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

Abstract

Precise tumor segmentation is fundamental to neuro-oncological planning and longitudinal monitoring. Existing interactive foundation models exhibit strong generalization capabilities, theoretically enabling the segmentation of diverse brain lesion types and sub-regions. However, they remain ill-suited for this domain due to their reliance on single-modality inputs and “trigger-happy” behavior — often hallucinating structures assuming that every user interaction targets a valid lesion. In this work, we propose SkepticalNet, a native 3D, multi-modal framework based on nnInteractive, trained with a skeptical prior. By integrating a novel background interaction sampling protocol during training, we enable our model to distinguish between valid and erroneous user guidance. We evaluate this approach on a diverse multi-cohort BraTS dataset encompassing adult/pediatric gliomas and meningiomas. Our method significantly outperforms the state-of-the-art nnInteractive foundation model, achieving a global mean Dice of 85.9% (vs. 76.3%) while reducing the annotation burden. Crucially, we demonstrate that the model learns a context-aware safety profile, effectively suppressing activations triggered on healthy tissue while preserving the ability to robustly correct imprecise clicks near tumor boundaries. Inference code and weights are available at https://github.com/stathisky-repo/SkepticalNet.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2357_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/stathisky-repo/SkepticalNet

Link to the Dataset(s)

BraTS Pre-Treatment Adult Glioma Dataset: https://www.synapse.org/Synapse:syn51156910/wiki/622351 BraTS Post-Treatment Adult Glioma Dataset: https://www.synapse.org/Synapse:syn53708249/wiki/627500 BraTS Meningioma Dataset: https://www.synapse.org/Synapse:syn51156910/wiki/622353 BraTS Pediatric Tumor Dataset: https://www.synapse.org/Synapse:syn51156910/wiki/622461

BibTex

@InProceedings{KyrEfs_SkepticalNet_MICCAI2026,
        author = { Kyriazis, Efstathios AND Kalliatakis, Grigorios AND Bisdas, Sotirios AND Tsiknakis, Manolis AND Marias, Kostas},
        title = { { SkepticalNet: A Safety-Aware Interactive Segmentation Framework for Neuro-Oncology } },
        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 introduced SkepticalNet, an interactive segmentation framework that specifically designed for lesion segmentation in multimodal images aiming at reducing the false positives from background healthy tissues. Through a background interaction sampling protocol, the model learns to distinguish between valid user guidance and false clicks on healthy tissue. It incorporates mpMRI volumes and a Topology-Constrained Stochastic Walk algorithm (to deal with the irregular shape of lesions) to improve the segmentation of irregular shape brain tumors. Evaluation was done on BraTs datasets for glioma, meningioma, and pediatric tumors showing its superior performance compared against standard existing methods such nnInteractive.

  • 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.

    For the proposed interactive segmentation pipeline, the authors introduces a “skeptical prior” through a hybrid training protocol that stochastically assigns iterations to either foreground or background supervision. This enables the model to validate user prompts against local anatomy and effectively suppress activations triggered on adjacent healthy tissue, reducing the distal Activation Rate to just 1.5%. In addition, to improve the realism of the training signal, the authors replace standard linear scribbles with a Topology-Constrained Stochastic Walk algorithm. This algorithm aims at generating scribbles which are constrained to the lesion mask and can deal with the irregular, non-convex morphologies of pathologies common in brain tumors.

  • 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 authors must clarify the binary segmentation procedure used during training, as it is currently unclear if the target class refers to individual sub-regions or combinations (e. g. , Whole Tumor or Tumor Core). This ambiguity makes it difficult to assess how the model resolves conflicting user intents. In addition, for the extremely small pediatric cohort (N=99) , the paper does not explain how the balanced dataloader avoids overfitting when operating with a batch size of only 2.Does it work by loading 1 sample per batch mostly from the pediatric set? A critical technical bottleneck is the model’s reliance on a fixed 11-channel input. The framework requires co-registered T1, T1c, T2, and FLAIR modalities to function. While the authors acknowledge this as a limitation, the current architecture lacks any mechanism for missing-modality resilience and/or miss-aligned imaging modalities. In real-world clinical settings, incomplete MRI protocols are common; the absence of even a single sequence led to failure of the whole workflow. Although the paper evaluates four distinct datasets, the experiments are limited entirely to standardized BraTS challenge datasets. These datasets are heavily pre-processed, co-registered, and skull-stripped. Moreover, size of the lesions in such tumor entities are usually large enough to be segmented well with conventional U-Net based models. Therefore, the “safety-aware” suppression has not been tested against “raw” clinical artifacts, such as patient motion, varying voxel resolutions, or extra-cranial structures and more importantly more challenging lesion such as brain metastasis. The model’s ability to suppress hallucinations in a noisy, non-standardized environment remains unproven. There is a concerning performance gap regarding safety near tumor boundaries (4 voxels). While the “skeptical prior” is highly effective at distal suppression, the global activation rate in the proximal zone remains at 36.3%, rising to 70.1% for sparse prompts. For infiltrative lesions with complex edges, this suggests the model still prioritizes user guidance over image evidence in ambiguous regions, potentially snapping to healthy peritumoral tissue and undermining the stated goal of hallucination suppression.

  • 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

    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 recommendation is based on the fact that while the proposed method is technically sound and clinically relevant, the paper exhibits some critical methodological gaps that limit its immediate applicability. the main concern is the model’s reliance on a fixed 11-channel input requiring four fixed MRI sequences to be always available. This dependency creates clinical limitation, as the current architecture provides no mechanism for missing-modality compatibility, which is a barrier in real-world (and SOTA method) environments where incomplete imaging protocols are common. Furthermore, the experimental validation is limited to standardized, pre-processed BraTS datasets, lacking extension to other challenging multi modal applications such as PET-CT for whole body lesions or multi-sequence MRI for Head and Neck cancers which would show true versatility. Methodological ambiguities also exist regarding the technical execution of the hierarchical binarization procedure for sampling target classes and the stability of the balanced dataloader.

  • 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 well-justified some of the raised comments by this reviewer, including the data loader, binary target class, and cut-off choice. One major comment by this reviewer remained un-justified: performance evaluation over challenging tiny lesions such as Brain METs, or other types of lesions such as those in whole body PET-CT, liver METs etc.



Review #2

  • Please describe the contribution of the paper

    The paper proposes an extension of the nnInteractive framework. The extended framework; 1) allows for inputs with multiple MRI modalities, 2) uses a “dual-mode” training strategy that stochastically selects between focusing on foreground segmentation and suppress the impact of invalid/inaccurate prompts, and 3) a topology-constrained stochastic walk (TCSW) algorithm to generate more realistic scribbles for use in training. The dual-mode training strategy discourages the model from generating hallucinated segmentations from inaccurate input prompts, thereby improving clinical robustness. The TCSW algorithm creates a more clinically accurate training signal to improve performance.

  • 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 provides a clinically focused approach to image segmentation and introduces some novel methods to reinforce model behavior that is conducive to clinical use as it targets common weaknesses in medical image segmentation models, specifically addressing hallucination from slightly inaccurate prompts, and generating realistic scribbles. The use of a “dual-mode” training strategy is an interesting and novel approach to add robustness to invalid prompts as it explicitly punishes the model for various “trigger-happy” outputs. The datasets and data splitting provide confidence in the presented results with its multiple dataset benchmarks. The provided safety analysis is an effective means of proving the advantage of the training protocol.

  • 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 topology-constrained stochastic walk (TCSW) algorithm is under-explored in the paper. The paper could benefit from a more direct ablation study of the TCSW algorithm. The methodology for implementing TCSW is could include more details, such as how the stochastic noise is injected into the neighbor selection process. The paper would also benefit from more baseline models for comparison, with the included baseline models (pre-trained nnInteractive model; single-modality fine-tuned variant) making it difficult to ascertain the true value of each component in the proposed framework. The datasets are all from the BraTS challenge, which despite using different benchmarks, still raises some potential concerns regarding generalizability.

  • 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

    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.

    (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?

    The clinically focused design decisions in the area of image segmentation make it an appealing paper to accept. Directly addressing the tendency of image segmentation models to over-predict based on inaccurate input prompts is a particularly interesting idea. Some aspects that hold the paper back are the number of comparative baseline models, the technical innovation being a bit limited, and the limited exploration on the TCSW algorithm. However, the merits mostly outweigh the flaws.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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.

    The problem is meaningful and the proposed methodology is an interesting means of addressing the problem. However, some of the major issues have not been addressed to the extent that I would hope for. The real-world applicability and generalizability are listed as limitations or future work, the justification for the 4-voxel proximal cutoff raises some questions such as given the justification, why was a 4 voxel cutoff selected over any other value between 2-5.Some of the methodological clarifications also leave more to be desired such as the authors only addressing stochastic noise injection for TSCW when the TSCW description as contained in the paper leaves other details uncertain. These indicate that the paper would need more a bit more time put in before it is ready for acceptance.



Review #3

  • Please describe the contribution of the paper

    The paper presents Skeptical Net. A model based on nnInteractive that uses both FG and BG interactions.

  • 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.
    • Strong experiments, results, and ablations section.
    • Safety Analysis and metrics to measure safety.
    • Code available- well-designed inference code is available (not training code)
  • 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.
    • terms and terminology are poorly defined, like safety, skeptical prior, snaps to valid pathology
    • What is $p$? It is not explicitly mentioned. Also, an ablation of $p$ values for different scenarios. What experiments are necessary to estimate $p$?
    • The claim for significant improvement over baselines needs to be backed by statistical testing
    • for the distal and proximal cutoff, assuming in the real world 1 voxel == 1mm, what are the clinical requirements for such an arbitrary cutoff? It might have been useful to see the NSD with the mentioned Dice in Table 1 to make more sense of the choice of 4 voxels.
    • Minor formatting issues - Table 1 Men. and Ped. not defined; Table 2, P_target is missing an uparrow
  • 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 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
    • Please refrain from using “” terms in the paper, as it reduces the readability of the scientific paper.
    • hallucination is a trendy term used; but generally valid for generative models and not discriminative models like segmentation.
    • Would the results still be valid if nnInteractive was domain-specific trained?
  • 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 feel the paper is addressing a real concern and is also doing good science. But I cannot verify the claims due to the unexplained choice of cutoffs and proposed metrics, which cannot be interpreted intuitively. Also, I found the paper very difficult to read and follow.

  • 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 feel the authors did a good job of adressing the concerns I raised. I think the authors’ promised changes would make the paper more readable and acceptable to the MICCAI audience. Moreover, I think it is an important topic that might help the community at large.



Author Feedback

We thank the reviewers and Area Chair for their constructive feedback and provide the following clarifications. Note: [CR] denotes confirmed updates to the camera-ready manuscript.

1.Methodological Clarifications (R1, R2, R4)

  • Target Sampling & Probabilities (p) (R1, R4): We clarify that for each training sample, the binary target (c_target) dynamically samples either an atomic sub-region or a composite, ensuring all possible classes per-sample are supervised throughout training, seamlessly resolving user intent. Variable p denotes branching probabilities (e.g., 75/25 Foreground/Background) in our stochastic tree. As stated, these ratios were experimentally optimized; internal testing via grid search on the validation set confirmed that altering them degrades segmentation accuracy or robustness to invalid prompts. [CR: clarified in Sec 2.2]
  • Dataloader (R1): Our dataloader uniformly samples two of four datasets per iteration (25% probability each), rather than forcing one pediatric sample per batch. Any re-sampled minority case is heavily modified via aggressive spatial/intensity augmentations (e.g. elastic deformations, gamma/Gaussian noise), preventing overfitting via input diversity. [CR: clarified in Sec 3.1]
  • TCSW Details (R2): We clarify that uniform stochastic noise is added to the geodesic distance map prior to each pathfinding step, enforcing different minimal connecting paths between the start and target points. [CR: clarified in Sec 2.3]
  • Terminology (R4): We will consolidate stricter definitions for “safety” (suppressing invalid prompts), “skeptical prior” (zero-output constraint), “snaps to” (adherence to valid pathology), and “hallucination” (used strictly within our context to denote unconstrained segmentation responses to invalid prompts). [CR: definitions added to Introduction]

2.Safety Metrics & The Proximal Gap (R1, R4)

  • 4-Voxel Clinical Cutoff (R4): We justify that assuming 1mm voxels, a 4mm margin aligns with neuro-oncology realities where radiotherapy margins utilize 2mm-5mm uncertainty buffers for microscopic extension and inter-observer variability (ESTRO-EANO 2023 Guidelines: Niyazi et al., Radiother Oncol), effectively defining a clinically ambiguous zone. [CR: justified via Niyazi et al.]
  • Proximal Activation is a Feature, not a Bug (R1): R1 notes a 70.1% Activation Rate for proximal sparse prompts. This is intentional. Because P_target (a precision metric over activated voxels) remains exceptionally high (88.6%), predicted activations successfully fall within true tumor boundaries instead of healthy peritumoral tissue, proving “near-miss” error correction (predicting nearby pathology) rather than unconstrained false-positives.
  • Statistical Significance (R4): Wilcoxon signed-rank tests confirm that baseline improvements are statistically significant (p < 0.01). [CR: amended Sec 4.1 to present p-values]

3.Scope, Generalizability & Baselines (R1, R2, R4)

  • Scope & Limitations (R1, R2): Evaluating PET-CT and Head/Neck cancers falls outside our explicit clinical scope (Neuro-Oncology), representing a deliberate boundary rather than a methodological flaw. While incomplete MRI protocols occur, 4-sequence mpMRI is the standard-of-care (ESTRO-EANO 2023 Guidelines); handling incomplete protocols is a distinct research problem explicitly acknowledged as a limitation. To address generalizability (R1, R2), our limitations section will explicitly include future testing on non-BraTS cohorts with unstandardized artifacts.
  • Baseline Choice (R2, R4): R4 asks if the results hold if nnInteractive is domain-specifically trained. Yes, our FT-T1c baseline is exactly this (nnInteractive fine-tuned on BraTS). Regarding R2’s request for additional baselines, each baseline isolates one added contribution: Native establishes the zero-shot floor; FT-T1c quantifies multi-modal fine-tuning gains; Standard Baseline isolates the Skeptical Prior’s impact — collectively providing a complete component-wise evaluation.



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 addresses a clinically relevant and practical workflow for 3D interactive brain tumor segmentation. Reviewers agree that the problem is meaningful and that the safety-aware training protocol is interesting. However, several major issues need to be addressed in the rebuttal: (1) with the fixed 11-channel input, it is unclear how the model can be applied to real-world interactive segmentation where raw clinical MRI sequences may be missing or misaligned (R1); (2) the experiments are limited to standardized BraTS datasets, so generalizability to other challenging segmentation tasks remains unclear (R1R2); (3) key methodological details require clarification, including the TCSW algorithm (R2), the binary target definition (R1) and the choice of the 4-voxel proximal/distal cutoff (R1R4).

  • 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 proposed valid-prompt check and interactive segmentation framework using multi-sequence MRI unlocks the potential clinical application of semi-automatic brain tumor segmentation. Most major concerns, including fixed input channels and the choice of thresholds, have been addressed in the rebuttal. Overall, the paper is well written and clearly presented.

    The work could be further strengthened through validation on more challenging brain tumor types, such as brain metastases. Including additional fully automatic baselines would also provide useful insight. In the current setting, fully automatic baselines already perform very well on the presented dataset, whereas they often fail more on harder tasks such as brain metastasis.



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 post-rebuttal opinions are mixed but slightly positive. Reviewer #1 recommends acceptance after the authors clarified the dataloader, binary target definition, and cutoff choice, although one major concern remains regarding validation on tiny lesions and other lesion types. Reviewer #2 maintains rejection, arguing that real-world applicability, generalizability beyond standardized BraTS datasets, the 4-voxel cutoff, and TCSW details remain insufficiently addressed. Reviewer #4 recommends acceptance, but with low confidence and suggests weighing the other reviews more strongly.

    Overall, the clinical problem is important and the safety-aware interactive segmentation framework is interesting, but several limitations remain. I recommend acceptance, provided that the final version clearly discusses the fixed 11-channel input limitation, BraTS-only evaluation, incomplete generalizability evidence, 4-voxel cutoff justification, TCSW implementation details, and lack of validation on tiny or non-Brain tumor lesions.



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.

    As mentioned in the original revision, the paper addresses a clinically relevant problem and provides a novel enough methodological contribution. Having read the paper, original comments and the rebuttal, I do agree with most of the reviewers in the sense that most major concerns have been addressed.

    I have mixed feelings about some of the omissions (why didn’t the authors talk about what happens when modalities are missing when that is a key part of the proposal) and I can see reviewer #2’s points for rejection. Having said that, and considering that we only have a binary choice during rebuttal, my decision is slightly positive as I still think the methodology is sound and the paper has enough merit. Nonetheless, my advice for the authors is to keep the unaddressed comments in mind for potential questions during the conference or future work extensions.



back to top