List of Papers Browse by Subject Areas Author List
Abstract
Ventriculostomy, or external ventricular drain (EVD) placement, is frequently performed freehand in urgent bedside settings, resulting in high catheter misplacement rates. Portable augmented reality (AR) guidance offers a low-cost alternative to conventional navigation systems, but mobile and standalone XR hardware impose strict latency, memory, and power constraints. A key enabling component is reliable real-time segmentation of the thin, low-contrast EVD catheter in unconstrained open-field environments. To address this challenge, we propose a real-time EVD catheter segmentation framework for mobile and XR deployment that includes a structure-aware loss promoting consistency along the catheter trajectory. We construct a scene-structured bedside dataset and complement it with a 3D-driven synthetic data generation pipeline to expand geometric and viewpoint variability. Lightweight backbones are evaluated under real-only, synthetic-only, and combined supervision, with systematic cross-domain and on-device analysis of robustness and accuracy–latency trade-offs under mobile/XR hardware constraints. Our results establish the feasibility of real-time catheter segmentation as a key component of future low-cost AR-guided ventriculostomy workflows. Code is available at: https://github.com/m-seddiqi/evd-catheter-segmentation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5965_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/m-seddiqi/evd-catheter-segmentation
Link to the Dataset(s)
https://huggingface.co/datasets/mussed/evd_catheter_segmentation_datasets
BibTex
@InProceedings{SedMus_DeviceConstrained_MICCAI2026,
author = { Seddiqi, Mustafa AND Castillo, Joshua AND Popa, Tiberiu AND Kersten-Oertel, Marta},
title = { { Device-Constrained Real-Time EVD Catheter Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16893},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper recognizes that while portable AR guidance can offer a low-cost alternative to conventional navigation systems, for external ventricular drain (EVD) placement, current standalone hardware imposes strict latency, memory, and power constraints. The main concern with these limited resources is whether the critical aspect of real-time segmentation of thin, low-contrast structures, such as the EVD catheters, can be achieved under the constraint of such low-power hardware. This paper formulates catheter segmentation as a thin-structure perception problem under resource constraints. It achieves this goal by introducing a loss function that favors longitudinal continuity of elongated predictions. Several lightweight segmentation backbones are evaluated, and their results support the feasibility of low-cost AR guidance for bedside ventriculostomy. The paper also emphasizes the value of complementing real training data with synthetic data to enhance the segmentation algorithm’s robustness.
- 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.
This paper recognizes that, to be accepted in clinical practice, an AR application must operate continuously at video rates to gain acceptance from the clinical community. It provides valuable insight into how formulating the problem under operational and hardware constraints can result in an effective solution that remains viable even under limited hardware capability. Two further contributions support the impact of this paper. In light of limited real-world training data, they construct a comprehensive scene modeling environment that yields 160K synthetic images, enabling the segmentation network to be trained under a wide range of expected clinical conditions. They demonstrate that, when combined with limited real-world data, this approach significantly enhances the algorithm’s performance, preventing overfitting. They also evaluate their algorithm using a range of segmentation backbones, as well as 32- and 16-bit FP GPUs and an 8-bit NPU, clearly illustrating the trade-off between latency and accuracy, as benchmarked on contemporary smartphone hardware.
The paper is well written and contains comprehensive relevent citations.
- 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.
Weaknesses. It is not entirely clear that the photometric enhancements of the real dataset are balanced against real-world experience. Is there a danger that such augmentation may bias the model? This could be clarified in a revision. The major weakness in this paper (which the authors acknowledge) is that the real dataset is modest and does not incorporate clinical workflow, so at this stage, it is not clear whether a more realistic training set would place more demands on the limited hardware available.
- 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?
A valuable example of how task-specific problems can be adapted to limited hardware resources. This is a particularly important consideration in adapting frugal solutions for low-resource environments.
- Reviewer confidence
Very confident (4)
- [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 provided satisfactory responses to the reviewers’ comments. While this is not an example of a mature, clinical-ready system, it nevertheless provides valuable insight into an approach to addressing a clinically relevant problem.
Review #2
- Please describe the contribution of the paper
Addressing the inherent limitations of conventional external ventricular drains (EVD), which can lead to an increased risk of catheter misplacement, the authors present a portable augmented reality (AR) guidance method aimed at providing a low-cost solution for mobile and extended reality (XR) platforms. More specifically, the paper considers computer vision-based catheter segmentation—an essential prerequisite for accurate AR-assisted guidance—as a thin-structure perception problem, due to the catheter’s thin, needle-like geometry. In this context, the authors claim three main contributions: A formalization of real-time EVD catheter segmentation under the hardware constraints of mobile and XR platforms. The creation of an image dataset for training generative AI models for catheter segmentation, consisting of both real annotated and synthetic images. A comprehensive cross-domain and on-device evaluation using the proposed dataset and multiple backbone architectures. The evaluation results support the authors’ hypothesis that accurate thin-structure catheter segmentation can be achieved in the context of AR-assisted EVD placement, even under real-time inference constraints on mobile and XR platforms.
- 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.
Overall, the paper is well thought out, with a strong motivation, a clearly articulated three-fold contribution, and a meaningful attempt to address a research gap by focusing on embedded mobile and XR platforms. In addition, the authors introduce a dataset comprising both real and synthetic data to train segmentation-focused generative AI models. The paper proposes a solid image data augmentation strategy to support domain generalization. In particular, the use of a task-aware photometric augmentation pipeline—rather than generic augmentation—effectively incorporates catheter-specific thin-structure constraints. The experimental evaluation is comprehensive. It includes multiple backbone architectures, three distinct training regimens (combined supervision with real and synthetic data, synthetic-only training, and real-only training), and reports both accuracy and on-device latency (inference time) under different numerical precisions. The inclusion of ablation studies and failure analysis further strengthens the evaluation.
- 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 authors motivate the creation of their dataset by highlighting the lack of publicly available datasets for EVD catheter segmentation. However, it is not stated whether the proposed dataset will be made publicly available, which limits its potential impact on the research community. 2.) There is no explicit evaluation of the intended surgical use case, namely portable AR-guided catheter placement. As a result, the practical applicability of the proposed method in a clinical setting remains unclear. The mobile phone (iPhone 15) and XR headset (Meta Quest 3) appear to be used primarily for data collection rather than for validating the AR-guidance scenario. Reproducibility is limited, as the availability of the code and dataset is not discussed. 3.) A figure illustrating the complete dataset generation process and catheter segmentation pipeline is missing. Including such a visualization would improve clarity and reproducibility. Additionally, some implementation details could be deferred to supplementary material to further support reproducibility. 4.) Figure 3: The text in the leftmost image is difficult to read. Adjusting the background (e.g., shifting it to emphasize the hands, while removing other parts of the background) could allow more space for the overlaid image, improving readability of the text.
- 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 provide sufficient information for 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
Dear authors, please consider the weaknesses outlined above, which I believe could help strengthen your work. In addition, I would like to raise the following questions: In Section 2.2, you state that ground-truth masks were generated using SAM 2 and SAM 3, followed by manual refinement. Could you elaborate on potential domain-specific limitations of these models? In particular, how do they perform on very thin structures such as catheters? Furthermore, what measures were taken to ensure that the manual refinement process resulted in accurate and reliable ground-truth annotations? Could you clarify your envisioned AR-assisted catheter segmentation scenario for mobile versus XR hardware? A more concrete description of potential surgical setups and how the system would be integrated into clinical workflows would be a valuable addition.
- 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?
Overall, this work represents a solid contribution. However, addressing the points raised in the weaknesses section is important to improve clarity, reproducibility, and practical relevance, and would significantly strengthen the paper.
- Reviewer confidence
Very confident (4)
- [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.
I believe the authors have addressed all reviewer comments sufficiently, and their plan for improving the paper appears feasible. Provided that all reviewer comments are adequately incorporated into the revised manuscript, the presented work could be an interesting contribution to the research community.
Review #3
- Please describe the contribution of the paper
1.A real‑time segmentation pipeline for EVD catheters, designed for low‑cost augmented reality guidance at the bedside.
2.A scene‑structured dataset and task‑specific augmentation/synthetic data generation to support cross‑domain evaluation.
3.On‑device inference analysis demonstrating feasibility for deployment in bedside ventriculostomy.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper addresses a clinically relevant problem (bedside ventriculostomy) with a clear practical motivation.
1.The work is framed in a deployment-oriented manner, focusing on real-world applicability rather than purely theoretical advances.
2.A scene-structured dataset is introduced, which is a valuable asset for the targeted task.
3.The evaluation includes cross-domain testing and on-device inference analysis, demonstrating feasibility beyond standard laboratory settings.
- 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 largely combines known components without substantial novelty.
2.The structure‑aware loss is not shown to be a clear improvement over a simple regularizer.
3.The real dataset is small and non‑clinical, baselines are narrow, and ablation is limited.
4.No end‑to‑end validation of AR guidance (e.g., trajectory accuracy, overlay stability, workflow benefit).
- 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 provide sufficient information for 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?
This manuscript studies real-time segmentation of external ventricular drain (EVD) catheters for low-cost augmented reality guidance in bedside ventriculostomy. The paper is clinically relevant and practically motivated, and its strongest aspects are the deployment-oriented framing, the scene-structured dataset, and the inclusion of cross-domain and on-device evaluation. However, in its current form, the submission is not yet fully convincing for MICCAI because the methodological novelty is moderate, the experimental validation remains incomplete, and the translational claims are stronger than the current evidence supports. I therefore recommend weak reject.
1.The methodological contribution is useful but not sufficiently strong. The overall system mainly combines known ingredients—lightweight backbones, an FPN decoder, task-specific augmentation, synthetic data generation, and a simple structure-aware loss. The practical framing is valuable, but the algorithmic advance over existing thin-structure or real-time segmentation methods is limited. In particular, the proposed structure-aware loss is intuitive, yet the paper does not fully demonstrate that it provides a substantial methodological improvement rather than a modest task-specific regularizer.
2.The experimental study is not yet strong enough to support the main claims. The results are promising, especially the cross-domain and on-device analyses, but the evaluation remains incomplete. The real dataset is small and laboratory-based rather than clinical, baseline coverage is narrow, and the ablation study is limited to a single backbone and scene. As a result, the paper supports the claim that the approach is promising and feasible, but not yet the stronger claim that it is a clearly established solution for robust bedside deployment.
3.The translational impact is not demonstrated end to end. The paper motivates the work in the context of AR-guided ventriculostomy, but the actual evaluation stops at segmentation. There is no direct analysis of trajectory estimation accuracy, AR overlay stability, workflow benefit, or clinical usability. This leaves a gap between the perception results and the broader conclusion that low-cost AR guidance is practically supported.
Overall, this is a relevant and carefully executed submission on an important clinical problem, and the deployment-oriented evaluation is a real strength. The main issue is not lack of value, but that the current manuscript draws somewhat stronger conclusions than the present evidence can fully justify. With stronger baselines, broader ablations, clearer validation of the structure-aware loss, and at least some end-to-end guidance evidence, the paper would become considerably more compelling.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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 partially addresses novelty and the structure‑aware loss with quantitative gains. However, the major experimental weaknesses I identified remain unresolved: the real dataset is small and non‑clinical, baselines are narrow, ablation is limited to a single backbone and scene, and there is no end‑to‑end validation of AR guidance (trajectory accuracy, overlay stability, workflow benefit). Reframing the scope as segmentation feasibility does not remedy the lack of evidence for the claimed AR application. Therefore, I maintain my recommendation of reject.
Author Feedback
We thank the reviewers and meta-reviewer for their constructive comments. We address the main concerns below.
Novelty (R3): To our knowledge, this is the first study to propose a thin-structure catheter segmentation method and evaluate it using external surgical videos with measured FP16/INT8 deployment on low-cost hardware. Thin-structure segmentation is challenging even outside clinical settings; in bedside ventriculostomy, low contrast, blur, occlusion, and extremely small catheter pixel occupancy further increase fragmentation and geometric instability. The paper introduces continuity-aware supervision, task-specific augmentation, scene-structured synthetic training, and cross-domain deployment evaluation for robust mobile inference.
Scope/MIC-CAI positioning (Meta, R2, R3): We believe that the contribution is primarily MIC with CAI application: a real-time, device-constrained EVD catheter segmentation under mobile/XR constraints for AR guidance. We agree that we do not propose a complete AR guidance system; the study focuses on the segmentation required for trackerless AR-guided EVD placement. The evaluation matches this scope: mobile/XR recordings test realistic acquisition, external videos test cross-domain robustness, and S23 FP16/INT8 benchmarking tests embedded deployment feasibility. Thus, the work establishes feasibility of real-time catheter segmentation under mobile/XR constraints, not end-to-end clinical AR validation. We will revise the paper to keep scope and claims aligned.
Structure loss (R3): Lightweight mobile models frequently produced fragmented catheter predictions under blur/low contrast. The proposed structure-aware term targets this failure mode by aligning predicted and GT responses along the catheter principal axis, encouraging longitudinal continuity rather than disconnected responses. Unlike a size/area prior, it encodes catheter-specific geometry and continuity. In FastViT-T8 ablation on surgeon-aligned Scene 2, adding the structure term improves mIoU from 50.8→56.1, while catheter-aligned blur further improves performance to 57.7.The +5.3 mIoU gain indicates complementary geometric information beyond augmentation alone. We will clarify that this is a practical catheter-oriented regularizer rather than a universal thin-structure loss.
Method clarity (R1–R3): We will add a schematic clarifying both real and synthetic data branches. The real branch uses mobile/HMD and external-test frames with SAM2/3-assisted initialization followed by manual correction and QC; the synthetic branch uses SAM-3D-inspired reconstruction with geometry-rendered images and masks. Training uses combined real/synthetic supervision, while surgical videos remain strictly test-only. Foundation-model masks were not accepted as final labels; blurred/low-contrast masks frequently required correction. Synthetic labels are rasterized directly from rendered 3D geometry, avoiding annotation noise. Validation/test data used only letterbox resizing. We will add augmentation/rendering ranges, split details, optimizer/export/timing settings, and improve Fig. 3. Augmentation bias (R1, R2): We agree that photometric augmentation can bias the model if not grounded in realistic catheter failure modes. All augmentations were training-only, with fixed ranges and unchanged masks. Importantly, the ablation shows that not all augmentation improves robustness: adding random exposure increased validation mIoU but reduced Scene-2 mIoU from 57.7→53.8, suggesting global exposure perturbations can suppress low catheter contrast. These results indicate that thin-structure augmentation must remain task-aware and be evaluated using cross-domain robustness rather than validation accuracy alone.
We will revise the manuscript and release code, training configurations, and data. We thank the reviewers again for recognizing the clinical relevance, deployment framing, cross-domain testing, and on-device evaluation.
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 are on the fence between weak reject and weak accept. It was not made clear by the reviewers whether this is considered a MIC contribution or CAI. An invitation for rebuttal is warranted.
- 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.
Although one reviewer has retained a “reject” ranking, two other reviewers are post-rebuttal ranking at a ‘weak accept’ on average, and so I am moving this to provisional accept.
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.
This submission addresses real-time segmentation of EVD catheters for low-cost AR-guided ventriculostomy under relevant hardware constraints. The main strengths consistently identified are the deployment-oriented framing, the synthetic+real dataset design, and the evaluation across multiple hardware settings.
Post-rebuttal, the key disagreement concerns the adequacy of validation for the intended AR use case. R1 and R2 accept the rebuttal and upgrade their decisions to accept, considering the clarifications sufficient and sufficient technical contribution. R3 maintains concerns that the work lacks end-to-end AR validation and views the method as incremental, leading to reject. The divergence is therefore mainly about whether the evaluation gives sufficient method support.
Taking these perspectives together, the work is best viewed as a solid but early-stage demonstration of a real-time, hardware-constrained catheter segmentation method, with promising engineering contributions. Although the validation is limited, the majority of reviewers agree that it is sufficient for acceptance in the context of a CAI contribution with a methodological focus.
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.
Reviewers appreciated the rebuttal’s acknowledgment of weaknesses, including the dataset size, but also pointed out that the limited evaluation is appropriate for a CAI paper. While not a mature, clinical ready system, the paper should generate interest at the conference.