List of Papers Browse by Subject Areas Author List
Abstract
Stereo-electroencephalography (SEEG) enables minimally invasive recording from deep and sulcal brain regions to localize epileptogenic networks for epilepsy surgery. However, traditional multiple SEEG electrodes planning remains largely manual, requiring experts to iteratively pick targets and search safe implantation paths while avoiding vessels, sulci, and other critical structures. To tackle this issue, we propose a coverage-aware planning method that integrates joint target localization and cooperative multi-electrode trajectory optimization into a unified reinforcement learning framework. Specifically, we introduce a capsule-based coverage model that represents functional recording sensitivity as a continuous region around each trajectory and maximizes overall region-of-interest (ROI) coverage through one-shot decision making. To ensure safety, we incorporate multi-layer constraints and inter-electrode spacing into deterministic projection with fixed constraint checking, enabling constraint-satisfying planning without manual penalty tuning. To validate the proposed method, we conducted planning on 8 retrospective temporal lobe epilepsy cases (84 electrodes) with CT/MRI-derived anatomical segmentations, achieving an 82% improvement in mean ROI coverage and an 88% increase in minimum inter-electrode spacing compared with existing methods while maintaining an 88% collision-free rate, supporting a versatile and safe planning framework that can adapt to diverse epilepsy treatment requirements and facilitate clinical deployment. Code is available at https://github.com/Joe-Galaxy/safe-seeg-multielectrode-planning.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2621_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Joe-Galaxy/safe-seeg-multielectrode-planning
Link to the Dataset(s)
N/A
BibTex
@InProceedings{QuZiq_Towards_MICCAI2026,
author = { Qu, Ziqiao AND Shi, Yangyang AND Wang, Yuanqing AND Xiao, Linxia AND Si, Weixin},
title = { { Towards Versatile and Safe Multi-electrode Path Planning for SEEG Implantation } },
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
The paper’s aim is to provide a method that would allow to implant multiple SEEG electrodes in such a way as to satisfy: maximal coverage of the ROI, avoidance of the forbidden zones, length of the electrode not exceeding given maximum, and proper angle of the skull entry. Authors achieve this using one-shot reinforcement learning using dedicated policy.
- 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.
Paper is clearly written and is accompanied with proper illustration explaining the goal and means to achieve it. Goals for the RL are clearly defined and described method has the advantage that is aims to achieve best ROI coverage and proper/safe stereotactic approach at the same time. The results section is detailed and explains the results of the approach.
- 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 teste were made on the retrospective data and ROIs in the tests were located in the temporal lobe. The Omega_k symbol is used inconsistently: in 3.1 it is a ROI’s mesh surface while in equation 1 it denotes ROI’s volume. As the paper has nine out of ten pages the feasibility projection could have been described in greater detail. In the equation 2 the minimal distance between trajectory and cortex/skull is not clear, the electrode has to pass through the cortex and its wiring must exist skull for recording purposes. Similarly the angle between trajectory and n_skull is not clear. It is not defined what n_skull is and I would expect the angle between trajectory and skull surface to be large, not small. Throughout the paper authors use abbreviation SDF without defining it. while it can be understood from the context it is still not defined. The same is with the R function in Fig. 1, can be understood from context but is undefined.
- 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
The paper would benefit from more detailed of the RL process. he policy and feasibility projection could be described in more detail. The Fig 1d while demonstrates the multiple electrode implantation is not a practical example - literally nobody would implant fusiform gyrus, hippocampsu, amylgada and insulta all at the same time.
- 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?
Well written paper the clearly states aims and methods for achieving them. There is no mentioned code source nor dataset but the description - while not detailed - allows one to attempt to reproduce the approach.
- 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.
paper shuould be accepted
Review #2
- Please describe the contribution of the paper
This paper proposes a reinforcement learning based framework for multi electrode SEEG implantation planning. The method jointly refines target locations and selects straight line trajectories using a one shot policy, followed by a deterministic projection enforcing anatomical and inter electrode constraints. A capsule based coverage model is used to encourage functional sampling. The approach is evaluated retrospectively on 8 cases. The method achieved an 82% improvement in mean ROI coverage and an 88% increase in minimum inter-electrode spacing compared with existing methods while maintaining 88% collision-free feasibility
- 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.Ambitious attempt to unify target refinement and trajectory planning. 2.Clear figures and well structured pipeline description. 3.Capsule based coverage model is intuitive and aligns with prior work on SEEG sensitivity (Dessert et al., 2023). 4.Deterministic projection for safety is conceptually reasonable.
- 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 reinforcement learning formulation is fundamentally flawed The paper does not define a Markov Decision Process (MDP): no state space, no transition dynamics, no horizon, no discount factor, no exploration strategy. The authors explicitly state that each electrode is a one shot decision, which contradicts the foundational definition of RL. This is not an RL problem; it is a regression problem. Using PPO in this context is unjustified and scientifically incorrect. 2.The deterministic projection performs all the actual planning The projection operator enforces vessel distance, cortex/skull distance, deep structure distance, entry angle, trajectory length and inter-electrode spacing. This means the RL policy does not learn safety, feasibility, or geometry. It only outputs a crude 5 D vector that is then corrected by projection. Thus, the “learning” component is superficial, and the projection is the real planner. 3.The coverage model is overly simplistic and not clinically validated Coverage is defined as the fraction of ROI points within a fixed radius of a line segment. This ignores: +electrode contact geometry +anisotropic sensitivity fields +tissue conductivity +ROI heterogeneity +segmentation uncertainty Dessert et al. (Brain Communications 2023) show that SEEG sensitivity is far more complex than a binary capsule threshold. The proposed metric is a coarse approximation and cannot support strong clinical claims. 4.Straight line trajectories are unrealistic The method assumes all trajectories are straight, which contradicts clinical practice. Real SEEG trajectories must respect: +skull thickness +burr hole placement +oblique entry angles +surgical workflow constraints Liakina et al. (Acta Neurochir 2025) demonstrated that straight line assumptions fail in many real cases. This severely limits the clinical validity of the proposed method. 5.Evaluation is insufficient and lacks statistical rigor For a clinical planning paper, Only 8 cases is far below the bar and this cannot demonstrate generalization. 6.Safety claims are overstated The paper repeatedly claims “safe planning,” but safety cannot be claimed based solely on SDF distances.
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
While the topic is important and the paper is well organized, the methodological foundation is too weak for acceptance. The use of RL is methodologically unjustified and contradicts standard RL definitions, the projection operator performs the actual planning, the evaluation is insufficient, and the clinical claims are overstated. Substantial rethinking of the core methodology is required before this work can be considered for publication.
- 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 paper addresses an important and clinically relevant problem, and I appreciate the authors’ efforts to clarify several aspects of the method. However, the remaining issues concern core scientific validity rather than presentation details. The reinforcement‑learning formulation remains technically unconvincing: the proposed MDP structure is introduced only in the rebuttal and still does not align with standard RL definitions. The projection operator continues to dominate the actual planning process, which raises questions about what the policy is truly learning. The coverage model remains a coarse geometric abstraction without clinical grounding, and the evaluation dataset is too small to support claims of generalizability or safety.
Although the rebuttal acknowledges that the dataset is a pilot study, eight patients is far below the threshold typically expected for evaluating a clinical planning system. The argument that 84 electrodes compensate for the small number of patients is not convincing, as electrodes within the same patient are not independent samples. For these reasons, the concerns raised in the initial review remain unresolved.
Review #3
- Please describe the contribution of the paper
This work presents a one-shot reinforcement learning framework for SEEG multi-electrode trajectory planning. The approach provides a unified treatment of target selection and trajectory optimization while incorporating a robust safety constraint mechanism to ensure clinical reliability. Experimental results on real-world datasets, alongside comparisons with state-of-the-art methods, validate the performance and efficacy of the proposed scheme.
- 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 experimental comparison is representative and robust, extensively covering both clinical gold standards and the most recent state-of-the-art (SOTA) methods from 2025.2.By directly addressing clinical requirements, the proposed method significantly reduces planning time compared to human experts while outperforming them across most metrics, demonstrating high practical persuasive power. 3.The integration of a deterministic projection layer for safety constraints ensures that the actions generated by the reinforcement learning policy network satisfy safety requirements without the need for additional parameter tuning.
- 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.Epilepsy focus detection using SEEG is based on models of neural signal propagation. While path planning methods using capsule-shaped coverage are well-suited for treatments like ablation, are they appropriate for the specific task of epileptic electrode detection? Is there a more accurate way to characterize the effective sensing range of instruments in electrophysiological signal detection? (In epilepsy, foci lack clear boundaries, and neural signals propagate via axons rather than through simple mean-field diffusion systems.) 2.The current dataset, comprising only 8 patients and 84 electrodes, remains statistically small. Consequently, the assessment of clinical usability remains questionable. 3.The manuscript does not specify the scale of the training dataset. While it mentions a test set of 8 cases, it is unclear whether these data were also included in the training phase. Given a potentially small training set, data augmentation strategies or simulation-based pre-training might be required; the lack of disclosure regarding these details raises concerns about the study’s reproducibility. 4.The authors should provide ablation experiments for the proposed key modules, such as the capsule coverage model and the deterministic projection layer. Quantifying the specific contribution of each module is essential to substantiate the novelty and effectiveness of the proposed solution. 5.Further interpretation of the metrics is required. Currently, the ROI metrics for all methods are below 10%, and the collision rate for manual physician planning is the highest. What do these results imply in actual clinical practice? A lack of explanation remains regarding whether these values are within acceptable clinical thresholds. 6.How was the radius r of the capsule model determined? The authors should provide a theoretical basis or a sensitivity analysis for the selection of this parameter. 7.Clinicians often have specific priorities based on the surgical context (e.g., sacrificing a degree of coverage to ensure greater distance from major blood vessels). How does the current RL reward function reflect the flexibility of these weights? 8.The introduction highlights the flaws of existing methods when planning for multiple targets simultaneously. However, that same section lacks a corresponding description of how the proposed method specifically handles multi-target path planning.
- 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.
(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?
This work provides a comprehensive solution to a practical clinical problem. Although the methodological innovation is somewhat limited—given the numerous existing examples of reinforcement learning in surgical planning—the manuscript thoroughly validates the feasibility of the results and confirms the effectiveness of the chosen approach. In terms of actual clinical application, I believe this level of rigorous validation is more valuable than merely proposing ‘fancy’ or complex methods.
- 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.
N/A
Author Feedback
We really appreciate the recognition from the reviewers, R1 for the ROI/safe-access objective, R2 for the ambitious attempt, clear pipeline, intuitive capsule model/reasonable projection, and R3 for clinical value. Specifically, (R2W1) We apologize for the unclear clinical-to-MDP mapping. Our MDP is patient-level K-step electrode planning: M=(S,A,T,R,gamma), horizon K, gamma=1.At step k, a target-direction pair is selected for a rigid electrode; accepted trajectories become spacing/collision constraints for later steps. Clinically, a rigid SEEG electrode comprises an ROI target/contact segment, a skull entry/burr-hole, and a straight shaft requiring clearance/spacing. Thus, each electrode’s choices are ROI sampling site and entry direction, encoded as a 3D target offset and 2D insertion direction. “One-shot” means this proposal determines one rigid electrode, not the patient plan. State=anatomy, ROI/corridor, target-ref geometry, structure distances, accepted trajectories; action = 5D target-direction proposal; transition=path projection ray-casts shaft/entry, checking validates distance/length/angle/corridor/spacing, then appends trajectory; reward= checked ROI coverage/path preference. This coupling differs from independent regression samples. PPO is used since no ground-truth target/direction labels exist. We will add this before Sec. 3.2.(R2W2&R3W8) The RL policy is the main decision-making component, selecting target-direction pairs for ROI sampling and multi-electrode coordination. Safety/feasibility/geometry preferences are learned from distance features, prior trajectories, and checked rewards; fixed checking prevents unsafe acceptance rather than replacing learning. Projection is not an independent shortest-path planner or post-processing optimizer. Given the policy proposal, the shaft is extended to the skull entry by path projection; checking performs accept/reject decisions under fixed constraints. The ROI subregion, 5D search, capsule optimization, and conversion of poor proposals into high-coverage ones are not handled by projection. We will separate these steps in Sec. 3.(R1&R2W3,4,6&R3W1,5,6,7) According to suggestions from our cooperating surgeons, the requested output is a straight SEEG trajectory from skull entry to intracranial target; burr-hole/workflow preferences are geometric constraints. Since a SEEG electrode records along its contact-bearing shaft rather than at a point, capsule coverage abstracts effective sampling range. Radius r follows electrode specifications/clinician-defined range; this abstraction does not replace patient-specific biophysical models (Dessert et al.). SDFs come from segmented anatomy and convert vessel/deep-structure/skull/cortex boundaries into distance queries; larger geometric margins are then preferred during optimization. (R1&R2W5&R3W2,3,4) We also think our dataset with 8 epilepsy patients is just a pilot study, despite the complexity of intricate vascular/skull reconstruction and faithful patient-specific environment modeling. Each electrode first has target-environment interaction, then relates to the overall plan as accepted trajectories constrain later electrodes through spacing/collision rules, making multi-electrode planning more challenging than single-electrode implantation. Thus, 84 electrode-implantation instances across brain maps are meaningful for this RL planner, without requiring a very large labeled dataset. We will clarify the patient-level 4/4 split so electrodes from the same patient are not treated as independent training/testing subjects. We are collecting cases for a 20-case journal version. (R1&R3W3) We define SDF, R in Fig. 1, Omega_k as ROI mesh points, n_skull as skull normal, and cortex/skull distance as forbidden non-entry intersections. Code will be released upon acceptance.
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.
Reviewer 2 raised serious methodological concerns. These should be addressed by the authors.
The other reviewers both felt this was a strong paper for MICCAI
- 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.
On balance two reviewers believe this is an important contribution to the SEEG trajectory planning space. One reviewer has remaining concerns over the RL formulation and writing describing how RL is applied to this problem.
The other open concern of a sample size is small remains, but similarly samples sizes for MICCAI publications on this topic have been accepted in the past. The nature of this dataset makes large cohorts difficult to obtain.
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.
The number of patients in the study, while small, is reasonable for a CAI paper. While the methodology could be improved (see reservations from R#3), I believe that the MICCAI audience would benefit from seeing this approach.
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.
Two reviewers showed enthusiasm for this work being presented at the conference. Several concerns remain from one reviewer, but the overall contributions are strong enough to justify acceptance.