List of Papers Browse by Subject Areas Author List
Abstract
Accurate tumor segmentation is essential for tumor-centric clinical modeling, yet most existing methods remain organ-specific or rely primarily on imaging and generic semantic prompts, neglecting incorporating patient-specific clinically recorded tumor burden into spatial inference. In routine practice, electronic health record (EHR) systematically record lesion count, longest diameter, and coarse anatomical subregion, providing structured but indirect spatial cues. Mapping structured clinical descriptors to voxel-level tumor delineation is challenging in pan-tumor settings, where cross-organ variability expands the spatial hypothesis space and sparse EHR descriptors alone cannot uniquely determine tumor location. To address this, we propose ECLIPSE (EHR-Constrained Localized Inference for Pan-Tumor Segmentation). Specifically, through a dedicated Clinical-Constrained Hypothesis Generation (CCHG) module, ECLIPSE reformulates pan-tumor segmentation as a feasibility-constrained localization–refinement process. First, a tumor-type prompt–conditioned spatial response map is estimated within the decoder space of a pretrained model to generate dense tumor hypotheses. Second, CCHG injects structured EHR descriptors as soft spatial constraints to enforce lesion count consistency, size plausibility, and regional coherence, producing clinically consistent bounding box prompts. Finally, the inferred bounding boxes are encoded as spatial prompts to refine voxel-wise delineation. Across 14 public and private tumor datasets comprising 5,360 cases for training, evaluation and testing, ECLIPSE achieves a mean DSC of 61.41%, outperforming the strongest baseline by +7.44%. It further improves performance on mixed tumor data (+8.91%), private clinical data (+3.88%), and attains the best DSC (71.23%) on an unseen dataset, demonstrating robust generalization.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1253_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{MenRun_ECLIPSE_MICCAI2026,
author = { Meng, Runqi AND Liu, Zonglin AND Ma, Qiong AND Teng, Lin AND Wang, Yulin AND Song, Sifan AND Jin, Pengfei AND Sun, Yiqun AND Shen, Dinggang},
title = { { ECLIPSE: EHR-Constrained Localized Inference for Pan-Tumor Segmentation } },
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
This paper proposes ECLIPSE, a pan-tumor segmentation framework that incorporates structured clinical tumor descriptors into spatial inference rather than treating them as auxiliary multimodal inputs. The key idea is to use a Clinically-Constrained Hypothesis Generator (CCHG) to transform coarse tumor-type-conditioned visual response maps into clinically admissible bounding box hypotheses under anatomical, cardinality, and scale constraints, and then use these automatically generated box prompts to refine segmentation. Overall, the paper contributes a clinically motivated coarse-to-fine formulation that connects image evidence, structured EHR descriptors, and prompt-based refinement in a unified framework.
- 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 problem formulation is clear and clinically meaningful. Instead of vaguely claiming that EHR information may help segmentation, the paper identifies a more specific issue: image-only coarse spatial hypotheses may be clinically implausible, and structured descriptors such as subregion, lesion count, and tumor size can serve as feasibility constraints during hypothesis generation. 2.The proposed pipeline is logically coherent. The method follows a clear sequence of coarse visual localization, clinically constrained hypothesis selection, and prompt-based refinement. This is much more convincing than a loosely assembled multimodal design. 3.The role of structured clinical information is better motivated than in many prior multimodal segmentation papers. Here, EHR descriptors are assigned explicit functional roles in spatial inference, rather than being simply concatenated as additional features. 4.The empirical gains appear substantial across datasets, suggesting that the overall framework is effective and that structured clinical priors may indeed improve pan-tumor segmentation 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.
1.The main concern relates to pseudo EHR descriptors in experiments. For public datasets without available EHR, the authors generate pseudo descriptors from the ground-truth masks is understandable given the scarcity of public datasets with structured EHR. However, it remains important to clarify whether the private Colon-FD cohort uses real EHR-derived descriptors or similarly generated pseudo descriptors. And it would also be helpful to explain how the pseudo descriptors are generated from the ground-truth masks for the public datasets. If Colon-FD indeed uses real structured clinical information, then this concern is largely alleviated and the paper would have a much stronger clinical grounding. 2.Several implementation details are not sufficiently specified. In particular, the definition of the parameter α in box generation is unclear, the role of “FC” in the framework figure is not explicitly explained, and in formula 11 may contain a notation inconsistency or requires clarification. These issues do not appear to undermine the overall method, but they reduce clarity and reproducibility. Manuscript only states that a combined Dice and BCE loss is applied, without clearly defining the exact formulation, weighting, or whether both stages follow the same supervision scheme. 3.The ablation section could also be clearer. In Table 2, the exact configuration of the “Baseline” is not clearly defined. In addition, since CCHG is the core methodological contribution and consists of multiple feasibility constraints, the paper would be strengthened by including internal ablations within CCHG. This is a relatively minor issue compared with the main methodological concern above, but it would still help readers better understand what is being ablated.
- 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
One minor suggestion is to further improve the clarity of the method description. In particular, please clarify whether Eq. (11) should use the original response map S or the clinically constrained response map S-tilde, since the latter appears to be introduced by CCHG for feasibility-aware hypothesis generation. It would also be helpful to clarify how the pseudo EHR descriptors are generated from the ground-truth masks for the public datasets. In addition, please clarify whether the descriptors used for the private Colon-FD cohort are derived from real EHR records or similarly generated pseudo descriptors. These clarifications would further improve the paper’s readability and reproducibility.
- 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?
I recommend this paper because it presents a clinically meaningful pan-tumor segmentation framework with a coherent methodological pipeline. The motivation is reasonable, and the overall formulation is fairly well developed. In particular, the progression from coarse visual localization, to clinically constrained hypothesis generation, to prompt-based refinement is logically consistent and methodologically sound. The reported improvements across multiple tumor types and datasets also suggest that the framework is effective in practice.
My main concern is pseudo EHR descriptors. In particular, it should be clarified whether the private Colon-FD cohort uses real EHR-derived descriptors or pseudo descriptors generated from the masks. If the Colon-FD evaluation is based on real EHR-derived descriptors, I would assess the paper more positively.
- 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.
After reading the rebuttal, my main concern regarding pseudo EHR descriptors has been sufficiently addressed. In particular, the authors clarified that the private Colon-FD cohort uses real clinical EHR-derived descriptors rather than pseudo descriptors generated from ground-truth masks, which substantially strengthens the clinical validity of the framework.
Review #2
- Please describe the contribution of the paper
The paper presents EHR-Constrained Localized Inference for Pan-tumor Segmentation (ECLIPSE), a framework that integrates tumor burden descriptors from electronic health records (EHR) into spatial inference via a Clinical Constrained Hypothesis Generation (CCHG) module. The framework generates localization hypotheses through tumor type prompt-conditioned spatial response maps. The CCHG module encodes structured EHR descriptors as soft spatial feasibility constraints, producing clinically informed bounding box proposals that are automatically generated as prompts and fed back into the decoder for refinement.
- 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 method demonstrates consistent and clinically meaningful improvements over four established baselines across all seven tumor types The ablation study clearly demonstrates the necessity of the CCHG module The integration of EHR descriptors as soft spatial constraints is a clinically motivated and novel 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.
Major: The evaluation relies exclusively on DSC. The dataset composition is insufficiently described. Are some tumor types severely underrepresented? No limitations section is present.
Minor: Figure 1 requires significant magnification to read and Figure 2 lacks numbers and descriptions that are only introduced later in the text The CCHG module is introduced and described multiple times throughout the paper unnecessarily. The DSC gains are ranging from 3.88% to 8.91%, but the authors do not discuss the clinical relevance of these margins or cite supporting literature. Runtime is not reported, which is relevant for clinical deployment assessment The bibliography does not conform to MICCAI formatting guidelines.
- 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
The evaluation relies exclusively on DSC. Complementary metrics such as HD95 would provide a more complete picture of clinical utility, particularly given the pan-tumor setting where tumor sizes and boundary complexity vary significantly across types.
- 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 clinically motivated and technically sound framework for pan-tumor segmentation, with a novel integration of EHR-derived constraints into spatial inference. The method demonstrates consistent and meaningful improvements over four established baselines across all tumor types. The large and diverse dataset of contrast-enhanced CTs further supports the validity of the results. The weaknesses are largely minor and addressable. The most significant being the exclusive use of DSC, insufficient dataset description, missing limitations discussion, and unclear annotation provenance. None of these undermine the core contribution, and most can be addressed through clarification in the rebuttal. A weak accept is recommended.
- 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.
The authors addressed my comments in the rebuttal.
Review #3
- Please describe the contribution of the paper
The paper proposes ECLIPSE, an EHR-constrained framework for pan-tumor CT segmentation. Its central contribution is to shift the role of structured clinical records from passive auxiliary metadata to active spatial feasibility constraints during inference. Specifically, the method first produces a tumor-type prompt-conditioned response map, then uses structured descriptors such as anatomical subregion, lesion count, and longest diameter to select clinically admissible tumor hypotheses and convert them into bounding-box prompts for segmentation refinement. This formulation is interesting because it treats segmentation not only as voxel-wise appearance recognition, but also as a constrained localization problem conditioned on how tumors are actually documented in clinical practice. The paper further evaluates this idea across multiple tumor types and datasets, including mixed-tumor, private clinical, and unseen tumor settings.
- 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 most appealing aspect of the paper is its attempt to connect clinical reporting logic with spatial segmentation. Many recent universal or prompt-based medical segmentation methods still rely on generic image prompts, while this work asks a more clinically grounded question: if radiologists already record lesion count, approximate size, and anatomical subregion, can these descriptors narrow the spatial search space for tumor segmentation? This is a meaningful perspective, especially for pan-tumor segmentation where purely image-driven models may produce diffuse or anatomically implausible responses.
Another strength is that the proposed CCHG module has an intuitive role in the pipeline. The module does not merely concatenate EHR features with image features; instead, it operationalizes clinical descriptors as constraints on where and how many lesions should be proposed. This makes the method easier to understand than many black-box multimodal fusion strategies. The improvement from the baseline to baseline+CCHG is also large, suggesting that the constraint mechanism is doing something substantial rather than serving as a cosmetic component.
The paper also makes a reasonable effort to evaluate generalization across different tumor sites. The inclusion of an unseen adrenal tumor dataset and a private Colon-FD cohort gives the work more clinical breadth than a single-organ study. The qualitative results are consistent with the stated motivation: ECLIPSE appears to reduce anatomically implausible spillover and produce more localized tumor masks than several universal segmentation baselines. This supports the broader claim that structured clinical priors can help suppress false positives in heterogeneous pan-tumor scenarios.
- 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 chosen baselines are mostly universal or prompt-based segmentation models, but the paper does not compare with strong CT tumor segmentation pipelines trained or adapted for the same organs, such as nnU-Net-style organ-specific or tumor-specific baselines. This matters because the proposed method receives highly informative clinical constraints, whereas some baselines may receive only point or text prompts. The comparison, therefore, mixes two questions: whether ECLIPSE is a better segmentation model, and whether extra localization information helps segmentation. A fairer evaluation would include strong 3D CT tumor baselines, organ-specific expert models, and a setting where the same EHR-derived boxes or constraints are provided to competing promptable models.
The formulation of the tumor-type prompt and class input is underspecified. The method assumes a tumor-type prompt, but the paper does not clearly define where this prompt comes from at inference time, whether it is extracted from EHR, manually selected, or assumed to be known. This is not a small implementation detail, because in pan-tumor CT segmentation, the model must know whether it is searching for a liver tumor, a kidney tumor, a lung tumor, an adrenal tumor, or multiple possible lesions. If the class is externally provided, the task becomes segmentation of known clinical findings rather than open pan-tumor detection. If the class is not externally provided, the paper needs to explain how the organ or tumor type is inferred.
Relatedly, the current CCHG descriptor E = (r, n, d) appears too simplistic for realistic multifocal and multi-organ disease. A patient may have lesions in different organs, with different diameters and different anatomical subregions. The paper evaluates a “mixed tumor” dataset, but it is not clear whether it handles a single CT volume containing simultaneous heterogeneous tumor types, such as liver and kidney tumors in the same scan. The current description suggests a single tumor-type prompt, a single subregion descriptor, and a single longest-diameter parameter for box expansion. This may not generalize to multi-lesion cases where each lesion has its own organ, size, and location descriptor.
The evaluation is also too narrow. Reporting only DSC does not sufficiently characterize tumor segmentation performance. Tumor segmentation is highly sensitive to small lesions, lesion-level detection, false positives outside plausible organs, and boundary accuracy. Metrics such as HD95, NSD, lesion-wise recall/precision, false-positive volume, count consistency, and small-lesion stratified performance would provide a much clearer picture. This is particularly important here because the method explicitly claims improvements in localization, count consistency, and anatomically constrained prediction, but the evaluation does not directly measure these properties.
- 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
I encourage the authors to more explicitly position the task. At present, the manuscript sometimes reads like a general pan-tumor segmentation method, but the actual setup appears closer to segmentation of known clinical findings under structured report-derived constraints. This distinction is important and should be stated clearly. A clinically useful version of this framework should be tested under realistic EHR uncertainty: missing subregion, approximate or wrong diameter, incorrect lesion count, multiple lesion descriptors, and contradictory image–EHR evidence. The method should also have a mechanism to express uncertainty or reject infeasible constraints rather than always forcing the top-n peaks into tumor boxes.
- 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 assign a weak reject because the paper presents a promising, clinically motivated idea, but the current evidence does not yet convincingly demonstrate that the proposed method works under realistic clinical conditions. The major factor behind my score is the potential oracle nature of the pseudo-EHR protocol, in which key descriptors are generated from ground-truth masks of public datasets. Since these descriptors directly constrain localization and bounding-box generation, the reported improvement may not reflect the model’s ability to infer tumors from CT and real EHR, but rather its ability to exploit annotation-derived spatial priors. In addition, the baseline comparison does not include strong dedicated CT tumor segmentation models; the evaluation relies only on DSC, and the handling of tumor-type prompts, multi-organ multifocal disease, and upstream anatomical segmentation errors remains unclear. I would be much more positive if the authors provided a stronger evaluation under noisy/real EHR descriptors, added lesion-wise and boundary metrics, compared with organ-specific 3D tumor baselines, and clarified the clinical inference protocol.
- 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 appreciate the authors’ rebuttal and clarifications. The response partially addresses my concerns, especially by clarifying that the tumor-type prompt and descriptors are derived from EHR findings, and that multi-lesion cases are processed finding-by-finding and then merged. This makes the task setting clearer: ECLIPSE is best understood as EHR-constrained segmentation of known/suspected clinical findings, rather than fully open pan-tumor detection.
However, several concerns remain only partially resolved. First, the baseline comparison is still not fully convincing. The authors explain why universal and prompt-based models were selected, but the absence of strong organ-specific or tumor-specific 3D CT baselines remains a limitation, especially because ECLIPSE receives clinically informative constraints. Second, the handling of heterogeneous multi-organ disease is clarified conceptually, but the manuscript should explicitly describe how multiple EHR findings are represented, processed, and merged, rather than leaving this implicit. Third, the evaluation remains narrow. I understand that new metrics may not be allowed during rebuttal, but the lack of lesion-level detection, boundary, false-positive, count-consistency, and small-lesion analyses weakens the empirical support for the claimed localization and clinical-plausibility advantages.
Overall, the rebuttal improves my understanding of the method and addresses some ambiguity in the task formulation. While the paper still has important limitations in evaluation breadth and baseline fairness, the clinical motivation is meaningful, and the EHR-constrained formulation is potentially useful. Therefore, I am willing to raise my recommendation to a weak accept, provided that the final version clearly states the task setting, clarifies the multi-finding inference pipeline, expands the implementation details, and explicitly acknowledges the limitations regarding baselines and evaluation metrics.
Author Feedback
We thank the AC and reviewers for their constructive feedback and for recognizing the clinical motivation (R2, R3, R5), methodological novelty (R2, R3, R5), and potential practical value (R3, R5) of our work. Below, we address the reviewers’ comments point by point. 1) Pseudo-EHR descriptors and Colon-FD clinical grounding (R2-Q1, R5): ECLIPSE targets EHR-constrained pan-tumor segmentation of suspected clinical findings. Colon-FD uses real clinical EHR, where tumor type, lesion count, maximal diameter, and coarse anatomical location are extracted from clinical records. For public datasets without released EHR, pseudo descriptors only simulate such clinical EHR information: lesion count is obtained by connected-component analysis, maximal diameter by the largest physical lesion extent, and anatomical subregion by the overlap with predefined organ subregions. 2) Baselines and fair comparisons (R5-Q1): For strong CT tumor baselines, we selected recent SOTA universal/prompt-based models because they are most relevant to ECLIPSE and have been benchmarked against nnU-Net-style pipelines in several prior studies, such as Universal Model or MA-SAM. Unlike point prompts that directly mark tumor locations, ECLIPSE infers feasible tumor regions from clinically available EHR descriptors without manually specified tumor points. Due to page limits and rebuttal policy, we will clarify this choice and provide more detailed comparisons in an extended version. 3) Tumor-type prompt and multi-organ disease (R5-Q2/Q3): The tumor-type prompt is obtained from the EHR, where tumor type and subregion have already been preliminarily identified, as described in Sec. Introduction and Fig. 1(b). Importantly, MSWAL is a public mixed-tumor dataset, with many scans containing multiple tumor types or lesions from different organs. For such cases, ECLIPSE processes each tumor finding separately with its own prompt and EHR descriptor, and then merges the masks. The MSWAL results support this strategy. 4) Evaluation beyond DSC (R3-Q1, R5-Q4): We agree that additional metrics may provide complementary perspectives. In the submitted manuscript, we used DSC as the primary metric because it is widely adopted in tumor segmentation benchmarks and enables consistent comparison across all baselines and datasets. The qualitative results in Fig. 3 provide complementary visual evidence. Following the rebuttal policy, we do not introduce new metrics here. A more comprehensive evaluation will be provided in the journal version. 5) Implementation details (R1-Q2): We clarify that α is a safety-margin coefficient for expanding the EHR-reported tumor size d into a loose segmentation box. “FC” denotes a Fully Connected layer, and S in Eq. 11 is the spatial response map from Eq. 5.The training loss is Dice+BCE with 1:1 weighting for both stages. 6) Ablation study (R2-Q3): The “Baseline” in Table 2 denotes the model before CCHG, namely the ViT and CLIP encoders with multimodal fusion followed by the decoder, as described in Sec. 2.1 and Fig. 2.CCHG further introduces three feasibility constraints: anatomical subregion for regional coherence, lesion count for cardinality consistency, and tumor size for scale plausibility. As these complementary constraints are methodological necessities for generating clinically plausible tumor masks, we omitted internal ablations due to page limits; but we will clarify the baseline configuration and the role of each CCHG component. 7) Dataset, limitations, and presentation (R3-Q2/Q3): The dataset composition is summarized in Fig. 1 with corresponding references; we did not include an additional dataset table due to page limits. We will add a concise limitations paragraph discussing pseudo-EHR simulation and reliance on anatomical priors. We will also enlarge Fig. 1, add clearer annotations in Fig. 2, remove repeated CCHG descriptions, and correct the bibliography. All revisions mentioned above will be updated in the final paper.
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 reviewers overall appreciate the motivation and formulation of the problem with the EHR-constrained pan-tumor CT segmentation. However, the reviewers also raised concerns regarding pseudo EHR descriptors, a single DSC-only metric for evaluation, insufficient dataset/implementation descriptions, unclear annotation process, unclear clinical inference protocol, and a lack of evaluation under noisy/real EHR descriptors. Authors are suggested to address these comments carefully.
- 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.
Authors’ rebuttal addressed the main concerns. There are some remaining concerns regarding evaluation. However, there is clear merit in the work and all three reviewers recommend acceptance. I strongly suggest the authors utilize the extra space in the camera-ready to discuss the limitations of this work.
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.
Important concerns raised by the reviewers, especially regarding the setup of specific datasets and prompt descriptors, have been answered in the rebuttal, leading to a positive rating from the reviewers, which I support. I strongly support the request to include at least a contour-based evaluation metric as recommended by the reviewers in a camera-ready version, which is also allowed under the rebuttal guidelines: “It is, however, allowed to amend the presentation of existing results.”, which includes additional metrics on same datasets/experiments already presented in 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 paper proposes a clinically meaningful and technically sound pan-tumor segmentation framework that integrates EHR-derived constraints into spatial inference. Its clear progression from coarse localization to refined segmentation is methodologically logical and supported by consistent improvements across multiple datasets.