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

Abstract

Precise instance segmentation of individual hair strands is fundamental to automated trichological analysis, yet it remains challenging due to their extreme thinness, frequent entanglement, and complex background interference. Existing weakly supervised approaches that rely on bounding-box annotations often suffer from geometric ambiguity, while the standard Segment Anything Model (SAM) exhibits severe over-segmentation when applied without task-specific guidance. To address these limitations, we propose R2P-SAM, a cost-effective framework that exploits sparse root annotations to guide a frozen SAM backbone. The core of our method is a geometric prior calibration mechanism that distills orientation and centrality fields to generate accurate point prompts, enabling effective separation of overlapping hair strands. Extensive experiments on three diverse datasets demonstrate that the proposed R2P-SAM consistently outperforms state-of-the-art weakly supervised methods, achieving superior boundary adherence while substantially improving annotation efficiency. The code is available at https://github.com/YYleafsheep/R2P-SAM.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2360_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/YYleafsheep/R2P-SAM

Link to the Dataset(s)

FDU_HairFollicleDataset: https://drive.google.com/file/d/1rMw8OzgxuTOSqKD8EBPN_BYXW18q9d3H/view?usp=sharing Hair Follicle_indi Dataset: https://universe.roboflow.com/mobal-mobal/hair-follicle_indi

BibTex

@InProceedings{HuHai_R2PSAM_MICCAI2026,
        author = { Hu, Haigen AND Yu, Yang AND Gong, Chenquan AND Zhang, Huihuang AND Ruan, Su AND Su, Yiping},
        title = { { R2P-SAM: Root-to-Prompt Hair Instance Segmentation with Geometric Priors under Point Supervision } },
        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

    This work introduces R2P-SAM, a cost-effective framework for hair strand instance segmentation that leverages sparse root annotations to guide a frozen SAM model. The method reformulates segmentation as a Root-to-Prompt generation task, transforming minimal point-level supervision into prompts. To address prompt sensitivity and off-center errors, a geometric prior calibration module distills orientation and centrality fields from SAM-derived pseudo-masks, enabling iterative refinement of prompts toward skeleton-aligned trajectories.

  • 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.Experiments across multiple datasets demonstrate that R2P-SAM achieves superior instance separation and boundary accuracy under dense overlap.

    2.Casting instance segmentation as a Root-to-Prompt generation task is conceptually clean and well-motivated. It aligns naturally with SAM’s prompting paradigm and avoids heavy redesign.

  • 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 highly relies on pretrained model like SAM, which is unfair when compared with methods in the point-supervision setting.

    2.The method is not compared against intuitive cascade methods. For example, if we train a detector (such as YOLO, FCOS, DETR, using a fixed box size if necessary), and use the detected box centers as the point prompts for SAM, what accuracy can be achieved?

    3.Based on the visualization, oriented object detection seems to be very suitable for hairs. It would be interesting to see if oriented boxes could be helpful in the task setting. [1] Wholly-WOOD: Wholly Leveraging Diversified-quality Labels for Weakly-supervised Oriented Object Detection [2] Point2RBox-v2: Rethinking Point-supervised Oriented Object Detection with Spatial Layout Among Instances

  • 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.

    (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 is overall good. While there are some weaknesses and questions, they can be addressed in a short period of time.

  • 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 paper is overall good.



Review #2

  • Please describe the contribution of the paper

    This paper proposes R2P-SAM, which formulates hair strand instance segmentation as a Root-to-Prompt generation task. The core of the method is a geometric prior calibration mechanism that distills orientation and centrality fields to generate accurate point prompts, thereby enabling more effective separation of overlapping hair strands. Extensive experiments demonstrate that R2P-SAM outperforms several state-of-the-art weakly supervised methods.

  • 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 idea of transforming sparse root signals into coherent, skeleton-aligned prompt trajectories is well motivated and technically interesting.

    2.The paper is clearly written and generally easy to follow.

    3.Extensive experiments are conducted to validate the effectiveness of the proposed method.

  • 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.R2P-SAM mainly samples a series of positive points as prompts for SAM-based segmentation. Have the authors considered incorporating additional prompt types, such as negative points or bounding boxes? It would be interesting to investigate whether these supplementary prompts could further improve segmentation performance.

    2.In the PGR stage, the method applies SAM to cropped image regions rather than to the full image. While cropping may improve segmentation granularity, it also increases inference cost and may make the comparison less fair when baseline methods do not use the same strategy. It would therefore be helpful for the authors to report the performance of the method without cropping. In addition, an ablation that removes PGR and directly applies coarse point prompts together with cropping would further clarify the contribution of the proposed design.

    3.Table 1 is currently not presented in a very polished format. I would recommend revising it using a standard three-line table style, consistent with the presentation of Tables 2 and 3.

  • 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 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.

    (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 idea of transforming sparse root signals into coherent, skeleton-aligned prompt trajectories is well motivated and technically interesting.

  • 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’ rebuttal has resolved my concerns.



Review #3

  • Please describe the contribution of the paper

    The paper proposes R2P-SAM, a framework for hair strand instance segmentation in trichoscopic images using only sparse root point annotations. The idea is to freeze SAM’s backbone and reformulate instance segmentation as a root-to-prompt generation task. The pipeline has two stages: first, an RPF-Net predicts a rootness probability heatmap from which candidate root points are extracted via spatial peak suppression. Second, a geometric prior calibration module (PriorNet) predicts three fields, a feasible region field, a tangent field, and a consistency field, which are used to iteratively refine the coarse root points into skeleton-aligned prompts that can be fed to frozen SAM. The consistency field acts as a gate at strand junctions to prevent the prompt trajectory from jumping to neighboring strands. They evaluate on three trichoscopy datasets and show improvements over general weakly supervised methods.

  • 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 problem formulation is sensible. Hair roots are spatially non-overlapping and clinically meaningful, so using them as anchor points for prompting SAM is a natural fit for this application. The root-to-prompt framing is a reasonable way to bridge the gap between sparse annotations and instance-level masks. The geometric prior calibration with three complementary fields is technically well thought out. The tangent field provides directional flow along the strand, the feasible region constrains where prompts can land, and the consistency field dampens updates at junctions where strand crossings create ambiguity. The ablation in Table 3 confirms that each component contributes meaningfully. The pixel-wise max heatmap generation instead of summation is a small but useful design choice. The intensity profile comparison in Fig. 3(d) clearly shows that summation merges adjacent roots while maximization preserves distinct peaks, which matters for dense strand regions. The qualitative results in Fig. 5 are convincing, especially on the Long-Hair dataset where overlapping strands are a real challenge. The method maintains strand continuity better than the baselines.

  • 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.

    One of the three evaluation datasets (Long-Hair) is private. This is a significant reproducibility concern, especially since it is the most challenging dataset in the paper and the one where the method shows the largest improvements. Without access to this data, the main results cannot be independently verified. All baselines are general purpose weakly supervised methods (DiscoBox, BoxLevelset, Box2Mask, SAPNet, P2Seg, PointSup) that were not designed for thin curvilinear structures. There is no comparison with methods specifically tailored for elongated objects like vessels or fibers, and no fully supervised upper bound is reported. Without knowing how far the results are from full supervision, it is hard to judge how well the method actually performs. The absolute performance numbers are still quite low. The best AP50 is 46.9% on Long-Hair and 50.0% on FDU. AP75 is around 23-31% across datasets. These numbers suggest the method is making progress but is far from being reliable enough for the clinical applications mentioned in the introduction. The paper claims clinical applicability for trichological analysis, but there is no validation with actual clinical metrics like hair density estimation, shaft diameter measurement, or terminal-to-vellus ratio computation. A downstream evaluation showing that the segmentations are good enough to extract clinically meaningful measurements would have strengthened the contribution. The geometric prior fields are derived from SAM’s pseudo-masks, which means the quality of the calibration is bounded by how good SAM’s initial coarse segmentation is. If SAM produces poor pseudo-masks for certain strand configurations, the geometric fields would also be unreliable. This dependency is not analyzed in the paper.

  • 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 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.

    (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?

    The paper presents a technically sound and well-motivated approach for hair strand instance segmentation using sparse root annotations to guide a frozen SAM backbone. The root-to-prompt formulation is appropriate for this application, and the geometric prior calibration with the three fields is well designed and supported by ablation studies.

    However, several limitations prevent a higher recommendation. The evaluation scope is limited, with reliance on a private dataset and comparisons primarily against general-purpose weakly supervised methods that are not tailored for thin structures. No fully supervised upper bound is provided, making it difficult to assess how far the method is from optimal performance. In addition, the absolute performance remains relatively low, raising concerns about practical usability. The claim of clinical applicability is not supported by downstream validation, and the method appears highly specialized to hair strands without evidence of broader generalization.

    Overall, while the approach is interesting and technically reasonable, the experimental validation and impact are not yet strong enough for acceptance in its current form.

  • Reviewer confidence

    Somewhat confident (2)

  • [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 rebuttal addresses my concerns but most of the responses point to things that will be done in the final version or future work, such as releasing the private dataset, adding fully supervised upper bounds, including failure mode analysis, and validating with clinical metrics. Per MICCAI guidelines, the decision should be based on the submitted manuscript. On W1 (reproducibility), the promise to release code and annotations is appreciated but cannot be verified at this stage. On W2, the explanation that creating full mask annotations is time-consuming is understandable, but it does not resolve the evaluation gap. On W3/W4, the argument about COCO IoU penalizing thin structures is a valid observation, but if the authors are aware of this mismatch, an alternative evaluation metric should have been included in the submission. On W5, the quality-aware weight q and the supporting ablation partially address the concern. I acknowledge that the other two reviewers found the method interesting and gave Weak Accepts. The root-to-prompt formulation is a reasonable contribution. However, the combination of a private dataset, no upper bound, low absolute numbers without an alternative metric to justify them, and unvalidated clinical claims leaves too many open questions for me to change my recommendation.



Author Feedback

We thank all reviewers and address questions below: Reviewer 1 W1 (Fairness): Weakly supervised fairness relies on annotation cost, not backbone size. Using frozen models as infrastructure is a standard paradigm. Due to severe domain shifts, SAM’s pre-training advantage is marginal. Table 1 shows zero-shot SAM performs poorly, indicating our gain stems mainly from our root-to-prompt and geometric calibration design, not just SAM. W2 (Detector cascades): The suggested “fixed box size” fails for curvilinear hairs. Geometric box centers often hit background or intersecting strands, cascading misdetections to SAM. Moreover, detectors merely replace root localization, yielding isolated endpoints. Lacking centerline trajectories, they fail on overlaps (Fig. 1c). R2P-SAM solves this via learned topological priors instead of geometric centers. W3 (Oriented boxes): While oriented boxes seem intuitive, they require higher annotation cost than point supervision. Rigid rectangles cannot fit curvilinear topologies, causing angle ambiguity and encompassing neighboring strands. Learning morphology directly via point features is more natural. We will explore oriented boxes in future work. Reviewer 2 W1 (Negative prompts): Bounding boxes violate point supervision, increase annotation costs, and suffer geometric ambiguity on crossing structures (Table 1). Furthermore, automatically defining reliable negative points without dense masks is challenging and introduces noise. Instead, our PriorNet provides implicit negative guidance by suppressing negative roots via Eq. 2.W2 (Cropping): The requested ablation (cropping without PGR) aligns with Table 3’s first row, yielding 0.188 AP75.Full PGR improves this to 0.235, proving geometric calibration, not cropping alone, drives the gain. Second, adapting SAM to high-res images causes severe scale mismatch for thin hairs and prohibitive VRAM limits. Processing uncropped full images squeezes strands into sub-pixel features, causing SAM to fail. Thus, cropping is a mandatory engineering step. We will detail no-cropping baselines in the final version. W3 (Formatting): We will revise Table 1 into a standard three-line format. Reviewer 3 W1 (Reproducibility): To facilitate future research, we will open-source our annotated instance test sets (long/short hair) and the complete R2P-SAM codebase in the final version. W2 (Upper bounds & Baselines): Public hair datasets lack instance masks. We manually annotated test sets for reliable evaluation. Since training set annotation requires drastically more effort, we only created point labels, precluding fully supervised bounds (e.g., Mask R-CNN). Regarding specialized baselines, most vessel/fiber methods target fully supervised or semantic segmentation, misaligned with our weakly supervised instance setting. As dense annotation requires months of expert labor, we will complete and release full masks for upper bounds in future work. W3 & W4 (AP & Clinical): Low AP partly reflects a known mismatch between COCO IoU and 1-pixel thin structures. A 1-2 pixel shift drastically drops IoU despite visually correct topology. As Fig. 5 illustrates, R2P-SAM preserves strand continuity compared with baselines. Moreover, accurate instance-level separation of entangled hairs is a prerequisite for reliable trichological analysis. By mitigating strand merging, R2P-SAM enables downstream measurements (e.g., density, shaft width), which we will explore in future work. W5 (Pseudo-masks): We mitigated this via a quality-aware weight q (Eq. 2) based on geometric consistency, down-weighting losses for unreliable pseudo-masks. Table 3 confirms this: our baseline yields 0.188 AP75, while full PGR reaches 0.235.This proves our framework actively calibrates unreliable masks rather than being bounded by them. In extreme blur/entanglement where SAM fails completely, prior derivation is constrained. We will discuss failure modes in the conclusion of the final version.



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 acknowledge a clean, application-targeted SAM adaptation with a well-motivated two-stage design and consistent gains across three trichoscopy datasets. Two are weakly positive overall, while one is moderately negative on grounds of reproducibility (the most challenging dataset is private), limited baseline scope (no thin-curvilinear-specialist comparison, no fully-supervised upper bound), and unsupported clinical-applicability claims. The split warrants a rebuttal phase.

    For the rebuttal, the most pressing items are the cascade-baseline experiment to isolate the contribution beyond a generic detector→SAM pipeline, and reproducibility steps for the private Long-Hair dataset. Further items raised by the reviewers — including a no-cropping ablation, a fully-supervised Mask R-CNN upper bound, and a discussion of SAM-pseudo-mask failure modes — should also be addressed.

  • 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 presents R2P-SAM, a coarse-to-fine pipeline for generating robust point prompts to drive a frozen SAM backbone for hair-strand instance segmentation, combining Rootness Probability Field learning with Geometric Prior-Guided Calibration. The root-to-prompt formulation is conceptually clean, well aligned with SAM’s prompting paradigm, and supported by extensive experiments across multiple datasets. After the rebuttal phase, two reviewers recommend acceptance, while R3 maintains a reject on the grounds of the private dataset, the absence of a fully supervised upper bound, low absolute performance, and unvalidated clinical claims. In my view, the conceptual strength of the root-to-prompt formulation and the breadth of the empirical evaluation merit acceptance on balance. That said, I share R3’s reproducibility concern, and would add that the manuscript lacks sufficient mathematical specification of the loss functions and the Geometric Prior Fields, which limits reproducibility independently of dataset access. I strongly urge the authors to revise the manuscript to clarify these points, and to release the code and dataset as committed in the rebuttal.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation.

    The authors provided additional clarifications, yet there are still some concerns about fully-supervised results and robustness / generalizability analysis such as evaluation on public datasets.



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 develops a root-to-prompt mechanism tailored to thin, overlapping hair strands, with geometric fields for feasible regions, tangent direction, and consistency at strand junctions. The camera-ready should substantially temper the claims. The clinical applicability claims are not yet supported by downstream hair-analysis metrics such as hair density, shaft diameter, or terminal-to-vellus ratio.



back to top