List of Papers Browse by Subject Areas Author List
Abstract
Radiology report generation (RRG) increasingly relies on rich evidence, most commonly retrieved similar reports from other patients and longitudinal history from the same patient. Prior approaches often concatenate evidence into a long prompt, which can degrade performance. Such long contexts are typically noisy and only partially relevant, making it challenging for the backbone decoder to reliably select the most relevant evidence for the current query. We propose HM-RRG, a hierarchical-memory framework that organizes and selects evidence instead of treating all context as a flat sequence. HM-RRG builds an image-conditioned retrieval memory bank by compressing each retrieved report into a compact memory vector grounded on the current image. It then processes the patient’s prior reports as a chronological stream, maintains segment-level longitudinal memories, and performs hierarchical retrieval over the union of cross-patient and within-patient memories to form an evidence prompt for each segment during decoding. This design enables selective recall of relevant information while suppressing irrelevant or conflicting details under long contexts. Experiments on MIMIC-CXR with Longitudinal-MIMIC show that HM-RRG improves both text quality and clinical accuracy. In particular, HM-RRG achieves the best factuality metrics among state-of-the-art RRG models, demonstrating that structured memory is effective for radiology report generation. Our code is available at https://github.com/QingyueJ-nd/HM-RRG.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1067_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/QingyueJ-nd/HM-RRG
Link to the Dataset(s)
MIMIC-CXR dataset: https://physionet.org/content/mimic-cxr-jpg/2.1.0/
BibTex
@InProceedings{JiaQin_Hierarchical_MICCAI2026,
author = { Jiao, Qingyue AND Zheng, Jiahao AND Zhuang, Jun AND Shi, Yiyu},
title = { { Hierarchical Memory for Radiology Report Generation with Rich Context } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16878},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes Hierarchical Memory Radiology Report Generation (HM-RRG), a novel framework designed to integrate both cross-patient retrieved reports and within-patient longitudinal history to generate factually grounded radiology reports. To solve the issue of long, noisy context degradation, HM-RRG compresses retrieved reports into image-conditioned long-term memory vectors. It then processes the patient’s prior reports as a chronological sequence to maintain short-term, segment-level longitudinal memories. Finally, it utilizes a hierarchical retrieval mechanism to form an evidence prompt during decoding, allowing the model to selectively recall relevant information while suppressing irrelevant noise.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The proposed method demonstrates clear motivation and clinical relevance by accurately mirroring the real-world workflow of radiologists, who selectively recall information to identify changes and refine diagnoses rather than processing patient history and similar cases as a flat sequence of text. To achieve this, the paper features a highly targeted architectural design that adeptly adapts the Hierarchical Memory Transformer. Specifically, the strategic choices to retain all segment memories—thereby preventing the loss of vital clinical temporal changes —and to utilize image-conditioned multi-modal soft prompts for filtering retrieval noise are particularly effective for the medical domain. Consequently, the framework yields strong empirical results, achieving state-of-the-art clinical accuracy on the MIMIC-CXR dataset and significantly outperforming prior multi-view and longitudinal baselines.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
Methodologically, the core algorithmic contribution is somewhat incremental; although integrating image-conditioned soft prompt tuning for radiology reports is beneficial, the fundamental memory structuring mechanism heavily relies on the existing HMT framework and lacks substantial theoretical innovation. Furthermore, the ablation studies reveal an unusually large, unexplained performance surge when transitioning from a flat concatenation baseline to the HM-RRG framework. Such a significant improvement requires deeper quantitative or qualitative investigations, such as attention visualizations, to definitively prove how the memory structure successfully suppresses noise. Additionally, the paper lacks a much-needed computational overhead analysis. Maintaining complex hierarchical memories and calculating cross-attention inevitably introduces processing burdens, necessitating a clear comparison of inference latency, memory footprint, and FLOPs against the flat baseline. Finally, the study’s reliance on a single primary dataset family limits the demonstration of the model’s generalization capabilities.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper tackles a highly relevant and timely problem in medical vision-language models: the dilution of critical signals and the introduction of noise when using long, flat contexts for retrieval-augmented and longitudinal generation tasks. The HM-RRG framework offers a logically sound, clinically inspired solution to this problem. The experimental methodology is highly rigorous, explicitly preventing patient-level data leakage , and the reported improvements on key clinical metrics are substantial. I have given a “Weak Accept” primarily due to the somewhat incremental algorithmic novelty (given the reliance on HMT ) and the lack of a computational efficiency analysis. If the authors can adequately address the computational overhead and provide further insights into the ablation performance gaps during the rebuttal, I am open to raising my score.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
I have reviewed the authors’ response and find it satisfactory. The rebuttal effectively clarifies the technical contributions over text-only HMT and provides strong justification for the hierarchical memory mechanism over the flat long-context baseline in terms of both performance and computational efficiency. Additionally, the explicit details regarding reproducibility and training stages resolve my previous doubts. Since the core concerns have been successfully addressed, I firmly maintain my recommendation of Weak Accept.
Review #2
- Please describe the contribution of the paper
HM-RRG addresses the performance degradation caused by noisy, long-context evidence in Radiology Report Generation (RRG) by replacing flat sequence concatenation with a structured hierarchical-memory framework. By building an image-conditioned memory bank and processing longitudinal patient history as a chronological stream, the model enables selective recall of relevant cross-patient and within-patient evidence while effectively suppressing irrelevant or conflicting details during decoding. Evaluated on the MIMIC-CXR dataset, HM-RRG achieves state-of-the-art results in clinical accuracy and factuality, demonstrating that structured memory management is a superior alternative to traditional long-context prompts for generating reliable medical reports.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
Selective Memory Recall over Flat Context: By replacing noisy, long-context concatenation with a hierarchical memory structure, the model effectively suppresses irrelevant information and selectively recalls only the most pertinent evidence during decoding.
Unified Cross-Patient and Longitudinal Evidence: The framework systematically integrates similar cases from other patients with the current patient’s chronological history, ensuring that both global medical knowledge and specific temporal changes are grounded in the query image.
State-of-the-Art Factuality: HM-RRG achieves superior clinical accuracy and the highest factuality metrics on the MIMIC-CXR benchmark, proving its capability to generate reliable, evidence-based radiology reports that align with expert standards.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
Limited Reproducibility: The paper lacks critical information required for independent verification. To ensure scientific transparency, the authors should have provided a public repository containing the implementation code and data processing pipelines.
Narrow Evaluation Scope: The framework was evaluated on only a single dataset (MIMIC-CXR), which raises significant concerns regarding its generalizability and reliability across different clinical settings, patient demographics, or imaging modalities.
Insufficient Model Diversity: The study would be more convincing if it included experiments with a wider variety of state-of-the-art LLMs and VLMs. Evaluating the framework across diverse backbones is necessary to demonstrate that its performance gains are architecturally robust and not dependent on a specific model configuration.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses a highly significant and practical challenge in Radiology Report Generation (RRG), the management of noisy, long-context evidence derived from both cross-patient similarities and within-patient longitudinal histories. The proposed hierarchical-memory framework offers a well-motivated and structured approach to selective information recall, providing a clear technical path to improving clinical factuality by suppressing irrelevant context. This methodological shift from flat sequences to structured memory is a timely and important contribution to the field.
However, the primary weakness lies in the limited scope of the experiments. Relying on a single dataset and a narrow selection of model backbones makes it difficult to fully assess the generalizability and robustness of the framework across diverse clinical settings. Despite these empirical limitations, the importance of the problem and the soundness of the methodology warrant a positive, albeit cautious, recommendation.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
This study demonstrated sufficient originality and technical merit. The authors’ responses have addressed a number of concerns.
Review #3
- Please describe the contribution of the paper
The paper proposes a hierarchical-memory framework for radiology report generation that jointly uses two clinically meaningful but structurally different evidence sources: retrieved reports from other patients and longitudinal reports from the same patient. Instead of flattening everything into one long prompt, HM-RRG builds an image-conditioned retrieval memory bank for cross-patient evidence and a chronological segment-level memory stream for within-patient history, then performs hierarchical retrieval during decoding to select the most relevant evidence for each segment of the output report. The interesting part is not just “using retrieval plus history,” but the explicit claim that these two sources should be modeled differently because one is an unordered noisy set and the other is a temporally ordered stream.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1,The key strength is the alignment between the paper’s motivation and its architecture: retrieved reports are compressed into image-conditioned memory vectors, longitudinal priors are modeled as a temporal memory stream, and decoding adaptively queries both sources. 2, The work is further strengthened by ablations showing that simple context concatenation is insufficient, while the proposed memory structure substantially improves clinical factuality metrics such as RadGraph and CheXbert F1.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1,The paper says it follows Longitudinal-MIMIC, but it also changes the construction of each example by aggregating all available prior studies into a chronological stream, whereas the original setup is more limited. At the same time, Table 1 states that many baseline results are cited from the original publications and MLRG, rather than rerun under the paper’s modified history construction. This weakens the fairness of the headline comparison, because the proposed model is evaluated under a richer input formulation while prior methods are largely compared using previously reported numbers. 2,There is some inconsistency for metric reporting. In the experiment text, the paper says it reports average BLEU, ROUGE-L, and BERTScore for NLG, plus CheXbert and RadGraph for clinical accuracy. But Table 1 reports BLEU, MTR, ROUGE-L, RG, P, R, F1, with no BERTScore;In addition, while the training section contains many hyperparameters, some implementation details remain unclear, such as the exact Stage-2 optimization behavior and how much of the decoder remains trainable during hierarchical-memory training. These issues do not make the paper irreproducible, but they reduce confidence.
3,Although the paper argues for a general hierarchical-memory mechanism for rich-context report generation, the empirical evidence is concentrated in essentially one benchmark ecosystem: MIMIC-CXR / Longitudinal-MIMIC. There is no external dataset validation, no robustness study across different retrieval settings or history lengths beyond the main ablation, and no evidence that the approach generalizes to other report-generation settings or modalities.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I recommend weak reject because the paper is well motivated and shows promising gains in clinical factuality, but the empirical support is not fully convincing. In particular, the comparison to prior work is weakened by differences in data construction and reliance on previously reported baseline numbers rather than controlled reruns. The validation is also relatively narrow, and some presentation/reproducibility inconsistencies reduce confidence. Overall, this is a thoughtful and potentially valuable idea, but the current evidence is not yet strong enough for acceptance.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
I am moving from my score of Weak Reject (3) to Weak Accept. My original review noted that I would not mind if the paper were accepted, dependent on rebuttal, and the rebuttal makes meaningful progress on two of my three concerns. While one issue remains incompletely resolved, on balance I believe the paper now meets the threshold.
Reproducibility concerns (my original W2) are substantively addressed. The authors explicitly acknowledge the BERTScore inconsistency as a presentation error and commit to correcting it. They also provide specific Stage-2 implementation details that resolve what I flagged: trainable components (decoder LoRA weights, image projector, memory/summary tokens, memory-retrieval projections) versus frozen components (base decoder weights, CXR-CLIP vision encoder); segment length L=128 with j=L/2=64; cache size equal to the number of longitudinal segments; and the Stage-2 initialization from Stage-1 checkpoint. Combined with the commitment to release preprocessing, retrieval, training, and evaluation code, this clarification meaningfully improves verifiability. Dataset scope (my original W3) has a defensible justification. The authors point out that MIMIC-CXR / Longitudinal-MIMIC is currently the only public resource combining paired chest X-rays, reports, and longitudinal histories, and that the closest comparable works MLRG and Diff-RRG (both MICCAI 2025) also evaluate only on this benchmark. This is factually accurate as a state of the field, and the authors commit to discussing external validation as a stated limitation. While broader evaluation would be ideal, this constraint reflects data availability in the RRG community rather than a methodological choice that could have been easily avoided. The comparison-fairness concern (my original W1) is only partially addressed. The authors defend against the argument that baselines should be modified to consume cross-patient retrieved reports, noting that this would require major architectural changes. However, this does not directly address the specific issue I raised: the paper aggregates all available prior studies into a chronological stream per example, whereas the standard Longitudinal-MIMIC setup uses only the most recent prior. For longitudinal baselines that already accept variable-length prior sequences, evaluating HM-RRG under the standard single-prior setup would have offered a cleaner fairness check. I would have welcomed at least an acknowledgment of this specific design choice. That said, the input modification is disclosed in the experimental setup of the submission, and the authors’ Q7 response commits to varying history length in future robustness studies.
On balance, the paper has a clear clinical motivation that mirrors radiologist workflows, a principled architectural separation of unordered cross-patient retrieval from ordered same-patient priors, and ablations that show structured memory meaningfully outperforms flat context concatenation on clinical accuracy metrics.
Author Feedback
We thank all reviewers for their valuable comments. Q1: Dataset/backbone scope (R#1, #2, #3). MIMIC-CXR was chosen because HM-RRG requires paired chest X-rays, reports, and longitudinal histories, and it is the largest public resource for this setting. Closest prior works, MLRG [8] and Diff-RRG [18] from MICCAI 2025, also evaluate only on MIMIC-CXR. We will discuss external validation as a limitation. We used BioMistral-7B with CXR-CLIP to enable comparison with Diff-RRG [18] and isolate hierarchical memory. Q2: Reproducibility details (R#2, #3). The BERTScore issue is a presentation error, and we will correct it. We will clarify Stage-2 optimization: decoder LoRA weights, image projector, memory/summary tokens, and memory-retrieval projections are trainable, while the base decoder and CXR-CLIP vision encoder remain frozen. Stage 2 loads the Stage-1 checkpoint and continues training with hierarchical retrieval enabled. We use one memory vector per retrieved report and one segment memory vector per longitudinal segment. Segment representations use L=128 and j=L/2=64.The cache size equals the number of longitudinal segments. We will release preprocessing, retrieval, training, and evaluation code upon acceptance. Q3: Improvement over flat concatenation (R#1). The ablation shows that richer evidence alone is insufficient: flatly concatenating priors and top-5 retrieved reports degrades performance, suggesting that unordered retrieved reports introduce irrelevant or conflicting findings. HM-RRG compresses retrieved reports into current-image-conditioned memory vectors, processes priors as chronological segment memories, and lets each segment query the unified memory to form a compact evidence prompt. This makes evidence selection explicit rather than leaving it to the decoder over a long noisy sequence. Fig. 2 shows HM-RRG preserving relevant disease-progression information while suppressing irrelevant retrieved-report content. We will expand this case discussion with internal evidence-selection steps. Q4: Computational overhead (R#1). Even though HM-RRG adds memory construction and querying, it greatly reduces decoder memory/computation cost compared with full flat long-context decoding. The flat baseline feeds all retrieved and longitudinal reports as one long sequence, whose self-attention memory grows with length. HM-RRG instead compresses retrieved reports into image-conditioned memory vectors, stores priors as segment memories, and attends over this compact memory set to form an evidence prompt. Thus, the decoder avoids processing the full concatenated evidence. We will discuss this with data in the final version. Q5: Contribution beyond HMT (R#1). HMT is text-only, while HM-RRG adapts hierarchical memory to multimodal RRG which is nontrivial: Retrieved reports are compressed into memory vectors conditioned on the current chest X-ray via projected CXR-CLIP features, not stored as generic text memories. HM-RRG also separates unordered cross-patient retrieval from temporally ordered same-patient priors, targeting the RRG-specific challenge of using rich context to avoid missed findings and filter irrelevant information. Q6: Fairness of comparison (R#3). Existing longitudinal RRG methods use different inputs: MLRG [8] uses multi-view longitudinal information and Diff-RRG [18] models disease progression, but neither uses cross-patient retrieved reports. Modifying them to consume additional data types/sources would require major architectural changes, thus no longer faithful reproductions. Note that these works also directly cite results from earlier baselines when the inputs are different. Q7: Robustness (R#3). We will discuss the fixed retrieval setting as a limitation and expand robustness analysis in future work by varying retrieval K, history length (latest prior vs. multiple/all priors), and retrieval noise to test stability as evidence becomes longer or noisier.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The paper proposes a hierarchical-memory framework for radiology report generation. It replaces flat sequence concatenation with a structured hierarchical-memory framework integrating cross-patient retrieved reports and within-patient longitudinal history. Experiments on MIMIC-CXR show strong gains on clinical accuracy and NLG metrics over prior work, supported by ablations. The paper is clinically well-motivated and mirrors radiologists’ reasoning over prior studies and similar cases via selective recall of relevant findings and changes rather than uniform processing of all contexts. A key conceptual contribution is the modelling of retrieved reports as a compressed, image-conditioned memory bank and longitudinal priors as a temporal stream. Experimental evidence on MIMIC-CXR demonstrates that structured memory management is a superior alternative to traditional long-context prompts for generating reliable medical reports. However, several concerns should be addressed: 1) A deeper analysis of results (e.g. qualitative examples, failure analysis, attention visualizations) would help explain the unusually large improvement from flat concatenation to HM-RRG observed in ablations. 2) A computational overhead analysis (e.g. latency, memory, FLOPs) is important, as hierarchical memory may be costly relative to a flat baseline. 3) Validation is restricted to the MIMIC-CXR dataset; experiments on external datasets would strengthen the generality of the claims. Evaluating across a broader range of state-of-the-art LLM/VLM backbones would also help demonstrate that gains are robust across architectures. 4) The paper should clarify exactly how each baseline is evaluated and ideally provide controlled reruns under the same setup. 5) Metric reporting should be made consistent (e.g., BERTScore mentioned but not shown; differing metric abbreviations). 6) There are missing implementation details in the paper leading to limited reproducibility. Also, providing a public repository would substantially strengthen verifiability and reproducibility.
Given the strong clinical motivation and empirical performance, I recommend inviting the paper for rebuttal and request authors to address the reviewers’ concerns.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
This paper addresses an important and clinically meaningful problem in radiology report generation, namely, how to integrate two heterogeneous sources of supporting evidence, cross-patient retrieved reports and within-patient longitudinal history. The proposed hierarchical-memory design is well aligned with this motivation, and reviewers broadly recognize the strength of the clinical intuition and the empirical gains on MIMIC-CXR. After the rebuttal, the reviewer consensus became more positive, with most major concerns addressed. The main remaining concern is comparison fairness, as this point was only partially resolved. I agree this is a legitimate limitation and should be acknowledged explicitly in the final version. That said, the rebuttal addresses the issues most critical to the acceptance decision.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The authors have addressed the reviewers’ comments very well. All reviewers have now agreed on acceptance of the paper.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The authors’ rebuttal successfully addresses the major concerns raised by the three reviewers. All reviewers upgraded their initial scores (from Weak Accept/Weak Reject to Accept) after reading the response. The rebuttal clarifies reproducibility details (Stage‑2 trainable components, segment length, cache size, code release) and corrects the BERTScore presentation error. It provides a convincing explanation for the large improvement over flat concatenation, highlighting how image‑conditioned compression and explicit evidence selection mitigate noise from unordered retrieved reports. Computational overhead is justified by noting that the hierarchical memory reduces decoder self‑attention costs compared to processing full long‑context sequences. The rebuttal also acknowledges external validation as a limitation. While Reviewer #3 noted that the fairness of comparison (using all prior studies vs. the standard single prior) was only partially resolved, the overall methodological soundness, clinical motivation, and strong empirical results outweigh this concern. Thus, the paper is voted for acceptance.