List of Papers Browse by Subject Areas Author List
Abstract
Intraoperative navigation in spine surgery demands millimeter-level accuracy. Currently, this is achieved through radiation-intensive intraoperative imaging and bone-anchored markers that are invasive and disrupt surgical workflow. Markerless RGB-D registration methods offer a promising alternative. However, existing approaches rely on weak segmentation labels to isolate relevant anatomical structures, potentially propagating errors through the registration process. We present End2Reg, an end-to-end deep learning framework that jointly optimizes segmentation and registration, eliminating the need for segmentation labels and manual steps. The network learns task-specific segmentation masks optimized for registration, guided solely by the registration objective without explicit segmentation supervision. End2Reg achieves state-of-the-art performance on ex- and in-vivo benchmarks, reducing median Target Registration Error by 32% and mean Root Mean Square Error by 61%, while maintaining robust performance under partial occlusions. Ablation results confirm that end-to-end optimization significantly improves registration accuracy. Overall, End2Reg advances towards fully automatic, markerless intraoperative navigation.
Code and interactive visualizations are available at: https://lorenzopettinari.github.io/end-2-reg/.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1921_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026/supp/1921_supp.zip
Link to the Code Repository
https://lorenzopettinari.github.io/end-2-reg/
https://github.com/lorenzopettinari/End2Reg
Link to the Dataset(s)
https://rocs.balgrist.ch/en/open-access/spinedepth/
https://huggingface.co/datasets/zcbecda/SpineAlign
BibTex
@InProceedings{PetLor_End2Reg_MICCAI2026,
author = { Pettinari, Lorenzo AND El Hadramy, Sidaty AND Wehrli, Michael AND Cattin, Philippe C. AND Studer, Daniel AND Hasler, Carol C. AND Licci, Maria},
title = { { End2Reg: Learning Task-Specific Segmentation for Markerless Registration in Spine Surgery } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16889},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The main contribution of the paper is development of an end-to-end framework for markerless registration to enable accurate intraoperative navigation, by bringing preoperative imaging to the intraoperative space and matching intraoperative RGBD data. The key technical contribution is a jointly optimization with both segmentation and registration.
- 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 critical need for initraoperative navigation without relying on radiation and markers. 2.A major strength is the joint optimization, where segmentation is learned for the purpose of downstream registration task. This is important in that segmentation accuracy itself does not necessarily reflect the utility in alignment. 3.Manual annotation was not involved. Combined with the joint optimization, It is particularly useful in differentiating soft tissue vs bone. 4.The method shows improvements over existing approaches across 2 publicly available datasets. The use of public datasets allows direct comparison to established baselines and supports the claim in performance improvements.
- 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.Sub-millimeter accuracy may not be necessary and challenging to achieve or prove using independently collected ground truth, given that measurement errors can be ~1 mm. The proposed method does not meet this level of accuracy. The authors may consider softening or clarifying the required accuracy thresholds for intraoperative navigation in spine surgery. 2.It is unclear how RMSE differs from TRE and should be explicitly defined and clarify how they relate to clinical accuracy and to each other. 3.The paper does not report computational efficiency, which is an important consideration for intraoperative use.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The work is clinically meaningful and technically sound. The joint optimization of segmentation and registration is smart and novel. The results are promising and show improvements over baseline approaches in public datasets. The weaknesses are minor and addressable.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
This paper presents the first directly learned registration pipeline for orthopedic markerless tracking, within the context of spine surgery. This is in contrast with previous approaches that relied on segmenting the depth surface and relying on traditional methods to register (i.e. ICP). The authors develop an architecture based on the GeoTransformer for this task.
- 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 authors effectively combine several existing architectures and methods in a novel manner for the task of orthopedic markerless tracking. The benchmarking of their method is very thorough, while the paper is well written with effective use of visualizations. Their method shows a large quantitative improvement over previous methods, even under occluded conditions.
- 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] Substantial text is devoted to re-explaining prior work rather than the authors’ own contributions. For example, the passage beginning “GeoTransformer builds upon KPConv…” through to “…between neighboring points” largely restates known material. Similarly, the Gumbel-Softmax estimator derivation, while central to the contribution, is a well-established technique and could be condensed, with readers directed to the original references for full detail. This space would be better used to expand on the authors’ specific design choices and ablations.
[2] It is unclear whether models were trained independently on each dataset and evaluated separately, or whether any form of cross-dataset training or transfer was employed. This should be stated explicitly, as it affects interpretation of the generalization claims.
[3] The authors restrict the task to single rigid transforms, citing dataset limitations. However, SpineDepth provides per-vertebra labels, which would allow artificial intervertebral articulation to be introduced to the preoperative CT scans. This would make the task non-rigid and substantially more clinically realistic, as the assumption of rigidity between preoperative imaging and intraoperative anatomy is rarely valid in practice. The authors should discuss why this was not explored.
[4] The method is framed broadly for orthopedic surgery, yet both datasets are spine-specific. Consideration should be given to how End2Reg would translate to other anatomies (e.g., hip, knee) where surface geometry, occlusion patterns, and overlap characteristics differ meaningfully.
[5] This method, as a transformer based architecture on the surface appears compute heavy. For a method targeting intraoperative use, runtime is clinically critical. No information is provided regarding inference speed or whether the pipeline could operate at rates compatible with real-time surgical navigation. This is a notable omission that should be addressed and again compared with classical methods.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The core contribution, end-to-end joint optimization eliminating segmentation labels, is well-motivated and yields clear empirical gains. However, the limited ablation scope, missing runtime analysis, and unexplored non-rigid extension etc. leave room for strengthening.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
Markerless RGB‑D registration methods provide a radiation‑free alternative to CT‑based registration in open‑spine procedures. These methods typically rely on extracting segmented surface features of the spine and then registering the CT segmentation masks to the intra‑operative RGB‑D surface mask. The authors propose End2Reg, a joint segmentation‑and‑registration network that mitigates segmentation noise and label‑uncertainty effects on the registration task. They report superior performance on publicly available datasets, especially when surface occlusion from surgical instrumentation is present.
- 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.
S1: Improved performance over state‑of‑the‑art models on public datasets, particularly under occlusion. In open‑spine procedures occlusion appears as (i) instrument blockage and (ii) soft‑tissue covering the bony anatomy, which both reduce available geometric features for CT‑to‑RGB‑D registration. Through end‑to‑end training the authors convincingly demonstrate the ability to overcome these challenges, achieving strong registration results. This approach is a novel methodological contribution with high potential to impact the spine‑navigation research community. S2: The authors provide a thorough comparison with existing methods and extensive ablation studies that clearly justify each component of their architecture.
- 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.
W1: The model is trained to minimize TRE/RMSE for the geometric registration task. Thus, the model is bounded by how good the underlying registraiton is between RGB-D imaging and preoperative/intraoperative imaging. If the ground‑truth alignment contains systematic bias, the learned segmentation may be optimised for a mis‑registered geometry, limiting generalisation. Discussion of this limitation and how to account for this should be expanded upon. While the authors state “Qualitative results show that our model highlights regions informative for registration, corresponding primarily to bony anatomy while suppressing surrounding soft tissue”, quantitative results are lacking that might further support that the correct anatomical features are being learned. W2: The limitation of a single global rigid transformation should be discussed further. The spine is highly mobile and deformation occurs between preoperative imaging and the intraoperative positioning of the patient such that a single rigid registration can still show surface misalignment. Discussion should be included of how this network could be adapted or combined with deformation or level by level registration approaches to account for this scenario experienced in clinical practice.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The authors propose a novel end-to-end training scheme for the RGB-D to volumetric imaging registration task. This scheme overcomes the segmentation challenges of RGB-D imaging in open spine surgery to enable downstream registration tasks. They show superior performance across benchmarking data sets. This is a strong methodological contribution to the community. However, further discussion of how to control the registration accuracy of CT/MRI to RGB-D is needed for generalizability and deployment of this approach particularly in a clinical scenario where intervertebral motion can reduce registration accuracy.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
Dear Reviewers, We sincerely thank you for your time, constructive feedback, and positive reception of our work. We have carefully considered your valuable suggestions and will incorporate them into the final manuscript as detailed below. 1.Rigid Transformation Assumption (Reviewer #2, #3) We fully agree that the spine is highly mobile and that intervertebral motion is a critical factor in clinical practice. We utilized a global rigid transformation as a simplifying assumption primarily due to the limitations of currently available datasets.
- SpineDepth: Although per-vertebra labels exist, the dataset features a small magnitude of intervertebral motion because it utilizes cadaveric specimens (which are stiffer) and lacks preoperative deformation, as the CT is acquired intraoperatively.
- SpineAlign: This dataset involves the registration of only a few vertebral levels, and the pose ground truth is limited to three collinear surgeon-placed landmarks. We explicitly acknowledge in the manuscript that future work must be directed toward compensating for both preoperative-to-intraoperative and intraoperative deformations. Furthermore, we emphasize that this limitation highlights a critical need within the community for new in-vivo datasets with highly accurate registration ground truths. 2.Computational Efficiency and Runtime (Reviewer #1, #2) We will add the information on the inference time of our pipeline to demonstrate its viability for real-time intraoperative surgical navigation. 3.Clinical Accuracy Thresholds and Metrics (Reviewer #1)
- Accuracy Claims: We agree that sub-millimeter accuracy may be overly strict and challenging to definitively prove given inherent measurement errors. We will soften our phrasing and clarify the required accuracy thresholds for intraoperative navigation in spine surgery.
- TRE vs. RMSE: Target Registration Error (TRE) is defined by surgeon-placed landmarks, whereas Root Mean Square Error (RMSE) is computed between the points of the point clouds. We chose to report these specific metrics on their respective datasets to allow for direct, fair comparisons with the established baselines set by Liebmann et al. and Daly et al.. 4.Training Setup and Generalization (Reviewer #2) To clarify, no cross-dataset training was performed. Training and testing were conducted independently for each single dataset. We strictly followed the validation splits of prior work to ensure fair benchmarking: we used an 8-fold cross-validation on SpineDepth (matching Liebmann et al.) and the single specified split on SpineAlign (matching Daly et al.). 5.Ground-Truth Bias (Reviewer #3) We agree that our model’s performance is inherently bounded by the quality of the underlying ground-truth registration. If the ground-truth alignment contains systematic biases, it will inevitably reflect on the registration results. This further supports the urgent need to acquire in-vivo registration datasets with accurate ground truths, which we highlight as a crucial direction for future work. We thank you again for your time and insightful comments, which will undoubtedly strengthen the final version of our manuscript.
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.
The paper describes a method to register intra-op RGB+D frames of spine with a pre-op 3D model from CT/MRI. While in previous methods the segmentation for discriminating bone from soft tissue in RGB+D and the 3D registration are carried in cascaded after independent training, in here a single model is used where the two operations of segmentation and registration are simultaneously optimized in an end-to-end fashion. The reviewers unanimously acknowledge novelty in the solution and recommend acceptance. I support the decision and encourage the authors to take into account the received feedback and add information about the computational efficiency of the approach.