acceptodds
Withdrawn from ICLR 2027

EMBER: Learning Hidden Lesion Dynamics from Intermittent Brain MRI

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Progression independent of relapse activity is a major source of disability in multiple sclerosis, yet routine brain MRI reveals lesion evolution only at intermittent visits. Assigning a lesion's birth to its detection date creates an artificial chronology; unobserved disease activity can also make unrelated regional lesions appear mutually exciting. We introduce EMBER, a learning framework that integrates interval-censored lesion histories with hidden Hawkes dynamics and longitudinal tissue measurements. Its first contribution is an observation-compatible branching representation that marginalizes lesion timing and latent common activity while allowing edge-specific mixtures of temporal kernels. Its second is a posterior-refinement objective: after scans are removed, the predicted risk must equal an average over compatible refinements, rather than remain identical to the prediction from one richer history. We establish exact coarsening identities and a Brier-risk decomposition that characterize the information lost through intermittent acquisition. Open longitudinal MRI tests lesion-change forecasting. In the clinical evaluation, EMBER predicts 24-month relapse-independent progression with held-out-site AUROC 0.846 and Brier score 0.123, and maintains discrimination under reduced scan availability. The resulting model links uncertain lesion chronology to clinically meaningful progression risk using routine T1-weighted and FLAIR MRI.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Multiple Sclerosis, Longitudinal Brain MRI, Latent Hawkes Processes, Interval-Censored Event Modeling

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_qtNuOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The authors want to tackle that MRI's do not show the full patient trajectory and multiple onset of lesions could have explained the current state of the patient. In this sense it is hard to estimate Progression independent of relapse activity. The authors designed a framework based on aggregating possible lesion histories in a unified framework to improve upon the state-of-the-art.

Critical strengths

The authors tackle an important problem as pseudoprogress is hard to spot. Thinking about the stochastic nature of the problem is a valuable effort and the theoretical framework developed is extensive.

Critical weaknesses

Dataset issues:

  • There are 37 pages in the paper and none of them show a single example from their datasets. No actual FLAIR images are shown whatsoever. Fig 1 uses drawn images and in Fig 2 I am not sure if these are even FLAIR images as v_pk looks like a T1w image. It would really help the reading flow to actually see what is predicted and how it is predicted, especially because they are working on a private dataset. Same with the formulas. The figures do not present well, how these formulas adapt the different quantities and how this merging works in principle.
  • Furthermore, the public datasets are not well suited for this task as they only have 2 visits per patient. How do you delete an intermediate visit, when the dataset only has 2? Eligibility on the private dataset are patients with at least 3 visits. How do these results correlate?
  • There is a publicly available longitudinal MS dataset with multiple visits given by Carass, Aaron, et al. "Longitudinal multiple sclerosis lesion segmentation: resource and challenge." NeuroImage 148 (2017): 77-102. albeit just 5 patients.
  • A single private testset with 240 patients is inconclusive for any kind of generalization claims.

Motivation of the individual components:

  • The authors failed to motivate any of their formulas in the entire paper. It is also not clear what they are doing differently from existing work and where they improve upon these works. There is no novelty statement and from the introduction it reads like they took a couple of components and stuck them together to improve upon the state-of-the-art in this application. Similar from the related work. Are you proposing a new model for the specific trajectories? Is the aggregation novel? Is it only the setup? This should be further specified?
  • In any case many of the equations, like Eq 1 are just stated, however they it is never explained why this is a good choice. It is not even taken from any citation, nor is it derived from anything. Same with their theoretical setup. Some introductory words what the authors are actually trying to prove and accomplish would significantly enhance this paper. The figures are also very basic and do not explain the content well. Why should anyone use these formula and where are they applicable? The paper reads like a technical report without any insight into why and how any of these components are used and where the limitations lie.
Questions for authors
  • Can you clarify how the results on your public benchmark relate with the evaluation on the private testset, given that the scan numbers are vastly different.

  • Could you visualize the actual workflow based on images to get a better understanding on how your framework operates?

  • Can you add motivation for the different components and detail the degree of novelty of these components?

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

EMBER models MS lesion events whose onset is only known to fall between MRI visits. A latent Hawkes process represents possible event histories and tissue evolution, and progression risk is averaged across those histories. A refinement objective makes predictions consistent when intermediate scans are missing. Validation uses public MRI tasks and a two-center clinical cohort.

Critical strengths
  • Modeling interval-censored lesion timing and hidden activity together is well motivated for irregular clinical MRI. Posterior-refinement addresses a real missing-scan problem.
  • Paper includes a held-out clinical center, masked-scan prediction, mechanism controls, and ablations relevant to its main claims.
Critical weaknesses
  • Main clinical result relies on only 80 test patients with 15 PIRA events.
  • AUROC gain over the strongest comparator is 0.064, with a paired confidence interval whose lower bound is only 0.007.
  • Public datasets validate imaging tasks, but do not independently validate PIRA prediction. That is limited evidence for a complex clinical model.
  • Figures 1 and 2 look AI-generated based on my visual inspection. I also tested both with several detection tools; all gave scores of about 99% AI-generated.
    • The AI-use statement mentions only language polishing and manuscript review, with no mention of figure generation.
    • I consider AI-generated scientific figures unacceptable, and this apparent omission raises a serious research-integrity concern.
  • Bibliography errors too: author order for Brox et al. (2016), and page information for Jin & Huang (2026).
Questions for authors
  • What are the ethics approval, consent or waiver, and data-use arrangements for the 240-patient clinical cohort?
  • How sensitive is the clinical gain to the small number of positive test outcomes?
Reviewer_wVnmOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes EMBER, a joint model that infers uncertain lesion timing and hidden common activity from intermittent brain MRI in multiple sclerosis. The model also keeps predictions consistent across different scan availability with a posterior-refinement learning scheme. Experiments were performed for Progression Independent of Relapse Activity (PIRA) prediction within 24 months, hidden-region prediction and hidden-visit reconstruction.

Critical strengths

• The paper addressed a realistic clinical problem that lesions may arise at unseen time points and proposed a novel method for modeling such issues and temporal irregularity between two observed time points, instead of simply processing missing visits.

• Extensive experiments were performed. For example, the experiments included hidden-region prediction, hidden-visit reconstruction, lesion-change forecasting, scan removal robustness, and synthetic mechanism controls. Moreover, the data were obtained from two independent clinical centers to enhance generalizability.

Critical weaknesses

• The dataset only 240 clinical patients, with relatively few positive outcomes (15 in the test set).

• It is unclear whether the model can reliably distinguish latent disease activity from regional lesion interactions.

• Most patients have only 3–4 MRI scans, which may be insufficient to learn complex temporal dynamics.

• The refinement-coherence theorem largely follows from standard conditional expectation properties, which limits the theoretical novelty of the proposed method.

• Given the high model complexity, the relatively small improvement from refinement learning raises questions about its practical benefit.

• Random scan removal may not reflect real clinical imaging patterns.

Questions for authors

• How can the model distinguish true cross-regional lesion interactions from shared latent disease activity when lesion onset times are unobserved?

• Given only 128 training patients and 15 positive outcomes in the independent test cohort, how robust are the reported improvements to model initialization and patient sampling?

• Although the refinement objective substantially improves prediction consistency, its contribution to clinical AUROC is relatively small. Can the authors demonstrate improvements in clinically meaningful risk stratification or decision-making?

• What are the computational and memory costs of posterior sampling compared with simpler temporal models, and how does performance change with fewer sampled histories?

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

Related papers

Discussion (0)

Sign in to comment.