List of Papers Browse by Subject Areas Author List
Abstract
Chronic obstructive pulmonary disease (COPD) is a leading cause of mortality worldwide. COPD causes irreversible lung damage, making early detection essential for slowing disease progression. However, current clinical assessments primarily rely on pulmonary function tests, which require respiratory expertise and cannot fully capture underlying structural abnormalities. Chest computed tomography (CT) provides rich anatomical information, yet manual interpretation remains challenging due to COPD’s subtle, diffuse, and heterogeneous disease patterns. Multiple instance learning (MIL) has demonstrated promising performance for weakly supervised COPD classification, but existing methods largely focus on prediction accuracy and offer limited anatomical interpretability, which reduces trustworthiness and phenotype-level insight. In this work, we propose AnatoMIL, an anatomy-aware hierarchical MIL framework for joint COPD detection, severity estimation, and phenotype analysis from 2D CT slices. Our method incorporates lung lobe information to guide hierarchical feature aggregation and employs a dual-head strategy for binary classification and ordinal severity modeling. Furthermore, we analyze the relationship between severity stages and lobe-level contribution profiles to characterize spatial disease patterns. The proposed model is trained and validated on a lung cancer screening dataset with 2,357 scans and externally evaluated on a dataset with 1,753 scans from an independent observational COPD study, demonstrating robust performance and strong generalization. This framework enables interpretable and robust COPD detection and phenotype analysis from CT imaging and has the potential to improve the ability to diagnose and understand the disease patterns of COPD.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6264_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/rahaahmadi/AnatoMIL
Link to the Dataset(s)
N/A
BibTex
@InProceedings{AhmRah_AnatomyAware_MICCAI2026,
author = { Ahmadi, Raha AND Sin, Don AND Lam, Stephen AND Eddy, Rachel AND Tam, Roger},
title = { { Anatomy-Aware Hierarchical Multiple Instance Learning for Interpretable COPD Diagnosis and Phenotype Analysis } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16886},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper presents a lung anatomy guided Multiple Instance Learning (MIL) framework designed for COPD classification and grading from 3D CT scans. The model is developed and evaluated on the Pan cancer dataset, then externally validated on the large ECLIPSE cohort. Interpretability is presented through a lobe specific contribution analysis, which reveals the relationship between spatial distribution of feature importance and COPD progress stages.
- 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.
Clinically grounded anatomical prior: The integration of lung lobe anatomy as a spatial prior is clinically well motivated, making the model’s reasoning more consistent with radiological practice. Clear and coherent technical presentation: The manuscript is well structured, and the method section is clearly written, making the overall model architecture easy to understand. The accompanying figures are clearly plotted for the proposed hierarchical MIL framework. Meaningful interpretability insights: The lobe level analysis adds practical interpretability by linking model attention to clinically meaningful lung regions, offering a clearer understanding of how the framework arrives at its COPD severity predictions.
- 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.
Limited encoder generalization: Relying exclusively on ResNet18 prevents a fair assessment of how well the proposed MIL framework generalizes across different feature extractors. The absence of comparisons with stronger or CT specific pretrained encoders (e. g. , DenseNet121, CT-CLIP) makes it unclear whether the reported improvements are a fundamental benefit of the hierarchical framework or merely an artifact of the specific interaction with a relatively shallow ResNet18. Unclear derivation of lobe fraction vectors: The formulation of the lobe fraction vectors is insufficiently described; it remains ambiguous whether these fractions are derived from slice level area ratios. Lack of transparency in benchmarking: The inclusion of a “Hierarchical MIL” baseline in the result table did not have clear citation or detailed description.
- 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?
This paper presents a clinically grounded MIL framework that incorporates anatomical priors for COPD grading. Its strongest contribution is the robust external validation on the ECLIPSE cohort, demonstrating promising generalizability, and the lobe level analysis adds clinically meaningful interpretability. However, the technical novelty is constrained by the exclusive use of ResNet18; this single encoder setup limits the assessment of generalizability, and the performance gains are marginal. A broader evaluation across commonly used and CT specialized pretrained encoders is needed to show that the improvements stem from the MIL framework itself. If the authors can provide such evidence and demonstrate consistent superiority over benchmark models under diverse encoder settings, the paper would be considerably strengthened and could be viewed more favorably for 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
Review #2
- Please describe the contribution of the paper
This paper proposes AnatoMIL, an anatomy-aware hierarchical Multiple Instance Learning (MIL) framework for joint COPD detection and severity staging. The core contribution lies in explicitly incorporating lung lobe anatomical priors to guide feature aggregation. By combining a dual-head prediction structure (binary classification and ordinal regression) with a gated attention mechanism, the model quantifies the pathological contributions of five lung lobe regions, thereby achieving strong interpretability for spatial COPD phenotyping.
- 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) Excellent clinical interpretability: This is the most significant highlight of the paper. Traditional MIL models lack anatomical constraints and operate as “black boxes.” This work cleverly utilizes anatomical priors (lobe-fraction vectors) and a gated attention mechanism to translate attention weights into clinically readable lobe-level contribution profiles. This greatly enhances the trustworthiness and transparency of the model in real-world medical scenarios. 2) Clinically aligned task modeling: Introducing an all-threshold ordinal regression head to predict the GOLD severity stage is highly logical. It aligns much better with the objective clinical reality of continuous disease progression compared to standard unordered multi-class classification. 3) Solid cross-cohort validation: Evaluated on a large-scale internal dataset (PanCan) and an independent external cohort (ECLIPSE) comprising over 4,000 scans, the model demonstrates highly consistent spatial phenotype shifts (disease contribution shifting from upper to middle/lower lobes with increasing severity). This proves the method’s strong generalization capability and biological plausibility.
- 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) Marginal performance improvement: While the model excels in interpretability, its improvements in core quantitative metrics (e.g., binary AUC, severity staging QWK) are very marginal compared to existing baselines like TransMIL (as shown in Table 1). It is difficult to fully justify the algorithmic superiority of the proposed framework based purely on predictive performance. 2) No code availability: The authors do not provide a link to open-source code or promise to release it. For a MICCAI-level paper proposing a novel MIL framework, the lack of code severely limits the reproducibility of the research. 3) Lack of robustness analysis for upstream segmentation errors: The framework heavily relies on pre-trained lobe segmentation masks. In advanced COPD patients (GOLD 3-4), severe emphysema often leads to blurred lung boundaries and a high risk of segmentation failure. The paper does not discuss the robustness or error propagation to downstream anatomical attention distributions when the upstream segmentation is inaccurate. 4) Insufficient ablation study: The main text lacks detailed ablation analyses to decouple the contributions of individual modules. For instance, what is the specific performance gain of the auxiliary ordinal loss on the primary binary classification task? How much does performance degrade if predicting directly without lobe-guided hierarchical aggregation? The absence of these experiments weakens the methodological necessity of the proposed architectural design.
- 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?
This is a study with exceptionally clear clinical motivation. Its core value lies in using anatomical priors to open the “black box” of medical image MIL, endowing the model with strong spatial interpretability. This design accurately addresses a major pain point in clinical AI deployment. However, the algorithmic superiority of the method is not fully demonstrated (quantitative improvements are marginal), and the absence of open-source code, along with lacking critical robustness and ablation analyses, are notable flaws. I strongly encourage the authors to release their code and provide the missing robustness and ablation analyses during the Rebuttal phase to strengthen their claims.
- Reviewer confidence
Somewhat confident (2)
- [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.
N/A
Review #3
- Please describe the contribution of the paper
The paper proposes an anatomy-aware hierarchical MIL framework for COPD diagnosis and severity assessment from CT. By introducing lobe-level aggregation into the MIL pipeline, the method aims to improve anatomical interpretability while maintaining competitive performance. The additional phenotype analysis based on lobe-level contributions is also an interesting attempt.
- 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 addresses a clinically relevant problem and the overall motivation is clear. The proposed framework is well organized, and incorporating lung lobe information into MIL is a reasonable and intuitive design choice. Another positive aspect is the inclusion of external validation, which makes the experimental study more convincing.
- 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 current experimental section is still not fully sufficient to support the claimed contribution. First, the paper lacks several key ablation studies, so it remains unclear which component is actually driving the gain. Second, the interpretability claim is mainly based on attention scores, but stronger supporting evidence is still needed; for example, more direct visual analysis or case-level qualitative comparisons would make this part much more convincing. Third, the baseline comparison is not strong enough, as several recent MIL methods are missing, such as RRT and MambaMIL. Finally, although the method is competitive, the improvement over strong baselines is relatively limited, and some reported standard deviations are not small, yet there is no further stability analysis to show the robustness of the method.
- 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.
(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?
Insufficient experimental design and results
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
The rebuttal has not addressed my main concerns that the experimental results presented in current paper are not sufficient enough to support their claimed contribution. Therefore, I recommend to reject this paper.
Author Feedback
We thank all reviewers for their constructive and thoughtful feedback. We address each point below but are not allowed to provide new results here. [R1, R2, R3] Marginal Quantitative Improvement. We agree the quantitative gains are modest. Our primary contribution is not a large jump in classification performance but the integration of lung anatomy into MIL to enable anatomically grounded interpretability and spatial phenotype analysis while maintaining high predictive performance. Notably, AnatoMIL matches a higher-complexity transformer (TransMIL) while adding clinically meaningful lobe-level contribution profiles. [R1, R2, R3] Ablation Study. While an explicit ablation table is not presented, our comparisons demonstrate the incremental benefit of two key design choices: hierarchical aggregation and anatomy-aware lobe priors. The gain from ABMIL to HierarchicalMIL reflects hierarchical pooling, while the gain from HierarchicalMIL to AnatoMIL reflects anatomical guidance. Regarding the ordinal head, its primary objective is not to improve binary AUC, but to model the clinically ordered nature of COPD progression, providing stage-aware supervision aligned with GOLD criteria and encouraging the shared embedding to reflect disease progression. [R1] Encoder Generalization. ResNet-18 was chosen as a lightweight and widely used backbone to evaluate all methods under the identical feature extraction conditions and isolate the effect of the proposed MIL aggregation strategy. Prior work by Tang et al. (Ref. [17]) showed on the same PanCan cohort that deeper ResNet variants (50/101/152) did not yield major improvements in COPD detection, suggesting encoder depth is not the primary performance driver. We agree evaluating CT-specific pretrained encoders would be valuable. [R1] HierarchicalMIL Baseline Clarification. HierarchicalMIL refers to Zhang et al. (MICCAI 2024, Ref. [25]), which introduced hierarchical aggregation for COPD grading. We will add this citation. [R1] Lobe Fraction Vector Derivation. Lobe segmentation is performed on the 3D volume. The vectors used by the MIL model are instance-level lobe fractions, computed for each selected 2D axial slice rather than as a single whole-lung 3D ratio. We will clarify this in the revision. [R2] Segmentation Robustness. This is a valid practical concern, with two mitigating factors already in the framework. First, TotalSegmentator (Ref. [20]) is a state-of-the-art segmentation model trained on large-scale CT datasets with demonstrated robustness across pathologies. Second, we use soft lobe-fraction vectors rather than hard binary slice-to-lobe assignments, making lobe representations less sensitive to imperfect segmentations. We agree explicitly analyzing severe segmentation failures is an important future direction. [R2] Code Availability. We will release the full source code upon acceptance. [R3] Missing Baselines. RRT and MambaMIL were originally developed for histopathology WSIs, where MIL bags contain much longer patch sequences than in our CT data. We expect our small bags of selected axial slices would reduce the relative advantage of long-sequence modeling architectures, but we agree evaluating additional MIL approaches would be worthwhile. [R3] Qualitative Interpretability Evidence. To propose our method, we prioritized cohort-level analyses to reveal systematic and reproducible spatial patterns generalizable across patients and independent datasets. Figure 2 demonstrates consistent lobe-level contribution patterns across GOLD stages in both cohorts, providing population-level evidence for the interpretability of the proposed framework. However, we agree case-level visualizations would complement this analysis and should be part of future work. [R3] Stability Analysis. The proposed method is similarly stable in a 5-fold CV compared to the baselines and produced similar improvements across both datasets, but we agree an explicit stability analysis should be done in extended work.
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.
The paper receives mixed reviews. Rebuttal is invited.
- 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 rebuttal, this paper receives mixed reviews with 2 positives (1 with low confidence) and 1 negative. After reading the comments and rebuttal, I find several major concerns unsolved. 1) All reviewers mention that the performance gain is marginal, and some standard deviations are not small. 2) Comparing method is relative old. 3) Technical novelty is constrained by the use of ResNet18.Hence, I lean to rejection of the work.
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.
Note: Despite my reminder to reviewer 1, a post-rebuttal evaluation was not provided.
Justification: Based on the available pre- and post-rebuttal evaluations and the rebuttal text, I follow reviewer 2’s recommendation for acceptance (which is also close to the weak-accept of reviewer 1 in the pre-rebuttal phase). The main reason is that in my view the authors eloquently clarify the concerns raised by all reviewers, despite reviewer 3 not seeing it like this. Assuming the authors implement the content of the rebuttal in the camera-ready paper - including an explicit mention of limitations, I feel that the needed quality is given, in particular due to the clinical interpretability and strong evaluation.
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.
One reviewer gave acceptance, the other one rejection. I agree with the acceptance one (given that the post rebuttal decision was missing from another reviewer who is giving weak acceptance) given the fact that overall trend is towards acceptance.
Apart from this analysis, to be honest, the rebuttal is sounding and addressing the primary concerns. The authors are not proposing something to jump in classification performance but something to match a higher-complexity transformer (TransMIL) while adding clinically meaningful lobe-level contribution profiles. Not we should always expect a method should beat every other method. Careful reading shows that the evaluations and the paper has some merit in it and I am advocating an acceptance for poster presentation as a fair conference paper.