List of Papers Browse by Subject Areas Author List
Abstract
Ultrasound is the primary imaging modality for evaluating fetal anatomy, and fetal standard ultrasound planes are essential for determining gestational age and assessing overall fetal health. Deep neural networks (DNNs) have achieved remarkable success in identifying fetal ultrasound standard planes. However, their black-box nature fundamentally lacks clinical interpretability and trustworthiness, limiting their real-world deployment. To bridge this gap, we propose SONIC (Sonographer-Inspired Conceptual Explanation for Neural Networks), an interpretable framework that aligns decisions with clinically defined anatomical concepts. Moving beyond pixel-level patterns, SONIC explicitly extracts anatomical concepts and models their spatial dependencies via a graph convolutional network. We further employ a graph-attribution method to quantify the importance of anatomical structures, ensuring transparent and trackable classification. We also analyze prediction–explanation consistency, showing that anatomically implausible evidence can serve as a practical safety signal for clinical verification. Extensive experiments on a public dataset and two private multi-center cohorts, including perturbation-based fidelity analysis and a blinded expert user study, demonstrate that SONIC achieves competitive or improved performance compared with pixel-based DNN baselines while providing concept-level interpretability.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2137_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{WanYin_SONIC_MICCAI2026,
author = { Wang, Yingni AND Sun, Yihua AND Khor, Hee Guan AND Dong, Licong AND Ye, Qiongyu AND Zhang, Huabin AND Sun, Desheng AND Yuan, Kehong AND Luo, Jianwen AND Chen, Fang AND Liao, Hongen},
title = { { SONIC: Sonographer-Inspired Conceptual Explanation for Neural Networks } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16882},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors propose an intrinsically interpretable framework for fetal ultrasound that grounds predictions in clinically defined anatomical concepts. A graph convolutional network models that spatial dependencies between concepts, and Graph Integrated Gradients attribute predictions to individual anatomical structures.
- 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.
Strong and consistent classification results across datasets and backbones. Method is well-suited for application. Very strong motivation. Perturbation-based fidelity analysis provides evidence on explanation faithfulness.
- 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.
No reference to ProtoPNet or prototype-based interpretability (https://arxiv. org/abs/1806.10574) despite the similarity matching being essentially this paradigm applied to ultrasound. User study evaluates SONIC explanations only. Experts were not asked to compare against other methods, so the ratings don’t show relative superiority. The term “incorrect explanation” is poorly defined and conflates anatomical implausibility with unfaithfulness to the model. No clear justification for training separately on public and private datasets rather than training on public and using both private cohorts as external validation, which would be a stronger experimental design. No code released or planned to be released.
- 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?
The method is well-executed and clinically motivated. The classification results are strong and consistent. The main weaknesses are: missing adequate prototype-based literature, and the user study design and lack of comparison with alternative approaches.
- 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 presents SONIC, an interpretable framework for fetal ultrasound standard plane classification that grounds predictions in clinically defined anatomical concepts. The pipeline constructs prototypes for 12 anatomical landmarks offline via superpixel-based proposals, manual verificaton, and K-means clustering on CNN embeddings. At inference, image sub-regions are matched to prototypes via cosine similarity, and a fully-connected concept graph is classified by an edge-gated graph convolutional netwrok, providing concept-level attribution. Evaluation spans one public and two private multi-centre datasets, including perturbation-based fidelity analysis, a blinded expert user study, and prediction-explanation consistency analysis. The authors report competitive or improved performance over CNN baselines while providing concept-level interpretability.
- 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.The clinical motivation is well-articulated and the core design choice of constraining explanations to a fixed vocabulary of named ISUOG-defined anatomical landmarks rather than unsupervised discovered concepts is useful for clinical interpretability, improving pixel-level saliency baselines.
2.The multi-centre design is sound, with one centre held out for external validation.
3.The combination of concept-based interpretability with explicit spatial relational modelling is a good methodological contribution compared to CBMs or PCBMs, which treat concepts as independent.
4.The evalutaion covers multiple complementary angles which is a good practice to evaluate the interpretablity of a model.
- 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.I feel like the paper misses a quantitative evaluation of the online concept-matching step. The authors could maybe report precision/recall of the matching against a verified expert agreement on a held out sample.
2.The user study would be more convincing with inter-rater agreement statistics, given the substantial variability across ratrs in Table 1.Having the experts also rate CRAFT or GradCAM explanations would help establish that the reported ratings relfect the quality of SONIC’s explanatios specifically rather than the general utility of anatomical highlighting.
4.The failure-awareness claim is currently only illustrated qualitatively and would benefit from quantitative support, such as th correlation between low concept-matching confidence and incorrect predictions.
- 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?
The paper is well written and the methodology is novel and sound. The evaluation protocol is rigorous. The paper would benefit from making the code publicly available upon acceptance.
- 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
The paper presents SONIC, an intrinsically interpretable framework for fetal ultrasound standard plane classification. To address the “black-box” limitation of CNNs, it extracts anatomical concepts via a prototype-based approach, models their spatial relationships using an edge-gated GCN, and enables concept-level attribution through Graph Integrated Gradients. Evaluated on one public and two private multi-center datasets, the method aims to enhance clinical interpretability and trustworthiness while maintaining competitive classification 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.
(1) Strong clinical motivation and concept-driven design. The paper addresses an important problem—improving interpretability in fetal ultrasound standard plane classification, where clinical trust is essential. The proposed framework introduces anatomically meaningful concepts and models their spatial relationships via a graph structure, aligning well with clinical reasoning and going beyond pixel-level explanations. (2) Comprehensive evaluation and clinical utility. The method is validated through a multi-level evaluation protocol, including multi-center datasets, perturbation-based fidelity analysis, and a blinded expert user study, which strengthens the empirical evidence. Additionally, the framework provides concept-level explanations and failure-aware signals, enhancing its potential clinical usability and trustworthiness.
- 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 novelty with respect to prior work. While the paper is well-motivated, the overall research direction is not entirely novel. Prior studies have explored concept-based interpretability and graph-based reasoning in medical imaging, including works that integrate anatomical concepts with graph neural networks for ultrasound plane analysis (e.g., A Cognitive Explainer for Fetal Ultrasound Image Classification Based on Medical Concepts and Concept-Induced Graph Perception Model for Interpretable Diagnosis). In this context, the core idea of combining concept extraction with GCN-based relational modeling has already been investigated, suggesting that the current work represents an incremental extension rather than a fundamentally new contribution. (2) A key limitation of SONIC is its strong reliance on manually defined anatomical concepts, which limits generalizability. First, it strongly depends on accurate ROI localization and segmentation. Since the pipeline begins with U-Net-based ROI extraction, failures at this stage—such as under poor image quality, extreme fetal poses, or abnormalities—can propagate through the pipeline, causing cascading errors in concept extraction and graph reasoning. Second, the concept extraction process requires manual intervention. As noted in Section 2.1, sub-region-to-concept pairings must be manually verified to remove incorrect matches, meaning the framework is not fully end-to-end. This reliance on expert input introduces scalability challenges and high annotation costs, making the approach less practical for large-scale or multi-task deployment. (3) Another limitation of SONIC is its complex multi-stage pipeline, which introduces cumulative error propagation. The framework consists of sequential components—ROI segmentation, superpixel generation, concept matching, and graph-based reasoning—where each stage is susceptible to its own errors. As these components are tightly coupled, inaccuracies in earlier steps can propagate and amplify through the pipeline, ultimately degrading the final prediction. This design increases the system’s sensitivity to intermediate failures and raises concerns about its robustness in real-world settings. (4) There is a lack of analysis on real-time performance and engineering complexity. The multi-stage pipeline—CNN, concept matching, GCN, and attribution—may introduce non-trivial inference overhead, yet no latency or resource metrics are reported. This raises concerns about its suitability for real-time ultrasound workflows, especially for continuous video 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 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?
The paper addresses an important clinical problem and proposes a concept-driven framework that aligns well with anatomical reasoning, supported by relatively comprehensive evaluation including multi-center data and expert studies. These strengths suggest practical relevance and potential clinical impact. However, the technical novelty is limited compared to prior concept-based and graph-based approaches, and the method relies heavily on manually defined concepts and accurate ROI segmentation, which restricts generalizability and scalability. In addition, the multi-stage pipeline introduces error propagation risks, and the lack of analysis on computational efficiency raises concerns about real-time applicability.
- 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
N/A
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.
Reviewers viewed this as a strong and clinically meaningful contribution, supported by solid methodology, comprehensive evaluation, and clear practical relevance. While some clarification and additional analysis would further strengthen the work, the overall reviewer consensus was strongly positive, and the paper is above the acceptance bar.