List of Papers Browse by Subject Areas Author List
Abstract
Recent years have seen substantial advances in radiology report generation (RRG), yet existing approaches predominantly adopt direct feature fusion when handling multi-view X-ray images. Such approaches overlook the potential clinical inconsistencies and inaccuracies arising when a single model processes different views, adversely impacting performance and clinical reliability. To this end, we introduce View-PNDF (View-specific Pattern Neuron Detection and Fine-tuning), a parameter-efficient framework that fosters view-consistent report generation from a neuronal perspective. Specifically, View-PNDF comprises: (i) a view-specific neuron detection module identifying neurons responsive to particular views, (ii) a verification module quantifying the existence of these neurons, and (iii) a selective fine-tuning strategy strengthening detected neurons while preserving view-agnostic representations. By updating only view-specific neurons, View-PNDF achieves consistent diagnoses across different views with reduced computational costs. Subsequently, we employ Large Language Models (LLMs) to consolidate the view-specific reports into a complete radiology report. Furthermore, we use traditional Natural Language Generation (NLG) metrics-based assessment on integrated reports for baseline comparison and employ LLM-based assessment (e.g., GPT-4o) on view-specific reports to capture clinical significance. Extensive experiments on two medical RRG benchmarks demonstrate that View-PNDF substantially improves view-specific chest X-ray report generation quality while maintaining robust general-view performance.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1298_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{CheYuc_Seeing_MICCAI2026,
author = { Chen, Yucheng AND Zhu, Jinjing AND Yu, Yang AND Shi, Yufei AND Naghshbandi, Hane AND Liu, Jinhua AND Koh, Angela S. AND Fen, Fang AND Ong, Kian Eng AND Yeo, Si Yong},
title = { { Seeing Through Multiple Views: Parameter-Efficient Fine-Tuning via Selective Neurons for Consistent Radiology Report Generation } },
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 tries to address the generalizability of VLM models for clinical report generation through a simple but effective method. Through VND and VNV, the model’s ability in dealing different views is detected and verified. Then through VNF, the specific neurons are finetuned and improved accordingly. The performance gain proves the effectiveness of the method.
- 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 idea is straightforward and effective. It is important to identify specific neurons and improve accordingly to improve the model without much training effort. 2.Solid ablation study according to the performance gain and hyperparameter.
- 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 usage and description of the LLM-based evaluation method is not clear. Why is this needed when other metrics (BLEU, METEOR) are employed? Table 3 and Table 4 are hard to understand. 2.It is important to note that the authors are NOT trying to deal with the problem when one single case has two views and summarized into one report. This is not clearly described throughout the paper.
- 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
More details are needed for reproducibility. For example, for attention layers, how do you decide the neurons? Do you select entire heads, specific projection channels (Q/K/V/O), or output dimensions? How is deactivation implemented in each case?
- 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 method provided is simple but effective. This method could be extended to other topics when part of a pattern works bad within the VLM and could be improved specificly. However, some part of the paper (especially about the LLM evaluation) is not clear and need improvement.
- 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.
The authors have addressed my concerns. Besides, regarding other reviewers, the rebuttal is clean and persuasive. There is no issue when applying on large dataset. However, I insist on the format problems in Table 3 and 4, and the space seems to be modified in section 3.1 and 3.4.I don’t want to reject this paper because of this format issue, but this format issue must be corrected in the final version, otherwise it may set an unfair example.
Review #2
- Please describe the contribution of the paper
The paper addresses the inconsistency problem in multi-view medical imaging for radiology report generation. The authors proposed a view-specific neuron detection, verification, and fine-tuning framework, which is able to capture view-specific neurons and improve the performance for multi-view RRG task.
- 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 paper addresses a relatively novel and meaningful problem in multi-view radiology report generation. The motivation is clear, and the hypothesis about view-specific neurons is reasonable and interesting. 2.The experimental evaluation is also fairly comprehensive, with validation across multiple models, datasets, and metrics. This provides useful empirical support for the proposed method.
- 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.
Relatively low visual quality of the figure. Several shapes and elements appear to be poorly aligned. Although this does not substantially affect the understanding of the proposed method, improving the figure is strongly recommended.
- Please rate the clarity and organization of this paper
Good
- 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 presents a reasonably novel idea for addressing inconsistency issue for multi-view radiology report generation task, and the experimental design is relatively comprehensive. Although there are some minor weaknesses, such as figure quality and the need for clearer experimental details, these issues do not significantly detract from the overall contribution of the paper.
- 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.
In my view, the manuscript demonstrates a certain degree of novelty and makes a meaningful contribution to the field. The authors have provided a well-prepared response to the reviewers’ comments and have adequately addressed the major concerns raised during the review process.
Although some imperfections remain in the formatting of the figures and tables, these issues do not undermine the overall quality and significance of the work. The revisions and explanations provided by the authors are satisfactory. Therefore, I recommend acceptance of the manuscript.
Review #3
- Please describe the contribution of the paper
This work introduces ViewPNDF (View-specific Pattern Neuron Detection and Fine-tuning), a parameter-efficient framework that fosters view-consistent report generation from a neuronal perspective. Specifically, View-PNDF comprises:(i) a view-specific neuron detection module identifying neurons responsive to particular views, (ii) a verification module quantifying the existence of these neurons, and (iii) a selective fine-tuning strategy strengthening detected neurons while preserving view-agnostic representations.
- 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.
Extensive experiments on two medical RRG benchmarks demonstrate that View-PNDF substantially improves viewspecific chest X-ray report generation quality while maintaining robust general-view performance.
- 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.
It seems that this work is simply a two-task model by decomposing the report generation with front-view and lateral-view report generation seperately. It also lacks of validation on large datasets compare to previous models(for example,MIMIC-X）.
- 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 mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure 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.
(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?
The novelty of this method seems limited.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
As the submission does not provide sufficient information for reproducibility, I still tend to reject this work.
Author Feedback
We thank R1, R2, R3, and the Meta-Reviewer for their constructive feedback on our work, particularly regarding the methodological contribution (R1, Meta), the clinically meaningful problem formulation (R2, R3), and the comprehensive experimental evaluation (R1, R2, Meta). We address the reviewers’ comments below, organized by topic. R3 W1, Meta: Novelty beyond a two-task model. We respectfully clarify that View-PNDF is fundamentally different from a two-task decomposition framework. Rather than training separate view-specific generators, the novelty of View-PNDF is different: it analyses cross-view consistency at the neuron-adaptation level within a single shared VLM. The same model is adapted to produce clinically coherent findings across both frontal and lateral views. The view-specific descriptions are then consolidated by an LLM into a unified report, mimicking the radiologist workflow. This design is particularly valuable in realistic scenarios where one view is missing or unclear due to trauma or positioning constraints. To our knowledge, this is the first neuron-level study of cross-view consistency in radiology report generation. R1 W2: Multi-view Setting. We apologize for the ambiguity in the manuscript. The objective is not to develop a conventional framework that simply combines frontal and lateral images from the same case into one report. Our work instead focuses on cross-view consistency within a shared VLM. The model generates view-specific findings for each available view, then subsequently consolidated into a unified report via an LLM. Our motivation is cross-view consistency within one shared model, together with robustness in realistic cases where one view is missing, unclear, or degraded due to trauma, positioning, or acquisition quality, so that coherent findings can still be produced from the remaining view without contradictory interpretations. We have made the revision to clarify this setting and motivation. R1 W1, Meta: LLM-based Evaluation and Tables. The two schemes are meant to be complementary. NLG metrics measure n-gram overlap with the GT report shared by both views. Tab. 3 shows that scores before and after fine-tuning are nearly identical, indicating that NLG metrics fail to capture subtle yet clinically important differences in view-specific quality. To address this, we employ LLM-as-a-judge for view-specific semantic fidelity. Tab. 4 uses three independent LLM evaluators, which produce highly consistent scores, showing that the evaluation is not biased toward any single judge. This will be clarified in the revision. R1 W3, Meta: Reproducibility. In our model, neurons are defined at the column level of the projection matrices, not at the attention-head level. For FFN layers, a neuron is one hidden activation channel, and deactivation zeros that channel. For self-attention layers, neurons are identified separately in the query, key, and value projections, each as one column of the respective projection matrix. Deactivation sets the corresponding output dimension of the query, key, or value tensor to zero, while the output projection remains unchanged. We will include more details in the revision and release the code upon acceptance. R3 W2: Validation on Large Datasets. Our experiments are reported on both IU-Xray (Tab. 1) and the large-scale MIMIC-CXR (Tab. 5), where View-PNDF consistently outperforms the strongest baseline MLRG on both NLG and CE across all four VLM backbones. Tab. 2 further reports view-specific GPT-4o scores before and after fine-tuning, with lateral-view performance improving substantially while frontal-view quality is preserved. We apologize that space constraints placed these tables far apart and left dataset labels missing in some captions, both of which we will clarify in the revised version. R2 W1: Figure Quality. We thank R2 for the suggestion and will improve Fig. 1 in the revision with better alignment, cleaner layout, and improved typography for better readability.
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.
This paper proposes View-PNDF, a parameter-efficient fine-tuning framework for improving view-consistent radiology report generation by detecting, verifying, and selectively fine-tuning view-specific neurons. The paper received moderately positive reviews with scores of 4, 4, and 3.Reviewers appreciated the meaningful problem setting, the simple and effective idea, and the relatively comprehensive experiments across models, datasets, and metrics. The main concerns are about limited novelty, unclear explanation of the LLM-based evaluation, insufficient reproducibility details, the need to better clarify the exact multi-view setting, and low visual quality of the figure. Overall, the paper presents a useful methodological contribution, and I thus invite the authors for rebuttal 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.
The paper received split post-rebuttal feedback (two Accepts, one Reject), with reviewers praising its clinical motivation, technical soundness, and comprehensive evaluations across four VLM backbones. In the rebuttal, the authors successfully addressed the main concerns: they clarified that View-PNDF is the first neuron-level study of cross-view consistency in a single shared VLM, not a basic two-task model; they justified their use of multi-judge LLM evaluations by showing that traditional NLG metrics fail to capture critical clinical anomalies; and they resolved ambiguities about neuron definition and deactivation at the column level of projection matrices. Validated on IU-Xray and MIMIC-CXR, with commitments to fix formatting and improve figure typography, the framework meets publication standards and is recommended for acceptance as a poster.
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 well. Two reviewers acknowledged the novelty and significance of the work. Therefore acceptance of the paper is concluded.
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 rebuttal clarifies the multi-view setting, LLM-based evaluation, large-dataset validation on MIMIC-CXR, and implementation details. With two reviewers supporting acceptance and remaining concerns mostly related to presentation, formatting, and clarity, I recommend weak acceptance but would not mind if rejected.