{"id":"5678042e-1dec-41fb-97d9-a1cdd1a5fa36","arxiv_id":"2411.16535","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"ADOBI combines a pretrained diffusion bridge with adaptive coil sensitivity calibration, delivering measurement-consistent blind parallel MRI reconstruction in 5 to 10 steps.","lead":"ADOBI adaptively re-estimates the unknown coil sensitivity maps while a pretrained diffusion bridge reconstructs the MR image, keeping the result consistent with the measured k-space data. This yields better parallel MRI reconstruction than fixed-map baselines, in only 5 to 10 sampling steps.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 2 as written never uses the data-consistent estimate x'_0|t in the resampling step, so the described method may not enforce measurement consistency at all.","rationale":"The reader's weakest assumption focused on distribution shift of edited states entering the pretrained DB backbone. That is a real risk, but there is a more basic flaw in the manuscript's own algorithm description: the data-consistent estimate x'_0|t computed in Eq (15) is never referenced after line 4 of Algorithm 2, so the described method may not implement measurement-consistent sampling at all. This directly undermines the strongest claim: being the first measurement-consistent DB for blind inverse problems. I do not charge the authors with dishonesty; the most likely explanation is a typographical error in the pseudocode (e.g., line 7 should pass x'_0|t). However, the paper as written cannot be checked without either a code release or a correction. Because the concern is about the algorithm specification rather than the underlying idea, the appropriate action is to keep the conditional verdict and add this concrete verification requirement. The distribution-shift concern remains relevant only if variant (B) is confirmed as intended.","tokens_in":17976,"tokens_out":5780,"duration_ms":47988,"concrete_test":"Run ADOBI inference on the same fastMRI test slices under two variants: (A) exactly as written in Algorithm 2, resampling with the unedited x0|t; (B) resampling with x'_0|t in Eq (11). Compare outputs to Tables 1-2. If (A) and (B) are identical, the published numbers must come from (B) or an unstated modification; if they differ, the variant matching the tables identifies the actual algorithm. Alternatively, release the inference code and trace whether the output of Eq (15) is an input to Eq (11).","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Algorithm 2 (§3.2), the image update (15) produces x'_0|t at line 4 and the CSM update (16) produces S_{t-1} at line 5, but the resampling step at line 7 (Eq 11) uses the unedited x0|t from line 3. x'_0|t is never used again, and S_{t-1} only affects the next iteration's line-4 gradient through the updated forward model. As written, the data-consistency image gradient step cannot influence the trajectory, so the method reduces to CSM-adaptive DDB without the core measurement-consistency mechanism. If this is a typo and the intended code passes x'_0|t to Resample, the paper must state that correction; if it is not a typo, the reported improvements over CDDB cannot be attributed to the described algorithm. Either way, the central claim of a 'measurement-consistent diffusion bridge for blind inverse problems' rests on an internal inconsistency in the algorithm specification.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ADOBI, an adaptive diffusion bridge method for blind inverse problems, applied to parallel MRI reconstruction. The forward model (coil sensitivity maps) is unknown, and ADOBI alternates between a data-consistency gradient step on the image estimate and a regularized update of the CSMs, using a pretrained diffusion bridge backbone that maps zero-filled or GRAPPA initializations to clean images. Experiments on fastMRI brain data at 4x and 8x acceleration, with and without noise, report state-of-the-art PSNR, SSIM, and LPIPS with 10 sampling steps, along with ablations on initialization, calibration, stochastic perturbation, and runtime.","tokens_in":18215,"tokens_out":3994,"duration_ms":38046,"significance":"If the method works as described, it is a useful contribution: it extends measurement-consistent diffusion bridges to blind inverse problems, achieves strong empirical performance with few sampling steps, and provides uncertainty quantification. The paper includes several ablation studies (Tables 3-5, Table 7, Fig. 8) that support the importance of calibration and initialization. However, the internal inconsistency in Algorithm 2 and the omission of key implementation details prevent the results from being attributed to the described method, and the paper cannot currently be reproduced.","major_comments":[{"comment":"The image update step is dead code: line 4 computes x'_0|t via Eq. (15), but the resampling step in line 7 uses x0|t, not x'_0|t. As written, the data-consistency image gradient never influences the trajectory, so the method reduces to CSM-adaptive CDDB without the image-domain measurement-consistency mechanism that the paper's central claim rests on. The pseudocode must be corrected to pass x'_0|t to Resample, or the text must explain how the update affects sampling; otherwise the reported improvements over CDDB cannot be attributed to the described algorithm.","section":"§3.2, Algorithm 2"},{"comment":"The CSM update is not specified sufficiently for replication: the text says a gradient descent algorithm iteratively minimizes Eq. (16), but the step size, number of inner iterations, initialization of S_t at each outer iteration, and the value of λ are never given. Since Table 4's claimed benefit of calibration depends on this inner loop, the method cannot be reproduced, and the alternating updates' convergence is not analyzed.","section":"§3.2.2, Eq. (16)"},{"comment":"The hyperparameters γ and λ are not reported in the final experiments. Figure 8 shows that γ is tuned on a 4x task, but the paper does not state whether this tuning is performed on a validation split or on the test set, nor what values of γ (and λ) are used for the 8x and noisy settings. The authors must specify the validation protocol and the exact hyperparameter values for all reported configurations.","section":"§8.6, Fig. 8; Tables 1-2"},{"comment":"The diffusion bridge backbone is trained on states xt obtained from a fixed initialization distribution (zero-filled or GRAPPA). At inference, the resampling states are edited by data-consistency gradients (once Algorithm 2 is corrected to use x'_0|t) and depend on adaptively updated CSMs. The paper does not analyze whether these edited states remain on the distribution seen during training; if they drift, the MMSE estimate x0|t degrades. Provide an analysis or at least empirical diagnostics (e.g., distance between edited states and the training distribution) to justify the method's validity.","section":"§3.1-3.2, Eqs. (13)-(16)"}],"minor_comments":[{"comment":"There are several typos and grammatical errors: 'from from' in the introduction, 'curial' for 'crucial', 'fine-turn' for 'fine-tune' in the appendix, 'a adaptive' for 'an adaptive' in the conclusion, and 'the object function' for 'the objective function' in §2.3.","section":"Throughout"},{"comment":"The list of ADOBI variants is duplicated: 'ADOBI w/o Calibration' appears twice, and the third variant is run together as 'ADOBIGroundtruth CSMs'.","section":"§4.3.1"},{"comment":"In the 8x ADOBI (ZF) row, LPIPS is reported as 0.125 ± 0.0200 with an extra decimal place; the other entries use two decimal places.","section":"Table 1"},{"comment":"The notation for β_t is introduced as β_t = (α_{t-1}/α_t), but the same symbol is used earlier for the DDPM noise schedule; the relationship between α_t in Eq. (9) and the DDPM α_t should be clarified.","section":"§2.3, Eq. (11)"},{"comment":"The caption says that the particular worst-case result is boxed in orange, but it is not immediately clear how to read the rows when each row corresponds to a different method's worst-case slice; please clarify the visualization.","section":"Figures 9-10"}],"recommendation":"major_revision","confidential_remarks":"The Algorithm 2 inconsistency is the main concern. It is likely a typo rather than a fundamental flaw, because the text and ablation studies suggest the authors intended to use x'_0|t in the resampling step. However, as written, the paper's central claim is not supported by the described algorithm. The missing hyperparameters and the lack of a validation protocol also need to be addressed before the paper can be considered for acceptance. Given the paper's applied focus, the absence of code is not fatal but is a significant reproducibility limitation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline message: the algorithm as written does not implement the method the paper advertises. In Algorithm 2, the image-domain data-consistency update (line 4, Eq. 15) computes x'_0|t, but the resampling step (line 7) uses the original x0|t. Nowhere is x'_0|t used again. So the gradient step that is supposed to enforce measurement consistency is dead code. The CSM update (line 5, Eq. 16) is similarly decoupled from the image trajectory; it only changes the model used in a gradient that is never applied. As written, the whole inference reduces to plain DDB sampling with no adaptive component affecting the output. That directly contradicts the abstract's claim of \"adaptively calibrates the unknown forward model to enforce measurement consistency throughout sampling iterations.\" This is not a minor typo in a tangential equation; it is the central mechanism of the paper. If the implementation actually passes x'_0|t to Resample, the authors need to state that clearly and correct the pseudocode. If it does not, the reported gains over CDDB cannot be attributed to the described algorithm, and the experimental results are unexplained.\n\nWhat is genuinely useful: grafting CSM estimation onto a pretrained diffusion bridge is a sensible idea, and the ablations (Tables 3-4) are the right checks: GRAPPA initialization helps, and CSM calibration helps when it is actually connected. The fastMRI numbers are strong and, if reproducible, would matter to the MR reconstruction community.\n\nOther soft spots, secondary but real: no code or data released; the CSM inner-loop iterations and step size are unspecified; lambda appears only in Eq. 16 with no value; gamma is tuned on the task. The citation practice is fine; references to the group's own work are not load-bearing.\n\nRecommendation: this deserves a serious referee, not a desk reject — a corrected algorithm plus released code could be a solid contribution. But the reviewers must ask for the fixed pseudocode, the missing hyperparameters, and confirmation that the experiments match the corrected method. I would not cite it in its current form.","headline":"Algorithm 2 never uses its own image-consistency update, so the described method is not measurement-consistent; major revision needed.","tokens_in":18754,"tokens_out":4355,"would_cite":false,"duration_ms":35603,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Adaptive diffusion bridge beats MRI baselines in 10 steps, even when coil maps are unknown.","keywords":["diffusion bridges","blind inverse problems","parallel MRI reconstruction","coil sensitivity maps","measurement consistency","adaptive forward model calibration","fastMRI","diffusion models"],"falsifier":"Run ADOBI on a set of slices with intentionally corrupted initial coil sensitivity maps, for example by shifting or scaling the ESPIRiT maps, and check whether the adaptive CSM update still drives the data-fidelity term to zero and preserves PSNR; a large drop would indicate the method relies on a good initialization rather than on the bridge dynamics. A more direct test is to compute a distribution-distance metric, such as FID or MMD, between the edited intermediate states produced after Eq. (11) and the states the bridge saw during training, and to check whether that distance grows with the number of steps and correlates with the reported performance drop.","tokens_in":17771,"feed_emoji":"🧠","tokens_out":6190,"duration_ms":55322,"temperature":0.7,"pith_summary":"The paper claims that blind inverse problems, where part of the measurement operator is unknown, can be solved by a diffusion bridge that alternates between two updates: one that pushes the current image estimate toward the measured data, and one that recalibrates the unknown forward model, here the coil sensitivity maps in parallel MRI. Existing diffusion-bridge methods enforce measurement consistency only when the forward model is fully known, so they fail in blind settings; ADOBI closes that gap without retraining the bridge or adding a second network. The result is a reconstruction method that reaches high fidelity and perceptual quality on fastMRI brain data in just 5–10 sampling steps, improving on both diffusion-model baselines and previous bridges. A careful reader would care because it makes fast, measurement-faithful MRI reconstruction practical in the realistic case where coil sensitivities are not known exactly.","feed_headline":"Adaptive diffusion bridge beats MRI baselines in 10 steps","feed_subtitle":"Joint image and coil-sensitivity updates enforce measurement fidelity with only 10 sampling steps.","key_machinery":"The central object is the pretrained diffusion bridge $R_\\theta$, which maps a degraded image $z$ (zero-filled or GRAPPA) to a clean image $x_0$ through the interpolation $x_t = (1-\\alpha_t)x_0 + \\alpha_t z + \\sigma_t \\epsilon$, trained to output the posterior mean $\\mathbb{E}[x_0|x_t]$. The mechanism that makes the method work for blind problems is the alternating update: Eq.\\ (15) applies a data-consistency gradient to the image estimate using the current coil sensitivities, and Eq.\\ (16) re-estimates those sensitivities by minimizing the same data-fidelity term plus a Tikhonov penalty anchored at the ESPIRiT initialization. This joint image-operator refinement is what distinguishes ADOBI from CDDB, which can only apply the data-consistency step when the operator is known.","core_discovery":"ADOBI's central claim is that measurement consistency can be enforced in a blind inverse problem by jointly optimizing the image and the unknown forward model during diffusion-bridge inference. At each of its few sampling steps, the method first obtains an MMSE estimate $x_{0|t}$ from the pretrained bridge, then takes one gradient step on the data-fidelity term $\\|y - P F S_t(x_{0|t})\\|_2^2$ with respect to the image, and then refines the coil sensitivity maps $S_t$ by solving a Tikhonov-regularized least-squares problem that keeps the updated maps close to the ESPIRiT initialization. These two alternating updates make the bridge's samples track the observed k-space measurements even though the forward model was not known at training time. On fastMRI brain data at $4\\times$ and $8\\times$ acceleration, this adaptive scheme reports the best PSNR, SSIM, and LPIPS among the compared methods with only 10 function evaluations, and the gain is largest when the backbone is initialized with GRAPPA reconstructions rather than zero-filled images. The paper also positions ADOBI as the first measurement-consistent diffusion bridge for blind inverse problems and the first image-domain diffusion bridge for parallel MRI.","pith_inferences":["The alternating update idea is not MRI-specific: any blind inverse problem with a parametric forward model and a differentiable data-fidelity term could use the same two-step refinement inside a diffusion bridge, so blind deblurring or joint reconstruction plus field-map estimation in MRI are natural next targets.","Training the backbone on the edited intermediate states, rather than only on the clean interpolation, should close the distribution-shift gap and would quantify how much the current method leaves on the table; this is a concrete experiment the paper does not run.","The variance maps produced by stochastic sampling could be calibrated against error to yield a quantitative confidence signal for clinical workflows, but the paper only demonstrates correlation visually, so establishing a calibration curve is an open step.","The method's reliance on a high-quality initialization suggests that integrating the CSM update with a more flexible prior, such as a learned sensitivity prior, could extend ADOBI to settings where ESPIRiT initial maps are unreliable."],"forward_implications":["With only 10 sampling steps, ADOBI outperforms diffusion-model baselines DPS and DDS, which require 1000 and 100 steps respectively, while also improving over the non-adaptive bridges I2SB and CDDB.","The adaptive forward-model calibration brings reconstruction quality close to that obtained with ground-truth coil sensitivities, while adding only about 0.7 seconds per image in the reported runtime.","Using GRAPPA as the bridge's initialization distribution yields consistently better reconstructions than zero-filled initialization, for both ADOBI and CDDB.","The stochastic (SDE) formulation of the bridge gives better reconstruction than the deterministic (ODE) version, preserving fine structure during the 10-step process.","The method also provides uncertainty quantification by computing pixel-wise variance across stochastic samples, which the paper reports correlates with reconstruction error in ill-posed regions."],"supporting_citations":[{"why":"CDDB establishes the consistent diffusion bridge with data consistency for known forward models, which ADOBI extends to blind inverse problems.","marker":"[42]"},{"why":"I2SB provides the image-to-image Schrödinger bridge backbone and training recipe that ADOBI reuses with zero-filled or GRAPPA initialization.","marker":"[39]"},{"why":"GRAPPA is used to generate the higher-quality starting distribution for the bridge and is shown to improve ADOBI's reconstructions.","marker":"[37]"},{"why":"ESPIRiT estimates the initial coil sensitivity maps that anchor the adaptive CSM update through the Tikhonov penalty in Eq. (16).","marker":"[38]"},{"why":"E2E-VarNet is the strongest model-based deep-learning baseline for parallel MRI that ADOBI outperforms, providing a key comparison for learned CSM estimation.","marker":"[20]"},{"why":"DDS is a diffusion-model baseline that ADOBI surpasses while using 10 steps instead of 100, serving as a speed and quality comparison.","marker":"[55]"},{"why":"GibbsDDRM is the blind diffusion baseline that jointly estimates image and forward model, which ADOBI beats in the reported experiments.","marker":"[35]"},{"why":"DPS supplies the measurement-consistency posterior sampling update for non-blind diffusion models and is one of the baselines ADOBI is compared against.","marker":"[28]"},{"why":"fastMRI provides the multicoil brain dataset with undersampling masks used for training and evaluation.","marker":"[56]"}],"fun_headline_variants":["Blind inverse problems solved by adaptive diffusion bridge in 10 steps","Adaptive bridge tunes MRI coils to enforce measurement consistency","Blind MRI reconstruction in 10 steps via adaptive diffusion bridge","ADOBI: adaptive diffusion bridge yields fast blind MRI reconstruction","Adaptive diffusion bridge achieves blind MRI reconstruction with 5-10 steps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pretrained bridge was trained on intermediate states that are convex combinations of a clean image and a fixed initialization, but at inference the data-consistency step rewrites those states using adaptively updated coil sensitivities; if these edited states drift outside the training distribution, the bridge's MMSE estimate degrades and the reported gains shrink.","fun_headline_variants_meta":{"raw":{"variants":["Blind inverse problems solved by adaptive diffusion bridge in 10 steps","Adaptive bridge tunes MRI coils to enforce measurement consistency","Blind MRI reconstruction in 10 steps via adaptive diffusion bridge","ADOBI: adaptive diffusion bridge yields fast blind MRI reconstruction","Adaptive diffusion bridge achieves blind MRI reconstruction with 5-10 steps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001072,"raw_usage":{"total_tokens":4484,"prompt_tokens":931,"completion_tokens":3553,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":547,"completion_tokens_details":{"reasoning_tokens":3467}},"tokens_in":547,"tokens_out":3553,"duration_ms":22497,"temperature":1.0,"reasoning_tokens":3467,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:59:06.490437+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run ADOBI on a set of slices with intentionally corrupted initial coil sensitivity maps, for example by shifting or scaling the ESPIRiT maps, and check whether the adaptive CSM update still drives the data-fidelity term to zero and preserves PSNR; a large drop would indicate the method relies on a good initialization rather than on the bridge dynamics. A more direct test is to compute a distribution-distance metric, such as FID or MMD, between the edited intermediate states produced after Eq. (11) and the states the bridge saw during training, and to check whether that distance grows with the number of steps and correlates with the reported performance drop.","supporting_citations":[{"cited_title":"Direct diffusion bridge using data consistency for inverse problems,","cited_arxiv_id":null,"evidence_quote":"CDDB establishes the consistent diffusion bridge with data consistency for known forward models, which ADOBI extends to blind inverse problems."},{"cited_title":"I2sb: image-to-image schr¨ odinger bridge,","cited_arxiv_id":null,"evidence_quote":"I2SB provides the image-to-image Schrödinger bridge backbone and training recipe that ADOBI reuses with zero-filled or GRAPPA initialization."},{"cited_title":"Generalized autocalibrating partially parallel acquisitions (GRAPPA),","cited_arxiv_id":null,"evidence_quote":"GRAPPA is used to generate the higher-quality starting distribution for the bridge and is shown to improve ADOBI's reconstructions."},{"cited_title":"ESPIRiT- an eigenvalue approach to autocalibrating parallel MRI: Where SENSE meets GRAPPA,","cited_arxiv_id":null,"evidence_quote":"ESPIRiT estimates the initial coil sensitivity maps that anchor the adaptive CSM update through the Tikhonov penalty in Eq. (16)."},{"cited_title":"End-to-end variational networks for accelerated MRI reconstruction,","cited_arxiv_id":null,"evidence_quote":"E2E-VarNet is the strongest model-based deep-learning baseline for parallel MRI that ADOBI outperforms, providing a key comparison for learned CSM estimation."},{"cited_title":"Decomposed diffusion sampler for accelerating large-scale inverse problems,","cited_arxiv_id":null,"evidence_quote":"DDS is a diffusion-model baseline that ADOBI surpasses while using 10 steps instead of 100, serving as a speed and quality comparison."},{"cited_title":"Gibbsddrm: A partially collapsed gibbs sampler for solving blind inverse problems with denoising diffusion restoration,","cited_arxiv_id":null,"evidence_quote":"GibbsDDRM is the blind diffusion baseline that jointly estimates image and forward model, which ADOBI beats in the reported experiments."},{"cited_title":"Diffusion posterior sampling for general noisy inverse problems,","cited_arxiv_id":null,"evidence_quote":"DPS supplies the measurement-consistency posterior sampling update for non-blind diffusion models and is one of the baselines ADOBI is compared against."}],"review_version":1}