List of Papers Browse by Subject Areas Author List
Abstract
Intracerebral hemorrhage (ICH) is a critical condition where rapid, objective brain injury assessment is vital for treatment planning and prognosis. In clinical practice, although the Glasgow Coma Scale (GCS) evaluates brain injury severity, it relies on indirect assessments by eye-opening, verbal, and motor responses, making its accuracy susceptible to the patient’s clinical status. Furthermore, this manual evaluation is hindered by significant subjectivity and limited efficiency. Consequently, automated multimodal evaluation via deep learning has become a research priority. However, existing methods struggle with feature redundancy and distributional heterogeneity across modalities, often failing to capture complex inter-patient variations via static fusion strategies. To address this problem, we propose ICH-DIANet, a novel framework integrating 3D Computerized Tomography (CT) images and clinical text for automated brain injury grading. The framework’s Modality Disentanglement Module (MDM) projects features into both a modal-consistent subspace and a modality-specific subspace via orthogonality constraints, effectively eliminating redundancy while preserving complementary information. Additionally, a proposed Attention-Based Mixture-of-Experts Module (AMoE) employs a dynamic gating mechanism to adaptively weight expert outputs based on individual patient characteristics, while integrating an attention mechanism to facilitate deep cross-modal semantic alignment. Extensive experiments on both private and public datasets demonstrate that ICH-DIANet significantly outperforms state-of-the-art methods in accuracy and robustness.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1049_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{LiuChe_ICHDIANet_MICCAI2026,
author = { Liu, Chengqing AND Xiong, Zicheng AND Rao, Yuxuan AND Lu, Guohui AND Li, Meihua AND Lee, Hoileong AND Niu, Jingzhen AND Chen, Shengbo AND Shu, Xujun},
title = { { ICH-DIANet: A Modality-Disentangled and Attention-Expert Interaction Network for Brain Injury Assessment in Intracerebral Hemorrhage } },
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
This paper addresses the brain injury classification by using the imaging and report more efficiently. Specifically, the MDM tries to separate the language and image features, which is followed by the AMoE fusion of different features. The performance gain shows the effectiveness of the method.
- 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 idea is simple but effective. The two components (MDM and AMoE) are combined to make a better performance. 2.Detailed ablation study on the loss functions.
- 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 word “disentangle” might not be suitable for image and text feature separation. To my understanding, it should be about separating two features that were originally twisted, which is not the case of different modalities. 2.The description of AMoE is very limited. I think the authors tries to use the figure to show the fusion branches but the design details are still needed in the main paper, not just from conceptual level. 3.BioClinicalBERT is used to process the reports, but details are missing. For example, does it involved in training? Are there any preprocessing of the reports?
- 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.
- 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 interesting and effective according to the present status. The quality should be improved by including more details and improve the reproducibility by either details in models or code release.
- 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.
I tend to accept as the authors have clarified my concerns and R2’s concerns, especially the key factors like the dataset design and usage of ResNet-10.However, one big concern is the reproducibility. The paper has missed a lot of details and promised to provide those details in the final version, including the whole model structure or the key parameters. It is hard to examine this and if the paper got accepted, the authors may also forget what they have promised. I hope the meta-reviewer and AC could take this into consideration.
Review #2
- Please describe the contribution of the paper
This paper proposes a multi-modal classification framework for predicting Glasgow Coma Scale (GCS) severity ratings from CT brain scans and clinical reports. The architecture introduces two principal components: a Modality Disentanglement Module (MDM) and an Attention-Based Mixture-of-Experts (AMoE) module. Following modality-specific encoding, the MDM decomposes learned representations into modality-invariant and modality-specific features, which are subsequently concatenated and passed to the AMoE module. The AMoE module employs multiple attention mechanisms with a gating mechanism to adaptively weight and aggregate features for final classification. The proposed method is evaluated on both a public and private dataset, demonstrating improvements over several baselines. An ablation study further validates the contribution of each proposed component and associated loss terms.
- 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.Performance gains over baseline methods are compelling, particularly relative to ICHPro, which similarly leverages both imaging and textual modalities. 2.The ablation study is complete, covering all proposed modules and individual loss terms. 3.Attention heatmaps and latent space visualizations provide qualitative evidence that the model attends to relevant regions.
- 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 choice of GCS as the prediction target is not well justified. The broader ICH outcome-prediction literature more commonly targets six-month functional outcomes such as the Glasgow Outcome Scale (GOS) or binary survival prognosis (e.g., Pease et al., and ICHPro itself, which the authors compare against). As stated by the authors, GCS reporting can also be unreliable in intubated or sedated patients, and the public dataset documentation does not clarify whether scores were collected under such conditions. 2.The clinical text inputs are not fully described. The public dataset used in this work contains features such as neurological examination findings and oxygen saturation, which can act as proxies for GCS. Without a complete list of clinical features fed to the model, it is impossible to rule out that such proxy variables were included as input. 3.The novelty of the methodological contributions is limited. Modality disentanglement and mixture-of-experts are both well-established paradigms, and the paper does not clearly articulate what is new beyond their combination and application to ICH. 4.The choice of a ResNet10 pretrained on Parkinson’s disease brain images as the CT encoder is not explained. Given the substantial domain gap between Parkinson’s disease and intracerebral hemorrhage, an explicit justification for this design choice is needed. 5.The baseline comparison against ICHPro is insufficiently described. The first row of Table 2 (the proposed backbone without MDM or AMoE) already outperforms ICHPro. A clearer description of how comparative methods were trained would help contextualize this finding. 6.The claim that the method “yields significant accuracy improvements” is ambiguous. If statistical significance is intended, the specific test and p-values should be reported. 7.Some details are missing: the dimensionality reduction technique used for latent-space visualizations (t-SNE, PCA, UMAP) is not identified, and the method used to generate the attention heatmaps is not specified. Figure 1’s caption is non-descriptive, and the figure also omits the reconstruction component of the MDM, which is a core part of the method. 8.Minor issues: two claims in the introduction (“unimodal data remains inherently limited…” and “unstripped intra-modal noise and redundancy…”) are stated without citations. Additionally, the nnU-Net citation is incorrect (the correct reference is Isensee et al., Nature Methods, 2021).
- 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.
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper shows consistent gains over baselines and a thorough ablation study are encouraging. However, the choice of prediction target would benefit from review or a more thorough justification, the clinical text inputs are underspecified, methodological novelty is somewhat limited, and a few experimental and presentation details are missing.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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.
Thank you to the authors for the rebuttal. On Q1 (GCS as target): reframing the task as “acute severity assessment” does not fully resolve the issue and the rebuttal also does not clarify how intubated or sedated patients were handled. On Q2 (proxy features): the authors state that no GCS surrogates were used but do not provide the actual feature list. Since the public dataset includes neurological examination findings, this is not sufficient. On Q3 (novelty): the response emphasizes the “synergistic” combination of disentanglement and MoE but does not identify a specific new mechanism beyond the combination itself. On Q4 (ResNet-10): the clarification that Parkinson’s data was not used for pretraining is helpful and resolves my domain-gap concern. It would be useful to specify the actual pretraining setup in the revision. On Q5 (ICHPro baseline): the note that ICHPro was re-implemented with its original hyperparameters helps clarify the setup, though it also suggests the comparison may not be fully tuned to the target dataset, which could partly explain the Table 2 results. On Q6/Q7/Q8: the planned corrections to Fig. 1, the nnU-Net citation, terminology, and visualization details are appreciated.
Review #3
- Please describe the contribution of the paper
The paper proposes a multimodal deep learning framework for automated brain injury grading in intracerebral hemorrhage by combining 3D CT images and clinical text. It introduces feature disentanglement to separate shared and modality-specific information and an attention-based mixture-of-experts module for adaptive cross-modal fusion. Experiments on private and public datasets show improved accuracy, AUC, and F1-score over prior methods, suggesting more effective and robust ICH severity assessment.
- 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 proposed Modality Disentanglement Module (MDM) helps separate shared and modality-specific features, which reduces redundancy while preserving complementary multimodal information.
-The Attention-Based Mixture-of-Experts (AMoE) design allows more adaptive, patient-specific fusion than fixed fusion strategies.
-The combination of 3D CT imaging and clinical text provides a meaningful and clinically relevant multimodal setting for ICH assessment.
-The method shows consistent improvements over prior approaches on both private and public 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.
-While ablation studies are provided, the individual contributions of the proposed modules appear relatively limited, with the main performance gains emerging only when they are combined. This raises the question of whether the increased architectural complexity is fully justified. -The ablation analysis of the loss functions suggests that some terms, particularly L_diff and L_rec, offer only marginal improvements when used individually, making their necessity less convincing. A clearer explanation of why each loss term is required would strengthen the paper. -In addition, all loss terms are combined through simple summation without weighting. It remains unclear whether alternative weighting strategies were explored, although such strategies could better balance their relative contributions and potentially reduce the need for multiple loss components. -The method strongly relies on the integration of clinical text for performance improvement, yet the paper does not discuss scenarios where clinical data is missing or incomplete. This limits the practical applicability of the model in real-world clinical settings.
- 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 presents a well-structured multimodal framework and shows improved performance over prior methods. However, the novelty remains moderate, as the main contribution lies in the combination of existing techniques rather than in a fundamentally new formulation. In addition, the model complexity may limit practical deployment, and the study does not address missing modalities or incomplete clinical text, which are common in real-world settings.
- 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 authors clarified the intended role of the MDM and AMoE modules, explained the motivation behind the loss-function design, and provided additional details regarding the training setup, visualization methods, and encoder usage. These clarifications improve the overall interpretability and coherence of the paper.
The proposed framework tackles a clinically meaningful problem by combining 3D CT imaging with clinical text for automated ICH severity assessment. The multimodal setting is relevant in practice, and the experimental results consistently show improvements over previous approaches on both the private and public datasets. The ablation studies are also reasonably comprehensive and support the contribution of the proposed components.
At the same time, I still think some limitations remain. The methodological novelty is moderate, since the main contribution mainly comes from the integration and adaptation of existing concepts rather than from a fundamentally new formulation. In addition, although the rebuttal explains the rationale behind the loss functions and the use of unweighted summation, these points would still benefit from clearer discussion in the final manuscript. The issue of missing or incomplete clinical text, which is highly relevant in real-world clinical environments, also remains largely unexplored and is currently acknowledged mainly as future work.
Overall, despite these remaining weaknesses, I believe the paper provides a meaningful and practically relevant contribution within the Clinical Translation/Application setting, and the rebuttal sufficiently addressed the concerns to support acceptance.
Author Feedback
Author Rebuttal (Submission ID: 1049) We sincerely appreciate the reviewers for their insightful and constructive feedback. In the following sections, we provide detailed responses to the specific concerns raised.
Response to Reviewer #1 (R1) R1-Q1: “Disentangle” Terminology. We clarify that disentanglement in our paper is applied within each individual modality to separate modality-consistent features from modality-specific features, rather than describing the separation between the image and text modalities themselves. R1-Q2/Q3: AMoE & BERT Details. Due to page limits, technical depths were condensed. In the final version, we will expand on the four specialized experts of AMOE. BioClinicalBERT was fine-tuned during training for semantic alignment; reports were standardized without manual feature engineering.
Response to Reviewer #2 (R2) R2-Q1: Choice of GCS & Data Reliability. We clarify that our goal is the automated assessment of acute brain injury severity rather than predicting long-term functional outcomes like GOS. GCS is used as the reference standard because it reflects the patient’s immediate clinical state, which is critical for urgent triage. Regarding reliability, the inherent subjectivity of manual GCS—especially in sedated or intubated patients—is exactly why an objective, AI-driven tool is necessary. R2-Q2: Clinical Feature & Proxy Variables. While the full feature list was omitted previously due to space constraints, we emphasize that no variables acting as GCS surrogates were included as inputs, thereby ensuring the model learns latent correlations between raw imaging and structured text. R2-Q3: Methodological Novelty. While disentanglement and Mixture-of-Experts (MoE) are established concepts, the core novelty of ICH-DIANet lies in the synergistic integration of these paradigms, specifically tailored to address the high pathological heterogeneity of ICH. MDM’s application in our framework is not a simple adaptation of existing methods; instead, it performs crucial front-end processing by specifically addressing the “entanglement” of inconsistent clinical text and 3D imaging features. Unlike conventional static fusion or standard MoE architectures that often rely on simple linear transformations of features, our AMOE module then implements a deep, non-linear interaction mechanism through four specialized experts to capture complex latent correlations. R2-Q4: ResNet-10 Justification. We apologize for the confusion. We did not use Parkinson’s disease data for pre-training. The cited literature was intended only to justify the architectural efficiency of ResNet-10 for 3D volumes; its mention of Parkinson’s was incidental. R2-Q5: Baseline Comparison (ICHPro). In fact, ICHPro demonstrated suboptimal adaptability to our specific dataset. We re-implemented ICHPro following its original architecture and hyperparameters for a fair comparison. R2-Q6/Q7/Q8: Missing Details & Clarity. We will: 1) To avoid ambiguity regarding statistical terminology, we will replace “significant” with more precise descriptors. 2) Update Fig. 1 to include the reconstruction encoder $D_m$. 3) Specify that t-SNE was used for visualization and Grad-CAM for heatmaps. 4) Correct the nnU-Net citation (Isensee et al., 2021) and add missing citations in the Intro.
Response to Reviewer #3 (R3) R3-Q1: Complexity & Synergy. Without MDM, experts struggle with noise; without AMOE, purified features aren’t optimally utilized. This synergy is crucial for capturing high ICH pathological heterogeneity. R3-Q2/Q3: Loss Functions & Weighting. $L_{diff}$ and $L_{rec}$ are essential structural constraints ensuring valid feature separation and information retention. Grid search showed that simple summation is robust, as loss magnitudes are naturally balanced by our architectural design, avoiding over-sensitivity to hyperparameter tuning. R3-Q4: Missing Data Scenarios. This is indeed one of the core directions of our next stage of research.
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.
Given the divergent reviewer comments, please draft a rebuttal letter to address their concerns. Your response should be concise and focused on clarifying the disagreements.
- 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.
agree with the majority of reviewers.
Meta-review #2
- 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.
Reviewers come to different conclusions post-rebuttal (2xAccept, 1xReject).
Considering the reviews, the rebuttal and the paper, I recommend to Reject this paper. The clinical relevance (prediction of GCS from 3D CT and clinical text) is not convincing. The rebuttal does not sufficiently specify the text or its availability in practice and it is not clear where in the patient trajectory (at what time) this model would add value. This concern coupled with limited novelty, and remaining concerns regarding experimental setup (e.g., competitiveness of reimplemented reference methods) are main reasons for the recommendation to reject this borderline paper.
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.
The rebuttal addressed several technical concerns and two reviewers now support acceptance. However, I share Reviewer 2’s remaining concerns: the clinical text inputs are still not fully specified, GCS reliability in sedated/intubated patients is not clearly handled, and reproducibility remains weak due to missing architectural and implementation details.