acceptodds
Under review as a conference paper at ICLR 2027

Learning How, Not Just What: Explicit Evolution Modeling for Longitudinal Radiology Report Generation

Abstract

Longitudinal radiology report generation (LRRG) aims to describe current findings and their temporal evolution by analyzing prior and current examinations. Existing methods largely rely on static comparisons to identify what has changed, but explicit modeling of how the disease evolves remains limited. To address this gap, we propose EvoRG, a framework for explicitly modeling disease evolution. We first aggregate visual features from both examinations using shared disease prototypes and Sinkhorn-based optimal transport. This alignment reduces interference from non-pathological spatial shifts and establishes semantically comparable disease-wise endpoints. Latent transitions between these endpoints are then parameterized as continuous geodesic trajectories on a joint visual–semantic product-sphere manifold. Logarithmic-map tangents encode evolution direction, while geodesic midpoints provide intermediate representations. Geodesic distances quantify evolution magnitude and modulate a product-of-experts module that fuses the midpoint and directed-movement descriptors. The resulting disease-specific transition tokens are used by a large language model to generate evolution-aware reports. Extensive experiments on Longitudinal-MIMIC show that EvoRG achieves state-of-the-art performance. Relative improvements over the strongest baseline for each metric are 12.5% in CIDEr and 9.6% in CheXbert F1.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
longitudinal radiology report generation, large-language models, chest X-rays

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_3oVrOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes EvoRG framework for longitudinal radiology report generation. Unlike prior work relying solely on static comparisons, EvoRG explicitly models disease evolution. It first aligns cross-visit visual features via shared prototypes and Sinkhorn optimal transport. Then, it captures disease transitions as geodesic trajectories on a product-sphere manifold, extracting tangents, midpoints, and distances as evolution descriptors. These then are fused to condition an LLM for report generation.

Critical strengths
  1. The paper focuses a clinically relevant problem in longitudinal radiology report generation: accurately describing the interval changes of persistent abnormalities, rather than solely identifying differences between two static exams. The change from detecting what has changed to characterizing the direction and magnitude of that change provides a strong and clear motivation for this framework.
  2. The proposed framework EvoRG has a coherent three-stage pipeline: disease-wise alignment, geometric modeling, and guided generation. The extraction of geodesic midpoints, tangents, and distances is particularly effective, as they offer complementary descriptions of intermediate states, movement directions, and transition magnitudes. Furthermore, the distance-aware fusion module sensibly balances these two types of information for report generation.
Critical weaknesses
  1. I have a concern regarding how the geodesic trajectory is constructed. It is fitted between two latent endpoints, but there is no clear evidence that this geometric path corresponds to how the disease actually evolves. The formulation ignores the actual time gap between exams or any intermediate observations. Hence, the method seems to model only the spatial displacement between endpoints, rather than true disease dynamics. The authors need to provide further justification.
  2. I am not convinced by the current clinical validation. Although the authors report Temporal F1 and provide qualitative examples, these evaluations do not sufficiently support whether EvoRG reliably identifies clinically meaningful changes, such as improvement, worsening, or stability. A more systematic evaluation is needed.
  3. The DTF module shows a mixed impact. It clearly helps with clinical precision, recall, and F1, which is great. However, we also observe a slight decrease in CIDEr (0.305 vs. 0.309 without DTF). This trade-off suggests that the fusion might not universally improve the generation quality. It would be helpful if the authors provides a more detailed analysis to show which aspects of report generation benefit from this module.
Questions for authors
  1. How do the authors justify interpreting the shortest geodesic between two latent endpoints as a meaningful representation of disease evolution, rather than merely a geometric interpolation?
  2. Beyond general report similarity and label agreement, can the authors directly evaluate disease evolution? Please report the accuracy of identifying improvement, worsening, stability, and new findings, preferably with radiologist annotations. This would confirm whether the performance improves truly because of better evolution modeling.
Reviewer_7qh5OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes EvoRG, a geometry-based framework for longitudinal radiology report generation. It uses shared disease prototypes and optimal transport to extract disease-specific features, constructs transition representations using geodesic midpoints, tangents, and distances on a product-sphere manifold, and fuses these descriptors through a distance-aware transition module to produce tokens that guide an LLM in generating the current report.

Critical strengths
  1. The paper addresses the clinically important task of describing changes in disease findings across longitudinal examinations, and the spherical geometric formulation offers an inductive bias for representing disease-state transitions within a technically coherent methodology.

  2. The results are promising, supported by comprehensive ablations comparing Euclidean and product-sphere transitions as well as supplementary clinical evaluation metrics.

Critical weaknesses

1. Insufficient clarity of ablation implementations and justification for key design choices

Tables 2 and 3 do not clearly specify how the w/o Midpoint and w/o DTF variants are implemented. For example, does w/o Midpoint remove the midpoint expert entirely, or replace its input with the prior state? Does w/o DTF remove only the distance modulation, or replace the entire fusion mechanism? These choices may change the model's architecture and parameter count, making it difficult to attribute performance differences to particular components.

Moreover, the geodesic midpoint is deterministically derived from the two endpoints and therefore introduces no additional independent information. Since the tangent already describes the transition from the prior state, why is a midpoint representation preferable to using the prior state together with the tangent? The reported w/o Midpoint ablation does not necessarily establish this advantage.

2. Missing independent evaluation of the auxiliary disease-state classifier

Section 3.3 uses a shared four-class classifier to predict positive, negative, uncertain, and unmentioned states for each of the 14 disease categories. Although the classifier is supervised with CheXbert-derived labels, its own classification performance is not reported.

Questions for authors
  1. How exactly are the w/o Midpoint and w/o DTF variants implemented? Are the corresponding networks and parameter budgets preserved?

  2. How does replacing the geodesic midpoint with the prior state affect performance? Does the midpoint's contribution vary with geodesic distance, as the midpoint can be very close to the prior point under some conditions?

  3. What is the independent classification performance of the four-state disease classifier used to construct the semantic sphere?

Reviewer_u2TZOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes EvoRG, a longitudinal radiology report generation framework that moves beyond static endpoint comparisons to explicitly model disease evolution. It aligns cross-temporal disease features using shared prototypes and Sinkhorn optimal transport, then parameterizes temporal transitions as geodesic trajectories on a joint visual-semantic product-sphere manifold. A distance-aware product-of-experts module fuses geometric descriptors to condition a frozen LLM for report generation. While the method reports state-of-the-art results on Longitudinal-MIMIC, it suffers from narrow scope and insufficient validation of its core design assumptions.

Critical strengths
  1. The work aims to reframe longitudinal report generation by modeling the disease evolution, introducing Riemannian geometric representations to the task and offering a novel perspective beyond conventional difference-based paradigms.
  2. The optimal transport-based disease alignment mitigates non-pathological spatial misalignment, and the product-sphere geodesic parameterization paired with distance-modulated PoE fusion forms a coherent.
Critical weaknesses
  1. The experiments are limited to pairwise chest X-rays on a single dataset. The claimed advantages of explicit evolution modeling are not tested on multi-visit sequences, other imaging modalities (CT, MRI), or diverse pathologies, severely limiting the method’s generalizability.
  2. The geodesic evolution hypothesis on the product-sphere manifold lacks theoretical justification. The incremental gains from the geometric design are not clearly disentangled from improved feature alignment.
  3. The performance relies almost entirely on automated N-gram and CheXbert metrics, lack of radiologist evaluation of report factuality or temporal description accuracy.
Questions for authors

Please refer to Weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.