acceptodds
Withdrawn from ICLR 2027

Phase as a Strong Prior: Surgical Phase Feature Modulation for Vessel Recognition in Complex Surgical Videos

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Surgical procedures unfold in stereotypical phases, and the anatomy exposed at any moment is strongly conditioned on the current surgical phase. This work investigates whether the surgical phase constitutes a sufficiently informative prior for fine-grained anatomical segmentation and, if so, how such a prior should be incorporated into a segmentation foundation model. We study this problem on key vessel segmentation in laparoscopic gastrectomy, using the LapGC-KVAD-30 dataset (30 procedures, 8 surgical phases, 15 vessel classes, case-level splits). A diagnostic experiment shows that the phase label alone predicts 15-class vessel presence with a macro-F1 of 0.810, whereas image-only prediction with a compact convolutional network remains near chance (macro-F1 0.136), indicating that the phase prior dominates appearance for vessel presence. Motivated by this observation, we inject the phase prior as feature-level modulation: the proposed Phase-FiLM maps the phase embedding to channel-wise scale and shift parameters that modulate the SAM2 image features before the mask decoder. On binary vessel segmentation, Phase-FiLM improves the test mean Dice from to over three seeds (mean gain points), outperforming the no-phase baseline on every seed, and yields consistent gains across two backbones (SAM2-L points; MedSAM2-T points). The improvements concentrate in the data-scarce dissection phases, where single-phase training collapses (S4: 0.320 vs. 0.428). Robustness analysis shows the method remains beneficial down to a break-even phase-recognition accuracy of 82.6%, within the 85-92% range reported for current surgical phase recognizers. Substituting an incorrect phase at test time collapses performance to 0.4212, indicating that the improvement arises from genuine exploitation of phase information rather than from increased model capacity. These results suggest that, in phase-structured procedures, workflow priors are most effectively incorporated as multiplicative feature modulation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
video segmentation, conditional feature modulation, FiLM, surgical data science, foundation model adaptation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_g2zsOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates surgical phase as a strong prior for vessel segmentation in laparoscopic gastrectomy. Experiments on the LapGC-KVAD-30 dataset show that Phase-FiLM improves binary vessel segmentation over a no-phase baseline. The authors also conduct additional ablations on feature injection strategy, backbone choice, wrong-phase inputs, and simulated phase-recognition errors.

Critical strengths

The research motivation is clear and has clinical significance. The paper utilises the ‘surgical phase’ as prior information to assist with vessel segmentation. This approach aligns with the characteristic of surgical procedures having a well-defined phasic structure, and phase-only classification experiments demonstrate a strong correlation between the phase and the appearance of blood vessels.

Critical weaknesses
  1. The methodological novelty appears relatively limited. The proposed Phase-FiLM primarily incorporates surgical phase embeddings into SAM2 features through the conventional scale-and-shift operations of FiLM. While the experimental improvements are encouraging, the architectural contribution is largely an adaptation of an existing conditioning mechanism. The authors should more clearly distinguish their technical contributions from established feature modulation approaches and provide stronger justification for the methodological novelty.

  2. The manuscript would benefit from further polishing and refinement. The abstract contains an extensive discussion of experimental results and numerical comparisons, which somewhat obscures the main research motivation and methodological contributions. In addition, several sections contain noticeable formatting and spacing inconsistencies, particularly around mathematical expressions and numerical results. A more concise abstract and careful proofreading would improve the overall presentation.

  3. The task formulation does not fully exploit the available annotations. Although LapGC-KVAD-30 provides annotations for 15 vessel categories, the primary segmentation task combines them into a single binary vessel mask. This formulation evaluates vessel localization but does not address the more challenging task of distinguishing individual vascular structures. Given that the paper emphasizes phase-dependent anatomical information and vessel recognition, additional experiments on fine-grained multi-class vessel segmentation would substantially strengthen the contribution.

  4. The experimental validation is limited to a single dataset. All experiments are conducted on LapGC-KVAD-30, which contains 30 surgical procedures collected at a single institution. Although the authors adopt case-level data splits and evaluate multiple backbones, these experiments do not fully establish generalizability across different surgical procedures, imaging conditions, or clinical institutions. Additional cross-dataset or external validation would provide stronger support for the broader applicability of the proposed approach.

  5. The figures could be improved to communicate the technical contributions more effectively. Figure 1 presents the overall architecture but offers limited insight beyond the standard FiLM conditioning mechanism described in the text. Similarly, several experimental figures primarily visualize numerical comparisons without providing additional analysis. The authors could improve the information density and clarity of these figures by highlighting the feature modulation mechanism more explicitly and incorporating more informative qualitative or diagnostic visualizations.

  6. The practical applicability of phase-conditioned segmentation requires further validation. The main experiments assume access to ground-truth surgical phase labels at inference time. Although the authors simulate phase recognition errors through random label corruption, this setting may not adequately reflect the temporally correlated and phase-dependent errors produced by real phase recognition models. An end-to-end evaluation using predicted surgical phases would provide more convincing evidence of the method's practical effectiveness.

Questions for authors
  1. Could the authors clarify the novelty of Phase-FiLM relative to standard FiLM-based conditioning? It would be helpful to better distinguish the methodological contribution from a straightforward application of existing FiLM conditioning.

  2. Can the authors further improve the presentation quality of the manuscript? The current abstract is unusually dense, several figures are visually simplistic.

Reviewer_tVruOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Phase-FiLM, which incorporates surgical phase information into SAM2 through feature-wise linear modulation for vessel segmentation in laparoscopic gastrectomy. Using LapGC-KVAD-30 dataset, the authors demonstrate that phase labels are capable to predict vessel presence (macro-F1 0.810). On binary vessel segmentation, Phase-FiLM improves mean Dice from 0.5892 to 0.6245 across three seeds. Ablations, cross-backbone comparisons, and simulated phase-label errors further investigate the benefits and limitations of phase conditioning.

Critical strengths
  • Meaningful use of surgical workflow information. The paper quantitatively demonstrates that surgical phase contains substantial information about vessel visibility, providing a motivation for phase-aware segmentation.

  • Consistent empirical improvements. Phase-FiLM improves Dice across three random seeds and two backbones. The concatenation and incorrect-phase controls provide useful evidence that phase conditioning contributes to the improvement.

  • Clean, parameter-light conditioning design with a control set.

Critical weaknesses
  • The diagnostic is not convincing. The image-only comparator is a compact CNN and is not capacity-matched to the SAM2 segmentation model; its near-chance 0.136 macro-F1 is hard to reconcile with the no-phase SAM2 baseline reaching 0.589 Dice for vessel-versus-background on the same images. The phase-only 0.810 is largely a learned per-phase presence prior/co-occurrence lookup. The experiment may therefore show that the compact CNN underfits, not that phase dominates appearance. This matters because the paper’s motivation and title rest on this claim.
  • The evaluation is too narrow for the stated claims. The 15-class problem is collapsed to binary union segmentation with only mean Dice; there are no per-class or boundary metrics and no false-positive analysis, so it is unclear whether phase improves localization or mainly suppresses vessels that are implausible in the current phase. A single 30-procedure dataset, correlated frames, and one concat control do not support the broad conclusion that multiplicative modulation is the right general principle.
  • Although the paper emphasizes fine-grained recognition of 15 vascular structures, segmentation is evaluated only using binary union masks. This does not demonstrate correct identification of individual vessels, limiting the claimed contribution to anatomical understanding and clinical decision support.
Questions for authors
  1. Could you report 15-class presence macro-F1/AP for an image-only classifier built on the same SAM2 features, or a comparable pretrained encoder? If the 0.136 gap shrinks substantially, how does that affect the “phase dominates appearance” claim?
  2. Can you provide per-class segmentation Dice, boundary metrics, and false-positive rates by phase, and clarify whether the gain comes from better localization or from suppressing phase-implausible vessels? Why was the primary task defined as binary union rather than 15-class segmentation?
  3. Can you evaluate the cascade with an actual phase recognizer, including errors near phase transitions, instead of uniformly random corruption? Does the 82.6% break-even still hold?
  4. Please reconcile the stated 17K versus 0.1M parameter counts.
  5. Do the improvements persist when evaluating the 15 individual vessel classes rather than their binary union? This would substantially strengthen the central claim.
  6. Can the authors evaluate Phase-FiLM using predicted phase labels from a phase recognizer trained on the target surgical dataset? The current random-error simulation may not reflect systematic or temporally correlated mistakes.
Reviewer_ocu7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a new method Phase-FiLM which uses the surgical phase as a prior for vessel recognition and segmentation in surgery videos. Phase-FiLM is a parameter-light mechanism that injects an 8-way phase embedding as channel-wise scale/shift modulation of SAM2 image features. On LapGC-KVAD-30, Phase-FiLM improves the performance of the no-phase baseline and transfers across backbones.

Critical strengths
  1. The paper is well-motivated and the writing is easy to follow.
  2. The proposed method is simple, light-weighted and can achieve real-time performance according to the reported results.
Critical weaknesses
  1. Inconsistent definition of main tasks. Section 3.1 defines the primary task as "binary segmentation of the union mask over the 15 classes (metric: mean Dice)". However, Figure 5 shows "Qualitative multi-vessel segmentation results" and Line 413-414 claims the model "accurately delineates multiple vessel classes against complex anatomical backgrounds". So it is not clear what is the primary task. Is it binary or multiple vessel segmentation?

  2. Line 239-240 states the improvement is "statistically robust rather than a single-seed artifact", but Table 2 reports only three seeds which is too small. It is suggested to try more seeds to support the claim.

  3. The nubmer of compared method in the experiments is limited, including only baselines and Camyla (Gao et al., 2026). It is suggested to compare the proposed method to more vessel segmentation methods in order to make the results more convincing.

  4. The introduction part is too short. It is suggested to clarify the motivation and add some short introduction of the proposed method before introducing the contributions.

Questions for authors
  1. Please clarify the primary task.
  2. Please add more seeds of the experiments and provide statistical analysis.
  3. Please add more compared methods if possible.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.