{"id":"245f331f-aafc-43fe-94a8-d3444cb17992","arxiv_id":"1908.05285","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A joint variational model with non-convex Bregman iteration reconstructs magnitude, velocity phase, and segmentation from undersampled velocity-encoded MRI, outperforming a sequential baseline on synthetic and real bubbly flow data.","lead":"This paper presents a joint optimization algorithm that reconstructs image magnitude, velocity phase, and segmentation simultaneously from undersampled velocity-encoded MRI data. The approach targets faster and more accurate flow imaging for medical and industrial bubbly flow applications.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1 reports MSE for only two of the four phases and no MSE for the velocity map, so the synthetic evidence does not establish the claimed velocity improvement.","rationale":"The paper's stated contribution is a joint method that improves velocity, magnitude and segmentation over a sequential approach. The only quantitative evidence for this is Table 1, which reporting MSE averages for two magnitudes and two phases out of the four unknowns in the model. Because the velocity is a difference of four phases, the reported phase MSEs are insufficient to infer a velocity improvement. This is a concrete, internal gap in the argument, separate from the reader's physical-model concern. The reader's concern about incomplete phase cancellation is real but would affect both the joint and sequential reconstructions similarly, so it is not the decisive issue for the comparative claim. A direct velocity MSE computation would settle whether the headline claim holds. The verdict remains CONDITIONAL: the manuscript is promising but needs this quantitative validation and the other details already noted by the reader before acceptance. No rejection is warranted because the missing evidence is obtainable and the approach is plausible.","tokens_in":13072,"tokens_out":12542,"duration_ms":130801,"concrete_test":"Re-run the synthetic 32-frame experiment of Section 5.1 with the same 11% undersampling and compute, per frame and averaged over frames, the MSE of the reconstructed velocity v = 1/2((φ1 − φ2) − (φ3 − φ4)) against the true velocity for both the joint and sequential methods. Also report per-frame MSE for all four u_j and φ_j, with mean and standard deviation. If the joint method's velocity MSE is not lower than the sequential baseline on the majority of frames, or if the error bars overlap, the central claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the joint model (Eq. 20) improves velocity, magnitude and segmentation over a sequential baseline. The only quantitative support is the synthetic 32-frame experiment, for which Table 1 reports average MSE for u1, u2, φ1, φ2 only. But the model in Eq. (20) reconstructs four images (j = 1..4), and the velocity is defined via Eq. (12) as a function of all four phases, v = 1/2((φ1 − φ2) − (φ3 − φ4)). No MSE is reported for v, nor for u3, u4, φ3, φ4. Improvement in φ1 and φ2 does not imply improvement in the phase difference that constitutes the velocity; the omitted terms could offset or dominate the gain. The real-data evaluation is visual only, and the reader's physical-model concern about phase cancellation would affect both methods similarly, so it does not by itself undermine the comparative claim. The missing velocity metric is the most load-bearing gap: without it, the paper has not actually demonstrated its headline improvement in velocity.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a joint variational approach for reconstructing magnitude, phase, and segmentation from undersampled velocity-encoded MRI. It models the acquired k-space signal as A(u e^{i\\phi}) and defines velocity as a phase difference over four acquisitions (two flow polarities and two zero-flow references). The authors introduce a non-convex Bregman iteration algorithm (Eq. 21) with TV regularization on magnitude/segmentation and H1-type smoothing on the phase-difference velocity. They report synthetic experiments on a rising-bubble phantom (32 frames, 11% sampling) comparing a joint model to a sequential CS-TV approach, and visual results on real bubble-burst data. The central claim is that the joint model improves velocity, magnitude, and segmentation.","tokens_in":13271,"tokens_out":6056,"duration_ms":59077,"significance":"If the claimed improvements hold, the approach offers a practically useful way to exploit structure in undersampled velocity-encoded MRI, and the extension of non-convex Bregman iteration to velocity-regularized phase reconstruction is methodologically interesting. The paper includes a physically motivated forward model and a concrete algorithmic scheme with synthetic and real-data demonstrations. However, the quantitative support for the headline velocity claim is incomplete: Table 1 reports MSE for only two of the four phases and two magnitudes, with no velocity MSE, no variance estimates, and no parameter values. The real-data evaluation is visual only. The significance is therefore conditional on substantial additional quantitative evidence.","major_comments":[{"comment":"The central claim that the joint method improves velocity is not quantitatively demonstrated, because Table 1 reports averaged MSE only for u1, u2, phi1, and phi2; no MSE is given for the velocity field defined in Eq. (12) as v = 1/2((phi1 - phi2) - (phi3 - phi4)), nor for the omitted u3, u4, phi3, and phi4. Please add per-frame velocity MSE, and ideally the full set of magnitude and phase errors, with standard deviations or confidence intervals; without this, the abstract's \"improves velocity\" assertion is unsupported by the synthetic experiment.","section":"Section 5.1, Table 1"},{"comment":"The numerical results are not reproducible because no parameter values are reported: alpha, beta, delta, eta, tau, c1, c2, the early-stopping iteration number, and the PDHG inner-loop settings are all unspecified. Please provide these values and either the tuning procedure or a sensitivity analysis; if space permits, releasing code would address this fully.","section":"Sections 4 and 5"},{"comment":"The sequential baseline is described only as a \"classic CS-TV-regularised approach\" followed by the phase method of [37] and [44], which are the authors' own prior works; no algorithmic details, regularization weights, or stopping rules are given. Because every quantitative comparison in the paper is against this baseline, a substantial description (or a reference to a standardized implementation) is needed to rule out that the reported gains are due to a weak or undertuned comparator.","section":"Section 5.1, baseline description"},{"comment":"The velocity interpretation depends on Eq. (12), which assumes exact cancellation of background, off-resonance, and eddy-current phase between flow-on and zero-flow acquisitions; the paper provides no validation of this cancellation on the real data, where no independent velocity ground truth exists. The comparative claim may survive this concern (both methods use the same phase combination), but the absolute velocity accuracy on real data is not established and should be stated as a limitation.","section":"Sections 2.3, 2.4, and 5.2"}],"minor_comments":[{"comment":"The fidelity term uses phi_i while the sum is over j; it should be phi_j.","section":"Eq. (20)"},{"comment":"Eq. (20) is presented as the joint cost, but the TV and H1 regularizers enter only through the Bregman distances in Eq. (21); please state explicitly that the overall model is E + J_u + J_v + J_phi and that Eq. (21) is an iterative regularization scheme for it.","section":"Eq. (20) and Eq. (21)"},{"comment":"The admissible set for the segmentation variable v (for example, v in [0,1] or binary) is not specified; please state it.","section":"Section 3.2"},{"comment":"The text says that visual differences are not significant while MSE improves; showing error maps would help the reader judge where the improvement occurs.","section":"Section 5.1, Fig. 1"},{"comment":"The caption contains a typo: \"Bottow\" should be \"Bottom\".","section":"Fig. 6 caption"},{"comment":"The figures would benefit from colorbars and explicit display windows for the velocity and phase maps.","section":"Figs. 3-6"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is in scope and the core idea is worth publishing if the quantitative gaps are filled. I would ask the authors to provide velocity MSE, full error metrics, parameter settings, and a clearer description of the baseline before further consideration. The self-cited baseline is not disqualifying by itself, but the paper should position it as one comparison rather than the only one."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThis paper does something genuinely new: it extends the authors' joint reconstruction-and-segmentation framework to velocity-encoded MRI by reconstructing four complex images (flow on/off, two polarities) and adding an H1 regularizer on the phase-difference combination that defines velocity. That is a sensible extension, clearly motivated by the physics, and the optimization via non-convex Bregman iteration is technically solid. The synthetic experiment at 11% sampling shows consistent MSE gains over the sequential baseline for the quantities it reports.\n\nThe soft spot is exactly what the stress-test flags: Table 1 only reports MSE for u1, u2, φ1, and φ2. It omits u3, u4, φ3, and φ4 and — critically — omits the velocity map v defined by Eq. (12). Since the paper's headline claim is that the joint model improves velocity, and velocity is a difference of phases, improvement in φ1 and φ2 does not establish improvement in v. The omitted terms could offset or dominate. So the central quantitative support for the velocity claim is missing. The real-data comparison is visual only, with no numbers. There are no error bars, and the regularization weights and c1, c2 are not specified. The sequential baseline is built from the authors' own prior methods, which is not disqualifying but does not help the comparison.\n\nNone of these issues look fatal. The modeling is thoughtful and the physical derivation in Section 2 is standard. The missing velocity metric is a fixable reporting gap; a revision that reports full MSE tables including velocity, adds error bars, specifies parameter choices, and ideally releases code/data would address the main concerns. I would send this to peer review rather than desk reject. A serious referee would ask for exactly those things, but the paper is coherent and the new coupling is worth evaluating.\n\nWho is this for: people working on undersampled phase-contrast or velocity-encoded MRI, especially in flow imaging and joint reconstruction-segmentation methods. I would not cite it in its current form until the velocity improvement is quantified, but I would want to see the revision.","headline":"A coherent extension of joint reconstruction-and-segmentation to velocity-encoded MRI, but the headline claim that joint reconstruction improves velocity is not actually supported by the reported quantitative results.","tokens_in":13821,"tokens_out":2154,"would_cite":false,"duration_ms":21162,"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":"Joint MRI reconstruction cuts velocity error versus sequential methods.","keywords":["velocity-encoded MRI","phase reconstruction","magnitude segmentation","joint variational model","Bregman iteration","non-convex optimization","total variation regularization","bubbly flow imaging"],"falsifier":"Take a straight-tube phantom with known steady flow and acquire the four-measurement velocity-encoded sequence twice, changing only the background magnetic field between runs. If the joint method returns different velocity maps while the physical flow is unchanged, the zero-flow cancellation of Eq. (12) is not complete, and the claimed improvement is in phase consistency rather than true velocity measurement.","tokens_in":12844,"feed_emoji":"🧲","tokens_out":11965,"duration_ms":118104,"temperature":0.7,"pith_summary":"Velocity-encoded MRI is magnetic resonance imaging that encodes flow motion into the phase of the measured signal, but recovering that phase from undersampled data is a nonlinear inverse problem, and conventional pipelines reconstruct magnitude and phase in separate steps. This paper argues that those steps should be solved together: one non-convex optimization that recovers the magnitude image, the motion-encoding phase, and a two-region segmentation of the image from the four measurements of a velocity-encoded acquisition. The payoff is that edges learned for segmentation sharpen the magnitude, which in turn improves the phase that becomes the velocity estimate. On synthetic bubble data the joint model lowers mean-squared error for phases and magnitudes relative to the sequential baseline, and on real bursting-bubble data it gives noise-reduced velocity fields with sharper fluid/air boundaries.","feed_headline":"Joint MRI model cuts velocity error versus sequential methods","feed_subtitle":"Combining magnitude, phase, and segmentation in one optimization improves velocity maps on synthetic and real bubbly-flow data.","key_machinery":"The central object is the joint variational energy of Eq. (20), which couples a Fourier-domain data-fidelity term $\\tfrac{1}{2}\\|A(u_j e^{i\\varphi_j})-f_j\\|_2^2$ for each of the four acquisitions with a two-region segmentation term $\\delta\\sum_n [v_{nj}(c_1-u_{nj})^2+(1-v_{nj})(c_2-u_{nj})^2]$ on the magnitude images. The optimizer is an alternating Bregman iteration in which total variation (a penalty on local intensity changes) regularizes magnitude and segmentation, while an $H^1$-type penalty smooths the phase difference, so the velocity—recovered afterwards from Eq. (12)—is regularized directly without assuming known phase structure. The Bregman distance, the difference between a functional and its linearization at a previous iterate, is what lets the scheme handle non-smooth and non-convex terms in an inverse-scale-space manner.","core_discovery":"The paper claims that the four measurements of a velocity-encoded MRI acquisition—two gradient polarities in the flow-on state plus their zero-flow references—can be inverted as one joint problem rather than as separate magnitude and phase reconstructions. The proposed model is the non-convex energy of Eq. (20), in which each magnitude image $u_j$ and phase map $\\varphi_j$ must explain the undersampled k-space data through $A(u_j e^{i\\varphi_j})$, while a binary segmentation $v_j$ pulls the magnitudes toward two intensity levels $c_1$ and $c_2$. Alternating Bregman iteration with total variation on magnitudes and segmentations and an $H^1$ penalty on the phase difference solves the system. Afterward the velocity component is computed from the four recovered phases via Eq. (12). The paper reports that this joint approach lowers mean-squared error for magnitudes and phases relative to the sequential baseline on simulated bubble-rise data and produces visually cleaner velocity fields, magnitudes, and segmentations on real bursting-bubble data.","pith_inferences":["Editorial inference: the largest joint-versus-sequential differences should occur at bubble edges, because that is where the segmentation term changes the reconstruction; the paper reports global MSE only, so a per-pixel error map on the synthetic data would make this prediction testable.","Editorial inference: the method reconstructs each time frame independently, and the paper notes that the lack of a 4D dataset prevents joint space-time reconstruction; a simulated dynamic phantom with known time-varying flow could test the expected gains before new real acquisitions.","Editorial inference: the fixed two-level segmentation constants tie the model to bimodal images; generalizing to multiple regions or data-driven intensity levels would extend the coupling idea to medical settings such as cardiac blood flow, where the same joint phase-magnitude structure holds."],"forward_implications":["At 11 percent k-space sampling on simulated data, the joint approach lowers mean-squared error for both magnitude images and phase maps compared with the sequential baseline, so the same reconstruction quality becomes available at lower sampling fractions.","Because the model regularizes the phase difference rather than each phase separately, smoothness is imposed directly on the quantity that becomes velocity.","On real bubble-burst data, the joint approach produces noise-reduced velocity fields and sharp fluid/air boundaries in magnitude and segmentation, compared with zero-filling and the sequential pipeline.","The four-measurement model, with velocity recovered from Eq. (12), applies to any single-component velocity-encoded MRI acquisition, not only bubbly flows.","Early stopping of the Bregman iteration acts as iterative regularization, so the algorithm does not need a global minimizer of the non-convex energy to produce useful reconstructions."],"supporting_citations":[{"why":"It supplies the joint reconstruction-and-segmentation model with non-convex Bregman iteration that this work adapts to velocity-encoded MRI.","marker":"[15]"},{"why":"It provides the sequential phase-estimation method used as the comparison baseline and one component of the hybrid Bregman scheme.","marker":"[37]"},{"why":"It provides the real bubbly-flow data, acquisition protocol, and the sequential pipeline to which the joint results are compared.","marker":"[44]"},{"why":"It supplies the standard phase-encoded velocity imaging formalism including the zero-flow compensation relation used to turn phase differences into velocities.","marker":"[16]"},{"why":"They define the generalized Bregman distance on which the alternating optimization algorithm is built.","marker":"[35, 36]"}],"fun_headline_variants":["Joint MRI inversion cuts velocity error over sequential pipeline","One joint solve for velocity MRI beats stepwise reconstruction","Joint magnitude and phase model improves MRI velocity maps","Non-convex joint MRI approach outperforms sequential baseline"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole velocity estimate rests on the assumption that subtracting the zero-flow reference from the two opposite-polarity measurements cancels every background phase contribution exactly, leaving only motion-induced phase; if any field imperfection fails to cancel, the reconstructed phase differences are not the true velocities.","fun_headline_variants_meta":{"raw":{"variants":["Joint MRI inversion cuts velocity error over sequential pipeline","One joint solve for velocity MRI beats stepwise reconstruction","Joint magnitude and phase model improves MRI velocity maps","Non-convex joint MRI approach outperforms sequential baseline"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000257,"raw_usage":{"total_tokens":1549,"prompt_tokens":886,"completion_tokens":663,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":502,"completion_tokens_details":{"reasoning_tokens":602}},"tokens_in":502,"tokens_out":663,"duration_ms":7382,"temperature":1.0,"reasoning_tokens":602,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:18:45.177035+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a straight-tube phantom with known steady flow and acquire the four-measurement velocity-encoded sequence twice, changing only the background magnetic field between runs. If the joint method returns different velocity maps while the physical flow is unchanged, the zero-flow cancellation of Eq. (12) is not complete, and the claimed improvement is in phase consistency rather than true velocity measurement.","supporting_citations":[{"cited_title":"Enhancingjointreconstructionandsegmentationwithnon-convexBregmaniteration","cited_arxiv_id":null,"evidence_quote":"It supplies the joint reconstruction-and-segmentation model with non-convex Bregman iteration that this work adapts to velocity-encoded MRI."},{"cited_title":"Signal sampling and processing inmagnetic resonance applications","cited_arxiv_id":null,"evidence_quote":"It provides the real bubbly-flow data, acquisition protocol, and the sequential pipeline to which the joint results are compared."},{"cited_title":"Velocity encoding and flow imaging","cited_arxiv_id":null,"evidence_quote":"It supplies the standard phase-encoded velocity imaging formalism including the zero-flow compensation relation used to turn phase differences into velocities."}],"review_version":1}