whole pipeline uses only 20 diffusion steps.","pith_inferences":["If self-consistency acts as a training regularizer rather than a new inference procedure, similar nested-consistency losses could be bolted onto other diffusion bridges (e.g., Schrödinger bridges) to stabilize few-step training.","The magnitude-image-only setup removes the need for raw k-space access, which could make diffusion-based reconstruction deployable on existing hospital DICOM workflows.","The CDEM's contourlet features are task-agnostic; one testable extension is to apply SC-NDB to other inverse problems dominated by high-frequency loss, such as super-resolution or deblurring of magnitude images.","Since the paper reports results with only 20 training steps, a systematic study of how few steps the nested bridge can tolerate (e.g., 4 or 8) would clarify the practical speed-up limits."],"forward_implications":["Reconstruction quality on magnitude-only MRI improves by roughly 0.6 to 1.0 dB PSNR over the previous best magnitude-image diffusion baselines, at both 4x and 8x acceleration.","Inference becomes much cheaper: 20 steps and a single denoising network, compared with 200-step DDIM sampling for the other diffusion baselines.","The method generalizes across anatomy: models trained on knee scans transfer to brain scans, with the largest gains over baselines appearing in the out-of-distribution setting.","The same training-time self-consistency regularizer could strengthen any bridge-based image translation model without adding inference cost."],"supporting_citations":[{"why":"Supplies the Brownian bridge formulation (Eqs. 1-4) that SC-NDB nests and extends.","marker":"Li et al. (2023)"},{"why":"Defines the DDPM denoising objective and is a primary baseline for comparison.","marker":"Ho et al. (2020)"},{"why":"selfRDB is the closest prior work on self-consistent recursive diffusion bridges for medical image translation and a key baseline.","marker":"Arslan et al. (2024)"},{"why":"I2SB provides the image-to-image Schrödinger bridge baseline against which SC-NDB is compared.","marker":"Liu et al. (2023)"},{"why":"HFS-SDE is a non-magnitude diffusion baseline and the source of the experimental preprocessing (first six slices excluded).","marker":"Cao et al. (2024)"},{"why":"The Laplacian pyramid used by the CDEM to separate low- and high-frequency components.","marker":"Burt and Adelson (1987)"},{"why":"The directional filter bank used to decompose high-pass subbands into orientation-selective features.","marker":"Bamberger and Smith (1992)"}],"fun_headline_variants":["Self-consistent nested diffusion bridge beats MRI baselines in 20 steps","Magnitude-only MRI: nested bridge gains 1 dB over diffusion models","20-step nested diffusion bridge outperforms k-space and image-space MRI methods","SC-NDB: nested self-consistency powers fast MRI reconstruction from magnitude","Bi-directional nested bridge sets new MRI reconstruction state-of-the-art in 20 steps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The training loop treats the model's own one-step estimate of the fully-sampled image as a trustworthy endpoint for the second diffusion bridge, even though early in training that estimate is far from a real image; if this assumption fails, the self-consistency loss can reinforce the model's errors instead of correcting them.","fun_headline_variants_meta":{"raw":{"variants":["Self-consistent nested diffusion bridge beats MRI baselines in 20 steps","Magnitude-only MRI: nested bridge gains 1 dB over diffusion models","20-step nested diffusion bridge outperforms k-space and image-space MRI methods","SC-NDB: nested self-consistency powers fast MRI reconstruction from magnitude","Bi-directional nested bridge sets new MRI reconstruction state-of-the-art in 20 steps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000539,"raw_usage":{"total_tokens":2636,"prompt_tokens":1047,"completion_tokens":1589,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":663,"completion_tokens_details":{"reasoning_tokens":1488}},"tokens_in":663,"tokens_out":1589,"duration_ms":15976,"temperature":1.0,"reasoning_tokens":1488,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T16:28:09.459989+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a variant in which the second bridge is built on the true ground-truth image rather than the model's estimate \\(\\bar{x}_0\\); if this variant achieves the same PSNR and SSIM as SC-NDB, the self-consistency mechanism is not the source of the improvement. A cheaper check is to measure early in training whether the one-step estimates \\(\\bar{x}_0\\) overlap the distribution of real fully-sampled images.","supporting_citations":[],"review_version":1}