List of Papers Browse by Subject Areas Author List
Abstract
Accurate collapse prediction is pivotal for joint-preserving treatment in osteonecrosis of the femoral head (ONFH). However, mainstream 3D-CNN-based methods struggle to capture decisive fine-grained pathological signs due to the intrinsic global receptive field dilution effect, where volumetric isotropic aggregation diffuses sparse pathological singularities. Conversely, clinical reports explicitly document key regional features indicative of collapse, providing critical semantic cues to guide visual focus and enhance prediction. Driven by this clinical insight and the critical need to mitigate the volumetric signal dilution effect, we propose COMPASS (Clinical Orthogonal Micro-volume Primitive Aligned Semantic Search), a novel text-guided fine-grained evidence identification framework for ONFH collapse prediction. First, we introduce the tri-planar orthogonal micro-volume primitive (OMVP), a topological manifold representation that concentrates singularity energy onto orthogonal bases to counteract volumetric smoothing. Second, we devise a dual-stage optimization strategy. It begins with semantic space alignment to proactively navigate from macro-ROIs to micro-lesions using textual priors. Furthermore, it incorporates counterfactual causal verification to eliminate spurious correlations via hybrid supervision. Independent evaluations across diverse clinical scenarios confirm COMPASS’s state-of-the-art accuracy, yielding 97.01% AUC on a cohort with fine-grained lesion masks and 84.91% on a public dataset relying solely on coarse regional annotations, alongside causally verified localization.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4518_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/HKUSTGZ-ML4Health-Lab/COMPASS
Link to the Dataset(s)
Private ONFH clinical cohort: not publicly available due to institutional data sharing and ethics restrictions.
Osteoarthritis Initiative (OAI), public dataset available via the NIH/NDA OAI data archive: https://data-archive.nimh.nih.gov/oai/
BibTex
@InProceedings{ZenQin_Text_MICCAI2026,
author = { Zeng, Qingyuan AND Guan, Zixin AND Chen, Yusen AND Wu, Zifeng AND Li, Hao AND Ma, Qian AND Lu, Zixiao AND He, Wei AND Chen, Leilei AND Zhou, Wu AND Chen, Jintai},
title = { { Text as a Compass: Semantic-Navigated 3D Bone Collapse Prediction via Orthogonal Micro-volume Primitives } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16896},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper presents a framework for ONFH collapse prediction that formulates the task as language-guided local evidence identification. Instead of predicting collapse directly from global 3D image features, the method uses clinical text to guide the model toward local image regions that may contain predictive pathological signs. To support this formulation, the paper introduces OMVP, which represents each local 3D neighborhood using three orthogonal planar views, and a two-stage training scheme that first aligns local image features with clinical text and then uses counterfactual text supervision to suppress mismatched text-image responses.
- 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 paper formulates ONFH collapse prediction as a language-guided local evidence identification problem, instead of a standard global 3D classification problem. This is an interesting formulation because it is well matched to the clinical motivation that predictive signs of collapse may appear as subtle local structural changes rather than only as global volumetric patterns. 2.Another strength of the paper is the counterfactual text constraint to refine the selection of informative local tokens. Rather than relying only on positive image-text alignment, the method also uses mismatched or semantically contradictory text to suppress responses from regions that are visually salient but not supported by the queried semantics. 3.Results demonstrate that the proposed method improves performance
- 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 paper presents OMVP as a projection-based representation, but from the figure and formulation it seems closer to extracting three orthogonal slices from a local 3D region. The manuscript does explain whether this operation is simple slice extraction or a more general projection process. 2.The counterfactual text constraint is an important part of the proposed framework, but limited detail on the counterfactual text generation makes it difficult to understand. Its unclear if the texts are created by template replacement, manual editing, or mismatching with other samples. 3.The paper states that a sparse loss is used when fine-grained masks are unavailable, but it does not provide the explicit form of this loss or explain how the sparsity is enforced. 4.The main novelty is the use language models but at the same time, using text priors are now routinely applied, and there is no innovation in the language model.
- 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?
This paper uses a private datasets, the authors should provide clinical trial registration number or acknowledge waiver consent.
- 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 makes a meaningful contribution by formulating ONFH collapse prediction as a language-guided local evidence identification task. Overall, the paper presents a coherent method with a clear motivation and a empirical study. However, the description of the OMVP operation is not clear and the construction of the counterfactual text is not described in sufficient detail.
- 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
The paper’s contributions lie in its clinically grounded integration of textual and imaging data, using report-derived semantic cues to guide fine-grained visual evidence identification. Introducing orthogonal micro-volume primitive (OMVP) is a technically innovative contribution that directly addresses a known limitation, the dilution of sparse pathological signals. The dual-stage optimization strategy, combining semantic alignment with counterfactual causal verification, further strengthens the framework by improving both localization and robustness against false correlations. Additionally, the method demonstrates strong performance across datasets with varying annotation granularity, highlighting its practical applicability. Overall, the approach is both conceptually novel and well-aligned with clinical reasoning, which enhances its potential impact.
- 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.
Evaluation of generalizability: The method is evaluated on a public dataset in addition to the internal cohort, providing some assessment of generalization beyond the training distribution. Well-justified problem formulation: The paper addresses a clearly motivated problem, particularly the challenges of detecting and localizing micro-lesions within volumetric data. Methodological novelty: The introduction of the tri-planar orthogonal micro-volume primitive (OMVP) offers an interesting topological manifold representation that aims to concentrate singularity energy along orthogonal bases, addressing issues related to volumetric smoothing. Additionally, the proposed dual-stage optimization strategy, combining semantic space alignment with textual priors to guide the transition from macro-regions of interest to micro-lesions is a novel design. Good empirical performance: The model achieves high performance, with an AUC of 97.01% on a cohort with fine-grained lesion annotations and 84.91% on a public dataset with coarse regional labels, along with reported causally verified localization.
- 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 the authors include text-only and vision-language baselines, in addition to label tokenization, the comparison remains limited, as other relevant multimodal fusion and geometry-aware approaches are not considered.
Clinical reports are often heterogeneous and inconsistently structured; however, the paper does not discuss how variability in the clinical reports affects the robustness of the proposed approach and how they have addressed .
The authors use BioBERT for textual embedding extraction. However, more recent approaches such as BioLinkBERT leverages links between documents, enabling a more comprehensive semantic finding across different reports, outperforming BERT. I recommend the authors to also consider this approach.
- 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?
This paper presents a well-motivated approach for integrating clinical reports with imaging data through semantic-guided modeling, addressing an important challenge in this application area. The idea of using clinical report semantics as a form of guidance signal represents a promising direction for improving performance. The experimental results demonstrate consistent improvements over baseline methods, including text-only and vision-language approaches, suggesting that the proposed framework effectively captures complementary information across modalities. The inclusion of interpretability components further strengthens the contribution by providing insight into how textual cues influence model behavior. While there are some limitations, such as a relatively limited comparison with other multimodal fusion and geometry-aware methods, lack of discussion on variability in clinical report structure, and the use of BioBERT instead of more recent alternatives like BioLinkBERT, these do not override the overall contribution. Overall, the paper introduces a compelling and impactful idea with solid empirical support.
- 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
1.The paper proposes a text-guided framework for ONFH collapse prediction from CT, using a two-stage vision-language optimization strategy.
2.The paper introduces a tri-planar representation of 3D volumes—Orthogonal Micro-Volume Primitives (OMVP)—to mitigate the receptive field dilution effect in 3D CNNs.
3.The paper incorporates counterfactual reasoning by constructing mismatched vision-language pairs to suppress inconsistent semantic activations.
4.Experiments on both a private dataset (175 cases) and a public dataset (191 cases) demonstrate improved performance over conventional methods.
- 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.Overall, the paper is generally well-written.
2.The use of tri-planar representations to alleviate volumetric dilution is well-motivated and effective for capturing sparse lesions.
3.The semantic alignment framework with MIL-based attention for selecting relevant OMVPs is novel and technically interesting.
4.The proposed method uses text input to guide the evidence identification and adds counterfactual grounding to reduce hallucinations. The results are qualitatively evaluated in Fig. 3. 5.The ablation study confirmed the positive effects of each component proposed.
- 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.Overstatement of OMVP novelty. OMVP is essentially a tri-planar (multi-planar) representation of a 3D volume, which has been widely used in prior works [*1, *2, *3, *4]. The novelty appears to lie more in the attention-based aggregation and top-k selection, and the use of tri-planar representation rather than the representation itself. I recommend revising the claims regarding the novelty of OMVP to better reflect this distinction.
2.Unclear use of 2D U-Net on OMVP for segmentation. The OMVP representation (H,W,3) consists of three orthogonal planes that are not spatially aligned. It is therefore unclear how a single 2D U-Net can produce a meaningful spatial output (e.g., a segmentation map) from such input. In standard practice, U-Net output channels correspond to aligned spatial features or segmentation classes, whereas in OMVP the channels represent different planes rather than consistent spatial locations. While using a 2D model may be reasonable for regression tasks (e.g., Stage I), where features are ultimately pooled into a vector, this design is questionable for segmentation. Segmentation requires pixel-wise spatial consistency, which OMVP does not provide. A more natural approach would be to process each plane independently with separate segmentation models (or at least forward passes) and aggregate the results afterward [*2, *3, *4]. The paper should clarify how the proposed U-Net handles OMVP inputs and produces spatially meaningful outputs.
3.The heuristic sparsity loss L_sarse in Eq (6) is not clearly defined, and its formulation or source is missing.
4.The parameter k used in Stage II is not specified, and the method for choosing it is unclear.
[1] A. Parsoon et al.: Deep Feature Learning for Knee Cartilage Segmentation Using a Triplanar Convolutional Neural Network. MICCAI 2013.[2] A. Mortazi et al.: Multi-Planar Deep Segmentation Networks for Cardiac Substructures from MRI and CT. STACOM 2017.[3] V. Sundaresan et al.: Triplanar ensemble U-Net model for white matter hyperintensities segmentation on MR images. Medical Image Analysis. (2021) [4] T. Chen et al.: 2.5D Multi-view Averaging Diffusion Model for 3D Medical Image Translation: Application to Low-count PET Reconstruction with CT-less Attenuation Correction. IEEE TMI. (2025)
- 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
The figure 2 Stage 2 losses can be confusing. It’s not clear how the “Consistency loss”, “Suppression loss”, and “Heuristic constraint” correspond to equations in section 2.4.I recommend explicitly labeling the loss terms (e.g., L_dice, L_cf, and L_sparse) in the figure for clarity.
- 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, this is a strong paper with solid motivation and promising results. However, the claims regarding OMVP novelty and the unclear handling of tri-planar inputs in the 2D U-Net raise important concerns. Therefore, I recommend weak acceptance.
- 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
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.
The three reviewers are all in favor of this work. This paper aims to predict bone collapse by using text from clinical reports to guide the model toward looking at the right spots in 3D CT scans, rather than just feeding the whole scan into a classifier. The idea is clever and deserves the accept.
Minor: my personal view is that the paper has some unclear parts and overclaims that could use some fixing (clarify the OMVP and U-Net implementation details, define the missing parameters as suggested by reviewers, and explain the text generation part).
Lastly, please address some ethics concern: regarding the use of private dataset (mention waiver or IRB, etc.).