{"id":"343dbc5c-aeb6-4e13-890e-f3bd9b899adc","arxiv_id":"2505.18365","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"BRITE uses a diffusion model to disentangle anatomy from fading tags and a physics-informed network to track motion in tagged MRI, with phantom evidence of improved accuracy.","lead":"BRITE, a new framework, separates the underlying anatomy from fading tag stripes in tagged MRI and estimates tissue motion from the clean decomposition. In gel-phantom tests with simulated deformations it reports better motion and strain accuracy than five established methods, but no code or in vivo validation is included.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"BRITE's brightness invariance is only global; local brightness changes can be absorbed as spurious motion and are not exercised by the homogeneous phantom experiments.","rationale":"The reader identified the material-constancy assumption as the weakest assumption, and I agree that it is the most load-bearing condition for the paper's broader implication. The narrow phantom-specific claim is plausible and internally coherent: for a homogeneous gel, the tag amplitude and offset really are global functions of time, so the two-scalar fading model is correct. But the paper's title and framing claim brightness invariance, and the reader's stated implication generalizes to tagged MRI tracking tolerating brightness changes without extra acquisitions. That generalization is not supported because the model has no mechanism for spatially varying brightness, and the experiments contain no spatially varying brightness. The concern is not that the forward model is wrong in the phantom regime; it is that the central claim's reach exceeds the tested regime, and the missing test is precisely a spatially nonuniform fading experiment. I therefore recommend keeping the reader's CONDITIONAL verdict rather than accepting or rejecting outright: the method should be conditionally accepted only after the local-brightness sensitivity is checked. The agreement is full because the reader's weakest assumption is the same concern, though I would frame it as a missing stress test rather than an internal inconsistency.","tokens_in":10937,"tokens_out":12893,"duration_ms":127342,"concrete_test":"Generate a tagged sequence from the same phantom forward model with known ground-truth B-spline deformation, but make tag fading spatially varying, e.g., A_t(x,y)=A_global(t)(1+\\alpha R(x,y)) with R a smooth random field, for \\alpha = 0, 0.2, and 0.5 while keeping SNR, tag period, and flip angle in the tested range (e.g., FA=5°, TP=18 mm). Run BRITE on all three cases and report EPE and eMPS as functions of \\alpha. If EPE rises monotonically with \\alpha and the error maps localize to regions of high brightness gradient, the global-fading assumption is the bottleneck; if EPE is unchanged for \\alpha up to 0.5, the concern is refuted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is that brightness evolution is captured by two global scalars per frame, A_t and B_t, in the Tag Fading Module (Eqs. 5-6). The gel phantom is homogeneous, so its T1-driven tag fading is spatially uniform and exactly matches this model. Real tissue has local T1, inflow, and off-resonance effects, so the true brightness change is spatially varying. Since BRITE fits motion by minimizing pixel-wise MSE between \\tilde g_t = \\tilde a_t \\otimes \\tilde p_t and the observed g_t, a local brightening or darkening that is not in the parametric tag model can be explained by locally warping \\tilde a_t and \\tilde p_t instead of by the two global scalars. This produces spurious deformation and strain. The phantom experiments never exercise this failure mode, so they support 'resistant to global tag fading' but not 'brightness-invariant' in the general sense implied by the title and by the conclusion that tagged MRI tracking can tolerate brightness changes without extra acquisitions. This is an external-generalization limitation rather than an internal inconsistency: under the global-fading model the formulation is coherent, but the central implication rests on the unverified assumption that global fading is the only relevant brightness variation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces BRITE, a two-stage tracking method for SPAMM-tagged MRI. In the first stage, the method disentangles the anatomical image from the horizontal and vertical tag patterns at the reference frame by jointly optimizing a DDPM latent code and a CNN-predicted sinusoidal tag model (Eq. 1). In the second stage, for each subsequent frame a PINN produces a stationary velocity field whose exponential map defines a diffeomorphic Lagrangian deformation, while a Tag Fading Module estimates two global scalars A_t and B_t per frame; all unknowns are fit by minimizing the reconstruction MSE in Eq. (6). The method is evaluated on a silicone gel phantom with eight SPAMM acquisitions (tag periods 9, 12, 18, 26 mm; flip angles 5 and 10 degrees) under simulated non-rigid deformations (N=20), one rigid rotation, and a static case, comparing EPE and eMPS against HARP, SinMod, DRIMET, SyN variants, and DeepTag. The paper claims that BRITE is more accurate than these baselines and is resistant to tag fading.","tokens_in":11176,"tokens_out":8914,"duration_ms":74912,"significance":"If the claims survive closer quantitative scrutiny, BRITE would be a useful contribution: it attacks tag fading and spectral overlap simultaneously, requires no extra acquisitions beyond the standard horizontal/vertical SPAMM pair, and the validation design has real merit because the phantom images carry natural T1-driven tag fading while the ground-truth deformations come from external simulation. The runtime and memory figures are modest. The main evidence, however, is global-fading-only phantom data, and the quantitative support is incomplete (no error bars or significance tests, and no numbers for the rotation/static cases), so the strength of the central claim is currently larger than the evidence.","major_comments":[{"comment":"The forward model represents brightness evolution by only two global scalars per frame, A_t and B_t, in the Tag Fading Module. The phantom is a homogeneous gel, so its T1-driven fading is spatially uniform and exactly matches this model. Real tissue has spatially varying T1, B1 inhomogeneity, inflow, and off-resonance; none of these are exercised by the experiments, and the manuscript's Limitations section does not discuss them. Under the pixel-wise MSE objective of Eq. (6), a spatially varying brightness change could be absorbed as spurious deformation, which is precisely the failure mode the title promises to avoid. Please either narrow the claim to global tag fading or add a validation with spatially varying brightness modulation (e.g., synthetic local intensity scaling on the acquired phantom data) and report EPE/eMPS for that case.","section":"§3, Eqs. (4)–(6); §5, Limitations"},{"comment":"The main quantitative evidence is reported only as curves without error bars, confidence intervals, or significance tests across the 20 non-rigid deformations. The claim that BRITE 'generally outperforms' all other methods cannot be verified statistically from the figures. Please provide per-sequence results, summary statistics with error bars, and, if appropriate, paired tests (e.g., Wilcoxon) on EPE and eMPS.","section":"§5, Fig. 5"},{"comment":"The rigid-rotation and static no-motion experiments are central to the tag-fading-resistance claim, yet the text states that 'quantitative results are omitted due to limited space.' In particular, the static case directly measures spurious motion caused by fading (cf. Fig. 2), so omitting its numbers weakens the main claim. Please include EPE and eMPS values for the rotation and static scenarios, alongside the qualitative images.","section":"§5, Fig. 7"}],"minor_comments":[{"comment":"Please specify how MPS is computed from the estimated displacement field (e.g., finite-difference strain and Lagrangian strain tensor), since eMPS is a central metric.","section":"§4, Evaluation Metrics"},{"comment":"Please report the B-spline control-point spacing and displacement magnitudes used to generate the 20 non-rigid deformations, as these parameters determine the difficulty of the test.","section":"§4, Datasets"},{"comment":"The text says 'a multilayer perception' (should be 'perceptron'), and the notation for the tag-pattern CNN and the PINN weights should be made consistent (f_{\\theta_{tag}} versus f_{\\theta_{pinn,t}}).","section":"§3, Implementation Details"},{"comment":"The figure captions contain garbled strings such as '10-12mm-15-18mm-2' and the legends in Fig. 5 are difficult to read; please clean up all figures and make the per-method plotting conventions explicit.","section":"§5, Figs. 5–7"},{"comment":"No code or data availability statement is present; please add one or state that the phantom data can be shared on request.","section":"Overall"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope and the method is original, but the title's 'brightness-invariant' promise is broader than what the homogeneous-phantom experiments can support. I would ask the authors to either restrict the claim or add the spatially varying brightness experiment, and to make the quantitative results for the rigid/static cases available. The missing error bars and deformation-simulation details are straightforward to fix."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The combination is genuinely new: a diffusion prior for anatomical disentanglement, a PINN for Lagrangian motion, and a per-frame tag-fading module with global amplitude and offset scalars. The phantom study is thoughtful, with several tag periods and flip angles, and the comparison set (HARP, SinMod, DRIMET, SyN, DeepTag) is appropriate. They tuned baselines and implemented SinMod themselves, which is more than most methods papers do. Under the global-fading model, the formulation is coherent and the results look credible.\n\nThe main soft spot is the gap between the title and the evidence. The forward model only accounts for global brightness changes via two scalars per frame, A_t and B_t. The gel phantom is homogeneous, so its T1-driven fading is spatially uniform and matches that model exactly. Real tissue has local T1 variation, inflow, and off-resonance effects; a local brightening or darkening that is not in the parametric tag model can be explained by locally warping the anatomy and tags instead, producing spurious displacement and strain. The phantom experiments never exercise this failure mode, so the results support 'resistant to global tag fading' but not 'brightness-invariant' in the general sense. The paper's own limitations section only mentions the sinusoidal tag model, not this local-brightness assumption.\n\nTwo smaller issues: no code or data are released, and the quantitative results appear only as figures with no error bars or significance tests across the 20 deformations. I cannot tell whether BRITE's edge over SinMod is meaningful. The DDPM is pretrained on synthetic ovals; that works for the phantom but adds uncertainty for real anatomy.\n\nFor readers in tagged MRI motion tracking, this is a useful contribution with a clear formulation. It deserves a serious referee, but the claims need to be scoped and the experiments need stronger reproducibility plus a test with spatially varying brightness, for instance an inhomogeneous phantom or simulated local intensity changes. I would send it out.","headline":"BRITE is a well-built tracking method, but 'brightness-invariant' is only shown for global brightness changes; local variations are not modeled or tested.","tokens_in":11780,"tokens_out":2197,"would_cite":false,"duration_ms":20773,"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":"BRITE separates anatomy from fading tags, yielding accurate motion and strain estimates in tagged MRI.","keywords":["MR tagging","tagged MRI","motion tracking","tag fading","spectral overlap","strain estimation","diffusion model","physics-informed neural network"],"falsifier":"Run BRITE on a tagged sequence with known deformation in which part of the field of view undergoes a local brightness change (for instance, a spatially varying T1 or flip-angle map, or an inflow of contrast agent) while the rest stays uniform; if the estimated displacement near that local region deviates from ground truth but is accurate elsewhere, the global-fading, constant-anatomy assumption is the reason.","tokens_in":10725,"feed_emoji":"🧲","tokens_out":18658,"duration_ms":134792,"temperature":0.7,"pith_summary":"Tagged MRI marks tissue with a striped saturation pattern that moves with the tissue, but the pattern fades and overall brightness drifts as the magnetization recovers, so both intensity-based optical flow and Fourier methods lose accuracy. The paper introduces BRITE, which treats each observed tagged frame as the product of an unknown anatomical image and a sinusoidal tag pattern, estimating both together with the motion of each material point from the first frame (the Lagrangian motion). Tag fading is absorbed by two per-frame scalars, the tag amplitude and offset, rather than being allowed to corrupt the motion estimate. On SPAMM-tagged gel phantom data with varied tag periods and flip angles, the paper reports that BRITE gives more accurate motion and strain estimates than HARP, SinMod, DRIMET, SyN variants, and DeepTag, and remains accurate as tags fade. If this holds, motion tracking from tagged MRI can tolerate brightness changes without the extra acquisitions used by methods such as CSPAMM and TruHARP.","feed_headline":"New MRI tracker stays accurate as tags fade","feed_subtitle":"BRITE separates anatomy from the fading tag pattern, beating HARP and others on displacement and strain.","key_machinery":"The load-bearing object is the factorized forward model $g^{h,v}_t = \\tilde a_t \\otimes \\tilde p^{h,v}_t$, where $\\tilde p^{h,v}_t = [A_t \\sin(2\\pi \\tilde\\mu s_{h,v} + \\tilde\\varphi^{h,v}) + B_t] \\circ \\phi_t^{-1}$, with $s_h=x$ and $s_v=y$. This factorization lets brightness change be absorbed by the two scalar parameters $A_t$ and $B_t$ instead of leaking into the motion field. The second essential piece is the deformation representation: a physics-informed neural network maps coordinates $(x,y)$ to a stationary velocity field $u_t$, and scaling-and-squaring exponential integration converts $u_t$ into the smooth, invertible deformation $\\phi_t$ and its inverse. Because $\\phi_t^{-1}$ warps both the anatomy and the tag pattern into the current frame, the same motion explains both image components, which is why no explicit smoothness penalty is needed.","core_discovery":"The central claim is that the ill-posed problem of jointly recovering the underlying anatomy, the fading tag pattern, and the Lagrangian motion (motion of each material point relative to the first frame) from a tagged MRI sequence has a tractable solution when the forward model is factored as $g^{h,v}_t = \\tilde a_t \\otimes \\tilde p^{h,v}_t$. The anatomy is required to be constant in material coordinates, $\\tilde a_t = \\tilde a_0 \\circ \\phi_t^{-1}$, while the tag pattern is a sinusoid whose frequency and phase are fixed from the first frame and whose amplitude $A_t$ and offset $B_t$ are the only time-varying quantities. A diffusion model pretrained on anatomical images provides the prior that makes recovering $\\tilde a_0$ from the product well-posed, and a physics-informed neural network outputs a stationary velocity field whose exponential map produces the deformation $\\phi_t$, a smooth, invertible warping. The paper reports that on gel phantom data, this disentanglement yields more accurate displacement and strain estimates than the compared methods, with the advantage largest when tag fading and spectral overlap are severe.","pith_inferences":["A test the paper does not run is a spatially nonuniform brightness change, since its fading model has only global $A_t$ and $B_t$; one would expect local T1 variation or inflow to be absorbed as apparent deformation, which would set the boundary of the method's applicability.","The diffusion prior is trained only on synthetic ovals, and the paper's own limitations section restricts the tag model to sinusoids and validation to a 2D static gel phantom with simulated deformations; applying the method to textured human organs with independent motion measurements will show how much the prior and tag model must be generalized.","The same factorization could be carried over to other amplitude-modulated tagging schemes, such as grid tags or higher-order SPAMM, by swapping the sinusoidal tag model while keeping the PINN tracking stage.","A direct comparison against CSPAMM and TruHARP on the same phantom with the same ground truth would quantify how much accuracy is gained or lost by avoiding their extra acquisitions."],"forward_implications":["If BRITE's reported accuracy holds, tagged MRI motion tracking can be made brightness-invariant without extra acquisitions, removing a major practical argument for CSPAMM or TruHARP.","Because the tag frequency is estimated from the data rather than assumed from the pulse sequence, the method can tolerate gradient imperfections that shift or blur the spectral peaks.","Initializing each frame's velocity-field network from the previous frame's solution lets the tracker follow motion larger than half a tag period without tag-jumping, and avoids error accumulation from composing pairwise registrations.","Resistance to tag fading should extend the usable time window of a single tagged acquisition, exactly where HARP and intensity-based methods degrade most.","The reported processing time of about 8.4 seconds per frame with modest GPU memory suggests 2D clinical workflows are feasible today, and the paper argues 3D extension is straightforward."],"supporting_citations":[{"why":"Defines HARP, the Fourier harmonic-phase method that is the principal baseline and the main source of spectral-overlap sensitivity BRITE addresses.","marker":"[21]"},{"why":"Introduces SinMod, a skewed band-pass filtering baseline whose accuracy on large tag periods shows the spectral-overlap problem.","marker":"[5]"},{"why":"Proposes variable brightness optical flow for tagged MRI, the prior image-domain approach BRITE extends by not requiring T1, T2, or spin density.","marker":"[24]"},{"why":"Supplies compressed sensing with generative models, the framework that motivates using a pretrained diffusion model as the anatomy prior.","marker":"[10]"},{"why":"Introduces physics-informed neural networks, which BRITE adapts to represent the stationary velocity field for motion.","marker":"[27]"},{"why":"Provides the fast exponential (scaling-and-squaring) mapping used to convert the velocity field into a diffeomorphic deformation and its inverse.","marker":"[4]"},{"why":"Presents DeepTag, the unsupervised deep-learning baseline that BRITE avoids by not needing in-domain tagged training data.","marker":"[31]"},{"why":"Analyzes why raw-tagged registration fails under tag fading, the specific failure mode BRITE is designed to overcome.","marker":"[8]"}],"fun_headline_variants":["Brightness-invariant MRI tag tracking survives fading","BRITE: separating anatomy from tags sharpens motion estimates","Tag fading no match for diffusion-physics MRI tracker","AI untangles MRI tags and tissue for accurate strain","Physics-informed AI keeps MRI tag tracking accurate"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's forward model assumes the anatomy is exactly carried along by the motion, with only a single global amplitude and offset changing in the tag pattern; if real tissue brightness changes locally rather than globally, the optimizer can translate that brightness change into a spurious apparent deformation.","fun_headline_variants_meta":{"raw":{"variants":["Brightness-invariant MRI tag tracking survives fading","BRITE: separating anatomy from tags sharpens motion estimates","Tag fading no match for diffusion-physics MRI tracker","AI untangles MRI tags and tissue for accurate strain","Physics-informed AI keeps MRI tag tracking accurate"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000197,"raw_usage":{"total_tokens":1389,"prompt_tokens":991,"completion_tokens":398,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":324}},"tokens_in":607,"tokens_out":398,"duration_ms":4338,"temperature":1.0,"reasoning_tokens":324,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:32:23.111778+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run BRITE on a tagged sequence with known deformation in which part of the field of view undergoes a local brightness change (for instance, a spatially varying T1 or flip-angle map, or an inflow of contrast agent) while the rest stays uniform; if the estimated displacement near that local region deviates from ground truth but is accurate elsewhere, the global-fading, constant-anatomy assumption is the reason.","supporting_citations":[{"cited_title":"Magnetic Resonance in Medicine42(6), 1048–1060 (1999)","cited_arxiv_id":null,"evidence_quote":"Defines HARP, the Fourier harmonic-phase method that is the principal baseline and the main source of spectral-overlap sensitivity BRITE addresses."},{"cited_title":"IEEE Transactions on Medical Imaging29(5), 1114–1123 (2010)","cited_arxiv_id":null,"evidence_quote":"Introduces SinMod, a skewed band-pass filtering baseline whose accuracy on large tag periods shows the spectral-overlap problem."},{"cited_title":"IEEE Transactions on Medical Imaging11(2), 238–249 (1992)","cited_arxiv_id":null,"evidence_quote":"Proposes variable brightness optical flow for tagged MRI, the prior image-domain approach BRITE extends by not requiring T1, T2, or spin density."},{"cited_title":"In: International Conference on Machine Learning","cited_arxiv_id":null,"evidence_quote":"Supplies compressed sensing with generative models, the framework that motivates using a pretrained diffusion model as the anatomy prior."},{"cited_title":"Journal of Computational Physics378, 686–707 (2019)","cited_arxiv_id":null,"evidence_quote":"Introduces physics-informed neural networks, which BRITE adapts to represent the stationary velocity field for motion."},{"cited_title":"In: MICCAI","cited_arxiv_id":null,"evidence_quote":"Provides the fast exponential (scaling-and-squaring) mapping used to convert the velocity field into a diffeomorphic deformation and its inverse."},{"cited_title":"In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR)","cited_arxiv_id":null,"evidence_quote":"Presents DeepTag, the unsupervised deep-learning baseline that BRITE avoids by not needing in-domain tagged training data."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Analyzes why raw-tagged registration fails under tag fading, the specific failure mode BRITE is designed to overcome."}],"review_version":1}