acceptodds
Desk rejected at ICLR 2027

Learning Continuous Echo Trajectories for Scan-Efficient Quantitative MRI

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Quantitative MRI derives tissue parameters from how signals evolve across echo time, yet acquiring these signal samples increases scan burden. Can a retained source image support a continuously queryable echo trajectory with an interpretable decay structure? We address this question through cross-repetition-time echo-train completion, synthesizing a long-TR target train from one short-TR source echo per flip angle. The Endpoint-Conditioned Trajectory Network modulates source-image features, while the Physics-Constrained Decoder uses a shared reference amplitude and integrated positive rates to turn queried features into nonnegative, nonincreasing signals. The rate represents local log-signal decay, connecting continuous image generation to an explicit signal constraint. The model learns from echo images alone and supports TE queries beyond the supervised echo coordinates without retraining. On a private healthy-brain cohort, interior and terminal held-out-TE predictions remain close to their fully supervised counterparts. At comparable echo-image fidelity, the Physics-Constrained Decoder improves downstream agreement with acquired-echo-fitted maps over the Unconstrained Decoder. The completion task targets omission of the long-TR block, corresponding to a nominal acceleration of the two-block repetition-time budget. These results connect continuous TE querying, interpretable decay, and acquisition-efficient synthesis within the evaluated protocol.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
quantitative MRI, echo-train completion, continuous representation learning, physics-constrained learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_JwPqOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies a quantitative MRI acquisition problem in which MR signal is measured at several different echo times (TEs). The signal decreases as TE increases, and this decay can subsequently be used to estimate the quantitative biomarkers. The proposed method keeps one source image acquired under a short repetition-time (TR) setting for each flip angle and predicts the corresponding set of multiple target images under a long-TR setting.

The model has two main components. First, an Endpoint-Conditioned Trajectory Network maps acquisition parameters (\alpha, TR, TE) and a continuous TE-related coordinate to a latent vector that modulates spatial features (by an image encoder) extracted from the source image. Second, a Physics-Constrained Decoder predicts a non-negative reference amplitude and a positive TE-dependent decay rate, and generates the target signal through exponential integration.

The method is evaluated on a private healthy-brain MULTIPLEX (MRx) dataset with 32 training and 10 evaluation subjects. The experiments include comparisons with existing MRI synthesis methods, several component ablations, prediction of acquired TE coordinates, off-grid TE querying within the target range, and downstream T2 (biomarker) estimation from the synthesized images.

Critical strengths
  1. The paper introduces a clean structural constraint that is directly relevant to downstream quantitative estimation by design. The positive-rate decoder guarantees non-negative and monotonically decreasing signal trajectories over TE by construction. I find this interesting since the unconstrained decoder in ablation achieves similar image-level fidelity but worse downstream T2 quantification.

  2. The experimental analysis is in-depth and relevant on a under-discovered problem. Quantitative MRI often requires parametric fitting, which combines information across multiple images. This requires more dedicated evaluation protocols. The acquired-TE hold-out experiments are well-designed completely excluded from image supervision while remaining available for final evaluation. The comparisons make the contribution of individual components relatively transparent and controllable.

Critical weaknesses
  1. The necessity of the proposed Endpoint-Conditioned Trajectory Network is not yet justified well. For subjects under the same protocol, the manuscript states that the latent path z(t)z(t) (conditioned on acquisition parameters) is shared, while the image encoder part takes all spatial information. The path is therefore mainly conditioned on the endpoints which are fixed within each FA group. This, somehow can degenerate the trajectory to a channelwise modulation on specific acquisition settings. It is then unclear why this requires the proposed endpoint-conditioned trajectory construction rather than a simpler continuous conditioning mechanism, such as FiLM or direct TE-conditioned feature modulation.

The Pointwise MLP ablation partially addresses this issue, but it changes the parameterization of acquisition conditioning and the use of endpoint representations as well. Therefore, it does not isolate whether the trajectory construction itself is responsible for the gain of performance.

This concern is closely related to the scope of the continuous-representation claim. The main experiments vary TE while keeping the acquisition protocol highly constrained. The manuscript explicitly notes that the held-out experiments do not evaluate unseen flip angles, TRs, or endpoint descriptors which should be of interest as well. The fully off-grid inferences are interesting, but still inside the target TE interval. I therefore think that the current evidence more clearly supports continuous in-range TE interpolation/resampling for a fixed acquisition setting than a general continuous acquisition-space representation where extrapolation is expected.

  1. The Physics-Constrained Decoder guarantees positivity and monotonic signal decay over TE by predicting a positive TE-dependent rate. However, an otherwise flexible positive rate can represent a broad class of monotonically decreasing signals. The manuscript itself explicitly states that rr is not a uniquely recovered tissue parameter, that a common relaxation parameter across flip angles is not enforced, and that cross-TR amplitude changes are learned from paired images. I therefore interpret this component mainly as a physics-motivated monotonic signal prior. The manuscript also acknowledges that a single source echo does not (and should not) uniquely determine an individual decay curve. For quantitative MRI, this raises the question of whether genuine subject-specific quantitative variation is preserved or whether the model primarily imputes cohort-level tissue statistics.
Questions for authors

Based on the weakeness, I would like to ask several questions that might help

  1. Could the authors better isolate the contribution of the Endpoint-Conditioned Trajectory Network from generic continuous feature conditioning? That could be, how would the method compare with a FiLM or direct TE-conditioned modulation baseline that retains the same source encoder and other setups? Since the endpoint descriptors are fixed within each flip-angle group in the current experiments, evidence that the proposed trajectory construction itself provides a benefit would substantially strengthen the representation-learning contribution.

  2. Can the authors provide more direct evidence that subject-specific quantitative variation is preserved than mainly inferred from cohort-level tissue priors? For example, within a fixed anatomical region, an across-subject correlation/ICC between T2 star estimated from acquired and synthesized echoes would be informative.

  3. For the off-grid TE set, how exactly was the T2∗T_2^* computed? Eq 20 uses a stride-two TE separation of 10.18 ms, which corresponds to the original grid, whereas the stride-two separation for the off-grid set is 8 ms. Was the factor in the T2∗T_2^* expression changed accordingly for the off-grid experiment? Also, how is the repeatability of the the mapping when generating more images off grid?

  4. Does the methodology works on other public datasets with quantitative mapping such as T1 mapping or on other organs? Since in this stage we can only infer from a private dataset on Brain MRI.

5 comments

Thank you for recognizing the explicit decay structure and the design of the hold-out experiments. Following your suggestions, we have added direct TE-conditioning and FiLM controls, across-subject agreement within fixed tissue regions, analyses of the actual off-grid separation and query density, and evaluation on public data. These additional results directly support the empirical benefits of the conditioning mechanism and individual predictive value at the level of tissue means in healthy subjects.

1. The contribution of endpoint conditioning relative to simpler continuous conditioning mechanisms

Within the same protocol and FA, the acquisition-conditioned path is shared across subjects, while encoded source-image features provide subject-specific spatial information. To directly address your proposed comparison, we added the following controls, retaining the source-encoder architecture, positive-rate decoder, training budget, and final-checkpoint rule, and training each model separately:

  • TE-Conditioned MLP: The conditioning network directly takes the continuous coordinate tt and FA identifier as input, without concatenating endpoint representations; source-feature factorization, multiplicative modulation, the boundary objective, and positive-rate integration are retained. The Full Model and this control each have 328,960 parameters in the conditioning module and 39,741,701 parameters in total.
  • FiLM: The same endpoint inputs and source-feature factorization are retained, while multiplicative modulation is replaced by scaling plus shifting; conditioning-module parameter counts are approximately matched.
Supervision settingFull Model MAE (ms)FiLM MAE (ms)TE-Conditioned MLP MAE (ms)
Full supervision14.12614.85915.094
TE=14.04 ms\mathrm{TE}=14.04\,\mathrm{ms} held out14.17315.02615.023
TE=24.22 ms\mathrm{TE}=24.22\,\mathrm{ms} held out14.27515.27915.134

The Full Model achieves lower mean joint T2∗T_2^* MAE in all three settings, as well as higher PSNR for both FAs. TE-MLP directly compares endpoint inputs with coordinate conditioning, while FiLM compares modulation forms under the same endpoint inputs. These additional results support the benefits of our endpoint-conditioned parameterization and multiplicative modulation relative to the alternative implementations tested; full implementation details and results are provided in Table 3 of the main text and Table 8 in Appendix B.1.

Here, trajectory refers to a conditioning function indexed by endpoints and a continuous coordinate. The shared path represents protocol-dependent TE variation, subject-specific information enters through source-image features, and positive-rate integration further couples outputs at different TEs into a single signal curve. This division of roles allows additional acquired inputs and continuous queries to be designed separately: a multi-input extension can fuse complementary observations at the encoder while continuing to use continuous conditioning and positive-rate integration to generate echoes. The current experiments directly compare the benefits of the conditioning mechanisms under a common single-source input budget.

2. Across-subject quantitative agreement within fixed tissue regions

Following your suggestion, we computed GM/WM tissue means separately for the ten test subjects, using the same fixed support, independent of predictions, for all methods and the acquired reference within each subject. Pearson rr and absolute-agreement ICC(A,1)\mathrm{ICC}(A,1) are reported below, with 95% CIs obtained by paired-subject bootstrap.

TissuePearson rr (95% CI)ICC(A,1)\mathrm{ICC}(A,1) (95% CI)Predicted/reference SD ratioRegression slope
GM0.654 (0.070–0.956)0.534 (0.035–0.828)0.6730.440
WM0.734 (0.502–0.941)0.596 (0.298–0.803)0.4790.352

Within-tissue association and absolute agreement support partial preservation of tissue-mean variation among healthy subjects. Relative to tissue constants determined from the 32 training subjects, GM/WM MAE decreases from 2.306/1.566 to 1.409/0.989 ms, with paired-improvement CIs of GM [−0.350,2.020][-0.350,2.020] and WM [0.120,1.043][0.120,1.043] ms. Compared with four external learning-based baselines, our method achieves the best point estimates for tissue-mean MAE, Pearson rr, and ICC. Complete tissue-mean, calibration, and per-subject results are provided in Table 2 of the main text, Tables 16–19 in Appendix D.3, and Figure 9.

Slopes and SD ratios below 1 also quantify the remaining compression of variation under the current input budget. Within-tissue error, absolute agreement, and the magnitude of variation together provide a more direct evaluation than pooled correlation, as well as clear criteria for assessing the contribution of additional measurements in future multi-input configurations. Our conclusions here concern tissue means in healthy test subjects.

Your interpretation of the decoder as a physics-inspired monotonic signal prior accurately characterizes its constraints. The positive rate represents the negative local log-signal slope of the synthesized signal, ensuring nonnegative, monotonically decaying signals within each FA; a shared T2∗T_2^{\ast} is an assumption of downstream joint fitting. Its quantitative contribution is tested directly through the Unconstrained Decoder control: at similar image quality, the paired improvement in joint T2∗T_2^{\ast} MAE is 0.454 ms (95% CI: 0.237–0.675), with improvement in 9/10 subjects. The separate cross-FA evaluation is reported in Appendix D.5.

3. The actual separation used for off-grid T2∗T_2^{\ast} fitting

The off-grid experiment uses the actual stride-two separation of 8 ms. Eq. 24 in Appendix C gives the expression for equally spaced echoes used in fitting:

q(x)=∑g∑j=1N−ℓSg(x,τj)Sg(x,τj+ℓ)∑g∑j=1N−ℓSg(x,τj)2,T^2∗(x)=−Δpairlog⁡q(x),Δpair=τj+ℓ−τj.q(x)= \frac{\sum_g\sum_{j=1}^{N-\ell}S_g(x,\tau_j)S_g(x,\tau_{j+\ell})} {\sum_g\sum_{j=1}^{N-\ell}S_g(x,\tau_j)^2}, \qquad \widehat T_2^{\ast}(x)=-\frac{\Delta_{\mathrm{pair}}}{\log q(x)}, \qquad \Delta_{\mathrm{pair}}=\tau_{j+\ell}-\tau_j.

Each term is computed from valid positive-signal pairs, with a default stride of ℓ=2\ell=2. The acquired grid corresponds to Δpair=10.18 ms\Delta_{\mathrm{pair}}=10.18\thinspace \mathrm{ms}, while the off-grid set {5,9,13,17,21}\lbrace 5,9,13,17,21\rbrace ms corresponds to Δpair=8 ms\Delta_{\mathrm{pair}}=8\thinspace \mathrm{ms}. Integration starts at the 3.86 ms anchor, and off-grid parameter fitting uses the five queried echoes, excluding this integration anchor. This experiment yields MAE 14.24±1.6214.24 \pm 1.62 ms and map R2=0.305±0.096R^2=0.305 \pm 0.096.

4. Fitting stability with more query points

Following your suggestion to generate more images, we queried the same model at 5, 9, 17, and 33 equally spaced points over 5–21 ms, fitting with the actual stride-two separation for each grid. All results use the same acquired-reference support.

Number of query pointsStride-two separation (ms)T2∗T_2^{\ast} MAE (ms)
5814.201
9414.209
17214.196
33114.187

The maximum difference in mean reference MAE across densities is approximately 0.023 ms. Direct comparison of the fitted maps gives a mean absolute difference of 0.305±0.0170.305 \pm 0.017 ms between 5 and 33 points, and 0.069±0.0070.069 \pm 0.007 ms between 17 and 33 points. These results support query stability for the same model under the tested interval, integration, and fitting rules; details are provided in Tables 12–13 in Appendix B.5.

The original five-point off-grid result uses a 213-node integration trajectory, whereas the density experiment uses the integration anchor plus query nodes, giving respective five-point MAEs of 14.24 and 14.201 ms. The appendix distinguishes these two numerical settings. The density experiment changes both integration nodes and observations used for fitting; a check that changes only the integration grid at common TEs is reported in Table 11 in B.4. This analysis evaluates query and fitting sensitivity: additional synthesized points do not add independent measurement information, and scan–rescan reliability requires evaluation with independent acquisitions.

5. The scope of continuous representation and the terminal hold-out

The interior hold-out at 14.04 ms evaluates interpolation within the range; the terminal hold-out at 24.22 ms uses image supervision only up to 19.13 ms and is evaluated against the excluded acquired target images. The latter tests prediction beyond the image-supervised TE range under a known protocol. At the terminal echo, the Full Model improves PSNR over TE-Conditioned MLP by 0.603/0.397 dB for the two FAs, respectively, with joint T2∗T_2^{\ast} MAE of 14.275 vs. 15.134 ms (Tables 8 and 10).

The numerical endpoint descriptors remain known; the current TE experiments support continuous querying within a fixed acquisition configuration, while direct transfer to unseen FAs, TRs, or endpoints has not been tested. The off-grid experiment evaluates parameter-map agreement after resampling, while the curves in Figure 5 beyond the acquisition endpoint demonstrate structural continuation. Their evaluation objectives are clarified in the main text and appendix.

6. Public data, T1T_1 mapping, and other organs

We added a public MPM experiment, splitting 108 healthy participants by subject into 78 for training, 15 for validation, and 15 for testing. The data contain eight echoes each for PDw (FA$6^\circ$) and T1w (FA$21^\circ$), both with TR of 25 ms. The Full Model, CCR, and D2Diff were all retrained on this dataset, retaining the first echo in each group and completing the remaining seven; the reference is obtained by jointly fitting the complete acquired echo trains from both groups.

MethodPDw PSNR (dB)T1w PSNR (dB)Joint T2∗T_2^{\ast} MAE (ms)
Full Model26.52±0.93\mathbf{26.52 \pm 0.93}28.06±0.85\mathbf{28.06 \pm 0.85}8.49±0.73\mathbf{8.49 \pm 0.73}
CCR26.29±0.9526.29 \pm 0.9528.03±0.8628.03 \pm 0.868.73±0.678.73 \pm 0.67
D2Diff25.89±0.9825.89 \pm 0.9827.79±0.8727.79 \pm 0.879.25±0.699.25 \pm 0.69

Image PSNR is computed only over the seven generated echoes in each group, and quantitative error uses the acquired-reference support shared by all methods. The Full Model achieves the highest mean PSNR and lowest mean quantitative error under another public brain-imaging protocol, supporting the framework's use in different echo-completion settings after retraining. Results and image comparisons are provided in Table 23 and Figure 10 in Appendix E.

Parameter-map evaluation in both sets of experiments focuses on T2∗T_2^{\ast}. Exploratory T1T_1 fitting of acquired MULTIPLEX echoes has not yet yielded reliable validation, so we do not report conclusions on T1T_1 mapping in this revision; experiments on other organs have not yet been conducted either. The public-data experiment extends evaluation to another brain cohort and protocol, providing an empirical basis for future task extensions.

Reviewer_nohEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes a method for synthesizing missing MRI echo trains from a single short-TR source echo per flip angle for qMRI acceleration. The propoed method use an endpoint-conditioned latent trajectory network with a decoder that generates continuous, nonnegative, monotonically decreasing signals across echo times.

Critical strengths
  • The paper addresses an interesting and clinically-important problem of reducing MRI acquisition time through missing echo-train synthesis.
  • The method achieves competitive echo-image synthesis quality and improves T2* map agreement over the evaluated baselines. The ablations also demonstrate measurable benefits from the proposed decoder and trajectory conditioning within the tested protocol.
  • The proposed decoder guarantees nonnegative, monotonically decreasing echo signals by construction.
  • The experimental evaluation includes several relevant synthesis baselines and component ablations.
Critical weaknesses
  • My biggest concern is the following sentence (intro) - "A single source echo does not uniquely determine an individual’s decay curve, so the missing information is inferred from the training population". The method attempts to infer complete T2* decay trajectories from two source images acquired at the same TE, differing only in flip angle. Under the standard gradient-echo signal model, these measurements do not uniquely determine T2* - proton density and relaxation remain confounded, and distinct decay trajectories can produce identical source observations. The authors acknowledge this issue and rely on a learned population prior, but I think this reveals an even deeper issue. A main use-case of qMRI is measuring individual tissue properties and detecting abnormal deviations. When the input does not contain sufficient information, the network may predict population-typical relaxation behavior that might ignore (or contradict) the individual's actual decay. This creates a systematic risk of suppressing clinically informative abnormalities and biasing outputs precisely against the abnormalities we are interested in, especially when the prior is learned exclusively from healthy subjects. As a concrete example, imagine a pathological abnormality that is invisible in both retained source images at TE = 3.86 ms, but becomes clearly visible at TE = 24.22 ms due to its different relaxation behavior. How could the proposed method recover this abnormality when no evidence of it is present in the input? I see no reason why the prior would guide us towards this abnormality. Priors are widely used in MRI and qMRI, but my concern is the unusually strong reliance on population priors to recover subject-specific information that is not directly measured.

  • The evaluation may reward population-average predictions over faithful quantitative recovery, and the reported results actually seem to raise this concern - The study uses only 42 healthy participants (32 training, ten testing), without pathological cases or expert assessment. Although the authors evaluate fitted T2* maps and do not rely only on PSNR/SSIM, the principal voxelwise agreement is modest. Moreover, the training objective optimizes echo-image reconstruction rather than quantitative-map accuracy or preservation of individual relaxation differences. Appendix D reports substantially reduced between-subject variation in predicted maps: GM standard deviation falls from 2.13 to 1.03 ms, and WM from 1.80 to 0.74 ms. This is consistent with regression toward population-typical tissue properties (although reference measurement noise may also contribute). The strong pooled GM/WM correlation (r2r^2 = 0.962) is also probably misleading - a predictor assigning a fixed mean T2* to each tissue, preserving no within-tissue individual variation, would achieve approximately r2r^2 = 0.93 using the reported statistics. The paper lacks a population-mean/anatomy-conditioned baseline, within-region subject-level calibration analysis, and tests of abnormal relaxation values. Consequently, the reported improvements may partly reflect metrics that reward typical population-common anatomical patterns, while failing to show preservation of the individual differences that make qMRI useful in the first place.

  • The authors motivate their method by saying that existing synthesis methods do not guarantee physically consistent echo decay, however this problem has a straightforward solution that is much more grounded in qMRI physics - predict the initial signal amplitude and T2* map, then generate the entire echo train using the known exponential decay model. This guarantees continuous, monotonic decay without requiring the proposed complex trajectory network and learned TE-dependent decay rate. Incorporating known relaxation models as priors into learning is also well established, as demonstrated by T1-PILOT, T2-PILOT, and Jacobs et al. (2024), the latter being particularly relevant to MRI contrast synthesis. Why do the authors need a more flexible learned decay model instead of this simpler, physically motivated prior? How do we know it performs better? Although the paper includes a Constant-Rate baseline, it also removes the trajectory network, making the comparison inconclusive. A matched baseline using the standard exponential decay model is needed to establish the benefit of the proposed formulation. This is especially important because a mono-exponential model would substantially restrict the set of possible decay trajectories, potentially reducing reliance on population priors, although it would not resolve the fundamental non-identifiability of the retained measurements. I think the usage of a prior should be compared to in order to alleviate certain concerns - but if not, it should at the very least be discussed for a more complete image of exisitng literature.

  • Even if the solution is convincing, I am not yet convinced there's a problem. The authors focus on reducing acquisition time by removing the entire long-TR echo train, but do not establish why this is preferable to existing faster acquisition strategies. Methods such as Joint MAPLE, Feng et al. (2024), and T1-PILOT demonstrate alternative approaches to accelerating quantitative MRI while preserving relaxation-sensitive measurements. Why discard the measurements that directly determine T2* rather than acquire fewer k-space samples across multiple echoes? The authors report a nominal 4.67× acceleration, but provide neither actual scan-time measurements nor an acquisition-time versus quantitative-accuracy comparison with these alternatives. Before demonstrating that the proposed method can synthesize a missing echo train, the authors should establish why synthesizing that train is a useful acceleration strategy in the first place, especially given the fundamental loss of subject-specific relaxation information discussed above. As things stand, the proposed method's merit seems to mostly apply to the specific scan approach presented in the paper.

  • The proposed physics constraint guarantees only positivity and monotonic decay. It does not enforce or evaluate consistency of T2* estimates across the two flip angles, despite both measuring the same tissue. The resulting curves could be individually monotonic yet quantitatively inconsistent. Why is cross-FA consistency not evaluated?

  • Experiments use fixed TR/FA settings, with no generalization to unseen acquisition configurations. Held-out experiments test only individual TEs, while off-grid queries lack corresponding acquired images and extrapolation beyond the TE range has no ground truth.

  • Regardless of issues with applicability, algorithmic contribution is narrow - the method largely combines well-established techniques: acquisition-conditioned feature modulation, continuous representations, and positive-rate integration for monotonicity (e.g., CCR, UMNN). While the ablations show some improvements, the paper does not demonstrate a substantial new algorithmic contribution or generality beyond the specific MRI protocol evaluated.

Questions for authors

My two most substantial concerns are whether there's a problem, and whether the proposed solution actually solves it, or does it generate good metrics that harm clinical practicality downstream? Specifically, I would consider raising my score if the authors could explain why their strong reliance on population priors does not bias model outputs against pathologies (that are probably not adundant in the training set), and why shouldn't I just use other qMRI acceleration methods over the proposed method.

While these are the two most major concerns, I am much more likely to change my score if the authors address other concerns I had raised.

9 comments

Thank you for your specific, testable suggestions. This work studies how to generate continuous target echo trains from retained shorter-TR acquisitions while making the cross-TE signal relationships suitable for subsequent quantitative fitting. A single source input per FA is the controlled setting used here to examine the generation structure; the framework can be extended to multiple inputs by fusing several acquired images at the source encoder, complementing k-space undersampling, as detailed in point 2 below.

Following your suggestions, we added a matched mono-exponential control that retains the conditioning network, a training-population baseline and within-tissue analysis of individual agreement, cross-FA evaluation, and experiments on the public MPM dataset. Results with another estimator and three training seeds further verify the stability of the quantitative gains. The main findings from these additions are:

  • The model provides individual predictive value beyond fixed tissue means. GM/WM tissue-mean MAE is reduced by 38.9%/36.8%38.9\%/36.8\% relative to the training-population constant, with within-tissue Pearson rr values of 0.654/0.734.
  • The Full Model still outperforms the matched mono-exponential control constructed following your suggestion in all three supervision settings. Under full supervision, the matched control and the Full Model have joint T2∗T_2^* MAEs of 14.96 and 14.13 ms, respectively, with the same direction of improvement in both TE hold-out settings.
  • The quantitative gains are repeatable and extend to another public-data setting. Gains over the Unconstrained Decoder persist with NLLS and across three seeds; on the public MPM test set, our method has lower mean quantitative error than CCR and D2Diff.

These results respectively examine your concerns about individual information, the necessity of the decay model, and the scope of applicability. Below, we address the two core questions and describe each piece of additional evidence.

1. Does the model merely predict typical population tissue values? How do we view bias in pathological cases?

We directly examine this question through across-subject analysis within the same tissue. The acquisition-conditioning path is shared under the same protocol, while each subject's source-image features contribute to amplitude and rate prediction; the key evaluation question is whether these input-dependent variations correspond to individual differences in the acquired reference. The additional experiment evaluates ten test subjects using common, prediction-independent GM/WM support and includes tissue-mean constants determined solely from the 32 training subjects.

TissueOur tissue-mean MAE (ms)Training-population constant MAE (ms)Pearson rr (95%95\% CI)ICC(A,1)\mathrm{ICC}(A,1) (95%95\% CI)
GM1.4092.3060.654 (0.070–0.9560.070\text{–}0.956)0.534 (0.035–0.8280.035\text{–}0.828)
WM0.9891.5660.734 (0.502–0.9410.502\text{–}0.941)0.596 (0.298–0.8030.298\text{–}0.803)

These results provide evidence beyond pooled tissue correlation: within the same tissue, predictions follow across-subject variation in the reference and achieve lower mean error than the training tissue constants. Compared with ResViT, D2Diff, CodeBrain, and CCR, our method achieves the best point estimates for tissue-mean MAE, Pearson rr, and ICC in both tissues. The complete comparisons have been added to Table 2 in §4.2 of the main text and Tables 16–19 and Figure 9 in Appendix D.3.

Your observation about pooled r2r^2 prompted us to use within-tissue analysis as the primary evidence for individual information preservation. In the added calibration statistics, the GM/WM prediction-to-reference SD ratios are 0.673/0.479 and the slopes are 0.440/0.352; paired improvements over the training constants are 0.897 ms (95%95\% CI: −0.350–2.020-0.350\text{–}2.020) and 0.577 ms (0.120–1.0430.120\text{–}1.043). These results quantify the tissue-mean differences that the current single-source configuration can recover, as well as the remaining compression of variation, with clearer evidence of error improvement in WM. This analysis also provides explicit metrics for assessing whether additional acquired inputs are beneficial: in addition to echo-image quality, within-tissue error, absolute agreement, and the magnitude of variation should be examined. The current constant control tests fixed tissue-value prediction; spatially anatomy-conditioned population prediction could further complement this evaluation.

We acknowledge that lesion visibility depends on TE; dual-echo T2∗T_2^* imaging has provided empirical evidence that a longer TE improves the visibility of some microbleeds (Conijn et al., 2010). Here, weak visual contrast at short TE should be distinguished from the complete absence of usable evidence in the retained inputs: the former may still contain prediction-relevant information, whereas recovery in the latter case cannot be guaranteed by a method relying only on those inputs. For lesion-sensitive applications, the framework's multi-input extension allows complementary acquired information at different TEs to be retained before generating the remaining echoes; the corresponding acquisition schemes are discussed in the next point. Bias in pathological cases arising from the healthy training prior still requires evaluation using pathological cases and expert assessment; the added within-tissue analysis directly evaluates individual predictive value in the healthy cohort studied here.

2. Why study echo generation from shorter-TR inputs? How can it complement multiple inputs and undersampling?

The research gap addressed here is how the structure of continuous echo generation from shorter-TR inputs affects downstream quantification. Existing acquisition acceleration and reconstruction methods have extensively studied how undersampled measurements can recover images and parameters; this work further investigates how to combine source images, continuous TE conditioning, and explicit decay structure after reducing target acquisition blocks to generate echo trains suitable for quantitative fitting. The Full Model's improvement in joint T2∗T_2^* over the Unconstrained Decoder at similar image quality, together with the added matched Mono results, directly demonstrates the value of studying this generation problem.

We use a clearly defined dual-TR task: for each FA, we retain a source image at TR=7.97 ms\mathrm{TR}=7.97\,\mathrm{ms} and TE=3.86 ms\mathrm{TE}=3.86\,\mathrm{ms} and complete the five target echoes at TR=29.23 ms\mathrm{TR}=29.23\,\mathrm{ms}. The five target echoes are read within a single long-TR repetition, so omitting the entire long block can reduce a distinct component of the acquisition budget. With equal repetition counts for the two blocks, the ratio of the full to retained protocol's TR budget is

7.97+29.237.97≈4.67,\frac{7.97+29.23}{7.97}\approx 4.67,

corresponding to a reduction of approximately 78.6%78.6\% in this budget component. This value illustrates the acquisition motivation for the current single-source task and represents a protocol-based nominal budget, without accounting for actual sequence overheads or prospective scan timing.

Multiple inputs provide an extension that retains additional relaxation information. The framework can fuse multiple acquired echoes at the source encoder, use the complementary observations to predict amplitude and decay rate, and continue to generate the target train through continuous querying and positive-rate integration. Retaining only one source image per FA is the input budget chosen for the current experiments; for lesion applications, acquired inputs with complementary information, such as short and intermediate TEs, can be selected and the model trained for that configuration.

Echo generation and k-space undersampling act on different dimensions of the budget and can be combined. One concrete input scheme is to retain acquired echoes at a small number of TEs or at a single TE, apply k-space undersampling and reconstruction to the retained acquisition, and then fuse these source images to generate the remaining echoes. When multiple distinct TEs are retained, these acquired echoes can provide direct observations of individual decay, while undersampling further reduces the spatial-encoding budget of the retained protocol. Work such as Joint MAPLE demonstrates the value of undersampled multi-echo reconstruction and quantitative estimation; the generation module studied here provides a completion mechanism along another dimension for reducing target-echo acquisition, making the two complementary within the workflow.

Combined acceleration can be decomposed using the same budget reference. Let the acquisition time of the full protocol be TfullT_{\mathrm{full}}, the time after selecting the acquired inputs be TkeepT_{\mathrm{keep}}, and the time after undersampling this retained protocol be Tkeep,kT_{\mathrm{keep},k}. Then

=\frac{T_{\mathrm{full}}}{T_{\mathrm{keep},k}} =\frac{T_{\mathrm{full}}}{T_{\mathrm{keep}}} \frac{T_{\mathrm{keep}}}{T_{\mathrm{keep},k}} =R_{\mathrm{keep}}R_k.$$ This multiplicative relationship connects acquisition-configuration reduction with effective undersampling acceleration within the retained protocol. $R_k$ is the effective acceleration within the selected retained protocol, while $R_{\mathrm{keep}}$ is determined by the actual TR and acquisition-block budget. After adding an intermediate TE, the retained protocol should be recalculated according to the longest retained TE, readout timing, and overheads; $4.67\times$ corresponds to the current single-source setting in this work. The experiments reported here remain single-source completion experiments. The multi-input and undersampling combinations above provide acquisition schemes that can be further validated, under which actual scan-time gains and lesion-information preservation can be directly evaluated. Accordingly, we use synthesis methods with the same source inputs and matched structural ablations as direct controls to isolate the effects of continuous conditioning and decay modeling. Undersampling methods address reconstruction from retained measurements and are appropriate for further evaluation within a combined system; an undersampling control alone cannot substitute for the current same-input generation controls in addressing this methodological question. Related discussion of Joint MAPLE, Feng et al., and T1-PILOT/T2-PILOT has been added to §2.
Reviewer_foLKOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The goal of the paper is to complete a long-TR echo train from one retained short-TR echo per flip angle. An endpoint-conditioned network modulates image features, while a decoder integrates positive decay rates to generate nonnegative, nonincreasing signals at queried echo times. The method is evaluated on 42 healthy participants, with 32 used for training and ten for evaluation. Besides echo-image quality, the authors assess held-out echo times and agreement with T2* maps fitted from acquired echoes. The proposed decoder improves map agreement at similar image fidelity.

Critical strengths
  1. The signal constraint is built into the decoder and has a clear interpretation through positive log-signal decay. The comparison with an unconstrained decoder is useful: similar echo-image PSNR gives different quantitative-map accuracy. The paired improvement of approximately 0.45 ms in map MAE is supported by subject-level analysis.

  2. The experiments examine more than interpolation at supervised coordinates. Interior and terminal hold-outs, an off-grid echo set, a matched pointwise-conditioning baseline, and quadrature checks support the continuous querying design. The paper also clearly distinguishes nominal acquisition-budget reduction from measured scan-time acceleration.

Critical weaknesses
  1. Preservation of individual quantitative differences remains the main concern. The source image does not uniquely determine the missing decay trajectory, and the regional results show reduced between-subject variation: predicted GM SD is 1.03 ms versus 2.13 ms in the reference, and WM SD is 0.74 versus 1.80 ms (Table 10). The high pooled tissue correlation can therefore coexist with regression toward typical tissue values. The authors acknowledge this limitation, but it remains consequential for the quantitative MRI motivation.

  2. The quantitative benefit is demonstrated against maps from one fitting implementation on a small healthy cohort. The fixed-mask voxelwise comparison is appropriate, but regional results additionally condition on prediction reliability, and the paired uncertainty covers evaluation subjects rather than retraining variability. It is not yet clear how stable the decoder’s relatively small MAE advantage is under another reasonable fitting procedure or training seed.

Questions for authors
  1. Can you report agreement across subjects within each tissue or region, and compare with a training-population prediction using the same anatomical support? This would help establish how much individual variation is recovered beyond typical tissue values.

  2. Does the constrained decoder retain its advantage when the same acquired and synthesized echoes are processed with another standard T2* fitting method? If repeated training runs are available, how large is the variation relative to the 0.45 ms paired MAE improvement?

2 comments

Thank you for your positive assessment of the explicit decay constraint, the continuous-query experiments, and the quantitative gains from the decoder. Following your suggestions, we have added an analysis of individual differences under fixed tissue support together with a training-population baseline, an alternative T2∗T_2^* fitter, and repeated training with three random seeds. These additional results further support the model's predictive value for individuals and the stability of the decoder's gains.

1. Within-tissue individual differences and training-population prediction

We computed GM/WM tissue means separately for each of the ten test subjects to directly evaluate between-subject variation within the same tissue. For each subject, the acquired reference and all methods use the same fixed tissue support, defined by tissue probability maps and the acquired-reference support, without excluding voxels based on prediction reliability. The training-population baseline is the average of the reference tissue means across the 32 training subjects and remains fixed at test time.

TissueOur tissue-mean MAE (ms)Training-population constant MAE (ms)Pearson rr (95%95\% CI)ICC(A,1)\mathrm{ICC}(A,1) (95%95\% CI)
GM1.4092.3060.654 (0.070–0.9560.070\text{–}0.956)0.534 (0.035–0.8280.035\text{–}0.828)
WM0.9891.5660.734 (0.502–0.9410.502\text{–}0.941)0.596 (0.298–0.8030.298\text{–}0.803)

Relative to the training-population constant, GM/WM tissue-mean MAE decreases by 38.9%/36.8%38.9\%/36.8\%, respectively. The positive within-tissue correlations and absolute agreement indicate that the predictions preserve variation among healthy subjects that is associated with the reference. Compared with the four learning-based baselines, ResViT, D2Diff, CodeBrain, and CCR, our method achieves the best point estimates for tissue-mean MAE, Pearson rr, and ICC(A,1)\mathrm{ICC}(A,1) in both tissues. The complete comparison is provided in Table 2 of the main text, Tables 16–19 in Appendix D.3, and Figure 9.

We have also added statistics on the magnitude of variation and paired uncertainty. The between-subject prediction/reference SD ratios are 0.673/0.479, and the regression slopes are 0.440/0.352, indicating that the current model partially preserves differences in tissue means, although the magnitude of variation remains compressed. The mean paired improvements over the training-population constant are 0.897 ms for GM (95%95\% CI: −0.350–2.020-0.350\text{–}2.020) and 0.577 ms for WM (0.120–1.0430.120\text{–}1.043); the improvement interval for WM excludes zero, while the magnitude of improvement for GM remains uncertain. All confidence intervals are calculated using subjects as the sampling unit.

Regarding the regional selection you noted, the original regional analysis used the intersection of the reference and prediction reliability masks. The new analysis uses fixed support independent of the predictions, directly addressing this concern. The selection criteria for the original analysis are retained in Appendix D.2, and the original and new results are presented separately. The main voxel-level results in the main text have consistently used fixed acquired-reference support.

2. Quantitative gains with an alternative fitting method

We refitted both acquired and generated echoes using magnitude-domain mono-exponential nonlinear least squares (NLLS), using the NLLS-fitted maps from acquired echoes as the corresponding references and retaining the original fixed evaluation support. The results are shown below; both the means and standard deviations are calculated across the ten test subjects.

FitterFull Model T2∗T_2^* MAE (ms)Unconstrained Decoder T2∗T_2^* MAE (ms)Mean paired improvement (ms)Subjects with improvement
Original MDI implementation14.13±1.6114.13 \pm 1.6114.58±1.7914.58 \pm 1.790.4549/10
NLLS14.337±1.62714.337 \pm 1.62714.664±1.63014.664 \pm 1.6300.32710/10

The constrained decoder continues to show improvements in a consistent direction under the alternative fitting objective. This supports benefits beyond the specific implementation of the original fitter. The fitting objective, reference construction, and subject-level results are provided in Appendix C.1 and Table 20 in Appendix D.4.

3. Stability across repeated training runs

We trained the Full Model and the Unconstrained Decoder separately with three random seeds, using the same data split, a training budget of 1000 epochs, and the final-epoch checkpoint selection rule.

Random seedFull Model MAE (ms)Unconstrained Decoder MAE (ms)Paired gain (ms)
202614.125514.57930.4538
202713.978214.49070.5125
202814.171814.64080.4690

The mean gain across the three pairs of training runs is 0.4784 ms, with a between-seed sample standard deviation of 0.0305 ms, which is substantially smaller than the mean gain. All three seeds support the originally reported improvement of approximately 0.45 ms. These results provide additional evidence regarding training randomness and are reported separately from the original subject-paired confidence interval; the complete results are provided in Table 9 in Appendix B.2.

Comments on the paper

2 comments

Thank you for your feedback. To clarify the scope of this discussion, we distinguish two claims: that clinical diagnostic reliability has not been established, and that the method therefore has no practical value for quantitative research in healthy cohorts. The former is explicitly acknowledged in the manuscript; the latter does not follow from it. This is the central distinction we wish to clarify.

1. This work studies acquisition efficiency and quantitative fidelity in healthy cohorts.

Reducing the acquisition burden of quantitative imaging in healthy cohorts motivates this work. Under the evaluated protocol, retaining one short-TR source echo per flip angle and omitting the long-TR block corresponds to a nominal 4.67× acceleration based on the two-block TR budget (Sec. 4.1). We examine how much individual quantitative variation remains predictable under this trade-off and whether explicitly modeling cross-echo decay improves downstream T2∗T_2^* estimation. Fully acquired echoes and maps fitted from them provide direct evaluation references.

The evaluation scope and limits of our conclusions are described in Secs. 4.1 and 5. The Ethics Statement also explicitly states that this study is not intended for clinical decision-making. Acquisition burden and quantitative information preservation in healthy cohorts are medical imaging research questions in their own right (e.g., for researchers working with public healthy-adult MPM cohorts such as the one evaluated in Appendix E). The value of this work should be assessed against these questions and the corresponding evidence.

2. We acknowledge the identifiability limit, but it does not justify dismissing the value of this research.

If a pathological change leaves no usable information in any retained input, the model cannot reliably recover it. This is not in dispute. Single-source-echo non-identifiability and the limits of recovering invisible pathological changes are explicitly discussed in Secs. 1 and 5, respectively.

The issue is how broadly this counterexample can be interpreted. It establishes that certain diagnostic tasks require acquired measurements sensitive to the lesion. However, the inability to recover information entirely absent from the inputs does not imply that quantitative prediction from retained inputs in healthy cohorts has no practical value. The latter question requires empirical evaluation of individual prediction errors, association, and calibration; a lesion counterexample alone cannot settle it.

3. Individual quantitative agreement is precisely a comparative strength of our method under matched inputs.

Our previous two responses provided direct evidence: relative to training-population constants, our method reduces GM/WM tissue-mean MAE by 38.9% and 36.8%, respectively. Against four same-input methods, it achieves the best point estimates for all six GM/WM comparisons of MAE, Pearson rr, and ICC, with calibration slopes closest to 1 (Table 2, Tables 17–18). The exploratory k-space comparison under the fixed protocol also provides evidence relevant to the acquisition strategy (Appendix F). We explicitly report residual compression of individual variation. Imperfect absolute calibration does not negate the comparative advantages established above.

4. Our quantitative contribution is supported by direct evidence beyond image similarity.

You note that high image similarity can conceal the loss of important information. Our decoder comparison examines quantitative differences beyond image quality: with identical inputs and comparable echo-image quality, the decoder that explicitly models decay produces T2∗T_2^* maps closer to the acquired-echo-fitted reference in 9/10 subjects, reducing mean MAE by 0.454 ms; with nonlinear least-squares fitting, all 10 subjects show lower errors (Appendices D.1, D.4). This demonstrates that, with the input information fixed, modeling cross-echo decay improves agreement between downstream parameter maps and the acquired reference—a benefit that image similarity alone cannot reveal.

We are grateful for your acknowledgment of the algorithmic contribution and for the identifiability argument, which helped us define the method’s limitations and scope more precisely.

AuthorsOpenReview ↗

Summary of Rebuttal and New Experiments for the Area Chair

Dear Area Chair,

Thank you for overseeing this discussion. We study whether explicit cross-echo decay modeling improves downstream T2∗T_2^* estimation in healthy-brain MRI and how much subject-specific quantitative information remains predictable from one short-TR source echo per flip angle (nominal 4.67× acceleration of the two-block TR budget).

1. All three reviewers acknowledge specific methodological contributions.

R1 (foLK) noted improved quantitative consistency at comparable image fidelity; R2 (nohE) confirmed the matched mono-exponential control addressed their question and stated “the algorithmic contribution stands”; R3 (JwPq) viewed the methodological contribution more positively after supplementary experiments.

2. Key Evidence Added

  • Controlled ablations: With comparable parameter counts and matched training budgets, the Full Model has lower T2∗T_2^* map MAE than Mono-exponential, TE-Conditioned MLP, and FiLM under full supervision and both held-out-TE settings (Table 3, Table 8, Appendix B.1).
  • Cross-estimator & seed consistency: The explicit-decay decoder’s advantage over the Unconstrained Decoder holds under the primary estimator (9/10 subjects, mean T2∗T_2^* MAE reduction 0.454 ms), NLLS (10/10, 0.327 ms), and three seeds (Appendices B.2, D.1, D.4).
  • Within-tissue analysis: We added training-constant MAE comparisons and within-tissue analyses against four same-input synthesis baselines (Table 2, Tables 17–18, Appendix D.3).
  • Public cohort & k-space comparison: After retraining on public healthy-adult MPM data (n=108n=108), our method has the lowest joint T2∗T_2^* MAE among three methods on 15 test subjects (Appendix E). We also compared against Joint MAPLE’s reconstruction module in zero-phase single-coil simulations at matched nominal budgets (~4.7×, ~9.3×; Appendix F, Tables 24–27).

3. Remaining Reservations and Our Position

Our supplementary experiments establish comparative gains in generative structure and subject-level quantitative prediction; remaining discussion concerns how strong an application claim these gains support.

(1) R2’s concerns on lesion identifiability and image similarity.

We accept that lesions leaving no usable information in the retained inputs cannot be reliably recovered (Secs. 1, 5; Ethics Statement: non-clinical scope), which restricts clinical lesion-diagnosis claims. However, extrapolating from this counterexample to dismiss our method’s value for healthy-cohort quantitative research goes beyond what it supports. Acquisition burden, predictable subject-level variation under reduced sampling, and how generative structure affects fitting are directly testable questions. Protocol-budget analysis and Appendix F address acquisition strategy, while held-out echo and subject-level analyses directly test accuracy and fitting (Appendix E also shows comparative gains on an independent public healthy-adult protocol); a lesion counterexample cannot replace evaluating these results.

On R2’s concern that image similarity may conceal quantitative loss, our matched decoder ablation directly tests performance beyond image quality: at identical inputs and comparable PSNR/SSIM, explicit decay modeling consistently reduces downstream T2∗T_2^* error (Appendices D.1, D.4), giving direct quantitative evidence beyond image-similarity metrics.

(2) R3’s concerns on quantitative fidelity and calibration compression.

R3 asks whether amplitude compression among healthy subjects compromises meaningful individual-level quantitative use. Under identical source images, fixed tissue supports, and the same acquired reference, our method lowers GM/WM tissue-mean MAE by 38.9%/36.8% over training constants and achieves the best point estimates across all six GM/WM comparisons of MAE, Pearson rr, and ICC(A,1) against four same-input synthesis baselines, with slopes (0.440/0.352) closest to 1 (Table 2, Tables 17–18). In the descriptive ~4.7× simulation, the undersampled k-space comparator also shows higher tissue-mean error and calibration deviation (GM/WM slopes 2.013/−0.320, Table 26).

Meanwhile, GM/WM slopes remain below 1, and in-vivo evaluation does not fully separate prediction bias from reference-fitting uncertainty. Sufficient individual fidelity for specific downstream applications and clinical reliability have not yet been established, restricting stronger claims (e.g., preserving full individual-difference magnitudes or predicting clinically relevant deviations). Yet these limits do not negate the comparative gains under matched inputs: the reported calibration results and evaluation conditions bound our application claims, while same-input comparisons and matched ablations directly support improvements in subject-level quantitative prediction.

We hope this helps the Area Chair weigh the verified methodological and comparative quantitative gains alongside the clearly stated boundaries.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.