List of Papers Browse by Subject Areas Author List
Abstract
Precise object localization in three-dimensional (3D) medical imaging is critical for clinical diagnosis, yet voxel-level segmentation remains labor-intensive and sensitive to boundary ambiguity. While object detection offers a more efficient alternative, current methods heavily rely on predefined anchors, which require extensive hyperparameter tuning and exponentially increase computational complexity within 3D search spaces. In this paper, we propose MedCenterDet, a novel anchor-free, center-based object detection framework tailored for 3D medical images. Our approach directly predicts object centers via heatmap regression and analytically decodes bounding box dimensions using an anisotropic Gaussian target generation strategy, inherently capturing morphological shape priors. To enhance feature discriminative power in dense 3D volumes, we introduce a counterfactual attention learning (CAL) module. By contrasting factual target-structure regions with counterfactual noise interventions, this module explicitly optimizes attention quality, forcing the network to focus on clinically relevant patterns. A subsequent two-stage refinement module leverages multi-scale features to correct spatial misalignments and ensure geometric precision. We comprehensively evaluated MedCenterDet on two public datasets. Experimental results demonstrate that our framework achieves state-of-the-art performance across both benchmarks, while maintaining a significantly lower parameter count compared to established methods. The code is available at https://github.com/WiDayn/MedCenterDet.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0759_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/WiDayn/MedCenterDet
Link to the Dataset(s)
BraTS2021: https://zenodo.org/records/19541844
TotalSegmentator: https://github.com/wasserth/TotalSegmentator
BibTex
@InProceedings{ZenQia_MedCenterDet_MICCAI2026,
author = { Zeng, Qiang AND Shen, Ao AND Du, Xiangtong AND Xu, Xiaoyu AND Zhi, Shaohua AND Xue, Wufeng AND Pan, Fei},
title = { { MedCenterDet: A Center-Based Object Detection Framework for 3D Medical Image } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16885},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes MedCenterDet, a novel anchor-free, center-based 3D medical image object detection framework. The methodology combines anisotropic Gaussian target generation for analytical bounding box decoding with a counterfactual attention learning module to explicitly optimize attention quality and suppress background noise. Experimental results show that the proposed method outperforms other state-of-the-art baselines while maintaining a significantly lower parameter count.
- 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) Novel and complete methodology framework: The proposed anchor-free, center-based framework (MedCenterDet) offers an innovative and complete solution for 3D medical image detection. Using an anisotropic Gaussian target generation strategy to replace traditional predefined anchors effectively reduces the massive computational burden typically seen in 3D search spaces.
2) Original application of causal inference: The Counterfactual Attention Learning module is an original design. By introducing random noise to create “counterfactual” attention maps, it forces the model to explicitly distinguish between true lesion features and background noise.
3) Comprehensive experimental design: The authors conducted solid evaluation across two well-known public datasets (BRaTS 2021 and Total Segmentator) and compared their method against recent baselines. Furthermore, the evaluation metrics are standard, and the ablation study convincingly demonstrates the effectiveness of the proposed modules.
- 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）Lack of experimental details on data splitting: Although the paper mentions using a “five-fold cross-validation strategy,” it lacks crucial details on how the data was partitioned. In medical imaging, it is critical to perform splits at the patient level rather than the scan or slice level to avoid data leakage. The authors must explicitly state their splitting strategy.
2) Abnormal baseline results and lack of significance tests: In Table 1 and Table 2, the MedYOLO-L model shows 0.00±0.00 across the board on the BraTS 2021 dataset. While the authors explain that the model “failed to converge,” simply presenting a completely failed baseline without attempting to tune its hyperparameters to achieve a reasonable performance undermines the fairness and credibility of the comparison. In addition, there are no statistical significance tests (e. g. , p-values) provided to confirm that the performance over the baselines are statistically meaningful rather than due to random variance.
3) Insufficient discussion on clinical significance and applicability: While the paper clearly demonstrates computational efficiency of the proposed method, it lacks a thorough discussion of its clinical feasibility. Specifically, the authors should explicitly discuss the specific clinical scenarios where their bounding box detection is highly valuable to connect the proposed algorithm with real-world clinical decision-making.
4) Reproducibility: The authors provided a GitHub link placeholder (github. com/**/) in the abstract instead of an anonymized link. So reproducibility cannot be verified.
5) Formatting and presentation issues: The visual layout of the paper needs improvement. The spacing between figures and text is not compact enough, which disrupts the reading flow. Additionally, the ‘Comparison with Existing Methods and Ablation Study‘ section structures could be better organized to improve overall readability.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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?
Although there are some issues in the manuscript regarding experimental details and formatting, the overall quality of the paper is good. The proposed method is novel, particularly in its original application of causal inference to 3D medical object detection. If the authors can sufficiently address the concerns raised in the review then I would be inclined to raising my rating to accept.
- 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 authors have sufficiently addressed my comments, and given that other reviewer comments were largely positive, I am happy to recommend acceptance as well
Review #2
- Please describe the contribution of the paper
The authors propose an anchor-free object detection method that employs an attention based on counterfactual learning, a strategy that considers anisotropic volumes and a geometric refinement for 3D medical images from BRATS and Total segmentator datasets. The article must be enhanced to better present the contributions and novelty of the proposed method against existing literature. The description of the methodology must include an overview that better describes the workflow of the proposed framework.
- 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 proposed framework utilizes an object detection method that specifically accounts for the anisotropic nature of medical images, addressing a common and critical property of this data domain. -The use of an anchor-free detection strategy effectively reduces computational complexity and eliminates the need for predefined anchor boxes, streamlining the detection process.
- 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 Introduction with the related works sections is insufficiently developed; the current review relies on older citations and lacks a comprehensive survey of recent advancements (2024–2025). A more thorough exploration of the state-of-the-art is necessary to properly contextualize the study and justify its claimed novelty.
-The methodological overview is somewhat ambiguous concerning the sequential or parallel execution of the proposed steps. The authors should include a clear workflow paragraph to better delineate the procedural stages of the study.
-The definition of relative target scales (Lower, Middle, and Upper tertiles) requires further clarification to ensure the reproducibility of the results. -Details regarding the parameters of the compared baseline methods are missing, which significantly hinders the transparency and reproducibility of the experimental benchmarking.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
The metrics in the final row of Table 1 for the TotalSegmentator results are merged and require proper alignment to ensure the data is clearly interpretable.
- 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 an anchor-free object detection method that employs a counterfactual-based attention combined with an anisotropic-aware strategy for 3D medical images. However, the manuscript is currently weakened by an underdeveloped literature review and a confusing methodological overview that lacks a clear procedural workflow. Additionally, the omission of some parameters hinders the reproducibility of the results. Given these concerns regarding clarity and experimental details, a Weak Accept is recommended to allow the authors to better contextualize their novelty and provide the necessary technical depth.
- 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 authors have adequately addressed my previous concerns, and I recommend the manuscript for Accept. Overall, the authors agreed to expand the literature review and clarify the methodological workflow-experiments details, in order to enhance results’ presentation and to highlight the paper’s key contributions: a computationally efficient anchor-free detection strategy for 3D medical images and a tailored strategy for anisotropic volumes.
Review #3
- Please describe the contribution of the paper
The paper proposes an object detection technique that derives the object centers from heatmap regression.
The main difference from traditional heatmap-landmark detection techniques is that anisotropic gaussians are used to model the target for the heatmaps (the sigmas are based on object size). So instead of using anchors, the dimensions of the detected objects’ bounding boxes are obtained from fitted Gaussians.
Contrafactuals are used to contrast the factual region vs noise interventions and are shown to significantly enhance the detection performance.
- 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.
Significant improvement in 3D object detection performance compared to other reported techniques for only a relatively small small additional compute.
It is interesting to see that this allows the use of feature backbones from even relatively small models (EffNet-L) to yield significantly better performance than that of competing methods. This increases the surface for applying these techniques to a broader part of our field where inference speed is important.
Presented technique are introducing interesting concepts that can be transferred this to other heatmap-regression-based approaches across the field.
Solid validation on BraTS and Total Segmentator datasets
- 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 use of the do(.) operator is a central component but lacks a formal definition within the context of the proposed architecture. Providing a brief theoretical background or a more explicit mapping of Pearl’s Causal Inference framework to your attention mechanism would make the paper more self-sufficient and accessible.
Does the isotropic Gaussian prior inherently bias the model against non-ellipsoidal structures? Could you comment on how this would apply for more curvy, elongated, or “spiky” structures? Please provide a performance breakdown for the already presented the organs, if possible.
The relationship between $L_{heatmap}$ and $L_{cal}$ is not particularly clear to me. For improved clarity, in equations 4 and 5, I suggest improving the definition of r, r_d and e_d. Also L_{heatmap} is not defined as an equation - is the difference between it and L_{cal} only that one is using the raw predicted heatmap and the other P_{heatmap}^{effect}, or is in that L_{cal} is applied per module? Where is CLS prediction from Fig. 1 entering into the loss?
It is unclear, which layers from the backbone networks are used for the feature extraction.
Dimensions of the Gaussians in the Pred Heatmap in Fig. 2 do not seem to correspond to the neighbouring images. These visual inconsistencies raise concerns about whether these are actual model outputs or illustrative approximations. Please clarify if these results are representative of the model’s raw predictions.
As for the future, have you already considered directly regressing also the sigmas directly and/or tried oriented /non-anisotropic Gaussians, or even multiple Gaussians as targets, or imposing structures between the centers of multiple organs (e.g., https://arxiv.org/pdf/2109.12848, https://arxiv.org/pdf/2510.16445, https://arxiv.org/pdf/1910.06278)?
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
“The CAL module guides the network to focus on factual lesion” -> as this is not only about lesions, but also organs, I suggest to replace “lesion” with “structure” or something similar
The reported MedYOLO-S on BraTS is very poor, is there a reason why?
- 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?
It is a solid paper with a clear performance improvement to the MICCAI community. The proposed methodological improvements have the potential to be adapted for other applications.
- 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 authors sufficiently covered the weaknesses in their rebuttal.
Author Feedback
We sincerely thank all reviewers and meta-reviews for their constructive feedback. We also acknowledge that some aspects of the menuscript were not as clear as they should have been. In below, we will provide explanations for the points that must be clarified immediately, and will revise the the manuscript based on all the suggestions from reviewers. R1: Data Split. We confirmed that all five-fold cross-validation experiments were split at the patient/subject level and the same partition was used for all methods, no data leakage. For BraTS 2021, we add that we used the official training cohort of 1,251 patient level FLAIR modality cases; For TotalSegmentator, each CT volume corresponds to a unique subject-level case. Clinical significance and applicability. The ROI boxes and center points produced by MedCenterDet can support downstream clinical analysis. For example, in detection-to-segmentation/classification cascades, accurate localization provides compact ROIs, reduces irrelevant background search space, and can improve downstream efficiency and accuracy. Center points can also serve as stable spatial descriptors for longitudinal follow-up, e.g., tumor displacement analysis. We will expand this discussion in the final version. MedYOLO Failure. We used the official MedYOLO implementation and recommended configuration. After observing degenerate predictions, we further tried to stabilize MedYOLO-L by adjusting the learning rate, batch size, training schedule, but it still failed to converge reliably on BraTS. Since this behavior is consistent with the convergence issue discussed in the original MedYOLO paper, we cited their explanation and reported the result transparently. We consider that our main conclusion does not depend on this failed baseline, as MedCenterDet consistently outperforms the normally converged baselines. Statistical Significance. Given the limited statistical power of 5 folds, we conducted subject-level bootstrap confidence analysis. The paired differences consistently favored MedCenterDet, and the 95% confidence intervals did not cross zero for the primary metrics. R2: Related works. We agree that the related-work discussion should be more comprehensive. The current version mainly focuses on representative 3D medical object detection baselines and the anchor-based/anchor-free distinction. We will expand the discussion, and we consider that the new discussion will help provide a better understanding rather than altering our novelty. Sequential/Parallel Execution. The framework is sequential at the stage level. The input volume is first processed by the backbone to extract multi-scale features, which are then fed into Stage 1 for feature fusion and center heatmap prediction. The heatmap decoder converts the predicted heatmap into K×C coarse proposals. Within Stage 2, the K×C proposals are processed in parallel, since each proposal can independently sample multi-scale features and pass through the refinement heads. Thus, the pipeline is sequential across stages but parallel across proposals in Stage 2.Relative target scales. Relative scales were defined by ground-truth box volume quantiles. For each dataset, we computed V=w×h×d, sorted all instances, and divided them into three equal groups using the 33.3% and 66.7% quantiles. For each group, AP50 and AR were computed only against ground-truth boxes in that group under the standard COCO-style protocol; thus, Table 2 reflects the targets by ground-truth volume, not predicted size. Baseline fairness. For each baseline, we used the official or publicly available implementation whenever possible and followed the recommended/default configurations from the original papers or repositories. This is a common practice in benchmarking (ASA, MICCAI 2024 supplementary; Swin-UMamba, MICCAI 2024). R3: We clarify that Fig. 2 shows original predictions, but the perspective was changed to show decoded RoIs in the predicted heatmap, which may have caused visual differences.
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 provide mixed evaluations of this paper (scores: 3, 4, 5), with overall positive feedback on the methodological novelty and empirical performance, but also several concerns regarding experimental rigor, clarity, and reproducibility . In particular, reviewers point out missing details on data splitting, unclear baseline fairness, lack of statistical significance testing, and insufficient methodological clarification. Given these differences and the identified concerns, I invite the authors to address these points in the rebuttal—especially by clarifying experimental protocols, strengthening the presentation of the method, and providing additional analysis where possible. A clear and thorough response will be important for the final decision.
- 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.
After carefully reviewing the rebuttal and the post-rebuttal reviewer comments, my recommendation is to accept the paper. The paper proposes a computationally efficient anchor-free, center-based framework for 3D medical image object detection, with anisotropic Gaussian target generation and a counterfactual attention learning module. The reviewers generally agreed that the method is technically interesting, computationally efficient, and validated on relevant public datasets. Although the initial reviews raised concerns regarding data splitting, baseline implementation, methodological clarity, literature review, and presentation details, the authors provided satisfactory responses and committed to improving the final manuscript. Importantly, all submitted post-rebuttal opinions support acceptance. Therefore, I recommend acceptance, with the expectation that the authors will further clarify the methodological workflow, experimental settings, statistical analysis, and clinical applicability in the camera-ready version.
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.
All the reviewers (R1, R2, R3) agree on the same recommendation of Accept.
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.
After the rebuttal, all reviewers provided positive comments on this work. The main contribution is an anchor-free framework for 3D object detection, and the experiments demonstrate the effectiveness of this design. Although the reviewers raised several concerns, the authors addressed them point by point in the rebuttal. Therefore, the authors are encouraged to further refine the paper in the camera-ready version, and a final acceptance recommendation is given.