List of Papers Browse by Subject Areas Author List
Abstract
Accurate segmentation of tissues and instruments in surgical scenes is annotation-intensive due to irregular shapes, thin structures, specularities, and frequent occlusions. While SAM models support point, box, and mask prompts, points are often too sparse and boxes too coarse to localize such challenging targets. We present SCISSR, a scribble-promptable framework for interactive surgical scene segmentation. It introduces a lightweight Scribble Encoder that converts freehand scribbles into dense prompt embeddings compatible with the mask decoder, enabling iterative refinement for a target object by drawing corrective strokes on error regions. Because all added modules (the Scribble Encoder, Spatial Gated Fusion, and LoRA adapters) interact with the backbone only through its standard embedding interfaces, the framework is not tied to a single model: we build on SAM 2 in this work, and by communicating only through standard embedding interfaces, the design is in principle compatible with other prompt-driven segmentation architectures. To preserve pre-trained capabilities, we train only these lightweight additions while keeping the remaining backbone frozen. Experiments on EndoVis 2018 demonstrate strong in-domain performance, while evaluation on the out-of-distribution CholecSeg8k further confirms robustness across surgical domains. SCISSR achieves 95.41% Dice on EndoVis 2018 with five interaction rounds and 96.30% Dice on CholecSeg8k with three interaction rounds, outperforming iterative point prompting on both benchmarks.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5377_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/ping-haonan/SCISSR
Link to the Dataset(s)
EndoVis 2018 Robotic Scene Segmentation: https://endovissub2018-roboticscenesegmentation.grand-challenge.org/
CholecSeg8k: https://www.kaggle.com/datasets/newslab/cholecseg8k
BibTex
@InProceedings{PinHao_SCISSR_MICCAI2026,
author = { Ping, Haonan AND Jiang, Jian AND Yuan, Cheng AND Sun, Qizhen AND Wu, Lv AND Ban, Yutong},
title = { { SCISSR: Scribble-Conditioned Interactive Surgical Segmentation and Refinement } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16884},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes SCISSR, a scribble-conditioned interactive segmentation framework built on SAM 2, introducing a Scribble Encoder, Spatial Gated Fusion, and LoRA-based adaptation for iterative refinement in surgical scenes.
- 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 idea of using scribble prompts for interactive segmentation is intuitive and practically relevant, especially for surgical scenarios with complex structures.
The proposed framework is well designed and modular, easy to implement and extend to other prompt-based segmentation models.
The method interacts with the backbone only through standard embedding interfaces, which improves generalizability and potential reuse across models such as SAM 3.
The experimental results are strong and consistent across both in-distribution (EndoVis 2018) and out-of-distribution (CholecSeg8k) datasets, demonstrating good generalization.
The paper is generally well 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 level of novelty is somewhat limited. While the use of scribble-based interaction is well motivated, similar ideas have been explored in prior work such as ScribblePrompt.
The paper could benefit from more analysis of limitations and challenging scenarios, such as performance under sparse or imperfect scribbles.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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 paper presents a practical and well-executed approach for interactive surgical segmentation by introducing scribble-based prompting within SAM 2.While the individual components are not highly novel, their integration is thoughtful, lightweight, and easy to adopt in existing prompt-based frameworks.
A key strength is the strong empirical performance and generalization to out-of-distribution data.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The paper builds on SAM2 and introduces SCISSR a scribble-conditioned interactive segmentation framework. The paper is well-motivated and clearly identifies the need for scribble annotation over traditional signals like clicks, boxes, masks, or text. In surgical settings scribbles can be valuable as they capture smaller non-smooth objects compared to other methods. The experiments clearly show the advantage of such supervision along with the proposed modules like SGF to SAM2.SCISSR enables iterative refinement through multi-round interactions and demonstrates strong performance on both in-distribution (EndoVis 2018) and out-of-distribution (CholecSeg8k) benchmarks.
- 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 presents a solid idea that is motivated by the limitation of existing prompting types in SAM models, especially in surgical data where objects are small and occluded. The paper writing was clear and compelling and the flow of ideas was smooth.
SCISSR is clean and easy to reproduce, making it a potential add-on to SAM-like architectures rather than a complete redesign.
The experiments section is strong and comprehensive. The paper includes multiple baselines and prompt types. The gains are consistent in later refinement rounds and for challenging object classes. The inclusion of out-of-distribution evaluation further strengthens the claims regarding generalization. Overall, the empirical results show that scribble prompts provide richer spatial information than clicks.
- 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.
While the paper claims that SCISSR is general for all segment anything models (SAM1-3), this claim was not empirically validated. The current evaluation is limited to SAM 2, even a minimal experiment or partial validation would be enough to show that.
The reliance on synthetic scribbles generated from ground-truth masks. Although the generation strategy attempts to mimic human annotations, it may introduce bias and may not fully capture the variability of real user interactions. This raises concerns about how well the method would perform in real annotation cases. Validation with human-generated scribbles especially from clinical experts would significantly strengthen the paper.
The evaluation is only focused on endoscopic surgical datasets. While the results are strong in this domain, the broader applicability of SCISSR to other modalities is not explored. This somewhat limits the generality of the conclusions, particularly given the claim of being a general prompting strategy.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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 proposed method SCISSR can be valuable for future works and is easily reproducible. There are limitations regarding the use of synthetic scribbles and the lack of validation on other modalities, however they do not undermine the core contribution. For these reasons, I recommend acceptance.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
all my points are addressed.
Review #3
- Please describe the contribution of the paper
This paper presents SCISSR, a scribble-promptable interactive segmentation framework built on top of SAM 2 for surgical scene segmentation. The core idea is to introduce scribble-based prompting as an alternative to point and bounding box prompts, which are argued to be either too sparse or too coarse for complex surgical structures (thin instruments, curved vessels, etc.). The main technical contributions include: (1) a lightweight Scribble Encoder that converts two-channel freehand scribble maps into dense prompt embeddings compatible with SAM 2’s mask decoder; (2) a dual-track scribble pathway where Track 1 accumulates all historical scribbles as a dense prompt and Track 2 injects only the latest correction into Memory Attention via a Spatial Gated Fusion (SGF) module; (3) a memory-driven iterative refinement loop that repurposes SAM 2’s temporal memory mechanism for multi-round correction on a single image; and (4) the use of toggleable LoRA adapters to adapt SAM 2 while keeping the image encoder frozen. The authors claim architecture-agnostic design transferable to SAM 3.Experiments are conducted on EndoVis 2018 (in-distribution) and CholecSeg8k (out-of-distribution), demonstrating that scribble prompting converges faster and achieves higher Dice than point-based and box-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.
S1.Well-motivated problem and prompt modality. The motivation for using scribble prompts in surgical scene segmentation is clearly articulated and convincing. The authors make a compelling argument that points are too sparse and bounding boxes too coarse for targets with irregular morphology (thin instruments, curved vessels), while scribbles offer a natural middle ground with dense spatial coverage. Figure 1 effectively illustrates this point.
S2.Clean and principled architectural design. The dual-track scribble pathway is an elegant design choice. Track 1 preserves the full interaction history via accumulated scribbles as dense prompt embeddings, while Track 2 focuses the model’s attention on newly corrected regions through SGF. The zero-initialization of both the LoRA adapters and the SGF scaling factor α ensures that the pretrained SAM 2 capabilities are preserved at the start of training, which is a thoughtful design decision.
S3.Strong quantitative results and OOD generalization. The results are compelling: 95.41% Dice on EndoVis 2018 (R4) and 96.30% Dice on CholecSeg8k (R2) without any CholecSeg8k training data. The OOD generalization is particularly noteworthy—exceeding bounding box baselines by over 14 percentage points on CholecSeg8k demonstrates robustness.
S4.Thorough convergence efficiency analysis. Table 2 provides a valuable analysis of convergence behavior across different Dice thresholds, showing that SCISSR reaches acceptable quality faster and with higher success rates than point-based baselines. The cumulative percentage breakdown per round is informative.
S5.Informative ablation study. The component ablation (Table 3) clearly demonstrates the individual contributions of SGF and Memory. The scribble generation strategy comparison across four types (contour, centerline, wave, adaptive) provides useful practical guidance. The qualitative visualization in Figure 4 showing SGF-induced query modifications and memory-induced updates is helpful for understanding the mechanism.
S6.Practical and lightweight design. Building on SAM 2 Tiny with only lightweight additions (Scribble Encoder, SGF, LoRA adapters) while keeping the image encoder frozen is practical. Training on a single RTX 4090 in 10 epochs demonstrates accessibility.
- 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.
W1.No validation of claimed architecture transferability to SAM 3.The paper repeatedly claims that the framework is “architecture-agnostic” and transferable to SAM 3 (Abstract, Section 1, Section 4). However, this claim is never experimentally validated. SAM 3 is only used as a point/box baseline, not as a backbone for SCISSR. Since this is listed as a core contribution (“architecture-agnostic design… making the approach transferable to architectures such as SAM 3”), the lack of experimental evidence significantly weakens this claim. Either the authors should demonstrate SCISSR on SAM 3 or substantially tone down this contribution claim.
W2.Lack of real user study with human-drawn scribbles. All experiments rely on synthetically generated scribbles from ground-truth masks. While the authors add “mild spatial perturbations to mimic freehand variability,” synthetic scribbles may not faithfully represent the noise, inconsistency, and variability of real human annotations. This is particularly relevant for a method whose core selling point is the naturalness and practicality of scribble interaction. A user study, even a small-scale one, would significantly strengthen the practical claims. The authors acknowledge this in the future work section but it remains a notable gap.
W3.Potentially unfair comparison with point-based baselines. The automated evaluation protocol generates corrective scribbles from ground-truth error regions, which inherently provides dense spatial information along boundaries and shapes. Point-click baselines, by contrast, receive only centroids of error-region connected components. While the authors partially address this by noting that even 10pt/ch degrades with more rounds, the information asymmetry between scribble and point prompts is substantial. A fairer comparison might normalize by annotation effort (e.g., time or number of annotated pixels) rather than by interaction rounds.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses a relevant and practical problem—interactive segmentation of surgical scenes—with a well-motivated choice of scribble prompts and a clean architectural design. The dual-track pathway and memory-driven refinement are principled contributions, and the quantitative results are strong, including impressive OOD generalization on CholecSeg8k. However, several factors prevent a stronger recommendation: (1) the claimed architecture-agnostic transferability to SAM 3, listed as a core contribution, is entirely unvalidated experimentally; (2) the reliance on synthetic scribbles without any real user evaluation limits the practical claims. The rebuttal could potentially address concerns with additional analysis or discussion. I therefore rate this as a borderline accept, contingent on a strong rebuttal addressing the key weaknesses.
- 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.
N/A
Author Feedback
We thank the reviewers and the Area Chair for their constructive feedback and address the major concerns below.
Response to R1.W1: The novelty may be limited given prior work such as ScribblePrompt.
SCISSR differs from ScribblePrompt in two key aspects.
First, our Scribble Encoder treats scribbles as dense spatial embeddings rather than converting each scribbled pixel into a click as in ScribblePrompt-SAM. This preserves the continuous shape and boundary information of the scribble as a whole.
Second, SCISSR introduces a dual-track pathway with Spatial Gated Fusion and memory-driven iterative refinement, repurposing SAM 2’s temporal memory for multi-round correction. ScribblePrompt does not incorporate such a memory mechanism.
As shown in Table 3, adding SGF improves mDice from 88.16 to 90.53 at R2, and adding Memory further improves it to 93.49, supporting the effectiveness of these SCISSR-specific components.
Response to R2.W1, R3.W1: Claims of transferability to other SAM variants are unsupported.
We agree that our current wording overstates the empirical scope of this claim.
To clarify our intended claim: SCISSR’s added modules (Scribble Encoder, SGF, and LoRA adapters) communicate with the backbone exclusively through standard embedding interfaces. This design is interface-compatible with SAM-like architectures.
However, interface compatibility alone does not guarantee empirical transferability. In particular, SGF and memory injection operate on spatial image tokens, and backbones with different token geometry or spatial resolution may require further adaptation.
We will revise the manuscript to replace the broad “transferable to SAM 3” wording with a more precise statement reflecting interface-level portability as a design principle, not an empirically validated property. Empirical validation on SAM 3 is outside the scope of the submitted manuscript.
Response to R1.W2, R2.W2, R3.W2: Real human-drawn scribbles and imperfect scribble tests are absent.
We agree that real human-drawn scribbles would be valuable for fully assessing clinical usability. Since large-scale real annotations are unavailable, synthetic scribbles enable controlled, reproducible evaluation across prompt types and baselines.
Sec. 2.7 reduces idealization by using four geometry-aware scribble types with mild spatial perturbations to mimic imperfect freehand input.
Preliminary feedback from domain experts during internal deployment preparation suggests that the scribble interaction is intuitive. However, this feedback is not used as formal evidence in the submitted evaluation; the current evidence remains the controlled synthetic-scribble benchmark.
Response to R3.W3: The comparison with point-based baselines may be unfair.
We acknowledge that scribbles and clicks carry different spatial information. Our comparison targets practical interaction paradigms under the same fixed-round correction protocol, not equal information per prompt.
To avoid underestimating click prompts, we include both 1pt/CC and 10pt/ch baselines. Table 1 and Sec. 3.4 show that increasing click density does not consistently improve multi-round refinement. This is consistent with our interpretation that isolated clicks provide less structured shape/boundary information than scribbles.
Preliminary expert feedback from ongoing deployment testing suggests that scribble-based correction is convenient in practice, but no expert feedback or time-based measurement is used as evidence for the submitted comparison. Rigorous time-effort evaluation remains future work.
Response to R2.W3: Generalization beyond endoscopic datasets is not validated.
SCISSR targets interactive surgical scene segmentation. Our submitted evaluation uses EndoVis 2018 and CholecSeg8k, which support cross-dataset and cross-procedure generalization within endoscopic/laparoscopic surgical scenes. Broader OOD validation is ongoing, and we will clarify this scope.
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 in general found the paper well-motivated and the proposed scribble prompted surgical scene segmentation convincing and practical. However, some concerns need to be addressed before the paper can be accepted. The authors are encouraged to clarify them in the rebuttal: 1) novelty compared to similar work such as ScribblePrompt, 2) unsupported claims of transferability to other SAM variants, 3) reliance on synthetic scribbles, 4) generalization to other modalities, and 5) performance under sparse or imperfect scribbles.
- 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 reviewers remained positive about accepting this paper. The rebuttal has mostly addressed the reviewers’ concerns.
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.
All 3 reviewers are overall satisfied with the rebuttal and recommend acceptance.
The authors have done a good job in the rebuttal clarifying all criticism.
I recommend acceptance.
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.
All reviewers were satistfied with the rebuttal. The authors should address the remaining concerns in the camera-ready version.