List of Papers Browse by Subject Areas Author List
Abstract
Detecting individual immune cells in time-lapse magnetic resonance imaging (MRI) enables the non-invasive study of inflammatory processes. However, automated detection of contrast-enhanced monocytes remains highly challenging due to their extremely small spatial footprint (2–6 pixels), heterogeneous appearance, temporal variability, and ambiguity with confounding phenomena such as micro-bleeding. In this work, we introduce the first publicly available time-lapse MRI dataset for contrastive monocyte detection. The dataset contains carefully annotated sequences capturing substantial appearance variability and realistic clinical artifacts. We establish baseline performance using state-of-the-art object detection architectures and demonstrate that standard detectors perform suboptimally in the sub-resolution regime, especially under domain shift. Furthermore, we systematically analyze the contribution of key components of the YOLO architecture by progressively pruning its detection heads, feature pyramid levels, and loss terms. Our results reveal which architectural elements are essential for reliable detection of ultra-small objects, with a focus on domain generalization. We believe this dataset and analysis provide a benchmark for future research in microscopic structure detection in medical imaging.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2334_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://zivgitlab.uni-muenster.de/ag-pria/sub-resolution-experiments
Link to the Dataset(s)
Monocyte MRI Dataset: https://www.uni-muenster.de/PRIA/en/forschung/index.shtml
BibTex
@InProceedings{RexRob_SubResolution_MICCAI2026,
author = { Rexeisen, Robin AND Havlas, Asli AND Wilken, Enrica AND Wachsmuth, Lydia AND Helfen, Anne AND Masthoff, Max AND Faber, Cornelius AND Jiang, Xiaoyi},
title = { { Sub-Resolution Monocyte Detection in Time-Lapse MRI: A Benchmark } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16895},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper introduces the first publicly available 4D time-lapse MRI dataset for sub-resolution monocyte detection, comprising 124 recordings and over 115,000 annotations. The authors establish a standardized benchmark protocol, including fixed data splits, a distance-based AUC evaluation metric, and baseline results using multiple YOLO variants and Faster-RCNN. On this basis, the authors conduct systematic ablation experiments on the detection heads, feature pyramid levels, and loss functions of the YOLO architecture, analyzing which components are essential for detecting 2–6 pixel objects. The dataset is of significant value to the field, with a clearly defined task and practical application background (inflammation and cancer research). However, this paper has notable shortcomings in annotation quality validation, domain shift analysis, temporal information utilization, and experimental design.
- 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) Clear dataset contribution: The first publicly available MRI dataset for sub-resolution immune cell detection, containing realistic clinical artifacts and diverse object appearances. 2) Clear problem definition: Reasonably distinguishes sub-resolution detection (SRD) from conventional small object detection and cell tracking tasks. 3) Comprehensive baseline experiments: Compares multiple YOLO variants and Faster-RCNN, evaluating domain shift on both k-fold cross-validation and an independent test set. 4) Insightful ablation study: Prunes detection heads, backbone depth, pooling operations, and loss terms, providing practical guidance for future method design.
- 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) Rows 2 and 3 in Table 1 have identical data. Is this because the experimental results are truly consistent, or is there a data duplication error? 2) The authors propose using a 4-pixel matching threshold for computing AUC, claiming that this threshold is based on “AUC stability and annotation uncertainty. ” However, no data or figures are provided in the paper to support this claim. 3) The specific sources of “domain shift” need to be clarified: individual mouse differences, scanning parameters, magnetic field inhomogeneity, artifact types, etc. It is recommended to add a simple comparison of domain adaptation methods. 4) The ground-truth bounding boxes use a fixed 10×10 size, which is much larger than the actual 2–6 pixel monocytes. This may cause the network to learn background patterns rather than the target signal, and indirectly leads to redundancy of coarse detection heads and degraded generalization of the CIoU loss. It is recommended to use tighter bounding boxes and supplement with a receptive field analysis.
- 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
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?
This paper addresses a clinically important and technically challenging task with a novel, valuable dataset and a complete benchmark. The overall work is solid and meaningful to the community. However, several issues in experimental design, result presentation, and validation need to be carefully revised. The paper is slightly above the acceptance threshold and can be accepted after proper revision.
- 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 introduces the first publicly available 4D time-lapse MRI dataset for sub-resolution monocyte detection, along with a standardized benchmark protocol. The dataset includes 124 recordings and over 115k annotations capturing immune cells at extreme small scales (2–6 pixels). The authors establish baseline performance using modern object detectors (YOLO variants, Faster R-CNN) and propose a distance-based evaluation metric tailored to sub-resolution detection. Additionally, they perform a systematic ablation study on YOLO architecture components to identify key factors influencing detection performance and domain generalization.
- 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.Important dataset contribution First publicly available dataset for this problem, addressing a clear gap in reproducibility and benchmarking. 2.Well-motivated problem Sub-resolution detection (2–6 pixels) is a genuinely hard and underexplored regime in medical imaging. 3.Comprehensive dataset analysis Strong statistical characterization (e.g., distribution of annotations and trajectories on page 4), which adds credibility and insight. 4.Appropriate evaluation redesign Replacing IoU with a distance-based matching metric is well-justified for this regime. 5.Systematic benchmarking Evaluation across multiple detectors and controlled experiments (Table 1, page 6) provides a solid baseline. Insightful ablation study The architectural analysis (Table 2, page 7) offers useful insights into what matters for ultra-small object detection. 6.Practical relevance Potential impact on non-invasive immune monitoring and inflammation research.
- 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.Limited methodological novelty The paper mainly evaluates existing architectures rather than proposing a new detection method. 2.Baseline-focused rather than solution-oriented While the benchmark is valuable, there is no strong new algorithm addressing the problem. 3.Generalization discussion is limited Domain shift is observed (Table 1), but deeper analysis or mitigation strategies are lacking. 4.No comparison to specialized small-object methods Missing comparisons to methods from IR small target detection or point detection literature. 5.Annotation quality concerns The authors acknowledge possible labeling noise but do not quantify its impact. 6.Clinical translation is indirect The work is preclinical (mouse data) and lacks discussion on human applicability. 7.No temporal modeling baseline Despite time-lapse data, models are mostly slice-wise; temporal models (e.g., RNN/transformers) are not explored.
- 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
1.Consider adding temporal modeling baselines, since the dataset is inherently time lapse. 2.Include comparison with specialized small-object detection methods beyond YOLO variants. 3.Provide quantitative analysis of annotation noise impact. 4.Expand discussion on domain adaptation strategies, given strong domain shift. 5.Clarify how this approach may translate to human MRI data.
- 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?
This paper provides a valuable dataset and benchmark, which is highly relevant for the community and fills a clear gap. The experimental analysis is thorough and offers useful insights into sub resolution detection challenges.
However, the work is primarily benchmark oriented with limited methodological novelty, and the absence of new algorithms or deeper modeling innovations reduces its impact compared to stronger MICCAI contributions.
Overall, the dataset contribution and careful evaluation justify acceptance, but the paper sits near the threshold due to limited novelty.
- 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
Detecting cells on MRI is challenging as MRI does not provide cellular-level resolution. The paper presents a “4D time-lapse MRI dataset for sub-resolution for monocyte detection” and evaluates Faster RCNN and YOLO variants. The paper further evaluates YOLO v8n with different loss and architectural changes (e.g. backbone down-steps, neck up-sampling, etc).
- 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 task is interesting, given the extreme scale of the target objects. The paper notes that the standard small object detection typically involves 32x32 pixel targets, whereas the monocytes on MRI span 2-6 pixels. The connection to the Infrared Small Target Detection and Segmentation and the reference to successful YOLOv8-P2 variants set up the paper well. Public datasets in medical imaging are always valuable.
- 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.
Ferucarbotran is mentioned as the contrast agent but the annotation process could be more clearly described. (e.g. It is unclear who performed the 115,570 annotations or whether any inter-observer variability analysis was done.) If the paper followed the protocol from reference [15] Maška et al. it would be helpful to state this explicitly, and clarify annotator expertise (e.g. radiologists/trained annotators). Considering the 4-pixel tolerance (assuming in all directions) is relatively lenient given the image size and the target size (2-6 pixels), inter-operator analysis could help justify this choice. It would be helpful to see other metrics (e.g. percentage of correctly predicted cells) aside from AUC.
- 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
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents a very difficult task and a valuable dataset. The paper highlights some valuable insights on YOLOv8n’s performance on their task (e.g. default setup vs different loss, upsampling layer, etc) and their dataset (e.g. ringing artifacts). The annotation process could be more detailed and the evaluation is mainly limited to AUC. Overall, the paper is promising but somewhat limited in terms of methodological depth and evaluation.
- Reviewer confidence
Somewhat confident (2)
- [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
We thank all the reviewers for their constructive feedback. As we find no major misunderstandings or inaccuracies in the reviews, we focus on addressing some concerns and questions.
1.Annotation Methodology and Threshold Justification (Reviewers 1 and 3):
The reviewers questioned the 4-pixel matching threshold used for computing the AUC and the fixed 10×10 bounding box size. We empirically determined the threshold by evaluating multiple trained networks under varying distance tolerances. Improvements beyond a radius of 4 were negligible, indicating AUC stability. Regarding the bounding box size, we tested smaller boxes and observed significantly worse performance. We hypothesize that background patterns are essential for distinguishing monocytes from similar phenomena, particularly in uncharacteristic regions. The acquisition and detection protocol follows the established procedures from [15] Masthoff et al. All annotations for this study were performed by a single experienced biologist. Given this setup and the considerable size of the dataset, a full interobserver analysis was infeasible. We therefore prioritized rigorous quality assurance through protocol standardization and expert-led consistency checks. The manuscript has been revised to clarify these points.
2.Domain Shift and Generalization (Reviewers 1 & 2):
The reviewers noted the limited discussion of sources of domain shift and the lack of domain adaptation methods. We clarify that the only source of variation in our dataset is natural biological variability among individual mice (e.g., healthy, tumor-bearing, or treated animals), which affects cell counts but not monocyte appearance or MRI characteristics. All scanning parameters and the measurement protocol were kept constant across acquisitions. While magnetic field inhomogeneities and artifacts cannot be fully excluded, they are comparable across all data. The observed domain shift is therefore attributable solely to expected biological variation. Due to page constraints, we could not include a detailed analysis or comparison of domain adaptation methods. We plan to address these techniques, including normalization methods and data augmentation, in a journal extension.
3.Comparison to Specialized Small-Object Methods (Reviewer 2):
The reviewer requested comparisons to the literature on IR small target detection. This paper is intended as a baseline to compare current state-of-the-art generic detection networks and establish a foundation for future work. Specialized small-object methods are a natural next step.
4.Clinical Translation (Reviewer 2):
We acknowledge the preclinical nature of this work. Time-lapse MRI is a versatile tool in preclinical research and can offer great insights into immune cell dynamics upon varying inflammatory stimuli. However, to the best of our knowledge, it is not yet feasible to track single cells in humans. Achieving clinical translation would require further development of MRI hardware to achieve sufficient spatial resolution, adjusting contrast agent dosing, and validating the pipeline in larger animal or human cohorts, alongside standard regulatory considerations.
Nevertheless, we anticipate that methods developed on this mouse dataset will serve as a foundation for monocyte detection in human MRI, as the appearance of single cells in humans will likely be similar, thereby enabling scalable applications in human immune system research.
5.Temporal Modeling (Reviewer 2):
Temporal models (RNN/transformers) are beyond the scope of this paper, which focuses on dataset release and baseline performance. We agree that temporal methods are promising and plan to explore them in follow-up work.
6.Additional Metrics (Reviewer 3):
Beyond the area under the curve (AUC), we intend to report additional metrics in a journal extension, as the page limit of the current manuscript does not permit their inclusion.
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 presents a benchmark dataset for sub-resolution monocyte detection in time-lapse MRI, along with a standardized evaluation protocol and comprehensive baseline experiments. The reviewers’ scores are 4, 4, and 4, indicating a clear consensus toward acceptance. The reviewers agree that the dataset is valuable and the experimental analysis provides useful insights for this challenging task, although the methodological novelty is limited. The AC therefore recommends early accept.
For the camera-ready version, the AC suggests addressing minor clarity issues noted by the reviewers, such as providing more detailed descriptions of the annotation process, clarifying the evaluation protocol and metrics, and improving the readability and presentation of figures and results. Please refer to the reviewers’ comments for more detailed suggestions.