List of Papers Browse by Subject Areas Author List
Abstract
Depression severity estimation is formulated as continuous regression, yet clinical decisions rely on categorical severity grades. This granularity mismatch leads to boundary-agnostic optimization and ordinally unstructured representations. We propose Multi-Granularity Supervision, jointly leveraging continuous scores and ordinal grades as complementary signals. Our key theoretical contribution proves that ordinal supervision induces boundary-aware gradient amplification, where gradient magnitude peaks at clinical thresholds. We instantiate this through Ordinal Consistency Regularization for boundary awareness in prediction space, and Prototype-based Severity Contrastive Learning for ordinal structure in feature space. On AVEC 2013, AVEC 2014, DAIC-WOZ, and E-DAIC, we achieve MAE of 5.06, 5.15, 4.33, and 5.02 respectively, establishing new state-of-the-art across all benchmarks. Code is available at https://github.com/aqlzh/BAMGL_Depression.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3817_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/aqlzh/BAMGL_Depression
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiZhi_BoundaryAware_MICCAI2026,
author = { Li, Zhihong AND Yang, Benqiang AND Wang, Yaowei AND Yang, Yun},
title = { { Boundary-Aware Multi-Granularity Learning for Depression Severity Estimation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16885},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
Designing a depression severity estimation by including novel loss functions to penalize more when crossing category boundaries and another loss based on learnable class prototypes for more structure in the feature space.
- 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 of using continuous predictions with ordinal classes are interesting and not much studied. Their proposed loss function is meaningful.
State-of-art results ranging from notable to significant in different datasets.
- 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 mention of the input modalities (face, voice. . . ) is a shortcoming. In general,
Fig 3 can be improved (remove Mamba encoder detail, and focus on the flow). Variables/dimensions may be used for clarity and redundancy about loss terms can be removed.
Datasets can be a little described for people not working in this same problem.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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?
Theoretical basis is good; results improve sota sufficiently.
- 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
The paper addresses depression severity estimation from audiovisual clinical interview recordings. The authors identify a granularity mismatch in existing approaches: standard regression methods predict continuous severity scores but ignore clinically meaningful categorical boundaries (e.g., none/mild/moderate/severe), treating errors that cross these boundaries the same as within-category errors. To address this, the paper proposes Multi-Granularity Supervision, combining three complementary loss functions: (1) standard MSE regression for continuous score prediction, (2) Ordinal Consistency Regularization (OCR), which amplifies learning at clinical severity thresholds, and (3) Prototype-based Severity Contrastive Learning (PSCL), which organises the feature space to reflect ordinal severity structure. The authors provide a theoretical proof that OCR induces boundary-aware gradient amplification, with gradient magnitude peaking at clinical thresholds. The method is evaluated on four standard depression benchmarks (AVEC 2013, AVEC 2014, DAIC-WOZ, E-DAIC) across visual, audio, and multimodal settings, achieving state-of-the-art results on all datasets. Ablation studies with statistical significance tests confirm that each component contributes independently.
- 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.
S1.Novel and well-motivated methodology: The multi-granularity supervision framework addresses a real clinical problem - that standard regression treats all prediction errors equally, regardless of whether they cross clinically meaningful severity boundaries. The combination of MSE, OCR, and PSCL is, to the best of this reviewer’s knowledge, novel for depression severity estimation. S2.Comprehensive comparison: The paper compares against 20 methods across four standard benchmarks (AVEC 2013, AVEC 2014, DAIC-WOZ, E-DAIC), covering visual-only, audio-only, and multimodal settings. State-of-the-art performance is achieved across virtually all configurations. S3.Rigorous ablation and statistical testing: The ablation study systematically removes each component and includes Wilcoxon signed-rank tests, paired t-tests, and bootstrap confidence intervals - exceeding the statistical rigour normally provided. S4.Theory-to-practice validation: Theorem 1 is empirically validated in Figure 4, where OCR loss peaks at the clinical thresholds as predicted. This connection between theoretical motivation and experimental evidence strengthens the paper’s narrative. S5.Additional analyses: The paper includes hyperparameter sensitivity analysis, Grad-CAM interpretability visualisations showing clinically relevant facial regions, and detailed component analysis of the contrastive learning module, providing a thorough understanding of the method’s behaviour.
- 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.
W1.Insufficient related work and baseline descriptions: The related work section discusses only two prior depression estimation methods, yet Tables 1 and 2 compare against 20 methods. The paper would benefit from briefly describing the key baselines, particularly the closest competitors (e.g., DepMGNN, FC-AEN, SIMMA), so the reader can understand what approaches the proposed method improves upon and why. Additionally, CORAL and SORD are discussed as motivation but not included as explicit experimental baselines. Given that the competing methods are not described, it is unclear whether any of them already employ ordinal regression approaches, making it difficult to assess whether the proposed method’s advantage comes specifically from the multi-granularity formulation. W2.Clarity and accessibility: The paper is difficult to follow for readers outside the area. The paper would benefit from a clearer description of the overall pipeline beyond the architecture diagrams. The input modality, preprocessing steps, and model architecture are largely left to be inferred from figures rather than described explicitly in the text. The four evaluation datasets are named but not described (number of subjects, recording conditions, label distributions). The evaluation protocol (e.g., whether the standard challenge partitions were used) is not explicitly stated. Readers unfamiliar with these benchmarks would need to check the original challenge papers to verify that comparisons are conducted under identical conditions. A few sentences describing the end-to-end pipeline, datasets, and evaluation protocol would significantly improve accessibility for the broader MICCAI audience. W3.No limitations discussion: The paper does not discuss any limitations of the proposed approach. A brief limitations paragraph in the conclusion would strengthen the paper.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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 paper includes helpful illustrations and tables that support the reader’s understanding. If additional space is needed (e.g., for expanding the related work), Figure 2 could be reduced in size, since Figure 3 represents the main contribution. Minor note: Figure 1 explains the two motivating problems effectively, but is visually crowded. For example, the MSE loss box in Figure 1a is placed inside the graph, which is initially confusing - this information could be incorporated into the Case A and Case B boxes instead. The emoji icons next to Challenge 1 and 2 are redundant and add visual clutter.
- 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?
The paper presents a novel and well-motivated approach to depression severity estimation, combining continuous regression with boundary-aware ordinal regularization and contrastive feature learning. The core idea - that prediction errors crossing clinical severity boundaries should be penalized more than within-category errors - is clinically meaningful and elegantly addressed through the multi-granularity supervision framework. The experimental evidence is strong: state-of-the-art results across four benchmarks, thorough ablation studies with proper statistical testing, hyperparameter sensitivity analysis, and empirical validation of the theoretical claims. The level of experimental rigour exceeds what is typically seen. The main weaknesses are presentational. The related work section is too brief, leaving the reader unable to understand what the 20 baseline methods do or how they relate to the proposed approach (W1). The paper assumes significant familiarity with the depression detection benchmarks, making it difficult to follow for the broader MICCAI audience (W2). These issues do not undermine the contribution but should be addressed in the camera-ready version. Overall, the strengths - particularly the novelty of the approach, the comprehensive evaluation, and the rigorous statistical analysis - clearly outweigh the presentational weaknesses.
- 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 addresses the challenging domain of severity predictions, where ground truth is often both continuous (e.g., between 0 and 5) and ordinal (e.g., low, moderate, high severity). The authors introduce additional regularization guided by the ordinal class labels and also use continuous values for regression loss, such as MSE.
- 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 theoretical justifications for the methodologies presented strengthen the contribution.
The authors also establish state-of-the-art results on several benchmark datasets for depression severity.
The paper was well written, with key contributions made clearer using diagrams.
Overall, the methods could be valuable for automated severity assessment, beyond the domain of depression.
- 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.
My main feedback concerns the decoupling of the model from the optimization process. It is not clear which one is the primary contributor to SOTA performance. Is it achieved via task-specific modeling, such as using carefully curated audio and video feature extractors and Mamba encoders? Or is it because of the loss functions and regularizations introduced by the authors?
It would further strengthen the study if the authors could examine what happens if the same (or similar) architecture is trained with
- simply MSE loss
- MSE loss + multi-class classification loss
- or other ordinal classification losses used in the literature
The authors mention comparing against several baselines. But ideally, one would want to show and separate the effects of the introduced loss function.
Some minor feedback: In Abstract: “This granularity mismatch leads to boundary-agnostic optimization and unstructured representations.” – unstructured representations sound a bit vague, especially when, in the abstract, the readers do not know what type of structure you are talking about.
Figure 2: Raw Input -> Are both top and bottom labels “Video”, or one should be “Video” and another should be “Audio”?
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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?
In Fig. 2, the data appear to be videos. Please discuss whether these are patient videos and whether they are subject to HIPAA compliance.
Also, do the authors have the subject’s consent to publish their photo (simply blurring the eyes does not make the data de-identifiable)?
- 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?
Please see the strengths and weaknesses above.
- 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
Dear Area Chair and Reviewers, We sincerely thank the reviewers for their careful reading, thoughtful evaluation, and highly constructive comments. We are truly encouraged by the positive feedback and deeply appreciate the time and expertise dedicated to improving our work. We will faithfully address all concerns in the camera-ready version. Our detailed responses follow. Q1.Architecture vs. loss contribution and ordinal-classification objectives [R3, R2-W1]. We sincerely thank R3 for this insightful question, which goes to the heart of our contribution. The ablation in Table 3 directly addresses the relative contribution of architecture and optimization. With the Mamba backbone fixed, moving from MSE-only to PSCL, OCR, and the full objective consistently reduces MAE on AVEC2013 and AVEC2014, validated by Wilcoxon signed-rank tests, paired t-tests, and bootstrap CIs. This confirms measurable gains from the proposed losses without changing the backbone. We fully agree with R3 that head-to-head comparisons against alternative ordinal losses under the same backbone would further isolate our objective’s contribution, and we gratefully take this excellent suggestion as a valuable direction for future work. Regarding CORAL and SORD, we will clarify their conceptual difference: they reformulate regression as ordinal classification, whereas our method preserves continuous regression and introduces ordinal information only as regularization, without altering the task formulation. Q2.Input modalities, pipeline, datasets, and evaluation protocol [R1, R2-W2]. We deeply appreciate R1 and R2 for this perceptive suggestion, which will substantially enhance the readability and reproducibility of our paper. In the Datasets and Implementation paragraph, we will clarify that experiments cover visual, audio, and audio-visual settings: visual inputs use HOG+SVM face features and audio inputs use STFT-based log-Mel features. We will also describe the four benchmarks, namely AVEC2013/2014 (BDI-II, 0–63) and DAIC-WOZ/E-DAIC (PHQ-8, 0–24), with severity thresholds and official splits. Q3.Related work and baselines [R2-W1]. We are very grateful to R2 for this constructive suggestion, which will strengthen the positioning of our work. We will explicitly position the closest competitors, including DepMGNN, FC-AEN, SIMMA, PointTransform, and AVA-DepressNet, as point-wise regression methods without explicit ordinal supervision, which is the gap our method addresses. CORAL and SORD will be described as ordinal-classification approaches that discard continuous scores. As they produce discrete grades rather than continuous BDI-II/PHQ-8 outputs, they are not directly comparable under our regression metrics (MAE/RMSE); we will make this explicit in §2.All revisions will remain within §2.Q4.Limitations [R2-W3]. We sincerely thank R2 for this valuable suggestion, which will further improve the rigor of our paper. We will add a concise statement in the conclusion noting that our evaluation relies on public benchmarks with limited scale and diversity, and that broader validation across larger and more diverse cohorts remains future work. Q5.Figures and wording [R1, R2, R3]. We greatly appreciate the reviewers’ meticulous attention to figures and wording. In the camera-ready version, we will: (i) correct Figure 2 by labeling the lower “Raw Input” branch as “Audio”; (ii) replace “unstructured representations” with “ordinally unstructured representations”; and (iii) following R1’s suggestion, streamline Fig. 3 by reducing internal Mamba details to highlight the overall flow, and trim redundant descriptions of loss terms. We again sincerely thank the reviewers for their careful reading and constructive suggestions. These targeted revisions will improve the clarity, rigor, and value of the paper for the MICCAI community.
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.
All three reviewers recommend acceptance, with strong consensus on the paper’s merits: a novel and clinically meaningful multi‑granularity supervision framework, state‑of‑the‑art results across four benchmarks, and exceptional experimental rigor including thorough ablations and statistical testing. Minor concerns, such as the need for clearer baseline descriptions, disentangling the contribution of the loss functions from the architecture, are addressed in a revision. Overall, the paper makes a solid and valuable contribution. Conclusion: Early Accept.