Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

In proton therapy planning, respiratory-gated non-contrast CT (NCCT) is commonly used for lesion segmentation; however, accurate delineation remains challenging due to low lesion-to-background contrast. Although learning-based methods have shown strong performance, they often struggle with non-contrast image segmentation. Inspired by clinical practice, where contrast-enhanced MRI is referenced to delineate lesions on NCCT, we propose ViPSAM, a visual prompting framework that leverages complementary cross-modality information. Built upon the Segment Anything Model (SAM), ViPSAM introduces a visual prompt encoder to extract guidance features from contrast-enhanced images and a visual-guided cross-attention module to integrate non-contrast and contrast-enhanced features, thereby enhancing lesion-relevant representations in low-contrast regions. The mask decoder is further adapted in a parameter-efficient manner to utilize visual prompts effectively. We evaluate the proposed method on liver lesion segmentation using NCCT acquired for proton therapy. Experimental results demonstrate that ViPSAM outperforms representative U-Net- and SAM-based methods, indicating that cross-modality visual prompting enables more robust and accurate segmentation in non-contrast images.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3107_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/torchViPSAM/ViPSAM

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LeeSan_ViPSAM_MICCAI2026,
        author = { Lee, San AND Kim, Nalee AND Yu, Jeong Il AND Park, Hee Chul AND Kim, Boah},
        title = { { ViPSAM: Visual Prompting Medical Image Segmentation Using Segment Anything Model } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16883},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes ViPSAM, a SAM-based medical image segmentation framework for lesion delineation on non-contrast CT by leveraging contrast-enhanced MRI as visual prompts. The method introduces a visual prompt encoder, a visual-guided cross-attention module to integrate cross-modality information, and LoRA-based parameter-efficient adaptation of the SAM decoder. The approach is evaluated for liver lesion segmentation in proton therapy planning.

  • 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 paper addresses an important and clinically relevant problem: accurate lesion delineation on non-contrast CT for proton therapy planning, where low contrast makes segmentation particularly challenging.
    • The proposed method is well motivated by real clinical workflow, where contrast-enhanced MRI is commonly referenced during contouring on planning CT. Translating this practice into a learning framework is both intuitive and impactful.
    • The technical contribution is coherent and well designed. The visual prompt encoder, cross-attention fusion module, and LoRA-based adaptation form a consistent framework rather than disconnected components.
    • Experimental results are strong, especially for liver lesion segmentation, with clear improvements over representative U-Net- and SAM-based baselines.
    • The paper includes useful ablation studies and statistical significance analysis, which strengthen the empirical evidence.
    • Parameter-efficient adaptation with a small number of trainable parameters is an additional practical advantage.
    • The manuscript is clearly written and easy to follow.
  • 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.
    • The dataset is relatively small and originates from a single institution, which may limit conclusions regarding broader generalization.
    • The held-out test cohort is limited in patient count, and validation on larger external datasets would further strengthen the study.
    • Since ViPSAM leverages additional MRI information, comparisons with other multimodal fusion approaches would be valuable to better isolate the benefit of the proposed visual prompting design.
    • The current formulation is slice-based; incorporating 3D contextual information could be an interesting future direction for volumetric lesion segmentation.
    • The framework depends on CT-MRI registration quality, and robustness to imperfect alignment was not explicitly analyzed.
  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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

    This is a strong and clinically meaningful submission. The proposed framework is well motivated, technically sound, and demonstrates clear benefits for a challenging segmentation task. Future work validating the method on larger multi-center datasets and comparing with additional multimodal baselines would further strengthen the impact.

  • 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.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    I recommend acceptance because this paper presents a well-motivated and technically solid contribution to an important clinical segmentation problem. The idea of leveraging contrast-enhanced MRI as visual prompts for non-contrast CT segmentation is novel, practical, and strongly grounded in real clinical workflow. The proposed architecture is coherent, the empirical improvements are substantial, and the study includes appropriate ablation and statistical analyses. Although the dataset size is limited and broader external validation would be desirable, the overall quality, relevance, and potential impact of the work place it clearly above the acceptance threshold.

  • 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 provided a clear and constructive rebuttal addressing the main concerns raised during review. The additional clarifications further strengthen the paper, and I maintain my recommendation for acceptance.



Review #2

  • Please describe the contribution of the paper

    This paper introduces ViPSAM, a SAM-based framework designed for lesion segmentation in non contrast medical images such as NCCT. The main idea is to use paired contrast-enhanced images (e.g., MRI) as guidance through a form of cross-modality visual prompting. The method includes a visual prompt encoder, a visual-guided cross-attention module, and LoRA based parameter efficient fine tuning applied to SAM. The experiments on a liver lesion dataset show improvements over both U-Net and SAM-based baselines.

  • 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 an important and clinically relevant problem, as segmenting lesions in low contrast modalities remains highly challenging. The idea of incorporating cross modality information through visual prompting is interesting and potentially impactful.

    2.Additionally, the use of LoRA for parameter-efficient adaptation is a practical and timely contribution, especially given the computational cost of fine-tuning large foundation models.

    3.The experimental section is generally solid, with both quantitative and qualitative evaluations. The inclusion of ablation studies is appreciated and helps demonstrate consistent improvements across components.

  • 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 description of the visual prompt encoder is somewhat brief, making it difficult to understand how it extracts “soft-tissue contrast cues.” It would be helpful to clarify its architecture (e.g., CNN, transformer, or hybrid) and the supervision signal used during training. 2.Although the formulation of Q/K/V is mentioned, the intuition behind the design is not fully explained. A more detailed discussion of why this specific attention formulation is chosen, and how it differs conceptually from standard cross attention, would improve clarity. 3.The distinction between token-to-image attention and image-to-token attention is introduced, but the motivation for using both is not clearly justified. It would be valuable to explain the role of each mechanism and how they complement one another. 4.The concept of a “visual prompt” is central to the paper, yet it is not rigorously defined. The relationship between visual prompts and sparse prompts is somewhat unclear, and the interactions among the different components (image prompt, sparse prompt, LoRA) could be described more explicitly.

  • 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?

    My overall recommendation of Weak Reject (marginally below the acceptance threshold) is based on a balance between the paper’s promising idea and several concerns regarding clarity and technical justification.

    1.The paper addresses lesion segmentation in low-contrast non-contrast imaging (e.g., NCCT), which is clinically important and genuinely challenging. This gives the work strong practical relevance. 2.The idea of leveraging paired contrast-enhanced modalities (e.g., MRI) as a form of visual prompting for improving segmentation is interesting and could be impactful if better developed and validated. 3.The use of LoRA for parameter-efficient fine-tuning is appropriate and practical, especially when adapting large foundation models like SAM in a medical setting. 4.The reported results, including ablations, show consistent improvements over relevant baselines such as U-Net and SAM variants, which suggests the method is directionally effective.

    5.Several key components (especially the visual prompt encoder and visual-guided cross-attention module) are not clearly or rigorously described. As a result, it is difficult to fully understand what is being computed and why certain design choices were made. 6.The concept of “visual prompt” is central to the paper but is not well-defined. The distinction and interaction between visual prompts, sparse prompts, and other inputs (e.g., LoRA-adapted components) remain ambiguous. 7.Important design decisions such as using both token-to-image and image-to-token attention are introduced without sufficient intuition or empirical justification. This weakens confidence in whether the architecture is optimally designed or somewhat adhoc. 8.The proposed visual-guided cross-attention is not clearly distinguished from standard multimodal cross-attention, making it difficult to assess novelty at the architectural level.

  • 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’ rebuttal satisfactorily addressed most of the concerns raised in the initial review process. In particular, the response clarified several methodological details, strengthened the explanation of the experimental setup, and provided convincing arguments regarding the validity and significance of the proposed approach. The rebuttal also improved confidence in the reproducibility and practical relevance of the work.

    The paper presents a meaningful contribution to the field, with a technically sound approach and empirical results that demonstrate clear advantages over relevant baselines. The novelty is sufficient for acceptance, and the overall presentation is reasonably clear. While a few minor limitations and open questions remain, they do not substantially weaken the main contributions of the paper.

    After considering both the original submission and the authors’ responses, I believe the work meets the standard for acceptance and would be valuable to the research community.



Review #3

  • Please describe the contribution of the paper

    Authors propose a segmentation model for lesions in the liver, employing the Segment Anything Model (SAM), which authors adapt to their particular segmentation task. While non-contrast CT images form the basis for segmentation, complementary information from paired contrast-enhanced MR images is additionally leveraged. To this end, authors (1) use SAM’s frozen encoder to encode the CT image, (2) train an additional encoder to encode the MR image, (3) fuse both image embeddings in a cross-attention module, (4) fine-tune SAM’s segmentation mask decoder via low-rank adaptation (LoRA). The model is evaluated on a manually annotated in-house dataset and compared to several (CT-only) baselines. An ablation study is provided to justify the individual components of the proposed model.

  • 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 paper is well-written and easy to follow.
    • The choice of methods for adapting SAM to the task at hand (in particular, using cross-attention for modality fusion, and using LoRA for finetuning) appears creative and adequate.
    • Likewise, the choice of baseline methods appears adequate.
    • The results look convincing.
    • Authors provide the code of their approach.
  • 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.

    My major issue with the paper is that authors did not manage to convince me of the need for both the separate visual prompt encoder and the cross-attention module. Since SAM accepts 3-channel input, both MR and CT image data are 1-channel, and authors use registration to align MR and CT images anyway, a very obvious alternative processing approach, as it appears to me, would have been to just concatenate MR and CT data as different input channels and then process them with SAM’s encoder – which would have obviated the need for both aforementioned components (separate encoder, cross-attention). The same approach could, likewise, have been used for the baseline models in the authors’ comparisons of Table 1; at present, if I understand correctly, neither of the baseline models receives the MR data (all are CT-only), which, to me, makes their comparative under-performance very unsurprising. Thus, while I see the choice of baseline models adequate, I think their utilization could have been set up differently to ensure greater fairness. Moreover, a corresponding setup (i.e. using channel concatenation) could have been used as a baseline in the ablation study, so as to actually demonstrate the benefit of separate MR encoding and cross-attention. There might be, however, an obvious reason (from the authors’ perspective), as to why they did not consider channel concatenation, or I might have missed or misunderstood something in this regard, rendering it unfeasible. So, I hope for the authors to provide an argument as to why channel concatenation was not considered – an argument, which, in my opinion, can still be added with the final iteration of the paper.

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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?

    Authors neither mention acquiring consent from the imaged subjects nor measures used for anonymization/pseudonymization of used data.

  • 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
    • p. 3: “using box prompts as input” – what boxes are actually used (i.e. what do they contain) and where do authors get the boxes for the box prompts from (e.g. interactive manual annotation, automated preprocessing, …)?
    • p. 4: I_nc, I_ce are defined as (single) images, yet their shape B×3×H×W includes a batch size. Maybe I_nc, I_ce should rather be defined as batches of images?
    • p. 4: Eq. (7) introduces an MLP. What is the architecture of this MLP?
    • p. 6: None of the \tilde{W} variables are used anywhere else, so why are they introduced here? Authors should consider leaving them out, or else using them elsewhere (e.g. in Fig. 2).
    • p. 6: (1) what slicing direction was used for 2D processing with SAM (axial, coronal, sagittal)? (2) How was intensity normalization achieved (min-max scaling, z-scores, …)? (3) How was resizing achieved (rescaling, cropping, padding, …)?
    • p. 6: In my opinion, details on how the visual prompt encoder (i.e. the MR image encoder) is trained is missing (it is marked as trainable in Fig. 1, in contrast to the CT image encoder, which is marked as frozen). In particular: Is its training happening end-to-end with cross-attention and LoRA finetuning, or is it a separate step?
    • p. 7: Table 1 and the last sentence of the page mention 0.83M trainable parameters for the proposed approach. This makes sense to me, if only LoRA and cross-attention is concerned. But then again, as already mentioned, Fig. 1 shows also the MR image encoder as trainable. In my understanding, this is ViT-B, so its number of trainable parameters should be on the order of tens of millions. This apparent discrepancy should be clarified.
    • p. 6/7: How many blocks/layers are used for the cross-attention (VGCA) module?
    • p. 7: Details on the training: How exactly were Dice loss and cross-entropy loss combined? How many epochs, iterations per epoch? What was the criterion for “best-performing checkpoint” (smallest loss, highest Dice/IoU, smallest HD95, …)?
    • p. 7: How did authors ensure fairness in their baseline comparison? In particular, how were hyperparameters chosen for training the baseline methods?

    Typos, formatting, etc:

    • The naming in mathematical equations is inconsistent: The section Segment Anything Model (SAM) (p. 3) introduces X_image = E_image(I) (thus X being the encoder embedding), while eq. (1) uses F_image = E_image(I) and eq. (4) uses X_image = LN(Flat(F_image)) (thus F being the encoder embedding and X being its flattened, normalized version).
    • Names in mathematical variables, such as “image” in “E_image” are not variables themselves and should thus not be italicized.
  • 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?

    Authors, so far, could not convince me of the need for their separate MR image encoder and cross-attention module, as mentioned above. I would nevertheless encourage them to provide a corresponding compelling argument during rebuttal. Apart from that, I think the proposed method is sound, reasonably novel, and striking through its efficiency. Likewise, it has been thoroughly evaluated.

  • 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 would like to thank the authors for their concise, yet in my opinion to-the-point rebuttal. All essential criticisms of all reviewers have been sufficiently refuted, as I see it. In particular, I would like to thank the authors for their clarifications regarding the need for the separate visual prompt encoder and the successive cross-attention module, which was my main concern in my initial review. I agree that the two modalities play very different roles in the segmentation process, which could probably not have been reflected in the CT–MR concatenation baseline that I hypothesized in my initial review – this is a crucial motivation for the choice of components that I did not take into account. Nevertheless, I would encourage the authors to also explicitly include the discussion and decision against such a baseline, together with the observed lower Dice scores that authors mentioned in their rebuttal, into the final version of the paper. This can prevent other readers from stumbling upon the very same aspect and question, especially those readers that, like me, are not immediately familiar with the underlying clinical workflows. Likewise, I would encourage the authors to explicitly discuss the aspect of registration dependence, as pointed out by reviewer 1, in the final version of their paper. I would also like to thank the authors for pointing out the visual prompt encoder erroneously being marked as trainable in Fig. 1, rendering further of my initial comments obsolete. For clarity, perhaps I would also explicitly mention in the main text, Sect. 2.1, that this encoder is kept frozen.

    In the light of the convincing arguments provided in the rebuttal, I would like to revise my decision to “accept”. Most importantly, I agree with the strengths pointed out by reviewers 1 and 2: the strong practical relevance of the method, reflecting actual clinical workflows and providing a computationally tractable setup. While I also agree with reviewer 1 on the weaknesses of using a small, single-center dataset and showing no comparisons to multimodal baselines, I likewise do not see them as a “showstopper” towards a conference paper. I would encourage the authors to cover the latter aspects in an extended journal version of the paper instead.



Author Feedback

We thank all reviewers for their valuable comments and positive feedback on our motivation, technical novelty, and strong experimental results. We address the major concerns below.

  • Small dataset & Generalizability (R1): Since ViPSAM is designed for lesion delineation in proton therapy planning, we first evaluated it on a dataset collected from our clinical workflow. We agree broader validation is important and will evaluate larger multi-center datasets in future work.
  • Comparison of multimodal fusion models (R1, R3): Our dataset only provides lesion labels on NCCT, and NCCT–MRI pairs are not perfectly aligned. Thus, multimodal fusion methods requiring aligned data were not directly applicable. We will include additional multimodal fusion model comparisons on publicly aligned datasets in a journal version.
  • Dependency on registration quality (R1): Despite using only rigid registration, ViPSAM achieved strong performance (Table 1), suggesting robustness to imperfect alignment. We plan to extend the framework to a 3D formulation to better leverage volumetric context and reduce alignment sensitivity.
  • Details of visual prompt encoder (R2, R3): The visual prompt encoder uses the same ViT-B architecture as the image encoder with SAM pretrained weights and remains frozen during training. Here, “visual prompt” refers to image-based guidance from contrast-enhanced images that provide clearer lesion contrast, unlike sparse prompts (boxes). Accordingly, the visual prompt encoder extracts contrast-enhanced visual cues, which are fused with non-contrast image features through the VGCA module. Meanwhile, sparse box prompts independently provide localization cues in the mask decoder, while trainable LoRA layers adapt the mask decoder for final mask prediction. We will clarify this in the final version.
  • Attention formulation in VGCA (R2): In our VGCA module, we set F_image as queries(Q) because segmentation is performed on non-contrast images, which naturally query complementary information from contrast-enhanced images. Accordingly, we set F_visual as keys(K) and values(V), as contrast-enhanced images provide clearer lesion boundaries and richer contrast cues, making them effective K/V sources. Unlike standard cross-attention, our VGCA further introduces gating (γ) and learnable weighting (λ) to selectively regulate cross-modal information transfer. We will clarify this in the final version.
  • Details of two-way attention layers (R2): Token-to-image attention allows sparse prompt tokens to retrieve spatially relevant cues for localization, while image-to-token attention enables image features to attend back to sparse prompt tokens for further refinement. Using both attention layers enables bidirectional interaction between sparse prompts and dense image features.
  • Ablation study on concatenation of CT&MRI (R3): While CT and MRI images can be concatenated and directly processed by SAM, this does not explicitly consider their different roles (non-contrast image: target modality, contrast-enhanced image: auxiliary guidance). We also observed inferior lesion segmentation performance with this straightforward design (3.5% lower Dice score). Instead, ViPSAM separately encodes each modality and selectively incorporates informative contrast-enhanced features through the VGCA module.
  • Implementation clarifications (R3): We will revise the final version to clarify notation inconsistencies (X_image/F_image, W ̃), I_nc/I_ce shapes, MLP architecture, VGCA configurations, training epochs/iterations, loss weighting, checkpoint selection, and hyperparameter settings for baseline methods. We will explicitly describe that box prompts were generated from manual masks with random perturbations (≤20 px) during training and fixed during inference, along with additional data preprocessing details, including axial slicing, resizing, and min-max normalization. We will also correct Fig. 1, where the visual prompt encoder was incorrectly marked as trainable.



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.

    Reviews showed that this is an interesting paper but with multiple issues to be clarified in the rebuttal, such as small and single-source dataset, more detailed method description and justification (R2 and R3).

  • 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’ clear and constructive rebuttal has successfully turned two weak rejects into acceptance thus three reviews agreed on acceptance. The paper represents a clinically well-motivated and technically solid contribution for accurate lesion delineation on non-contrast CT for proton therapy planning.



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.

    ViPSAM introduces a highly innovative and clinically grounded SAM-based framework that mimics real-world proton therapy workflows by leveraging contrast-enhanced MRI as cross-modality visual prompts to segment low-contrast lesions on non-contrast planning CT scans. The authors’ rebuttal successfully defended the architecture against a simple channel-concatenation approach by demonstrating that treating the modalities symmetrically fails to capture their asymmetric roles, leading to inferior Dice scores, while also fixing a schematic error by confirming that the visual prompt encoder remains frozen to maintain the 0.83M parameter low-rank adaptation (LoRA) efficiency.



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.

    This paper proposes a SAM-based framework that leverages contrast-enhanced MRI as visual prompts for lesion segmentation on non-contrast CT. The problem is clinically important and the proposed approach is well aligned with real clinical workflows. After considering the reviews and rebuttal, I find that the authors have satisfactorily addressed the major concerns from all reviewers. The method is technically sound, demonstrates consistent improvements over relevant baselines, and is supported by ablation studies and statistical analysis. Therefore, I recommend Accept.



back to top