List of Papers Browse by Subject Areas Author List
Abstract
Instance-level quantification of kidney functional units is essential for morphometric analysis, yet most publicly available pathology datasets provide only semantic segmentation annotations, where adjacent structures of the same class are merged into single regions. This prevents reliable instance-level analysis and limits downstream quantitative studies. Existing heuristic post-processing methods often yield suboptimal instance separation, particularly in crowded and adherent regions, while deep learning-based instance segmentation approaches typically require intensive instance-level annotations that are costly and labor-intensive to obtain. We propose MORI-Seg, a deep learning framework that enables instance segmentation without requiring instance-level annotations. Instead of heuristic splitting or instance supervision, MORI-Seg learns morphology-aware geometric representations directly from semantic masks by jointly modeling object-centric distance fields and boundary-band representations to encode interior structure and contact interfaces. A class-conditioned feature disentanglement module further promotes intra-instance coherence and inter-instance separation. Under semantic-only supervision, MORI-Seg decomposes connected semantic regions into distinct instance masks in an end-to-end manner. Experiments demonstrate improved instance separation accuracy and more reliable morphometric quantification compared with classical post-processing pipelines and representative semantic-to-instance learning approaches. The official implementation is publicly available at https://github.com/ddrrnn123/MORI-Seg.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3815_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/ddrrnn123/MORI-Seg
Link to the Dataset(s)
KPMP dataset: https://atlas.kpmp.org
KI dataset: http://haeckel.case.edu/data/KI_data
BibTex
@InProceedings{ZhaLei_MORISeg_MICCAI2026,
author = { Zhao, Leiyue AND Shi, Tianyu AND Reisenbüchler, Daniel AND He, Xinzi AND Zhu, Junchao AND Yao, Tianyuan AND Yang, Yuechen AND Zhu, Yanfan AND Guo, Junlin AND Xu, Gelei AND Yang, Haichun AND Huo, Yuankai AND Sabuncu, Mert R. AND Yang, Yihe AND Deng, Ruining},
title = { { MORI-Seg: Learning Morphological Geometry for Instance Segmentation Without Instance Annotations } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16883},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes MORI-Seg, a new deep learning framework, for the segmentation of renal pathology instances using only semantic-level supervision. Its core contribution lies in learning morphologically aware geometric representations directly from semantic masks, eliminating the need for time-consuming instantiation-level annotations.
- 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.Unlike heuristic methods or general seman-to-instance methods, MORI-Seg explicitly combines kidney-specific morphological priors. 2.This method provides a new idea for the instance annotation of pathological images and does not require a lot of time to annotate the images. 3.The model was trained and tested on different datasets, verifying its generalization ability on different stained images. 4.A detailed comparative experiment was conducted, and the selected comparative methods included different frameworks and were compared with multiple latest methods from 2026.
- 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.This work employs a specific category of magnification strategies, which can further discuss the sensitivity of the morphologically aware branch to changes in image resolution, or whether fixed-size patches would limit the capture of extremely large or slender structures. 2.The provided network framework diagram is not conducive to reading. A clear version needs to be provided.
- 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.
(6) Strong Accept — must be accepted due to excellence
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This method can perform reliable instance segmentation without instance-level labels. The introduced boundary band mapping and distance field provide a multi-faceted geometric signal, logically resolving the failure mode of the previous embedding method. Furthermore, the design choice of removing auxiliary branches during reasoning ensures that the increased morphological intelligence does not come at the cost of computational speed, making it suitable for large-scale clinical full-slice imaging.
- 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 #2
- Please describe the contribution of the paper
This paper proposes MORI-Seg, a framework for instance segmentation in pathology images under semantic-only supervision, aiming to address the practical limitation that most public datasets provide only semantic masks without instance-level annotations.
The key idea is to bridge the gap between semantic segmentation and instance segmentation by explicitly incorporating morphology-aware geometric priors into the learning process. Specifically, the method introduces a morphological geometry branch that jointly models:
1、object-centric distance fields, capturing interior structure of instances 2、boundary-band representations, encoding contact interfaces between adjacent objects
In addition, a feature-level instance disentanglement module is designed to promote intra-instance compactness and inter-instance separability. The framework is trained end-to-end using only semantic masks, and all auxiliary branches are removed during inference, preserving the original instance segmentation pipeline.
The method is evaluated on renal pathology datasets and compared with heuristic post-processing methods, fully supervised instance segmentation models, and semantic-to-instance approaches, demonstrating improved instance separation and morphometric reliability.
- 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 tackles a very important and realistic problem in computational pathology: the lack of instance-level annotations. Most prior works either rely on heuristic post-processing (e.g., watershed, morphology) or require expensive instance masks. This work directly targets this gap and proposes a solution that operates under semantic-only supervision, which is highly valuable for large-scale clinical deployment.
The introduction of morphology-aware geometric modeling is a strong and well-justified idea. In particular:
The distance field modeling captures object interior structure in a continuous manner.The boundary-band representation explicitly models contact regions, which are critical for separating touching structures
This combination is both intuitive and tailored to pathology images, where structures are often irregular and densely packed. Compared to prior semantic-to-instance approaches (e.g., pixel embedding [Chen et al., MICCAI 2019], boundary-aware methods such as DCAN [Chen et al., MedIA 2017]), this formulation explicitly incorporates domain-specific morphology priors, which is a meaningful advancement.
An important practical advantage is that all auxiliary branches are used only during training and removed at inference, preserving the original RTMDet-Ins pipeline.
This makes the method attractive for real-world deployment, as it improves performance without increasing inference complexity.
- 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、Although the overall framework is well designed, its components are related to existing lines of work:
Distance map supervision has been used in boundary-aware or regression-based segmentation (e.g., InstaBound, distance regression methods) Boundary modeling is a common strategy (e.g., DCAN, boundary-aware segmentation) Instance embedding and feature disentanglement are also well studied (e.g., pixel embedding methods)
The novelty lies mainly in how these components are combined and adapted to pathology morphology, rather than introducing a fundamentally new formulation. While this is acceptable, the paper could better emphasize what is truly new compared to recent works such as InstaBound (MedIA 2026) or SIM (CVPR 2023).
2、Limited comparison with the most recent semantic-to-instance methods
Although several baselines are included, comparison with very recent methods (2024–2025) is still somewhat limited. For example:
Recent morphology-guided or binary-to-instance methods (e.g., DyMorph-B2I, 2025) More advanced foundation-model-based segmentation pipelines
Including such comparisons would further strengthen the empirical claims.
3、Ablation analysis could be deeper
While the ablation study shows the contribution of distance and boundary priors, the analysis remains somewhat limited:
The interaction between feature disentanglement and geometry modeling is not deeply explored Sensitivity to hyperparameters (e.g., λ weights, boundary width) is not analyzed It is unclear how robust the method is to different semantic mask qualities
More detailed analysis would help better understand the behavior of the model.
4、Generalization beyond renal pathology is not fully demonstrated
The method is designed with renal morphology in mind, which is reasonable, but:
It is unclear how well the approach generalizes to other organs or tissue types The morphology priors may be domain-specific
A brief discussion or additional experiment would strengthen the generality of the method.
5、Interpretability and error analysis could be improved
Although qualitative results are provided, the paper lacks:
quantitative evaluation of instance separation quality (e.g., split/merge errors)
- 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 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.
(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?
Overall innovations and improvements in experimental results
- 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 #3
- Please describe the contribution of the paper
The paper proposes MORI]Seg, a deep learning framework for instance segmentation in renal pathology images that operates under semantic-only supervision. Its main contributions are: L1) enabling instance segmentation without any instance-level annotations; L2) incorporating explicit morphological priors of renal structures by jointly modeling object-centric distance fields and boundary-band representations to guide instance separation; and L3) demonstrating through cross-dataset evaluation that this approach consistently outperforms heuristic post-processing and existing semantic-to-instance methods in terms of instance separation accuracy and morphometric reliability.
- 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.It addresses the critical and common challenge in medical imaging where only semantic annotations are available, bridging the gap between annotation cost and analysis need. 2.The integration of morphology-aware geometric priors (distance fields and boundary bands) with feature disentanglement provides a principled and effective way to decompose connected semantic regions into instances. 3.The method demonstrates state-of-the-art performance on a challenging renal pathology dataset under semantic-only supervision, significantly outperforming both classical heuristics and recent learning-based baselines. 4.The auxiliary geometric branches are only used during training and removed at inference, ensuring no extra computational overhead in deployment.
- 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 experiments are exclusively conducted on renal pathology datasets LKI for training, KPMP for evaluation). The generalizability of the proposed morphology-aware priors to other medical image domains (e.g., cell nuclei, neurons, or organs in radiology) is not demonstrated. 2.The framework is built upon the RTMDet-Ins architecture. The paper does not sufficiently explore how the choice of the base model affects the final performance or the relative gain from the proposed components. 3.While the inference cost is the same as the base model, the paper lacks a discussion on the increased training complexity and time due to the additional geometric branches and loss terms.
- 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.
(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 presents a well-motivated, technically sound, and empirically validated solution to a significant problem in medical image analysis. The core idea of leveraging explicit morphological geometry for instance separation under weak supervision is novel and effectively executed. The results are compelling and clearly demonstrate the method’s superiority over relevant baselines. Despite the limited scope of validation, the work is solid, clearly presented, and offers a valuable contribution to the field of weakly-supervised instance segmentation in medical imaging.
- 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
Author Feedback
We thank the reviewers and AC for the positive and insightful feedback and summarize the comments into three major and two minor concerns. [Major concerns] C1.Clarification of novelty beyond existing semantic-to-instance methods (R2): We thank the reviewer for pointing out similarities to prior distance-map, boundary-aware, and embedding-based methods. We would like to clarify that the novelty of MORI-Seg lies in introducing a morphology-aware geometric learning framework specifically designed for renal pathology under semantic-only supervision. (1) MORI-Seg enables reliable instance segmentation without any instance-level annotations, without requiring instance-level annotations. (2) MORI-Seg jointly models object-centric distance fields and boundary-band representations to explicitly capture renal pathological morphology in densely packed regions. (3) We introduce geometry-guided feature disentanglement to improve intra-instance compactness and inter-instance separability. We will further clarify the differences from DCAN, InstaBound, SIM, and embedding-based methods in the revised manuscript. C2.Generalization, backbone dependency, and magnification strategy (R1, R2, R3): We thank the reviewer for raising the concern regarding the additional training complexity introduced by the geometric branches. Compared with the baseline model (57.309M parameters), MORI-Seg introduces only a marginal parameter increase of 0.024M. The training time increases from 10:02:08 to 12:53:32 due to the additional geometric supervision and optimization objectives. The auxiliary branches introduce minimal memory overhead and no additional inference cost since all geometry branches are removed during deployment. Multi-scale sampling is necessary because renal pathological structures vary significantly in size, with some categories containing very small or thin structures while others are much larger. Our multi-scale sampling strategy follows OmniSeg. RTMDet-Ins was selected because it achieved the best overall performance in our preliminary experiments. In future work, we plan to further investigate the generalizability of MORI-Seg by integrating the proposed morphology-aware modules into other segmentation frameworks and backbones. C3.Comparison with recent methods and foundation model (R2, R3, AC): We will include more comprehensive discussion. Some recent approaches such as DyMorph-B2I still rely on heuristic post-processing and manually designed morphology rules, which differs from our end-to-end learning-based framework. Regarding foundation-model-based methods, we have included a DINOv3-style feature extractor combined with Mask2Former as a strong baseline, and MORI-Seg still achieved superior performance under semantic-only supervision. [Minor concerns] C4.Limited validation on additional datasets and organs (R2, R3): We agree that broader validation would strengthen the paper. Due to current dataset limitations, experiments focused mainly on renal pathology. We agree broader validation would strengthen the paper and will clarify this limitation in the manuscript. C5.Additional metrics and deeper ablation analysis (R2): We additionally performed statistical significance analysis on all existing results to validate the observed improvements. The p-values from the Wilcoxon signed-rank test show statistically significant improvements, with all p-values between the proposed methods with other methods less than 0.001, including both the baseline comparison and the ablation study settings. More detailed split/merge statistics and sensitivity analyses will be included in future work.
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.
This paper received two weak accepts and one strong accept, and all reviewers recommended acceptance. The final recommendation is to accept the paper. The paper addresses a clear practical need in biomedical image segmentation evaluation, where instance-wise assessment is often more clinically meaningful than global semantic metrics, while existing tools remain fragmented, limited, or inconvenient for 2D/3D biomedical use cases. The proposed panoptica package provides a modular and extensible pipeline covering instance approximation, instance matching, and instance-level metric computation, enabling unified evaluation of semantic and instance segmentation outputs. Its open-source implementation, support for multiple metrics, applicability to both semantic and instance maps, and demonstrations on two biomedical tasks make it a useful contribution to the community. Although the experimental comparison with existing tools and runtime analysis could be strengthened, these limitations do not outweigh the practical value of the framework. Overall, the paper provides a reusable and timely evaluation tool for instance-based segmentation assessment; therefore, the final recommendation is acceptance.
