Abstract

Automated Surgical Error Detection is critical for intraoperative safety and objective skill assessment in robotic-assisted surgery. Although major errors are visually apparent, subtle execution-level errors reflect early control degradation and often resemble normal maneuvers, posing a significant challenge for reliable detection. Although existing methods explore various spatio-temporal and multi-modal features, they formulate error detection as implicit discrimination. Consequently, they struggle to resolve the visual ambiguity of subtle errors without an explicit reference of correct execution. In this paper, we propose KIDA, a Kinematic-Intent Dual-path Alignment framework that enhances the discrimination of subtle surgical errors by explicitly aligning actual execution with clinically grounded intents. KIDA leverages kinematic information to more comprehensively capture the current motion state, and incorporates predefined textual descriptions based on clinical standards to serve as an explicit reference for correct execution. Specifically, the Kinematic Execution Branch (KEB) processes lightweight kinematic motion dynamics to provide subtle execution details about the ongoing surgical action. The Clinically-Guided Intent Branch (CIB) utilizes semantic intent derived from clinical textual descriptions to provide explicit guidance, dynamically evaluating whether the motion-aware visual features conform to the intended standard maneuver. To further improve error discrimination, we introduce an Intent-Execution Alignment (IEA) optimization strategy, which explicitly enforces consistency for normal actions while actively decoupling subtle errors from the semantic intent to enlarge the feature margin. Experiments on the SAR-RARP50 dataset demonstrate that KIDA achieves state-of-the-art performance, improving AUC by 2.28% and AP by 4.86% while demonstrating generalizability on the JIGSAWS benchmark with a 0.84% AUC improvement. Code is available at https://github.com/yxluo8/KIDA.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3626_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/yxluo8/KIDA.

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LuoYux_KIDA_MICCAI2026,
        author = { Luo, Yuxuan AND Yang, Kaixiang AND Li, Yuxi AND Fang, Wei AND Li, Qiang AND Wang, Zhiwei},
        title = { { KIDA: Kinematic-Intent Dual-path Alignment for Surgical Error Detection } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16892},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes a Kinematic-Intent Dual-path Alignment (KIDA) framework to enhance the discrimination of subtle surgical errors through explicit motion guidance and semantic supervision. The key contribution is to reformulate surgical error detection from implicit appearance-based discrimination into execution-consistency evaluation against clinically grounded intent.

  • 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.Repositioning SED as intent-execution consistency measurement is conceptually well-grounded. 2.Asymmetric IEA loss is a technically interesting contribution. IEA explicitly handles the asymmetry between the two operating regimes, which is appropriate given the severe class imbalance in SED datasets. 3.The proposed method achieved the best performance in SAR-RARP50 dataset, improved SEDMamba from 71.20 to 73.48 AUC and from 44.87 to 49.73 AP.

  • 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 paper describes the method as “kinematic-intent” alignment, but the motion signal used in KEB is derived from OpenCV-based optical flow with a low-dimensional representation. This makes the terminology somewhat imprecise. The branch is closer to a video-derived motion cue than actual robotic kinematics.

    1. Table 3 removes components in a strictly cascaded manner (KIDA->IEA->KEB->CIB), meaning the individual contributions of KEB and CIB cannot be isolated 3.The paper’s motivation centers on detecting subtle, short-duration errors but the AUC improvement for short errors (<3s) is only +1.58%, compared to +2.72% for long errors. The paper claimed: KEB captures micro level instabilities missed by temporal smoothing should yield disproportionate gains precisely on short errors. This discrepancy is neither acknowledged nor explained. 4.Minor but notable presentation issues: 4.1 Figure 1 Motion Encoder” is misspelled as “Mosion Encoder” 4.2 The paper’s section numbering is inconsistent, e. g. , main experimental comparisons (Section 2.7) are inappropriately embedded within the Method section, while Section 3 contains only the ablation study 4.3 Conclusion section should discuss limitations as well
  • 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.

    (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?

    Please see the above comments.

  • 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 have addressed my previous concerns. I recommend acceptance.



Review #2

  • Please describe the contribution of the paper

    This paper addresses the problem of automatically detecting subtle execution-level surgical errors in robot-assisted surgery (RAS). The authors argue that such errors, arising from degraded motion control, repeated ineffective attempts, or unstable manipulations, are highly similar to normal operations in visual appearance and are therefore difficult to detect with the naked eye or with conventional algorithms. To address this issue, the paper reformulates surgical error detection as measuring the consistency between actual motion execution and clinically intended operation, and proposes the KIDA framework for motion–intention dual-path alignment. Specifically, the method includes a motion execution branch to preserve fine-grained kinematic deviations without excessive temporal smoothing, a clinically guided intention branch that uses guideline-based textual descriptions as semantic anchors, and an asymmetric intention–execution alignment loss to enlarge the feature gap between normal and erroneous samples without requiring additional contrastive pairs. Experiments on the SAR-RARP50 clinical dataset and the JIGSAWS dry-lab dataset show consistent improvements over prior methods, including gains over the previous state of the art on both long and short errors, and ablation studies indicate that all three core modules contribute independently.

  • 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 introduces a reformulation of the task at the problem-paradigm level. Instead of following the conventional paradigm of implicit classification based on appearance features, it redefines surgical error detection as measuring the consistency between motion execution and clinically normative intent, which directly targets the core ambiguity caused by the strong visual overlap between subtle errors and normal operations.

    2) The proposed KIDA framework is well aligned with the task structure. Its decoupled dual-branch design models motion dynamics and clinical semantic constraints separately: the motion execution branch captures fine-grained motion information through optical flow, while the clinically guided intention branch provides normative semantic guidance from clinical text, matching the natural two-level structure of execution characterization and compliance checking.

    3) The asymmetric intention–execution alignment loss is a customized optimization objective rather than a generic auxiliary regularizer. It explicitly encodes task priors by encouraging normal samples to align with clinical intent while forcing erroneous samples to decouple from that intent, thereby enlarging the representation gap between normal operations and subtle errors without constructing extra contrastive sample pairs.

    4) The temporal modeling design is well adapted to the task. By combining convolutional modules with Mamba, the method retains both local dynamic modeling capability and the ability to capture long-range temporal dependencies, which suits surgical error detection where temporal context is essential beyond single-frame judgment.

  • 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 method relies on a strong assumption and has a high adaptation cost. It requires manually written clinical guideline text for each surgical step as intention anchors, which makes it mainly suitable for a single standardized surgical stage. For complex procedures with many steps, the cost of writing full specifications is high, and the method may not flexibly adapt to differences across surgeons or centers.

    2) The ablation study is incomplete and does not verify whether the chosen components are optimal. The paper does not examine the effects of the granularity and phrasing of the clinical intention text, the replacement of optical flow with native robot joint-motion data, the independent gain of the convolution-plus-Mamba hybrid temporal modeling design, or the impact of different pretrained language models for encoding intentions.

    3) The evidence for generalization is limited. Clinical validation is conducted only on the suturing stage of prostatectomy, without evaluation on other common surgical actions such as cutting, hemostasis, or dissection, and without cross-procedure or cross-center generalization experiments.

    4) Deployment-related evaluation is missing. The paper reports only accuracy-related metrics and does not provide key deployment indicators such as parameter count, inference latency, or memory usage, making it unclear whether KIDA, with added optical-flow computation and a dual-branch structure, can satisfy the real-time requirement for intraoperative warning.

    5) The current clinical utility is limited. The method outputs only frame-level error probabilities and cannot identify specific error types or provide concrete corrective guidance, which reduces its direct practical usefulness for surgeons.

  • 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.

    (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?

    This paper studies an important and practically meaningful problem in robot-assisted surgery, namely the automatic detection of subtle execution-level surgical errors that are difficult to distinguish from acceptable operational variation yet may lead to serious postoperative complications if missed. The motivation is clear, and the paper’s main value lies in its reformulation of the task from implicit appearance-based classification to consistency measurement between motion execution and clinically normative intent. The proposed KIDA framework, together with the motion branch, clinically guided intention branch, and asymmetric alignment loss, forms a coherent method tailored to the ambiguity of fine-grained surgical error detection and is supported by improved results on both clinical and dry-lab datasets. At the same time, the paper has several limitations, including strong dependence on manually specified clinical intention text, incomplete ablation analysis, limited coverage in generalization experiments, lack of deployment-related evaluation, and restricted clinical practicality in its current output form. Overall, I view the paper as a meaningful and technically interesting contribution with demonstrated empirical gains, while the above concerns keep my recommendation at the level of weak accept.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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 #3

  • Please describe the contribution of the paper

    This paper proposes a framework for surgical error detection that reformulates the task as measuring the alignment between observed execution and clinically defined intent. The method, termed KIDA (Kinematic-Intent Dual-path Alignment), integrates two main components: a kinematic execution branch that captures motion dynamics and a clinically guided intent branch that encodes textual descriptions of correct surgical behavior.

  • 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.A key strength is the conceptual reformulation of error detection as intent-execution alignment, which provides an intuitive perspective grounded in clinical reasoning. 2.The integration of kinematic cues via optical flow into the learning pipeline is also reasonable, as motion patterns are known to be informative for fine-grained surgical analysis.

  • 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.

    First, the methodological novelty is limited. The proposed framework combines existing ideas multi-modal fusion (vision + text), kinematic feature integration, and alignment-based objectives without introducing a fundamentally new learning paradigm.

    Second, the use of textual intent is relatively simplistic and potentially restrictive.

    Third, the performance improvements are relatively modest. The reported gains (e.g., ~2.28% AUC improvement over SEDMamba) are incremental and may not justify the added model complexity

  • 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

    The paper tackles an important problem and presents a reasonable approach, but it would benefit from stronger positioning and deeper validation. Specifically, I encourage the authors to:

    1.Clarify the novelty relative to existing multimodal and alignment-based methods 2.Discuss limitations and failure cases more explicitly

  • 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.

    (2) Reject — should be rejected, independent of 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 problem and proposes a reasonable framework, the overall contribution is limited by incremental novelty, modest performance gains, and limited validation. The idea of aligning execution with intent is conceptually appealing, but its implementation does not sufficiently advance beyond existing multimodal alignment approaches.

  • Reviewer confidence

    Very confident (4)

  • [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



Author Feedback

We thank all reviewers for their constructive feedback and for recognizing our work as “conceptually well-grounded” (R1), a “coherent method aligned with task” (R2), and an “intuitive perspective based on clinical reasoning” (R3). We address the major concerns below.

1.Novelty (R3) Existing SED methods, including multi-modal variants, still follow an appearance-driven formulation: visual/motion features are fused and directly mapped to error labels through implicit discrimination. In contrast, KIDA reformulates SED as an intent-execution consistency evaluation problem by introducing clinically grounded intent as an explicit normative reference. This shifts the objective from recognizing error patterns to measuring deviation from intended execution. Accordingly, the proposed alignment objective explicitly pulls normal executions toward clinical intent while separating subtle errors from it.

2.Ablation Studies (R1, R2) (1) Only KEB: Using only KEB yields 71.51% AUC and 45.47% AP, improving upon the baseline. This confirms the independent effectiveness of KEB. (2) Intent Text: We validated our method using irrelevant (72.06% AUC), coarse-grained (73.30% AUC), and medium-grained text (73.36% AUC). The comparisons to our standard clinical intent (73.48% AUC) directly demonstrate that the domain-specific clinical intent is highly effective and robust to varying granularities. (3) Language Model: Replacing the text encoder with CLIP maintains a 73.40% AUC, demonstrating robustness to different language models. (4) Conv-Mamba: The hybrid Conv-Mamba design yields an 71.20% AUC and 44.87% AP.

3.Kinematics Terminology and Real Robotic Data (R1, R2) (1) We use “kinematic” because optical flow effectively represents the dynamic information changes across video frames, serving as a reliable visual proxy for surgical motion. (2) We validated KIDA on JIGSAWS using real robotic kinematics, achieving 73.92% AUC and 75.62% AP, outperforming the optical flow (73.53% AUC, 74.73% AP). This demonstrates that ours can effectively utilize direct motion information to enhance surgical error detection.

4.Subtle vs. Short-Duration Errors (R1) We clarify that “subtle” (exhibiting high visual similarity to normal maneuvers) and “short” (brief duration) are independent concepts. Subtle errors can also be long-duration errors, such as prolonged unstable manipulations. Short errors (~5 frames) are heavily diluted by normal contexts during temporal smoothing. Thus, even modest gains on them are highly valuable.

5.Efficiency, effectiveness, and generalization (R2, R3) (1) KIDA contains 7.76M parameters and 4.94 GFLOPs while achieving 32.0 FPS, satisfying intraoperative real-time requirement. (2) The performance gains over SEDMamba are consistent across three independent video-level trials. Paired two-sided t-tests further confirm statistical significance, with p-values of 0.0398 (AUC) and 0.0452 (mAP), indicating that the improvements (+2.28% AUC and +4.86% AP) are reliable rather than arising from random variation. Importantly, these gains are achieved on a strong baseline, where even small improvements are clinically meaningful for safety-critical error detection. (3) The clinical intent guidelines are derived from standardized medical knowledge, allowing the proposed textual anchors to generalize naturally across surgical centers and procedures without task-specific redesign.

6.Failures and Limitations (R1, R3) (1) While KIDA may occasionally over-smooth dense errors, similar behavior is also observed in SEDMamba due to subtle temporal boundary ambiguity. KIDA nevertheless shows more stable optimization and faster convergence. (2) KIDA currently supports only binary error detection, mainly limited by existing annotation granularity. Future work will explore finer-grained surgical error annotations for multi-class warning systems. Lastly, per R1’s advice, we have fixed the typo in Fig. 1 and corrected the section numbering.




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 authors should better justify the main assumption on the surgical error classification. Also, the paper novelty wrt SoA should be better clarified.

  • 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 managed to provide complete and satisfactory answers to the reviewer comments.



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.

    KIDA reformulates surgical error detection from implicit appearance-based discrimination into execution-consistency evaluation against clinically grounded intent — a reformulation reviewers found “conceptually well-grounded” and an “intuitive perspective grounded in clinical reasoning.” The proposed dual-branch design and asymmetric alignment loss provide a coherent methodological framework for this task. The initial concerns mainly related to component attribution, the clinical grounding of the intent representation, validation scope, efficiency, and the strength of the empirical evidence. The rebuttal clarified the authors’ intended design choices and addressed several reviewer concerns sufficiently to support acceptance.

    R1 updated to Accept post-rebuttal, while R2 remained Weak Accept and R3’s concerns remain relevant but not decisive. On balance, the task-level reformulation, coherent method design, empirical results, and post-rebuttal reviewer discussion support acceptance. For camera-ready, the authors should fold in the rebuttal clarifications; substantiate the “clinically grounded intent” claim more carefully, since performance changes little when the clinical text is replaced with generic or irrelevant text, leaving the gains not clearly attributable to the clinical content itself; acknowledge that validation is limited to a single surgical stage; and add a clear discussion of limitations.



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 is interesting, and the idea is well-suited to the task. The rebuttal successfully addressed the raised points.



back to top