List of Papers Browse by Subject Areas Author List
Abstract
Accurate estimation of wound dimensions during the main incision step of cataract surgery is a critical indicator of surgical proficiency, since a properly made wound that allows the widest part of the incision knife to pass through minimizes the risk of bleeding. In this work, we propose a novel, training-free framework that integrates Segment Anything Model 2 (SAM2) with automatic prompt generation via the Hough Transform to segment incision trajectories and estimate wound dimensions directly from surgical videos. By leveraging geometric priors to guide zero-shot segmentation, our approach eliminates the need for task-specific segmentation annotations or model finetuning.
We evaluated the method on the public Cataract-1K dataset, using expert surgeon annotations as ground truth. Quantitative results demonstrate high balanced accuracy in predicting incision architecture and size ratings as indicators of skill, using only the features derived from the predicted segmentation masks. In general, our findings highlight the potential of combining zero-shot generalizable segmentation models with classical geometric priors for fine-grained surgical video analysis and skill assessment.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2588_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Octobserver/hough-sam.git
Link to the Dataset(s)
N/A
BibTex
@InProceedings{DaiWei_HoughSAM_MICCAI2026,
author = { Dai, Weina AND Paranjape, Jay AND Sikder, Shameema AND Patel, Vishal AND Vedula, Swaroop},
title = { { HoughSAM: Geometry-Guided Zero-Shot Segmentation of Main Incision in Cataract Surgery Using SAM2 for Skill Assessment } },
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 HoughSAM, a geometry-guided pipeline that combines Hough-based automatic prompting with SAM2 to segment the main incision in cataract surgery videos without manual annotations or finetuning, and demonstrates that segmentation-derived wound geometry features can support interpretable skill assessment. To be specific, the authors avoid direct wound segmentation in this pipeline because the wound in cornea is transparent and unable to provide enough semantic information for segmentation models. HoughSAM first generate limbus and pupil prompt points through Hough circle detection and pass them to SAM2 for limbus and pupil segmentation excluding the knife. The final wound mask is acquired by applying convex hull completion to the pupil mask and retain the knife area only.
- 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 addresses a practically meaningful surgical problem. Rather than performing a general full-length skill assessment, the authors focus on the segmentation of the main incision in cataract surgery, which is directly related to surgical proficiency. The paper also attempts to validate whether the extracted wound segmentation is useful for downstream skill assessment, which strengthens the practical relevance of the study.
2.The paper presents a zero-shot strategy for a task where dense annotations are difficult to obtain. The method does not rely on pixel-level wound annotations and finetuning. Instead, it leverages anatomical and geometric priors of the eye, in particular the limbus and pupil structure, to automatically generate prompts for SAM2 and infer the wound region indirectly, which is an original way to handle a challenging transient target in a low-annotation setting.
- 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 introduction makes an overly broad claim that existing skill assessment methods lack sufficiently granular and context-specific evaluation. This statement is not fully supported by the cited literature, which is entirely focused on cataract surgery. The related work should also include more recent cataract-specific studies, such as CatSkill (Giap B D, Ballouz D, Srinivasan K, et al. CatSkill: artificial intelligence-based metrics for the assessment of surgical skill level from intraoperative cataract surgery video recordings[J]. Ophthalmology Science, 2025, 5(4): 100764.), which performs phase-wise skill assessment across 10 surgical phases.
2.The overall pipeline is largely implementation-driven, and many of its design choices appear heuristic. For example, the circle selection procedure after Hough detection relies on several manually chosen thresholds without sufficient justification, and it is also unclear why the valid pupil/limbus pair should be selected by minimizing the radius difference. These choices may substantially affect robustness.
3.The critical preprocessing module is insufficiently specified. This stage aims to clean the SAM2 masks so that the missing region caused by knife occlusion can later be interpreted as a wound-related region. However, it is unclear whether all missing regions can indeed be attributed to the wound, rather than segmentation errors, illumination artifacts, or other occlusions. The paper also does not describe the key details of the preprocessing operations, including morphological closing, connected component filtering, and polygonal approximation. These steps inevitably require parameter choices to determine which holes should be removed as noise and which concavities should be preserved as knife-induced structures. Likely, it is unclear how the authors ensure that the difference between the limbus convex hull and the original limbus mask corresponds specifically to knife-induced occlusion.
4.Although the proposed method appears qualitatively better than the compared baselines in Figure 2, the predicted masks seem to overfill the knife region, which could in turn affect the downstream skill-assessment features. By contrast, while the other methods show severe over-segmentation, their predictions tend to remain within the knife region.
5.The connection between the derived mask and the intended clinical target remains unclear. Conceptually, the true wound area may be better understood as the union of all knife trajectories through the limbal region over time, rather than the knife-related mask in a single frame. How are the frame-level measurements aggregated into case-level descriptors with this potential mismatch?
6.The quantitative results are not strong for a binary classification task. Moreover, this downstream task can be leveraged to validate the proposed segmentation pipeline itself, rather than trying to prove this task-specific features can be more useful than generic pretrained embeddings. A more informative quantitative result would be to extract the same downstream features from masks produced by direct SAM2 and MedSAM prompting and evaluate them under the same classifier setting.
- Please rate the clarity and organization of this paper
Poor
- 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.
(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?
1) The paper does not clearly establish the relationship between the knife region, the derived per-frame mask, and the actual wound area. This is important because the whole method and the downstream skill-assessment features are built on the assumption that the proposed geometric proxy meaningfully represents the wound. 2) The downstream skill-assessment results are relatively weak, which raises substantial questions about the practical utility of the approach 3) The method contains many insufficiently specified and heuristic design choices, particularly in the circle selection, preprocessing, and wound extraction stages. Several key operations are not described in enough detail, and some assumptions appear ad hoc rather than well justified, which makes it difficult to assess the robustness, reproducibility, and generalizability of the pipeline.
- 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 #2
- Please describe the contribution of the paper
This paper proposes HoughSAM, a training-free pipeline for segmenting the main incision wound in cataract surgery videos. The method uses Hough Transform circle detection to automatically generate point prompts for SAM2, which segments the limbus and pupil. Since SAM2 excludes occluded regions (where the knife disrupts the tissue), a convex hull completion and mask subtraction step isolates the wound area. Geometric features extracted from the wound masks are then used to predict expert ratings of incision architecture and size as indicators of surgical skill.
- 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.
Clever Problem Framing: The core insight here is great—instead of fighting SAM2’s sensitivity to occlusions, the authors actually weaponize it. Using the “missing” parts of anatomical landmarks to indirectly map the wound is a very elegant workaround for a structure that is notoriously hard to segment directly due to a lack of labels.
Zero-Overhead Deployment: The fact that this pipeline is both training-free and annotation-free is a huge practical win. By relying on geometric priors (the limbus and pupil) and a foundation model, the method bypasses the “data scarcity” trap entirely. It’s a “plug-and-play” approach that’s much more realistic for clinical settings than a model requiring heavy fine-tuning.
Strong Clinical Grounding: The paper doesn’t just focus on the math; it builds a solid case for why wound geometry matters. Linking the segmentation results to real-world surgical risks—like bleeding and self-sealing capability—gives the technical work a lot of credibility and purpose.
Intuitive Feature Engineering: The seven geometric features (like area growth rates and max/min ratios) are well-chosen. They manage to capture both the shape and the timing of the incision in a way that’s actually interpretable for a surgeon, rather than just being “black box” numbers.
- 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.
Single-Rater Reliability Issues: The skill labels depend entirely on one surgeon. In surgical skill research, a single expert’s opinion is usually considered too subjective to serve as ground truth. Without a second rater to establish inter-rater agreement, the validity of the labels—and therefore the model’s performance—is on shaky ground.
Rigid Geometric Assumptions: The pipeline relies heavily on the limbus and pupil being nearly circular and viewed from the front. The Hough Transform parameters also seem fine-tuned for this specific setup. Without a sensitivity analysis or testing on off-axis views, it’s unclear if this method would hold up in a real clinical environment where camera angles and eye shapes vary.
Narrow Baseline Comparisons: While comparing against SAM2 is a good start, the paper ignores a lot of relevant “prior art,” such as classical iris segmentation from biometrics or active contour models. On the skill assessment side, comparing only against frame-level features (CLIP/CSMAE) feels a bit thin—temporal models or standard video transformers should have been included as baselines.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
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 presents a creative idea — exploiting SAM2’s occlusion sensitivity combined with geometric priors to indirectly segment a transient surgical structure without training. This is a novel and practically motivated approach. However, the evaluation is the main weakness: the segmentation is only assessed qualitatively on three examples, the skill assessment results are modest and based on unreliable statistics (single split, ~16 test cases, single annotator), and the baseline comparisons are narrow. For a venue like MICCAI, the experimental rigor needs to match the ambition of the claims. A rebuttal that provides quantitative segmentation metrics (even on a small annotated subset), cross-validated skill prediction results, and inter-rater reliability would substantially strengthen the paper.
- Reviewer confidence
Not confident (1)
- [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 HoughSAM, a training-free framework for segmenting the main incision wound in cataract surgery videos. The key insight is to avoid segmenting the transient wound directly. Instead, the method uses the Hough Transform to automatically generate point prompts for SAM2 to segment stable anatomical structures, the pupil and limbus, then exploits SAM2’s occlusion sensitivity: regions disrupted by the incision knife are excluded from the predicted masks. Convex hull completion and mask subtraction are used to isolate the wound as the missing region. From the resulting wound masks, geometric features of wound ratios, area growth rates, max area, etc. are extracted and used to train lightweight classifiers that predict expert ratings of incision architecture and size as indicators of surgical skill. The method is evaluated on 80 clips from the public Cataract-1K dataset, with code and annotations to be released.
- 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.
Directly segmenting a transient tool-tissue interaction region like a wound is extremely difficult inlcuding a lack of available labelled data. The paper suggests using a different strategy of segment the pupil and limbus, and derive the wound from what is missing. This is a clever and well-motivated solution. Using SAM2’s occlusion sensitivity as a feature rather than treating it as a limitation is an insightful design choice.
The segmentation component requires no task-specific annotations or finetuning, which is a significant practical advantage given that no publicly available pixel-level wound annotations exist for cataract surgery. This makes the approach immediately applicable to other datasets.
The paper makes a strong case for context-specific skill assessment over global procedure-level assessment, grounding the work in the clinical significance of wound geometry for patient outcomes such as self-sealing, infection risk, instrument passage. Section 2.1 clearly defines what constitutes a well constructed main incision and is imperative for readers without prior knowledge.
The authors use a public dataset Cataract-1K and state that code and annotations will be released. This is appreciated and lowers the barrier for follow-up work.
Despite using only hand-crafted geometric features, HoughSAM features outperform both generic CLIP and domain specific CSMAE learned features for skill prediction, demonstrating that task-specific geometric reasoning can be more informative than general purpose representations for this problem.
- 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 skill ratings incision architecture and size come from a single expert. This means there is no inter-rater agreement statistics, the rating scale itself is also not clearly defined, the paper mentions around 70% of cases are labelled “good” but does not specify whether ratings are binary, ordinal, or what the full set of categories is. For a skill assessment paper, this is the main gap. Without establishing that the ground truth labels are, the downstream classification results are hard to interpret.
With 80 video segments and a 0.6:0.2:0.2 split, the test set contains approximately 16 cases. Combined with a class imbalance where 70% are “good”, the reported balanced accuracies could be quite noisy. Confidence intervals or results across multiple random splits would help establish whether these numbers are stable.
The comparison with SAM2 and MedSAM baselines is purely qualitative. While the absence of pixel-level ground truth annotations is acknowledged, this means the entire downstream evaluation rests on the assumption that the segmentation is accurate, supported only by visual inspection of selected examples. Even a small manually annotated subset for IoU/Dice evaluation would strengthen confidence.
The text refers to CSMAE features, but Table 1 labels the same row as VMAE features. This should be corrected to avoid confusion.
The paper does not discuss where the pipeline fails in its current setting. The Hough Transform assumes approximately circular limbal geometry and that the eye is roughly centred in frame. What happens under extreme viewpoint changes, non-standard eye geometries, or heavy instrument occlusion unrelated to the wound? A failure case analysis or at minimum reporting the fraction of clips where the pipeline produces reasonable output would be valuable.
- 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.
(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?
My recommendation is driven primarily by the paper’s originality, clarity, and practical relevance. The core idea of recovering the wound indirectly by segmenting stable anatomical structures and exploiting SAM2’s occlusion behaviour is simple, creative, and well matched to the annotation constraints of this problem. I found the paper well motivated clinically: it makes a convincing case that incision quality is an important, localised indicator of surgical skill, rather than treating skill assessment only at the whole procedure level. The method is easy to follow, appears lightweight and practical, and the use of a public dataset with planned code and annotation release further strengthens the papers value.
At the same time, the current evaluation leaves some important questions unresolved. In particular, the skill labels come from a single expert, the label definition is not fully specified, and the dataset is small enough that the reported classification results may be unstable. In addition, the segmentation quality is assessed mainly qualitatively, so confidence in the downstream analysis depends on visual examples rather than quantitative validation. These issues do not undermine the novelty of the approach, but they do limit how strongly the results can be interpreted.
Overall, I view this as an interesting and promising paper with a genuinely novel idea and a clear practical contribution. My score reflects that positive view, while also accounting for the fact that the validation is still somewhat preliminary. With stronger label validation, a more robust quantitative evaluation, and some analysis of failure cases, this work could become a significantly stronger paper.
- 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
Author Feedback
N/A
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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.
Overall, the paper has been positively received. All reviewers highlight the intuitiveness of the approach and applicability for zero-shot deployment. Reviewers also agree on key weakeness.
1) A single rater for skills annotation which increases subjectivity of ground truth.
2) Preliminary stage validation with purely qualitative comparison, naturally limiting the strength of conclusions regarding improved behaviour.
3) A limited comparison to segmentation pipelines for the task. The authors must justify the selected pipelines and why these are representative-enough to support the conclusions.
4) Hard geometric and segmentation assumptions that are not fully discussed and absence of failure cases.
My recommendation is for the paper to be accepte and the authors are strongly advised to address the reviewers’ point in the final submission