List of Papers Browse by Subject Areas Author List
Abstract
Diffusion MRI (dMRI) tractography enables non-invasive reconstruction of white-matter pathways, but its accuracy is fundamentally limited by indirect, low-resolution measurements of axonal organization. Tracer injection studies in non-human primates provide a gold standard for validating dMRI tractography. This, however, requires time-consuming manual annotation of fiber bundles in histology sections. We propose a synthetic-data augmented framework for automated fiber bundle segmentation in macaque tracer histology. Our approach uses ex vivo dMRI tractography as a generative prior to synthesize 2D image patches for training. This provides us with sufficiently realistic foreground texture, which we compose with backgrounds from blockface photos and diversify via domain randomization. A 2D U-Net is trained on mixed real and synthetic patches. Experiments on held-out brains demonstrate improved generalization across brains and fiber bundle densities compared to training with real data only. Training with synthetic data only leads to poor performance, underscoring the need for real supervision. Overall, our approach achieves performance comparable to the state-of-the-art while requiring 3x less manually annotated data.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4001_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/lincbrain/synthetic-tracer-histology-segmentation
Link to the Dataset(s)
N/A
BibTex
@InProceedings{BinKyr_TractographyDriven_MICCAI2026,
author = { Bintsi, Kyriaki-Margarita AND Makharia, Sparsh AND Balbastre, Yaël AND Romero Avila, Joselyn AND Lehman, Julia F. AND Haber, Suzanne N. AND Yendiki, Anastasia},
title = { { Tractography-Driven Synthetic Data Generation for Fiber Bundle Segmentation in Tracer Histology } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16881},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The article addresses segmentation of fiber bundles from tracer histology, which represents the gold standard for validating ambiguous information from lower resolution tracings from diffusion MRI. Since manual segmentation of fiber bundles in histology is time consuming, the authors propose a synthetic data generation approach, which aims to synthesize realistic tracer histology based on diffusion MRI as generative prior. The authors validate their approach on held-out brains, demonstrating that mixing synthetic data with real data leads to improved performance, while training on synthetic data alone achieves inferior 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.
The work addresses segmentation of tracer histology as a practically relevant problem. The use of synthetic data generation for improving segmentation performance of machine learning algorithms is reasonable and based on existing work from applying similar approaches in other domains. The synthesis pipeline is well motivated and explained, and the effect of using synthetic data alone and in combination with real data is appropriately covered in ablation studies. The method and parameters are well described, sufficiently detailed, and easy to follow. I particularly like that the paper presents multiple baselines and evaluates on hold-out brains, which is important to accurately assess the methods performance under realistic conditions. The paper is well written.
- 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.
I don’t see maor weaknesses with the paper, but I have several suggestions for improvements.
1.The yellow trace in Figure one (top, left yellow square under “Foreground + Ground truth mask Generation”) is very difficult to see. It would be good to enlarge the traces to improve visibility, with an appropriate description of the modification in the caption. 2.The notation in equation one seems to be inconsistent: The minimum runs over p, but p_y is used in the difference. Since p_y is not defined elsewhere, I assume p_y should be replaced with p (or vice versa). 3.p is used twice, once in equation 1, and once as probability (p = 0.3) for adding injection-size artifacts. I would replace one of them to avoid confusion. 4.The method involves many free parameters (e. g. , sampling intervals, opacity fading factor, artifact injection rate, . . . ). I assume these were selected empirically. It would be beneficial to briefly comment how these parameters were selected (e. g. , through visual inspection or based on prior knowledge). 5.The pipeline is applied on-the-fly during training, which is impressive. It would be interesting to provide some estimates of how long the synthesis takes for one patch (including used hardware). 6.In 2.2, the parameters for the connected component filtering should be specified. 7.The source of the dataset is unclear from the text. Is this an in-house dataset? If so, this should be briefly mentioned.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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?
The paper presents a solid new method, with clear practical relevance, clear presentation, and detailed evaluation. I recommend accept.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
I thank the authors for their feedback. Given the provided additional information, I would like to keep my initial recommendation and recommend accept.
Review #2
- Please describe the contribution of the paper
The primary contribution of this work is the introduction of a synthetic-data augmented framework that addresses the label-scarcity problem in macaque tracer histology. By elegantly repurposing dMRI tractography as a generative prior, the authors demonstrate that a U-Net can be trained to achieve SOTA-level segmentation performance with significantly reduced manual supervision, showing particular strength in generalizing to challenging sparse fiber bundles across different brain samples.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
-
The paper uses ex vivo high-resolution dMRI tractography as a geometric prior to synthesize histology training data. This allows the AI to learn from computer-generated data instead of relying only on slow, manual human labeling, which bypasses the “gold standard” bottleneck of manual annotation.
-
The authors purposefully use sparse fibers in training images but dense masks for the labels. This teaches the U-Net to correctly identify and “fill in” fiber bundles even when the real signal is very faint or hard to see.
-
The method reaches state-of-the-art results using 3x less manual data than previous methods. It works well even when the AI is trained on just one brain or a single annotated section.
-
- 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.
-
Insufficient technical details limit the reproducibility of the study. For example, while the authors mention using multi-shell constrained spherical deconvolution and a b_max of 40,000 s/mm^2 , they do not provide the specific number of shells, the b-values for each shell, or the number of gradient directions per shell. Moreover, the paper states that seeded probabilistic tractography was used , but it does not specify the software (e.g., MRtrix3, FSL), the specific algorithm (e.g., iFOD2), or the tracking parameters (such as curvature threshold, step size, or minimum FOD amplitude).
-
While the method improves detection sensitivity, it suffers from a higher rate of false positives compared to the state-of-the-art. This suggests that the synthetic data may not yet fully capture the complexity of background textures or injection-site artifacts found in real brains.
-
The approach struggles with certain domain shifts, such as different tracer characteristics or background appearances that the current synthesis pipeline does not replicate. This limits how well the model can generalize to entirely new types of tissue preparations or tracers.
-
- 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
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?
The major weakness (lack of details, high false positive rate, etc) dampen my enthusiasm to this manuscript
- 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
We thank the reviewers and meta-reviewer for the constructive feedback and positive assessment. We are pleased that the practical relevance and evaluation of our method were recognized.
We appreciate the reviewers’ suggestions to improve the readability of our paper. Most requested clarifications can be incorporated in the camera-ready version. As suggested by R1 and R2, we will: improve Fig. 1; clarify Eq. (1); add details including the dMRI protocol (three b-shells at 4k/8k/12k s/mm² with 64/64/128 directions), tractography details (MRtrix, MSMT-CSD, max bending angle=30°, minimum FOD amplitude=0.1), synthesis runtime, parameter selection, connected-component filtering; and dataset origin. The dataset was provided by one of the authors, was used in prior work, and is not yet public. We also clarify that the codebase will be publicly released upon acceptance.
The main concern raised by R2 and the meta-reviewer relates to the sensitivity–specificity trade-off. The proposed method achieves higher sensitivity, particularly for sparse bundles, at the cost of increased false positives (FPs).
First, the proposed method aims to compensate for limited real training data and therefore uses images from a single macaque. The most relevant comparison is thus the real-only baseline trained with data from one macaque. Relative to this baseline, the proposed mixed training substantially improves sparse-bundle sensitivity across held-out brains. In contrast, baselines #3/#4 additionally used data from three macaques and benefited from substantially more annotations.
Importantly, increased FPs do not necessarily imply a worse false discovery rate (FDR), which depends on the TP/FP balance. For example, on M3 FPavg increases while FDR decreases relative to the real-only baseline, indicating TP gains outweigh FP increases. This behavior suggests a more favorable sensitivity-specificity trade-off relative to the real-only baseline, rather than a systematic failure mode.
The meta-reviewer raises the concern that tractography priors may transfer tractography FPs into the segmentation model. We appreciate this important concern. However, the definition of an FP is fundamentally different in tractography studies (where the task is to detect whether a pair of distant gray-matter regions are connected) and in our segmentation problem (where the task is to detect whether a cluster of voxels in a thin histological section has been labeled by a fluorescent tracer). Importantly, our segmentation model does not “see” information about the long-range trajectory of tractography streamlines. We use the streamlines only to generate local image patches with foreground structure resembling a group of axons traveling through a 50μm-thick brain section. In this sense, the model learns local texture/structure rather than global connectivity from the synthetic patches. In addition, the best-performing model uses mixed training with real histology and manual annotations, which constrains learning toward biologically valid tracer appearance. Therefore, we believe the risk of simply transferring tractography-specific false positives into the segmentation model is limited. The availability of both dMRI and tracing in our data would allow us to further validate this point in future work, e.g., by generating a “clean” tractogram containing only axon bundles confirmed by tracer injection.
Finally, we agree with R2 that domain shift across tracer preparations and tissue appearance remains challenging. However, compared to the real-only baseline, the proposed method consistently improves cross-brain generalization, suggesting that the synthesis framework already improves robustness to unseen domains despite the current synthesis not fully capturing all appearance variations. Future work will investigate generative model approaches to produce more diverse tracer preparations and tissue appearances, with the goal of further improving generalization across domains.
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.
This manuscript explores the use of synthetic data, based on diffusion MRI tractography, to reduce the required annotation effort for tract segmentation in tracer histology. Reviewers had a number of suggestions that should be possible to account for in the final version of the paper, and ask to include additional information which also seems feasible. The main point that I would like to see answered in the rebuttal is the following: The fact that the proposed mixed training substantially increases the number of false positives makes its net benefit unclear. Authors attribute this to domain shifts and state that future more realistic synthesis should reduce this problem. However, I see the following potential alternative explanation, which would fundamentally question the proposed approach: It is a known limitation of diffusion MRI tractography that it creates many false positives. Therefore, the proposed use of a tractography-derived prior for training a segmentation network whose main purpose is validation of tractography carries the risk that the segmentation network starts to copy the false positives present in the tractography result, and therefore becomes useless as a “gold standard” for validation. How can we exclude this possibility?
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
After reading the rebuttal, I still have fundamental concerns regarding the safety and soundness of the proposed approach: Using tractography results to train a network that should in turn be used for validating tractography still appears circular to me. Even though I understand that tractography is used to generate local textures rather than global connectivity, the authors’ assertion that they “believe the risk … is limited” does not convince me that augmenting the training as proposed here will not introduce bias, and I still do not see a practical way to control that bias other than not using the proposed approach at all.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
This paper proposes a synthetic data generation framework that repurposes diffusion MRI tractography as a geometric prior to synthesize tracer histology training patches, reducing the manual annotation burden for fiber bundle segmentation in macaque histology. The reviewers agree that the task is practically important, the synthesis pipeline is well motivated and clearly described, the evaluation on held-out brains is appropriate, and mixed training achieves competitive performance with substantially less manual supervision than prior methods. The rebuttal addressed the requested clarifications on the dMRI protocol, tractography parameters, synthesis runtime, parameter selection, and dataset provenance. One substantive concern is potential circularity, namely that a tractography-derived prior could transfer tractography-specific false positives into a segmentation network intended to validate tractography. The rebuttal responded with a principled distinction: the model sees only the local axon-bundle texture a streamline produces within a thin section, not long-range trajectories, so the dominant failure mode of tractography does not propagate directly into the segmentation task, and mixed training with real annotated histology further grounds the model in biologically valid tracer appearance. This provides a coherent conceptual response, while higher false-positive rates and the single-dataset evaluation are honestly acknowledged as residual limitations. The authors are encouraged to incorporate the requested details and to discuss the local-versus-global distinction and the residual circularity risk in the limitations.
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.
Innovative and on a very hard to acquire data modality. Experiments are a strong point.