List of Papers Browse by Subject Areas Author List
Abstract
Deep learning-based models have significantly advanced clinical ultrasound tasks by detecting anatomical structures within vast ultrasound image datasets. However, their remarkable performance inherently requires extensive training of annotated medical datasets. Few-shot learning addresses the challenge of limited labeled data for model training. Currently, few-shot learning in the field of medical image analysis mainly focuses on classification and semantic segmentation, with relatively fewer studies on object detection. In this paper, we propose a novel few-shot anatomical structure detection method in ultrasound images called TRR-CCM, which consists of Circular Channel Mamba (CCM) and Topological Relationship Reasoning (TRR) based on human anatomy knowledge. CCM, as a new Mamba variant, performs contextual modeling of anatomical structures and captures long- and short-term dependencies. TRR learns spatial topological relationships between human anatomical structures to further improve the accuracy of detection and localization. Experimental results on two fetal ultrasound datasets demonstrate that TRR-CCM outperforms 9 state-of-the-art baseline methods.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0344_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/yuyizhilian/TRR-CCM
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ZhuYin_Anatomical_MICCAI2025,
author = { Zhu, Ying AND Liang, Bocheng AND Li, Ningshu AND Zhao, Lei AND Li, Xi AND Li, Hao AND Yang, Fengwei AND Pu, Bin},
title = { { Anatomical Structure Few-Shot Detection Utilizing Enhanced Human Anatomy Knowledge in Ultrasound Images } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15964},
month = {September},
page = {35 -- 45}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper presents a few-shot anatomical structure detection method in US images by interpreting the topological relationship reasoning. the idea is very interesting.
- 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 idea to combine the physiological knowledge for few-shot leaning is very interesting and novel.
- 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 explanation of the main methodology is difficult to follow. For example, Figure 2 is not very easy to understand, and I find the details of each module somewhat unclear. If the authors could improve the clarity and explanation of the methodology, it would significantly enhance the overall quality of the paper.
Based on the ablation study, the final model does not show significant improvement over the version using only the CCM module. Therefore, a more detailed analysis of the results is necessary to justify the additional components.
Additionally, I could not find any theoretical design or justification aimed at improving the generalization of the model. Including further explanation on this aspect would be important for understanding the broader applicability of the proposed method.
Minor points: It will be important for evaluating the reproducibility if the authors can release the code.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- 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/2025/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 — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
see above
- 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.
The paper is interesting, but the method and details are still not properly justified and explained after the rebuttal. Since this is MICCAI, I do appreciate the new method and try. So I still suggest accepting, but more clearer explanation will help to further enhance the paper quality.
Review #2
- Please describe the contribution of the paper
The paper proposes a few-shot detection framework for anatomical structure localization in fetal ultrasound images. The main contribution is the introduction of a Circular Channel Mamba module, designed to capture both long-range and short-range dependencies across multiple anatomical structures. In addition, the method incorporates topological relationship reasoning by encoding anatomical knowledge through a human anatomy graph, which models spatial and structural relationships between different organs or regions. The approach is evaluated on two fetal ultrasound image datasets, where it demonstrates improved performance over existing methods. These results suggest potential for clinical translation.
- 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.
-
Tackles few-shot object detection in ultrasound, a domina with limited prior work and high clinical relevance
-
Proposed and implemented a method that combines contextual feature modeling with anatomical relationship reasoning
-
The proposed TRR architecture seems to be a great idea to incroprate anatomical priors using graph-based modeling
-
- 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 method is only evaluated on fetal ultrasound data; its applicability to other anatomical sites or imaging modalities remains untested.
- 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.
- 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/2025/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 — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I recommend weak accept for this paper. The work addresses a relevant and underexplored problem in few-shot anatomical structure detection in ultrasound imaging. The proposed TRR-CCM framework, which combines contextual modeling through the Circular Channel Mamba module with anatomical topology reasoning via a graph-based approach, is interesting. The method is evaluated on two fetal ultrasound datasets and demonstrates improved performance. However, its applicability to other anatomical sites or imaging modalities remains untested beyond ultrasound. Despite this limitation, the core technique is promising.
- 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
Review #3
- Please describe the contribution of the paper
This paper presents a novel few-shot anatomical structure detection method for ultrasound images, TRR-CCM. The proposed approach combines two key components: Circular Channel Mamba (CCM), which is designed to capture both long- and short-term contextual dependencies while preserving channel-specific information, and Topological Relationship Reasoning (TRR), which encodes human anatomical knowledge through graph-based spatial topological relationships. The method’s effectiveness is validated on two fetal ultrasound datasets, TT and 3VT, where it achieves notable performance improvements over nine state-of-the-art few-shot object detection (FSOD) baselines.
- 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.
Novelty in Model Architecture:
- Circular Channel Mamba extends the Mamba framework to the channel dimension, addressing a known limitation in previous Mamba-based methods.
- TRR leverages consistent anatomical spatial relationships, an underexplored area in FSOD for ultrasound imaging. Domain-specific Insight: Incorporating anatomical topology and contextual knowledge as priors directly address domain shift challenges in ultrasound.
Empirical Rigor:
- A comprehensive evaluation of two real-world datasets.
- Ablation studies clearly show the individual and joint impact of CCM and TRR.
- Substantial performance gains over baselines.
Clinical Applicability: Not much explanation is provided in the paper. Based on the results of these two datasets, the author mentioned that it can become a clinical application.
- 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.
While presenting a novel and practical framework, the paper lacks theoretical analysis to substantiate the design choices of CCM and TRR. These modules are intuitively motivated, yet a deeper theoretical grounding from a signal processing or information-theoretic perspective would enrich an understanding of their performance advantages. Additionally, there is no discussion on the scalability of the TRR module. Specifically, the computational cost of its graph convolutional network as the number of anatomical structures or image resolution increases remains unexplored, which raises concerns about its practical deployment. Moreover, the paper focuses solely on few-shot object detection without providing any comparative results against fully supervised detection models. Such a comparison would offer valuable context regarding the trade-offs between annotation effort and model accuracy. Finally, the work does not include any detailed qualitative or quantitative error analysis regarding the model’s performance limitations, such as occluded structures, rare cases, or smaller anatomical features.
- 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.
- 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/2025/en/REVIEWER-GUIDELINES.html
-
TRR uses graph convolutional networks with a learned adjacency matrix and top-k nearest nodes. How does this scale with increasing region proposals or classes? Is there a computational bottleneck?
-
In the ablation study (Table 3), the CCM and TRR modules are shown to contribute individually. Could you isolate the effect of different components inside CCM (e.g., Mamba vs. convolutional mixing) to verify which part contributes most?
-
TRR relies on stable anatomical topology. How would the model perform in pathological or deformed anatomical cases where topology might deviate? Can the model handle such variations?
-
Have you compared the TRR-CCM model with a fully supervised detector trained with large labelled datasets? It would be helpful to contextualize the performance drop (if any) vs. annotation savings.
-
Could you provide examples or statistics on failure cases?
-
Why did you choose undirected graphs with Gaussian-weighted kernels for topological reasoning? Did you explore alternatives like attention-based or directed graphs?
-
Few-shot learning is known for high variance across runs. Did you report the average of multiple runs or provide variance/error bars?
-
Given that your training data comes from one hospital, how confident are you in the generalizability across populations, devices, and imaging protocols? Were there any observed biases?
-
- 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 — could be accepted, 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 novel and technically robust approach to the challenging problem of few-shot anatomical structure detection in ultrasound imaging. The proposed integration of Circular Channel Mamba and Topological Relationship Reasoning is original and practically well-motivated, addressing key limitations in current few-shot detection models. The empirical results are compelling, with extensive evaluations demonstrating consistent improvements over baselines. However, the current version of the paper exhibits several limitations. It lacks theoretical justification for the proposed modules, does not provide a public implementation or sufficient details to ensure reproducibility, and falls short of offering a comprehensive failure analysis or demonstrating the model’s generalization capabilities across varied clinical scenarios.
- 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.
Authors have addressed the concerns in their rebuttal.
Author Feedback
To AC and Reviewers: We appreciate the reviewer’s constructive feedback and recognition of the TRR-CCM. They pointed out that our idea is very novel (R1, R2, R3), that topological relationships are valuable in the field of Few-Shot Object Detection (FSOD) in ultrasound imaging (R2, R3), and that CCM addresses the limitations of existing methods (R3). Additionally, significant results have been achieved on two real-world datasets (R2, R3), which suggest the potential for clinical translation of this research (R2). All codes will be available.
To R1: Q1: Methodology unclear. A: Due to the space limit, we will clarify our statements in the final version. Q2: Ablation study. A: The original ablation only demonstrated split 1 of the TT and 3VT datasets (TRR improvement is limited). On splits 2 and 3 of the 3VT dataset, the accuracy of TRR is higher than that of CCM in most cases, thus proving that TRR is effective. We will provide the complete results of the ablation experiments in the final version.
To R2: Q1: Applicability to other anatomical sites or imaging modalities. A: Ultrasound is widely used in the medical field, and our method has achieved significant results on several real ultrasound datasets, demonstrating the potential of our method for generalized ultrasound. The other modal datasets are not available, it may be future work. We have also added a comparison FUSH dataset [ICML2024]. Compared with 9 other comparative methods, our method achieves the best performance in most cases, thereby demonstrating its applicability. Update the results on the final version.
To R3: Q1: Computational bottleneck. A: In TRR, the adjacency matrix is dynamically learned, with its size increasing as the number of region proposals or classes grows, while maintaining sparsity. The GCN’s time complexity, which depends on the feature matrix dimension F and adjacency matrix sparsity, is kept at O(∣E∣F) through sparse matrix optimization and a top-k mechanism, where ∣E∣ is the number of edges. Potential memory bottlenecks are managed via sparse matrix representation and efficient storage strategies. Q2: Ablation study in Tab. 3. A: The ablation experiments on Mamba and convolutional mixing indicate that convolutional mixing contributes more to CCM. We will update these results in the final version. Q3: Pathological or deformed anatomical. A: The TRR-CCM refines and propagates the contextual and spatial information of anatomical structures into the graph, and updates node features through the GCN by aggregating information from neighboring nodes. This process ensures effective information propagation and fusion even when the topological structure changes. Additionally, it enhances the identification of difficult regions in pathological or deformed anatomical cases by leveraging enhanced contextual information from related areas. Q4: Fully supervised detector. A: In a fully supervised way, testing split 3 of TT, with mAP@50 of 92.5%. Under the 10-shot few-shot setting, TRR-CCM achieved 86.8% (very small gap), which shows the effectiveness of TRR-CCM. Update the results in the final version. Q5: Undirected graphs with Gaussian-weighted kernels. A: In anatomical topology, adjacent structures share similar features, modeled effectively by the Gaussian kernel with efficient computation. In most scenarios, bidirectional node relationships make undirected graphs ideal for capturing intrinsic topological structure. Q6: The average of multiple runs. A: Yes (three runs), we will report on standard deviation in the final version. Q7: Any observed biases. A: Despite the training data being sourced from one hospital, the diversity in gestational age (18–32 weeks), maternal ages, acquisition devices, and operators endows the dataset with diversity and broad representativeness, thereby strongly demonstrating the robust adaptability and generalizability of our model.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
N/A
- 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A